diff --git a/.changeset/codex-login-denied-cancelled.md b/.changeset/codex-login-denied-cancelled.md
new file mode 100644
index 000000000..1cc20b812
--- /dev/null
+++ b/.changeset/codex-login-denied-cancelled.md
@@ -0,0 +1,5 @@
+---
+"@pymodel/pythinker-code": patch
+---
+
+Report a denied OpenAI Codex sign-in as cancelled instead of asking for the redirect URL.
diff --git a/.changeset/subagent-turn-prompts.md b/.changeset/subagent-turn-prompts.md
new file mode 100644
index 000000000..8a93531dc
--- /dev/null
+++ b/.changeset/subagent-turn-prompts.md
@@ -0,0 +1,5 @@
+---
+"@pymodel/pythinker-code": patch
+---
+
+Show the prompt that started a subagent turn in the transcript.
diff --git a/.changeset/task-detach-action.md b/.changeset/task-detach-action.md
new file mode 100644
index 000000000..9bbaaa541
--- /dev/null
+++ b/.changeset/task-detach-action.md
@@ -0,0 +1,5 @@
+---
+"@pymodel/pythinker-code": minor
+---
+
+Add a task detach action to the server API. Call `POST /api/v1/sessions/{session_id}/tasks/{task_id}:detach` to move a running foreground task to the background.
diff --git a/.changeset/vscode-duplicated-stream-events.md b/.changeset/vscode-duplicated-stream-events.md
new file mode 100644
index 000000000..7a08bf1c4
--- /dev/null
+++ b/.changeset/vscode-duplicated-stream-events.md
@@ -0,0 +1,5 @@
+---
+"pythinker": patch
+---
+
+Fix duplicated streaming output when a session is opened twice at the same time.
diff --git a/.changeset/wire-journal-repair-retry.md b/.changeset/wire-journal-repair-retry.md
new file mode 100644
index 000000000..870894a98
--- /dev/null
+++ b/.changeset/wire-journal-repair-retry.md
@@ -0,0 +1,5 @@
+---
+"@pymodel/pythinker-code": patch
+---
+
+Retry a failed session journal repair before writing new records, so no message is appended behind a corrupted tail.
diff --git a/apps/pythinker-code/dist-web/.web-bundle-manifest.json b/apps/pythinker-code/dist-web/.web-bundle-manifest.json
index 234acb0c3..2f1bc773c 100644
--- a/apps/pythinker-code/dist-web/.web-bundle-manifest.json
+++ b/apps/pythinker-code/dist-web/.web-bundle-manifest.json
@@ -1,4 +1,4 @@
 {
-  "sourceHash": "d84e17f04092f5fb9afa9f4d323b614d3945ec8e33abbbc92f5793ad2c30959e",
+  "sourceHash": "aa9ad464f74ddd042d5435d29ed27bac364c25274ede0eade5d39cc7593d0eb6",
   "sourceFileCount": 404
 }
diff --git a/apps/vscode/src/runtime/pythinker-runtime.ts b/apps/vscode/src/runtime/pythinker-runtime.ts
index 013c97be7..fd325310c 100644
--- a/apps/vscode/src/runtime/pythinker-runtime.ts
+++ b/apps/vscode/src/runtime/pythinker-runtime.ts
@@ -48,6 +48,7 @@ export class PythinkerRuntime {
   private readonly log: PythinkerRuntimeOptions["log"];
   private readonly sessions = new Map<string, SessionRuntime>();
   private readonly sessionByView = new Map<string, string>();
+  private readonly viewChains = new Map<string, Promise<void>>();
   private readonly pendingPermissionByView = new Map<string, PermissionMode>();
   private closed = false;
 
@@ -101,6 +102,10 @@ export class PythinkerRuntime {
   }
 
   async openSession(options: OpenSessionOptions): Promise<SessionRuntime> {
+    return this.serializeView(options.webviewId, () => this.openSessionInner(options));
+  }
+
+  private async openSessionInner(options: OpenSessionOptions): Promise<SessionRuntime> {
     this.ensureOpen();
     const current = this.getSessionForView(options.webviewId);
     const requestedId = options.sessionId ?? current?.id;
@@ -119,7 +124,7 @@ export class PythinkerRuntime {
     if (runtime !== undefined) {
       assertSessionWorkDir(runtime.session, options.workDir);
       await applySessionPermission(runtime.session, runtime.permissionMode);
-      await this.detachView(options.webviewId);
+      await this.detachViewInner(options.webviewId);
     } else {
       const seedMode = defaultPermissionMode(options.yoloMode);
       const session =
@@ -135,7 +140,7 @@ export class PythinkerRuntime {
       try {
         assertSessionWorkDir(session, options.workDir);
         const mode = await restorePermissionMode(session, seedMode);
-        await this.detachView(options.webviewId);
+        await this.detachViewInner(options.webviewId);
         runtime = this.wrapSession(session, mode);
       } catch (error) {
         await session.close().catch((closeError: unknown) => {
@@ -156,6 +161,16 @@ export class PythinkerRuntime {
     webviewId: string,
     session: Session,
     yoloModeSetting = false,
+  ): Promise<SessionRuntime> {
+    return this.serializeView(webviewId, () =>
+      this.attachResumedSessionInner(webviewId, session, yoloModeSetting),
+    );
+  }
+
+  private async attachResumedSessionInner(
+    webviewId: string,
+    session: Session,
+    yoloModeSetting: boolean,
   ): Promise<SessionRuntime> {
     const existing = this.sessions.get(session.id);
     if (existing !== undefined && this.sessionByView.get(webviewId) === session.id) {
@@ -163,7 +178,7 @@ export class PythinkerRuntime {
       await existing.announceStatus(webviewId);
       return existing;
     }
-    await this.detachView(webviewId);
+    await this.detachViewInner(webviewId);
     let runtime = existing ?? this.sessions.get(session.id);
     if (runtime === undefined) {
       try {
@@ -184,6 +199,10 @@ export class PythinkerRuntime {
   }
 
   async detachView(webviewId: string): Promise<void> {
+    return this.serializeView(webviewId, () => this.detachViewInner(webviewId));
+  }
+
+  private async detachViewInner(webviewId: string): Promise<void> {
     const id = this.sessionByView.get(webviewId);
     if (id === undefined) return;
     this.sessionByView.delete(webviewId);
@@ -196,6 +215,23 @@ export class PythinkerRuntime {
     }
   }
 
+  // A view attaches to at most one session, so opens/detaches for one view
+  // must never overlap: concurrent callers that both miss `this.sessions`
+  // would wrap the same SDK session twice and double every streamed event.
+  private serializeView<T>(webviewId: string, work: () => Promise<T>): Promise<T> {
+    const prev = this.viewChains.get(webviewId) ?? Promise.resolve();
+    const run = prev.then(work, work);
+    const next = run.then(
+      () => undefined,
+      () => undefined,
+    );
+    this.viewChains.set(webviewId, next);
+    void next.finally(() => {
+      if (this.viewChains.get(webviewId) === next) this.viewChains.delete(webviewId);
+    });
+    return run;
+  }
+
   async closeSession(id: string): Promise<void> {
     const runtime = this.sessions.get(id);
     if (runtime === undefined) {
diff --git a/apps/vscode/test/pythinker-runtime.test.ts b/apps/vscode/test/pythinker-runtime.test.ts
index acbb761b4..9c013ced2 100644
--- a/apps/vscode/test/pythinker-runtime.test.ts
+++ b/apps/vscode/test/pythinker-runtime.test.ts
@@ -382,6 +382,91 @@ describe("Pythinker runtime (owns shared SDK sessions for Webviews)", () => {
     expect(boundary.handlerInstallations).toEqual({ approval: 1, question: 1 });
   });
 
+  it("does not double-wrap the SDK session when two opens race for it", async () => {
+    const sdk = createFakeHarness();
+    const broadcasts: { event: string; data: unknown; webviewId?: string }[] = [];
+    const runtime = new PythinkerRuntime({
+      version: "0.6.0",
+      harness: sdk.harness,
+      broadcast: (event, data, webviewId) => {
+        broadcasts.push({ event, data, webviewId });
+      },
+      captureBaseline: () => undefined,
+      log: () => undefined,
+    });
+    const boundary = sdk.addSession("saved-1", "/workspace");
+
+    const [first, second] = await Promise.all([
+      runtime.openSession(openOptions({ sessionId: "saved-1" })),
+      runtime.openSession(openOptions({ sessionId: "saved-1" })),
+    ]);
+
+    expect(second).toBe(first);
+    expect(boundary.subscriptionCount()).toBe(1);
+
+    boundary.emit({
+      type: "assistant.delta",
+      sessionId: "saved-1",
+      agentId: "main",
+      turnId: 1,
+      delta: "Hello",
+    });
+
+    const parts = broadcasts.filter(
+      ({ data }) => (data as { type?: string }).type === "ContentPart",
+    );
+    expect(parts).toHaveLength(1);
+  });
+
+  it("coalesces two concurrent new-session opens for one view onto one session", async () => {
+    const { runtime, sdk } = createRuntime();
+
+    const [first, second] = await Promise.all([
+      runtime.openSession(openOptions()),
+      runtime.openSession(openOptions()),
+    ]);
+
+    expect(second).toBe(first);
+    expect(sdk.createInputs).toHaveLength(1);
+    expect(first.subscribers).toEqual(["view-1"]);
+  });
+
+  it("does not double-wrap the SDK session when two attaches race for it", async () => {
+    const sdk = createFakeHarness();
+    const broadcasts: { event: string; data: unknown; webviewId?: string }[] = [];
+    const runtime = new PythinkerRuntime({
+      version: "0.6.0",
+      harness: sdk.harness,
+      broadcast: (event, data, webviewId) => {
+        broadcasts.push({ event, data, webviewId });
+      },
+      captureBaseline: () => undefined,
+      log: () => undefined,
+    });
+    const boundary = sdk.addSession("saved-1", "/workspace");
+
+    const [first, second] = await Promise.all([
+      runtime.attachResumedSession("view-1", boundary.session),
+      runtime.attachResumedSession("view-1", boundary.session),
+    ]);
+
+    expect(second).toBe(first);
+    expect(boundary.subscriptionCount()).toBe(1);
+
+    boundary.emit({
+      type: "assistant.delta",
+      sessionId: "saved-1",
+      agentId: "main",
+      turnId: 1,
+      delta: "Hello",
+    });
+
+    const parts = broadcasts.filter(
+      ({ data }) => (data as { type?: string }).type === "ContentPart",
+    );
+    expect(parts).toHaveLength(1);
+  });
+
   it("preserves the resumed session's model instead of reapplying the configured default", async () => {
     const { runtime, sdk } = createRuntime();
     const session = sdk.addSession("saved-1", "/workspace", { model: "old-model" });
diff --git a/docs/reference/server-api.md b/docs/reference/server-api.md
index aba7eaca4..9bebdc4c1 100644
--- a/docs/reference/server-api.md
+++ b/docs/reference/server-api.md
@@ -177,6 +177,7 @@ Endpoints are grouped by resource below. A `:{action}` suffix in a path is the a
 | `GET /api/v1/sessions/{session_id}/tasks` | List background tasks |
 | `GET /api/v1/sessions/{session_id}/tasks/{task_id}` | Read a task (optional output preview) |
 | `POST /api/v1/sessions/{session_id}/tasks/{task_id}:cancel` | Cancel a task |
+| `POST /api/v1/sessions/{session_id}/tasks/{task_id}:detach` | Move a foreground task to the background |
 
 ### Skills, tools, and MCP
 
diff --git a/packages/agent-core-v2/docs/state-manifest.d.ts b/packages/agent-core-v2/docs/state-manifest.d.ts
index 71b9be85f..0d7d65d3c 100644
--- a/packages/agent-core-v2/docs/state-manifest.d.ts
+++ b/packages/agent-core-v2/docs/state-manifest.d.ts
@@ -1353,6 +1353,7 @@ export interface AgentStateSnapshot {
     readonly command: string;
     readonly pid: number;
     readonly exitCode: number | null;
+    readonly parentToolCallId?: string;
     readonly taskId: string;
     readonly description: string;
     readonly status: /* AgentTaskStatus — packages/agent-core-v2/src/agent/task/types.ts */ 'completed' | 'failed' | 'running' | 'timed_out' | 'killed' | 'lost';
@@ -1400,6 +1401,7 @@ export interface AgentStateSnapshot {
     readonly command: string;
     readonly pid: number;
     readonly exitCode: number | null;
+    readonly parentToolCallId?: string;
     readonly taskId: string;
     readonly description: string;
     readonly status: /* AgentTaskStatus — packages/agent-core-v2/src/agent/task/types.ts */ 'completed' | 'failed' | 'running' | 'timed_out' | 'killed' | 'lost';
diff --git a/packages/agent-core-v2/src/agent/loop/turnEvents.ts b/packages/agent-core-v2/src/agent/loop/turnEvents.ts
index aa6ab6630..eaca52b92 100644
--- a/packages/agent-core-v2/src/agent/loop/turnEvents.ts
+++ b/packages/agent-core-v2/src/agent/loop/turnEvents.ts
@@ -69,6 +69,7 @@ export function turnPromptAttachments(
 
 export function isDisplayablePromptOrigin(origin: PromptOrigin): boolean {
   if (origin.kind === 'user') return true;
+  if (origin.kind === 'system_trigger' && origin.name === 'subagent') return true;
   return (
     (origin.kind === 'skill_activation' || origin.kind === 'plugin_command') &&
     origin.trigger === 'user-slash'
diff --git a/packages/agent-core-v2/src/agent/tools/agent/agentTool.ts b/packages/agent-core-v2/src/agent/tools/agent/agentTool.ts
index 714400af0..b68ad9fb7 100644
--- a/packages/agent-core-v2/src/agent/tools/agent/agentTool.ts
+++ b/packages/agent-core-v2/src/agent/tools/agent/agentTool.ts
@@ -467,14 +467,14 @@ export class SubagentTool implements ISubagentTool {
 
       if (runInBackground) {
         return {
-          output: formatBackgroundAgentResult(taskId, handle, args.description, allowBackground),
+          output: formatBackgroundAgentResult(taskId, handle, args.description, allowBackground, false),
         };
       }
 
       const release = await this.tasks.waitForForegroundRelease(taskId);
       if (release === 'detached') {
         return {
-          output: formatBackgroundAgentResult(taskId, handle, args.description, allowBackground),
+          output: formatBackgroundAgentResult(taskId, handle, args.description, allowBackground, true),
         };
       }
       return await this.formatForegroundResult(taskId, handle, timeoutMs);
@@ -560,7 +560,11 @@ function formatBackgroundAgentResult(
   handle: SubagentHandle,
   description: string,
   allowBackground: boolean,
+  detachedByUser: boolean,
 ): string {
+  const nextStep = allowBackground
+    ? `next_step: The completion arrives automatically in a later turn — do NOT wait, poll, or call TaskOutput on it; continue with other work or hand back to the user. (If you have nothing to do until it finishes, run such tasks in the foreground next time.)`
+    : 'next_step: The completion arrives automatically in a later turn.';
   return [
     `task_id: ${taskId}`,
     'status: running',
@@ -570,9 +574,7 @@ function formatBackgroundAgentResult(
     '',
     `description: ${description}`,
     '',
-    allowBackground
-      ? `next_step: The completion arrives automatically in a later turn — do NOT wait, poll, or call TaskOutput on it; continue with other work or hand back to the user. (If you have nothing to do until it finishes, run such tasks in the foreground next time.)`
-      : 'next_step: The completion arrives automatically in a later turn.',
+    detachedByUser ? `note: The user moved this subagent to the background.\n${nextStep}` : nextStep,
     `resume_hint: To continue or recover this same subagent later, call Agent(resume="${handle.agentId}", prompt="..."). The parameter is agent_id ("${handle.agentId}"), NOT task_id ("${taskId}") or source_id from a later <notification>. Recovery cases: a later <notification type="task.lost" | "task.failed" | "task.killed"> for this subagent — its conversation history is preserved across session restarts and resume will pick it up.`,
   ].join('\n');
 }
diff --git a/packages/agent-core-v2/src/agent/tools/ask-user-question/askUserQuestionTool.ts b/packages/agent-core-v2/src/agent/tools/ask-user-question/askUserQuestionTool.ts
index 38375003c..e53a0b04b 100644
--- a/packages/agent-core-v2/src/agent/tools/ask-user-question/askUserQuestionTool.ts
+++ b/packages/agent-core-v2/src/agent/tools/ask-user-question/askUserQuestionTool.ts
@@ -146,7 +146,7 @@ export class AskUserQuestionTool implements IAskUserQuestionTool {
         'next_step: Continue your current work; the answer will arrive automatically when the user responds.\n' +
         'next_step: Use TaskOutput with this task_id for a non-blocking status/answer snapshot.\n' +
         'next_step: Use TaskStop only if the question should be cancelled.\n' +
-        'human_shell_hint: The pending question is also visible in /tasks.',
+        'human_shell_hint: The pending question is also visible in the client UI.',
     };
   }
 
diff --git a/packages/agent-core-v2/src/agent/tools/os/bash/bash.md b/packages/agent-core-v2/src/agent/tools/os/bash/bash.md
index 6b3ec9c4b..63599e3b8 100644
--- a/packages/agent-core-v2/src/agent/tools/os/bash/bash.md
+++ b/packages/agent-core-v2/src/agent/tools/os/bash/bash.md
@@ -13,11 +13,11 @@ The dedicated tools render in the per-tool permission UI and keep raw stdout out
 **Output:**
 The stdout and stderr will be combined and returned as a string. The output may be truncated if it is too long. If the command exits non-zero, the output ends with a `Command failed with exit code: N` line; a command killed by its timeout or interrupted by the user ends with its own message instead.
 
-If `run_in_background=true`, the command will be started as a background task and this tool will return a task ID instead of waiting for command completion. When doing that, you must provide a short `description`. Background commands default to a ${DEFAULT_BACKGROUND_TIMEOUT_S}s timeout and `timeout` is capped at ${MAX_BACKGROUND_TIMEOUT_S}s; set `disable_timeout=true` only when the task should run without a timeout. You will be automatically notified when the task completes. After starting one, default to returning control to the user instead of immediately waiting on it. Use `TaskOutput` only for a non-blocking status/output snapshot — do not wait on a task you just launched, since its completion arrives automatically. Use `TaskStop` only if the task must be cancelled. If a human user wants to inspect background tasks themselves, point them to the `/tasks` command, which opens an interactive panel; it has no subcommands.
+If `run_in_background=true`, the command will be started as a background task and this tool will return a task ID instead of waiting for command completion. When doing that, you must provide a short `description`. Background commands default to a ${DEFAULT_BACKGROUND_TIMEOUT_S}s timeout and `timeout` is capped at ${MAX_BACKGROUND_TIMEOUT_S}s; set `disable_timeout=true` only when the task should run without a timeout. You will be automatically notified when the task completes. After starting one, default to returning control to the user instead of immediately waiting on it. Use `TaskOutput` only for a non-blocking status/output snapshot — do not wait on a task you just launched, since its completion arrives automatically. Use `TaskStop` only if the task must be cancelled. If a human user wants to inspect background tasks themselves, point them to the background-task panel.
 
 **Guidelines for safety and security:**
 - Each shell tool call will be executed in a fresh shell environment. The shell variables, current working directory changes, and the shell history is not preserved between calls. To run a command in a particular directory, pass the `cwd` argument (or use absolute paths) rather than relying on a `cd` from an earlier call.
-- The tool call will return after the command is finished. You shall not use this tool to execute an interactive command or a command that may run forever. For possibly long-running foreground commands, set the `timeout` argument in seconds. Foreground commands default to ${DEFAULT_TIMEOUT_S}s and allow up to ${MAX_TIMEOUT_S}s. When a foreground command hits its timeout it is moved to the background instead of being killed, and you will be automatically notified when it completes.
+- The tool call will return after the command is finished. You shall not use this tool to execute an interactive command or a command that may run forever. For possibly long-running foreground commands, set the `timeout` argument in seconds. Foreground commands default to ${DEFAULT_TIMEOUT_S}s and allow up to ${MAX_TIMEOUT_S}s. When a foreground command hits its timeout it is moved to the background instead of being killed, and you will be automatically notified when it completes. The user can also move a running foreground command to the background at any time.
 - Avoid using `..` to access files or directories outside of the working directory.
 - Avoid modifying files outside of the working directory unless explicitly instructed to do so.
 - Never run commands that require superuser privileges unless explicitly instructed to do so.
diff --git a/packages/agent-core-v2/src/agent/tools/os/bash/bashTool.ts b/packages/agent-core-v2/src/agent/tools/os/bash/bashTool.ts
index 961ffd903..41090010c 100644
--- a/packages/agent-core-v2/src/agent/tools/os/bash/bashTool.ts
+++ b/packages/agent-core-v2/src/agent/tools/os/bash/bashTool.ts
@@ -69,13 +69,17 @@ function renderBashDescription(shellName: string): string {
 function withoutBackgroundDescription(description: string): string {
   return description
     .replace(
-      /\r?\n\r?\nIf `run_in_background=true`,[\s\S]*?point them to the `\/tasks` command, which opens an interactive panel; it has no subcommands\./,
+      /\r?\n\r?\nIf `run_in_background=true`,[\s\S]*?point them to the background-task panel\./,
       '\n\nBackground execution is disabled for this agent. Do not set `run_in_background=true`.',
     )
     .replace(
       ` For possibly long-running foreground commands, set the \`timeout\` argument in seconds. Foreground commands default to ${String(DEFAULT_TIMEOUT_S)}s and allow up to ${String(MAX_TIMEOUT_S)}s. When a foreground command hits its timeout it is moved to the background instead of being killed, and you will be automatically notified when it completes.`,
       ` For possibly long-running commands, set the \`timeout\` argument in seconds. The default is ${String(DEFAULT_TIMEOUT_S)}s; foreground commands allow up to ${String(MAX_TIMEOUT_S)}s; a foreground command that hits its timeout is killed.`,
     )
+    .replace(
+      ' The user can also move a running foreground command to the background at any time.',
+      '',
+    )
     .replace(
       /\r?\n- Prefer `run_in_background=true`[\s\S]*?conversation to continue before the command finishes\./,
       '\n- Do not set `run_in_background=true`; background task management tools are not available.',
@@ -145,8 +149,8 @@ export class BashTool implements IBashTool {
       },
       approvalRule: literalRulePattern(this.name, args.command),
       matchesRule: (ruleArgs) => matchesGlobRuleSubject(ruleArgs, args.command),
-      execute: ({ signal, onUpdate, onForegroundTaskStart }) =>
-        this.execution(args, signal, onUpdate, onForegroundTaskStart),
+      execute: ({ signal, onUpdate, onForegroundTaskStart, toolCallId }) =>
+        this.execution(args, signal, toolCallId, onUpdate, onForegroundTaskStart),
     };
   }
 
@@ -171,6 +175,7 @@ export class BashTool implements IBashTool {
   private async execution(
     args: BashInput,
     signal: AbortSignal,
+    toolCallId: string,
     onUpdate?: (update: ToolUpdate) => void,
     onForegroundTaskStart?: (taskId: string) => void,
   ): Promise<ExecutableToolResult> {
@@ -226,7 +231,7 @@ export class BashTool implements IBashTool {
     let taskId: string;
     try {
       taskId = this.tasks.registerTask(
-        new ProcessTask(proc, command, description, onProcessOutput, () => lease.dispose()),
+        new ProcessTask(proc, command, description, onProcessOutput, () => lease.dispose(), toolCallId),
         {
           detached: startsInBackground,
           timeoutMs,
@@ -266,8 +271,8 @@ export class BashTool implements IBashTool {
                 brief: `Backgrounded ${taskId} after timeout`,
               }
             : {
-                title: 'Task moved to background',
-                brief: `Backgrounded ${taskId}`,
+                title: 'Task moved to background by the user',
+                brief: `Backgrounded ${taskId} by the user`,
               };
         return this.backgroundStartedResult(
           taskId,
@@ -275,7 +280,7 @@ export class BashTool implements IBashTool {
           description,
           labels,
           builder,
-          'foreground_detached',
+          release === 'timeout_detached' ? 'foreground_detached' : 'foreground_detached_by_user',
         );
       }
 
@@ -375,17 +380,19 @@ export class BashTool implements IBashTool {
     description: string,
     labels: { title: string; brief: string },
     builder = new ToolOutputAccumulator(),
-    scenario: 'background_started' | 'foreground_detached' = 'background_started',
+    scenario: 'background_started' | 'foreground_detached' | 'foreground_detached_by_user' = 'background_started',
   ): ExecutableToolResult {
     const status = this.tasks.getTask(taskId)?.status ?? 'running';
+    const detachedByUser = scenario === 'foreground_detached_by_user' ? 'detached_by_user: true\n' : '';
     const metadata =
       `task_id: ${taskId}\n` +
       `pid: ${String(proc.pid)}\n` +
       `description: ${description}\n` +
       `status: ${status}\n` +
+      detachedByUser +
       `automatic_notification: true\n` +
       this.nextStepLines(scenario) +
-      'human_shell_hint: Tell the human to run /tasks to open the interactive background-task panel.';
+      'human_shell_hint: The task is visible in the background-task panel.';
 
     const foregroundResult = builder.ok('');
     const foregroundOutput = foregroundResult.output.length > 0 ? foregroundResult.output : '';
@@ -403,14 +410,18 @@ export class BashTool implements IBashTool {
   }
 
   private nextStepLines(
-    scenario: 'background_started' | 'foreground_detached',
+    scenario: 'background_started' | 'foreground_detached' | 'foreground_detached_by_user',
   ): string {
-    if (scenario === 'foreground_detached') {
+    if (scenario === 'foreground_detached' || scenario === 'foreground_detached_by_user') {
       const avoid = this.allowBackground()
         ? 'do NOT wait, poll, or call TaskOutput on it'
         : 'do NOT wait or poll';
+      const moved =
+        scenario === 'foreground_detached_by_user'
+          ? 'The user moved this task to the background.'
+          : 'The task now runs in the background.';
       return (
-        'next_step: The task now runs in the background. You will be automatically notified ' +
+        `next_step: ${moved} You will be automatically notified ` +
         `when it completes — ${avoid}; continue with your current work.\n`
       );
     }
diff --git a/packages/agent-core-v2/src/agent/tools/os/bash/process-task.ts b/packages/agent-core-v2/src/agent/tools/os/bash/process-task.ts
index bddf4c160..fe66797c7 100644
--- a/packages/agent-core-v2/src/agent/tools/os/bash/process-task.ts
+++ b/packages/agent-core-v2/src/agent/tools/os/bash/process-task.ts
@@ -16,6 +16,7 @@ export interface ProcessTaskInfo extends AgentTaskInfoBase {
   readonly command: string;
   readonly pid: number;
   readonly exitCode: number | null;
+  readonly parentToolCallId?: string;
 }
 
 declare module '#/agent/task/types' {
@@ -44,6 +45,7 @@ export class ProcessTask implements AgentTask {
     readonly description: string,
     private readonly onOutput?: ProcessTaskOutputCallback,
     private release?: () => void,
+    readonly parentToolCallId?: string,
   ) {}
 
   async start(sink: AgentTaskSink): Promise<void> {
@@ -101,6 +103,7 @@ export class ProcessTask implements AgentTask {
       command: this.command,
       pid: this.proc.pid,
       exitCode: this.exitCode,
+      parentToolCallId: this.parentToolCallId,
     };
   }
 
diff --git a/packages/agent-core-v2/src/app/codexLogin/codexLoginService.ts b/packages/agent-core-v2/src/app/codexLogin/codexLoginService.ts
index b98cc2795..c30b35ba3 100644
--- a/packages/agent-core-v2/src/app/codexLogin/codexLoginService.ts
+++ b/packages/agent-core-v2/src/app/codexLogin/codexLoginService.ts
@@ -118,6 +118,10 @@ export class CodexLoginFlow {
       void callback
         .waitForCode({ timeoutMs: LOGIN_TTL_MS })
         .then((result) => {
+          if (result !== null && 'denied' in result) {
+            this.deny(attempt);
+            return;
+          }
           if (result === null) {
             if (
               attempt.state === 'pending' &&
@@ -222,6 +226,13 @@ export class CodexLoginFlow {
     this.attempt = undefined;
   }
 
+  private deny(attempt: Attempt): void {
+    if (attempt.state !== 'pending' || attempt.committing === true) return;
+    attempt.state = 'cancelled';
+    attempt.message = 'OpenAI Codex authorization was denied.';
+    this.cleanup(attempt);
+  }
+
   private expire(attempt: Attempt): void {
     if (attempt.state !== 'pending' || attempt.committing === true) return;
     attempt.state = 'failed';
diff --git a/packages/agent-core-v2/src/wire/repair.ts b/packages/agent-core-v2/src/wire/repair.ts
index 801793bcf..19fe4e87f 100644
--- a/packages/agent-core-v2/src/wire/repair.ts
+++ b/packages/agent-core-v2/src/wire/repair.ts
@@ -23,7 +23,7 @@ export async function repairWireJournal(
   key: string,
   records: readonly unknown[],
   truncation: AppendLogTruncation,
-): Promise<void> {
+): Promise<'repaired' | 'failed'> {
   const { appendLog, storage, log, telemetry } = services;
   let backupCreated = false;
   let outcome: 'repaired' | 'failed' = 'repaired';
@@ -60,6 +60,7 @@ export async function repairWireJournal(
     dropped_count: droppedCount,
     backup_created: backupCreated,
   });
+  return outcome;
 }
 
 function countJournalLines(data: Uint8Array): number {
diff --git a/packages/agent-core-v2/src/wire/wireService.ts b/packages/agent-core-v2/src/wire/wireService.ts
index 7dbdf75b8..aeeffab26 100644
--- a/packages/agent-core-v2/src/wire/wireService.ts
+++ b/packages/agent-core-v2/src/wire/wireService.ts
@@ -39,6 +39,10 @@ export class WireService extends Service implements IWireService {
 
   private readonly wireScope: string;
   private persistQueue: Promise<void> | undefined;
+  private pendingRepair:
+    | { readonly records: WireRecord[]; readonly truncation: AppendLogTruncation }
+    | undefined;
+  private persistError: Error | undefined;
 
   constructor(
     @IAgentScopeContext scopeContext: IAgentScopeContext,
@@ -63,7 +67,11 @@ export class WireService extends Service implements IWireService {
   }
 
   appendRecord(record: WireRecord, dehydrate?: RecordDehydrator): void {
-    if (dehydrate === undefined && this.persistQueue === undefined) {
+    if (
+      this.pendingRepair === undefined &&
+      dehydrate === undefined &&
+      this.persistQueue === undefined
+    ) {
       try {
         this.appendRecordLow(record);
       } catch (error) {
@@ -77,6 +85,9 @@ export class WireService extends Service implements IWireService {
       ) as Promise<readonly unknown[]>;
     const queued = (this.persistQueue ?? Promise.resolve())
       .then(async () => {
+        if (this.pendingRepair !== undefined) {
+          await this.repairPendingJournal();
+        }
         const output = dehydrate === undefined ? record : await dehydrate(record, transform);
         this.appendRecordLow(output);
       })
@@ -165,7 +176,7 @@ export class WireService extends Service implements IWireService {
         records.push(record);
       }
     }
-    await repairWireJournal(
+    const outcome = await repairWireJournal(
       {
         appendLog: this.log,
         storage: this.storage,
@@ -177,10 +188,38 @@ export class WireService extends Service implements IWireService {
       records,
       truncation,
     );
+    this.pendingRepair = outcome === 'failed' ? { records, truncation } : undefined;
+  }
+
+  private async repairPendingJournal(): Promise<void> {
+    const pending = this.pendingRepair;
+    if (pending === undefined) return;
+    await this.repairJournal(pending.truncation, pending.records);
+    if (this.pendingRepair !== undefined) {
+      const error = new WireError(
+        WireErrors.codes.RECORDS_WRITE_FAILED,
+        'Wire journal repair did not complete; record was not appended',
+        {
+          details: {
+            scope: this.wireScope,
+            key: AGENT_WIRE_RECORD_KEY,
+            lineNumber: pending.truncation.lineNumber,
+          },
+        },
+      );
+      this.persistError = error;
+      throw error;
+    }
   }
 
   async flush(): Promise<void> {
     await this.persistQueue;
+    if (this.pendingRepair !== undefined && this.persistError === undefined) {
+      await this.repairPendingJournal().catch(() => undefined);
+    }
+    const persistError = this.persistError;
+    this.persistError = undefined;
+    if (persistError !== undefined) throw persistError;
     await this.log.flush();
   }
 
diff --git a/packages/agent-core-v2/test/agent/fullCompaction/fullCompaction.test.ts b/packages/agent-core-v2/test/agent/fullCompaction/fullCompaction.test.ts
index 88c4296d3..833f22974 100644
--- a/packages/agent-core-v2/test/agent/fullCompaction/fullCompaction.test.ts
+++ b/packages/agent-core-v2/test/agent/fullCompaction/fullCompaction.test.ts
@@ -658,7 +658,7 @@ describe('FullCompaction', () => {
       event: 'compaction_finished',
       properties: expect.objectContaining({
         source: 'manual',
-        tokens_before: 17_935,
+        tokens_before: 17_943,
         retry_count: 1,
         trace_id: 'trace-compact-1',
       }),
@@ -1125,7 +1125,7 @@ describe('FullCompaction', () => {
       properties: expect.objectContaining({
         agent_id: 'main',
         source: 'manual',
-        tokens_before: 17_935,
+        tokens_before: 17_943,
         duration_ms: expect.any(Number),
         round: 1,
         retry_count: 0,
@@ -1350,7 +1350,7 @@ describe('FullCompaction', () => {
       event: 'compaction_failed',
       properties: expect.objectContaining({
         source: 'manual',
-        tokens_before: 17_935,
+        tokens_before: 17_943,
         duration_ms: expect.any(Number),
         retry_count: 4,
         error_type: 'APIConnectionError',
diff --git a/packages/agent-core-v2/test/agent/loop/loop.test.ts b/packages/agent-core-v2/test/agent/loop/loop.test.ts
index 41ad96dae..5d77e5f54 100644
--- a/packages/agent-core-v2/test/agent/loop/loop.test.ts
+++ b/packages/agent-core-v2/test/agent/loop/loop.test.ts
@@ -142,8 +142,8 @@ describe('Agent loop', () => {
       [emit] turn.step.started           { "time": "<time>", "agentId": "main", "turnId": 0, "step": 1, "stepId": "<uuid-1>" }
       [emit] agent.activity.updated      { "time": "<time>", "lifecycle": "ready", "turn": { "turnId": 0, "origin": { "kind": "user" }, "phase": "running", "step": 1, "ending": false, "pendingApprovals": [], "activeToolCalls": [], "since": "<time>" }, "background": [], "agentId": "main" }
       [wire] context.append_loop_event   { "agentId": "main", "event": { "type": "step.begin", "uuid": "<uuid-1>", "turnId": "0", "step": 1 }, "time": "<time>" }
-      [wire] llm.tools_snapshot          { "agentId": "main", "hash": "d3e22e7861e40632ba8e85ac56a6f0adea6ca3acf2f00b249b2b656657b59473", "tools": [ { "name": "Agent", "description": "Launch a subagent to handle a task. The subagent runs as a same-process loop instance with its own context and wire file. Delegating also keeps the bulk of intermediate file contents out of your own context — you get a conclusion back instead of a pile of dumps.\\n\\nWriting the prompt:\\n- The subagent starts with zero context — it has not seen this conversation. Brief it like a colleague who just walked into the room: state the goal, list what you already know, hand over the specifics.\\n- Lookups (read this file, run that test): put the exact path or command in the prompt. The subagent should not have to search for things you already know.\\n- Investigations (figure out X, find why Y): give the question, not prescribed steps — fixed steps become dead weight when the premise is wrong.\\n- Do not delegate understanding. If the task hinges on a file path or line number, find it yourself first and write it into the prompt.\\n\\nUsage notes:\\n- When the task continues earlier work a subagent already did, prefer resuming that agent (pass its \`resume\` id) over spawning a fresh instance — the resumed agent keeps its prior context.\\n- A subagent's result is only visible to you, not to the user. When the user needs to see what a subagent produced, summarize the relevant parts yourself in your own reply.\\n- Subagents use a fixed 2-hour timeout. If one times out, resume the same agent instead of starting over.\\n\\nWhen NOT to use Agent: skip delegation for trivial work you can do directly — reading a file whose path you already know, searching a small known set of files, or any task that takes only a step or two. Delegation has a context-handoff cost; it pays off only when the task is substantial enough to outweigh it.\\n\\nOnce a subagent is running, leave that scope to it: do not redo its searches or reads in parallel, and do not abandon it midway and finish the job manually. Both undo the context savings the delegation was meant to buy.\\n\\n\\nWhen \`run_in_background=true\`, the subagent runs detached from this turn. The completion arrives in a later turn as a synthetic user-role message containing its result — you do not need to poll, sleep, or check on its progress. Continue with other work or respond to the user. Never fabricate or predict what the result will say.\\n\\nDefault to a foreground subagent (omit \`run_in_background\`) when your next step needs its result — foreground hands the result straight back. Reach for \`run_in_background=true\` only when you have other work to do while it runs and do not need its result to proceed. Never launch in the background and then immediately wait on it (by polling \`TaskOutput\`, sleeping, or otherwise): that just blocks the turn for no benefit — run it in the foreground instead.\\n\\n\\nAvailable agent types (pass via subagent_type):\\n- plan: Read-only implementation planning and architecture design. Use this agent when the parent agent needs a step-by-step implementation plan, key file identification, and architectural trade-off analysis before code changes are made.\\n  Tools: Read, ReadMediaFile, Glob, Grep, WebSearch, FetchURL\\n- coder: General software engineering agent — the only subagent type with file-editing tools; use it for any delegated task that must modify code. Use this agent for non-trivial software engineering work that may require reading files, editing code, running commands, and returning a compact but technically complete summary to the parent agent.\\n  Tools: Bash, CronCreate, CronDelete, CronList, Edit, EnterPlanMode, ExitPlanMode, Glob, Grep, Read, ReadMediaFile, Skill, TaskList, TaskOutput, TaskStop, TodoList, WaitFor, WebSearch, FetchURL, Write, mcp__*\\n- explore: Fast codebase exploration with prompt-enforced read-only behavior. Fast agent specialized for exploring codebases. Use this when you need to quickly find files by patterns (e.g. \\"src/**/*.yaml\\"), search code for keywords (e.g. \\"database connection\\"), or answer questions about the codebase (e.g. \\"how does the auth module work?\\"). When calling this agent, specify the desired thoroughness level: \\"quick\\" for basic searches, \\"medium\\" for moderate exploration, or \\"thorough\\" for comprehensive analysis across multiple locations and naming conventions. Use this agent for any read-only exploration that will clearly require more than 3 search queries. Prefer launching multiple explore agents concurrently when investigating independent questions.\\n  Tools: Bash, Read, ReadMediaFile, Glob, Grep, WebSearch, FetchURL", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "prompt": { "type": "string", "description": "Full task prompt for the subagent" }, "description": { "type": "string", "description": "Short task description (3-5 words) for UI display" }, "subagent_type": { "description": "One of the available agent types (see \\"Available agent types\\" in this tool description). Defaults to \\"coder\\" when omitted.", "type": "string" }, "resume": { "description": "Optional agent ID to resume instead of creating a new instance. When set, do not also pass subagent_type — the resumed agent keeps its own type, and supplying both is rejected.", "type": "string" }, "run_in_background": { "description": "If true, return immediately without waiting for completion. Prefer false unless the task can run independently and there is a clear benefit to not waiting.", "type": "boolean" } }, "required": [ "prompt", "description" ], "additionalProperties": false } }, { "name": "AgentDynamicWorkflow", "description": "Launch multiple subagents from one prompt template, existing agent resumes, or both.\\n\\nUse AgentDynamicWorkflow when many subagents should run the same kind of task over different inputs. The placeholder is exactly \`{{item}}\`. For example, with \`prompt_template\` set to \`Review {{item}} for likely regressions.\` and \`items\` set to \`[\\"src/a.ts\\", \\"src/b.ts\\"]\`, AgentDynamicWorkflow launches two new subagents with those two concrete prompts. For a few differently-shaped tasks, make separate \`Agent\` calls in one message instead.\\n\\nUse \`resume_agent_ids\` to continue subagents that already exist from earlier work, such as ones that failed or timed out: map each agent id to the prompt for that resumed subagent (usually \`continue\` if no extra information is needed). You may combine \`resume_agent_ids\` with \`items\` in the same call to resume existing subagents and launch new ones. Do not duplicate resumed work in \`items\`.\\n\\nEach of these is enforced — a violation is rejected before any subagent starts: provide at least 2 \`items\` unless you pass \`resume_agent_ids\`; whenever \`items\` are present, \`prompt_template\` is required and must contain \`{{item}}\`; and the filled-in prompts must be distinct (two items that expand to the same prompt are rejected).\\n\\nUse enough subagents to keep the work focused and parallel. AgentDynamicWorkflow supports up to 128 subagents, and launches are queued automatically, so it is safe to split large tasks into many clear, independent items.\\n\\nIf \`AgentDynamicWorkflow\` is called, that call must be the only tool call in the response.\\n\\nThis capability is called a \\"dynamic workflow\\" (or simply \\"workflow\\"). Never use the word \\"swarm\\" in the \`description\` field, in subagent prompts, or when talking to the user about this tool.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "description": { "type": "string", "minLength": 1, "description": "Short description for the whole dynamic workflow. It is shown to the user as the workflow and subagent title, so word it as a workflow or run and never use the word \\"swarm\\"." }, "subagent_type": { "description": "Subagent type used for every new subagent spawned from items; defaults to coder when omitted. Resumed subagents always keep their original type, so passing subagent_type together with resume_agent_ids is allowed — it only affects the item-based spawns.", "type": "string", "minLength": 1 }, "prompt_template": { "description": "Prompt template for each subagent. The {{item}} placeholder is replaced with each item value.", "type": "string", "minLength": 1 }, "items": { "description": "Values used to fill {{item}}. Each item launches one new subagent.", "maxItems": 128, "type": "array", "items": { "type": "string", "minLength": 1 } }, "resume_agent_ids": { "description": "Map of existing subagent agent_id to the prompt used to resume that subagent. These resumed subagents are launched before new item-based subagents.", "type": "object", "propertyNames": { "type": "string", "minLength": 1 }, "additionalProperties": { "type": "string", "minLength": 1 } } }, "required": [ "description" ], "additionalProperties": false } }, { "name": "AskUserQuestion", "description": "Use this tool when you need to ask the user questions with structured options during execution. This allows you to:\\n1. Collect user preferences or requirements before proceeding\\n2. Resolve ambiguous or underspecified instructions\\n3. Let the user decide between implementation approaches as you work\\n4. Present concrete options when multiple valid directions exist\\n\\n**When NOT to use:**\\n- When you can infer the answer from context — be decisive and proceed\\n- Trivial decisions that don't materially affect the outcome\\n\\nOverusing this tool interrupts the user's flow. Only use it when the user's input genuinely changes your next action.\\n\\n**Usage notes:**\\n- Users always have an \\"Other\\" option for custom input — don't create one yourself\\n- Use multi_select to allow multiple answers to be selected for a question\\n- Keep option labels concise (1-5 words), use descriptions for trade-offs and details\\n- Each question should have 2-4 meaningful, distinct options\\n- Question texts must be unique across the call, and option labels must be unique within each question\\n- You can ask 1-4 questions at a time; group related questions to minimize interruptions\\n- If you recommend a specific option, list it first and append \\"(Recommended)\\" to its label\\n- The result is JSON with an \`answers\` object keyed by question text; each value is the chosen option's label (comma-separated labels for multi_select, or the user's own words if they picked \\"Other\\"); if \`answers\` is empty and a \`note\` says the user dismissed it, they chose not to answer — do not treat this as selecting the recommended option; decide based on context and do not re-ask the same question\\n- Set background=true when you can keep working without the answer. This starts a background question task and returns a task_id immediately. The answer arrives automatically in a later turn — you do not need to poll, sleep, or check on it. Continue with other work; never fabricate or predict the answer.", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "questions": { "minItems": 1, "maxItems": 4, "type": "array", "items": { "type": "object", "properties": { "question": { "type": "string", "minLength": 1, "description": "A specific, actionable question. End with '?'." }, "header": { "default": "", "description": "Short category tag (max 12 chars, e.g. 'Auth', 'Style').", "type": "string" }, "options": { "minItems": 2, "maxItems": 4, "type": "array", "items": { "type": "object", "properties": { "label": { "type": "string", "minLength": 1, "description": "Concise display text (1-5 words). If recommended, append '(Recommended)'." }, "description": { "default": "", "description": "Brief explanation of trade-offs or implications.", "type": "string" } }, "required": [ "label" ], "additionalProperties": false }, "description": "2-4 meaningful, distinct options. Do NOT include an 'Other' option — the system adds one automatically." }, "multi_select": { "default": false, "description": "Whether the user can select multiple options.", "type": "boolean" } }, "required": [ "question", "options" ], "additionalProperties": false }, "description": "The questions to ask the user (1-4 questions)." }, "background": { "default": false, "description": "Set true to ask in the background and return immediately with a background task_id; you are notified automatically when the user answers — do not poll with TaskOutput while the question is pending.", "type": "boolean" } }, "required": [ "questions" ], "additionalProperties": false } }, { "name": "Bash", "description": "Execute a \`bash\` command. Use this for shell semantics — pipes, env, processes, git, package managers, build/test runners, anything genuinely interactive or multi-step.\\n\\n**Translate these to a dedicated tool instead:**\\n- \`cat\` / \`head\` / \`tail\` (known path) → \`Read\`\\n- \`sed\` / \`awk\` (in-place edit) → \`Edit\`\\n- \`echo > file\` / \`cat <<EOF\` → \`Write\`\\n- \`find\` / recursive \`ls\` to locate files by name pattern → \`Glob\` (plain \`ls <known-directory>\` is fine for listing a directory)\\n- \`grep\` / \`rg\` (search file contents) → \`Grep\`\\n- \`echo\` / \`printf\` (talk to the user) → just output text directly\\n\\nThe dedicated tools render in the per-tool permission UI and keep raw stdout out of the conversation; that is why they are worth reaching for whenever one fits.\\n\\n**Output:**\\nThe stdout and stderr will be combined and returned as a string. The output may be truncated if it is too long. If the command exits non-zero, the output ends with a \`Command failed with exit code: N\` line; a command killed by its timeout or interrupted by the user ends with its own message instead.\\n\\nIf \`run_in_background=true\`, the command will be started as a background task and this tool will return a task ID instead of waiting for command completion. When doing that, you must provide a short \`description\`. Background commands default to a 600s timeout and \`timeout\` is capped at 86400s; set \`disable_timeout=true\` only when the task should run without a timeout. You will be automatically notified when the task completes. After starting one, default to returning control to the user instead of immediately waiting on it. Use \`TaskOutput\` only for a non-blocking status/output snapshot — do not wait on a task you just launched, since its completion arrives automatically. Use \`TaskStop\` only if the task must be cancelled. If a human user wants to inspect background tasks themselves, point them to the \`/tasks\` command, which opens an interactive panel; it has no subcommands.\\n\\n**Guidelines for safety and security:**\\n- Each shell tool call will be executed in a fresh shell environment. The shell variables, current working directory changes, and the shell history is not preserved between calls. To run a command in a particular directory, pass the \`cwd\` argument (or use absolute paths) rather than relying on a \`cd\` from an earlier call.\\n- The tool call will return after the command is finished. You shall not use this tool to execute an interactive command or a command that may run forever. For possibly long-running foreground commands, set the \`timeout\` argument in seconds. Foreground commands default to 60s and allow up to 300s. When a foreground command hits its timeout it is moved to the background instead of being killed, and you will be automatically notified when it completes.\\n- Avoid using \`..\` to access files or directories outside of the working directory.\\n- Avoid modifying files outside of the working directory unless explicitly instructed to do so.\\n- Never run commands that require superuser privileges unless explicitly instructed to do so.\\n\\n**Guidelines for efficiency:**\\n- Use \`&&\` to chain commands that genuinely depend on each other, e.g. \`npm install && npm test\`. Independent read-only commands (separate \`git show\`, \`ls\`, or status checks) should be issued as separate parallel Bash calls in one response, not chained into a single call — chaining serializes their execution and mixes their output. Do not stitch outputs together with \`echo\` separators.\\n- Use \`;\` to run commands sequentially regardless of success/failure\\n- Use \`||\` for conditional execution (run second command only if first fails)\\n- Use pipe operations (\`|\`) and redirections (\`>\`, \`>>\`) to chain input and output between commands\\n- Always quote file paths containing spaces with double quotes (e.g., cd \\"/path with spaces/\\")\\n- Compose multi-step logic in a single call with \`if\` / \`case\` / \`for\` / \`while\` control flows.\\n- Prefer \`run_in_background=true\` for long-running builds, tests, watchers, or servers when you need the conversation to continue before the command finishes.\\n\\n**Commands available:**\\nThe following common command categories are usually available. Availability still depends on the host, so when in doubt run \`which <command>\` first to confirm a command exists before relying on it.\\n- Navigation and inspection: \`ls\`, \`pwd\`, \`cd\`, \`stat\`, \`file\`, \`du\`, \`df\`, \`tree\`\\n- File and directory management: \`cp\`, \`mv\`, \`rm\`, \`mkdir\`, \`touch\`, \`ln\`, \`chmod\`, \`chown\`\\n- Text and data processing: \`wc\`, \`sort\`, \`uniq\`, \`cut\`, \`tr\`, \`diff\`, \`xargs\`\\n- Archives and compression: \`tar\`, \`gzip\`, \`gunzip\`, \`zip\`, \`unzip\`\\n- Networking and transfer: \`curl\`, \`wget\`, \`ping\`, \`ssh\`, \`scp\`\\n- Version control: \`git\`; for GitHub-hosted work (PRs, issues, CI runs, API queries) prefer the \`gh\` CLI when installed — it carries the user's GitHub auth and can return structured JSON\\n- Process and system: \`ps\`, \`kill\`, \`top\`, \`env\`, \`date\`, \`uname\`, \`whoami\`\\n- Language and package toolchains: \`node\`, \`npm\`, \`pnpm\`, \`yarn\`, \`python\`, \`pip\` (use whichever the project actually relies on)\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "command": { "type": "string", "minLength": 1, "description": "The command to execute." }, "cwd": { "description": "The working directory in which to run the command. When omitted, the command runs in the session's working directory.", "type": "string" }, "timeout": { "default": 60, "description": "Optional timeout in seconds for the command to execute. Foreground default 60s, max 300s. Background default 600s, max 86400s. Ignored for background commands when disable_timeout=true.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 }, "description": { "description": "A short description for the background task. Required when run_in_background is true.", "type": "string" }, "run_in_background": { "description": "Whether to run the command as a background task.", "type": "boolean" }, "disable_timeout": { "description": "If true, do not apply a timeout to the command. Only applies when run_in_background is true.", "type": "boolean" } }, "required": [ "command" ], "additionalProperties": false } }, { "name": "CreateGoal", "description": "Create a durable, structured goal that the runtime will pursue across multiple turns.\\n\\nCall \`CreateGoal\` only when:\\n\\n- the user explicitly asks you to start a goal or work autonomously toward an outcome, or\\n- a host goal-intake prompt asks you to create one.\\n\\nDo NOT create a goal for greetings, ordinary questions, or vague requests that lack a\\nverifiable completion condition. A goal needs a checkable end state.\\n\\nWhen the request is vague, ask the user for the missing completion criterion before creating\\nthe goal. If the user clearly insists after you warn them that the wording is vague or risky,\\nrespect that and create the goal.\\n\\nInclude a \`completionCriterion\` when the user provides one, or when it can be stated without\\ninventing new requirements. Keep \`objective\` concise; reference long task descriptions by file\\npath rather than pasting them.\\n\\nCreating a goal fails if one already exists, so use \`replace: true\` only when the user explicitly\\nwants to abandon the current goal and start a new one.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "objective": { "type": "string", "minLength": 1, "description": "The objective to pursue. Must have a verifiable end state." }, "completionCriterion": { "description": "How to verify the goal is complete. Include when the user provides one.", "type": "string" }, "replace": { "description": "Replace an existing active, paused, or blocked goal instead of failing.", "type": "boolean" } }, "required": [ "objective" ], "additionalProperties": false } }, { "name": "CronCreate", "description": "Schedule a prompt to be enqueued at a future time. Use for both recurring schedules and one-shot reminders.\\n\\nUses standard 5-field cron in the user's local timezone: minute hour day-of-month month day-of-week. \`0 9 * * *\` means 9am local — no timezone conversion needed.\\n\\n## One-shot tasks (recurring: false)\\n\\nFor \\"remind me at X\\" or \\"at <time>, do Y\\" requests — fire once then auto-delete.\\nPin minute/hour/day-of-month/month to specific values:\\n  \\"remind me at 2:30pm today to check the deploy\\" → cron: \\"30 14 <today_dom> <today_month> *\\", recurring: false\\n  \\"tomorrow morning, run the smoke test\\" → cron: \\"57 8 <tomorrow_dom> <tomorrow_month> *\\", recurring: false\\n\\nOne-shots are best for near-term reminders. A task only fires while its session is still alive (see Session lifetime below), so favor near times — within hours or a few days — rather than scheduling weeks or months ahead.\\n\\n## Recurring jobs (recurring: true, the default)\\n\\nFor \\"every N minutes\\" / \\"every hour\\" / \\"weekdays at 9am\\" requests:\\n  \\"*/5 * * * *\\" (every 5 min), \\"0 * * * *\\" (hourly), \\"0 9 * * 1-5\\" (weekdays at 9am local)\\n\\n## Avoid the :00 and :30 minute marks when the task allows it\\n\\nEvery user who asks for \\"9am\\" gets \`0 9\`, and every user who asks for \\"hourly\\" gets \`0 *\` — which means requests from across the planet land on the API at the same instant. When the user's request is approximate, pick a minute that is NOT 0 or 30:\\n  \\"every morning around 9\\" → \\"57 8 * * *\\" or \\"3 9 * * *\\" (not \\"0 9 * * *\\")\\n  \\"hourly\\" → \\"7 * * * *\\" (not \\"0 * * * *\\")\\n  \\"in an hour or so, remind me to...\\" → pick whatever minute you land on, don't round\\n\\nOnly use minute 0 or 30 when the user names that exact time and clearly means it (\\"at 9:00 sharp\\", \\"at half past\\", coordinating with a meeting). When in doubt, nudge a few minutes early or late — the user will not notice, and the fleet will.\\n\\n## Coalesce semantics\\n\\nFires are delivered only while the session is idle: a fire that comes due during an active turn is held and delivered at the next idle moment, never injected mid-turn.\\n\\nIf the scheduler slept past multiple ideal fire times (laptop closed, long-running turn, etc.), only **one** fire is delivered when it wakes up. The origin carries \`coalescedCount\` showing how many ideal fires were collapsed into this single delivery. You should treat \`coalescedCount > 1\` as \\"I missed some checks; only the latest state matters\\" rather than running the prompt that many times.\\n\\n## Cron-fire envelope\\n\\nWhen a cron task fires, the prompt you scheduled is re-injected wrapped in an XML envelope that exposes the fire context:\\n\\n\`\`\`\\n<cron-fire jobId=\\"...\\" cron=\\"...\\" recurring=\\"true|false\\" coalescedCount=\\"N\\" stale=\\"true|false\\">\\n<prompt>\\nyour original prompt text, verbatim\\n</prompt>\\n</cron-fire>\\n\`\`\`\\n\\nThe envelope is parseable. Use \`coalescedCount > 1\` to know multiple ideal fires were collapsed into a single delivery (treat as \\"only the latest state matters\\"), and \`stale=\\"true\\"\` as a cue that the task is past its 7-day threshold.\\n\\n## 7-day stale behavior\\n\\nRecurring tasks that have been alive for more than 7 days fire one\\nfinal time with \`stale: true\` on the envelope, and the system then\\nauto-deletes the task. The flag is the model's notice that this is\\nthe last delivery. If the schedule is still wanted, call \`CronCreate\`\\nagain with the same \`cron\` and \`prompt\` — that resets \`createdAt\` and\\nstarts a fresh 7-day window. One-shot tasks are never marked stale.\\n\\n## Jitter behavior\\n\\nAnti-herd jitter is applied deterministically per task id:\\n  - Recurring: ideal fire time is shifted **forward** by an offset ≤ min(10% of the cron period, 15 minutes). A \`*/5 * * * *\` task can drift up to 30s; a \`0 9 * * *\` task can drift up to 15 minutes.\\n  - One-shot: only when the ideal fire lands on \`:00\` or \`:30\` of the hour, the fire is pulled **earlier** by ≤ 90 seconds. Other minutes pass through unchanged.\\n\\n## One-shot vs recurring — when to pick which\\n\\nUse \`recurring: false\` for \\"remind me at X\\" style requests, single deadlines, \\"in N minutes do Y\\", and any task that should not repeat. Use \`recurring: true\` for periodic polling (CI status, build watchers, scheduled reports), workday rituals, and anything the user explicitly described as recurring.\\n\\n## Session lifetime\\n\\nCron tasks live in the current session. When you exit, they\\nare persisted under the session homedir; resuming the same session\\nreloads them and the scheduler resumes from each task's \`createdAt\`. Fire times that fell during the offline window are\\ncollapsed into a single delivery via \`coalescedCount\` (and recurring\\ntasks past their 7-day window arrive with \`stale: true\` as their final\\ndelivery).\\n\\nTasks do **not** carry over into a brand-new session — they are scoped\\nto the resumed session id, not to the working directory.\\n\\n## Limits\\n\\nA session holds at most 50 live cron tasks; creating one beyond that is rejected. (The \`prompt\` body is also capped — see its parameter description.) Expressions that never fire within the next 5 years (e.g. \`0 0 31 2 *\`, an impossible date) are rejected at create time.\\n\\n## Returned fields\\n\\n\`id\` (ULID), \`cron\` (the normalized expression), \`humanSchedule\` (English summary), \`recurring\`,\\n\`nextFireAt\` (local ISO timestamp with numeric offset, or null). \`id\` is needed by \`CronDelete\`.\\n\\n## Tell the user how to cancel or modify\\n\\nAfter successfully creating a task, proactively tell the user how they can cancel or modify it later. Users have no direct \`/cron\` command or self-service UI to manage reminders themselves; they must ask the model to make changes (e.g. \\"cancel my 9am reminder\\" or \\"change my daily check to 10am\\"). Include the task \`id\` in your message so the user can reference it.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "cron": { "type": "string", "description": "5-field cron expression in local time: \\"M H DoM Mon DoW\\" (e.g. \\"*/5 * * * *\\" = every 5 minutes; \\"30 14 28 2 *\\" = Feb 28 at 2:30pm local — a pinned date like this repeats yearly unless you also pass recurring: false)." }, "prompt": { "type": "string", "minLength": 1, "maxLength": 8192, "description": "The prompt to enqueue at each fire time. Limited to 8 KiB (UTF-8)." }, "recurring": { "default": true, "description": "true (default) = fire on every cron match until deleted or auto-expired after 7 days. false = fire once at the next match, then auto-delete. Use false for \\"remind me at X\\" one-shot requests with pinned minute/hour/dom/month.", "type": "boolean" } }, "required": [ "cron", "prompt" ], "additionalProperties": false } }, { "name": "CronDelete", "description": "Cancel a scheduled cron job by id.\\n\\nUse this tool to remove a cron task previously scheduled with\\n\`CronCreate\`. The \`id\` is the ULID value returned by \`CronCreate\`, or\\nshown in the \`id:\` column of \`CronList\` — quote it verbatim, no\\nprefix.\\n\\nBehaviour by task kind:\\n\\n- **Recurring task** (\`recurring: true\`): stops all future fires\\n  immediately. The scheduler picks up the deletion on its next tick.\\n- **One-shot task** (\`recurring: false\`): cancels the pending fire if\\n  it has not happened yet. One-shots that have already fired\\n  auto-delete themselves, so calling \`CronDelete\` on a fired one-shot\\n  returns \\"no cron job with id ...\\".\\n\\nNot-found is reported as an error (not a silent no-op) so you can\\ncorrect yourself — typically by calling \`CronList\` to see which ids\\nare actually live, rather than re-trying with the same stale id.\\n\\nRefresh pattern (use when you want a stale recurring schedule to\\ncontinue):\\n\\nStale recurring tasks are auto-deleted by the system after their final\\nfire — there is nothing for \`CronDelete\` to remove at that point. To\\nkeep the schedule running, just call \`CronCreate\` with the same \`cron\`\\nand \`prompt\`. Use \`CronList\`'s \`prompt\` field to recall the original\\ntext after a context compaction.\\n\\n\`CronDelete\` remains the right call when you want to cancel a task\\nthat is still live (recurring not yet stale, or a one-shot still\\npending).\\n\\nGuidelines:\\n\\n- Users have no direct \`/cron\` command or self-service UI to delete\\n  tasks themselves; they must ask the model to cancel a reminder.\\n  When deleting on behalf of a user, confirm the action and report\\n  the result plainly.\\n- Cron deletion is irreversible — there is no undo. If you delete the\\n  wrong task, you must re-create it with \`CronCreate\`.\\n- If the model is unsure which id is current (e.g. after a context\\n  compaction), call \`CronList\` first rather than guessing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "id": { "type": "string", "description": "The cron job id (ULID) returned by CronCreate / CronList." } }, "required": [ "id" ], "additionalProperties": false } }, { "name": "CronList", "description": "List all cron jobs currently scheduled in this session.\\n\\nUse this tool to see every pending cron task — both recurring jobs and\\none-shot reminders — that you (or the user) have scheduled with\\n\`CronCreate\`. The output is the entry point for inspecting scheduled\\nwork: it returns a stable id, the original cron expression, a human\\nrendering, the next post-jitter fire time, the recurring flag, the\\ntask's age in days, and a stale indicator.\\n\\nEach record carries:\\n\\n- \`id\` — the task id (a ULID). Pass this to \`CronDelete\` to remove the\\n  task, or quote it in user-facing messages when asking for\\n  confirmation.\\n- \`cron\` — the verbatim 5-field cron expression as scheduled.\\n- \`humanSchedule\` — plain-English rendering (e.g. \`every 5 minutes\`).\\n- \`prompt\` — the scheduled prompt text, JSON-encoded so embedded\\n  newlines stay on one line. Truncated to 200 UTF-8 bytes with\\n  \`…(truncated)\` if longer. Use this to recall what a task is for\\n  after a context compaction, and as the source for the\\n  \`CronCreate\` refresh ritual.\\n- \`nextFireAt\` — local ISO timestamp with an explicit numeric offset\\n  for the next fire **after jitter has been applied**. The actual fire\\n  may land slightly before or after a round \`:00\` / \`:30\` minute mark\\n  due to herd-avoidance jitter; this is the value the scheduler will\\n  compare against, so it reflects what will really happen. \`null\` if\\n  the expression has no fire in the next 5 years (should not happen\\n  for tasks created through \`CronCreate\`, which validates).\\n- \`recurring\` — \`true\` for cadenced jobs, \`false\` for one-shots.\\n- \`ageDays\` — \`(now - createdAt) / day\`, two decimal places. Useful\\n  when deciding whether a long-running cron is still relevant.\\n- \`stale\` — \`true\` when a recurring task is older than 7 days. The\\n  system **auto-deletes the task after this fire** to bound session\\n  lifetime; the \`stale: true\` flag is the model's notice that this is\\n  the final delivery. To resume the same schedule, call \`CronCreate\`\\n  again with the original \`cron\` and \`prompt\` (the \`prompt\` row above\\n  carries it for exactly this purpose). One-shots are never marked\\n  stale — they fire at most once by construction.\\n\\nGuidelines:\\n\\n- This tool is read-only and never mutates state, so it is always\\n  safe to call (including in plan mode).\\n- Users cannot directly manage cron tasks themselves; if they want to\\n  cancel or modify a schedule, route the request through the model\\n  (i.e. call \`CronDelete\` or \`CronCreate\` on their behalf).\\n- The empty case returns \`cron_jobs: 0\\\\nNo cron jobs scheduled.\`. Cron\\n  tasks survive a resume of the same session but do not bleed into new\\n  sessions.\\n- After a context compaction, or whenever you are unsure which cron\\n  jobs are live, call this tool to re-enumerate them rather than\\n  guessing ids from earlier in the conversation.\\n- Records are separated by a line containing just \`---\`, in the\\n  insertion order they were scheduled.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "Edit", "description": "Perform exact replacements in existing files.\\n\\n- Edit is mandatory for every incremental change, especially small edits. DO NOT use Write or Bash \`sed\`.\\n- Read the target file before every Edit. DO NOT call Edit from memory, stale context, or a guessed \`old_string\`.\\n- Take \`old_string\` and \`new_string\` from the Read output view.\\n- Drop the line-number prefix and tab; match only file content.\\n- \`old_string\` must be unique unless \`replace_all\` is set.\\n- If \`old_string\` is ambiguous, add surrounding context. Use \`replace_all\` only when every occurrence should change — for example, renaming a symbol throughout the file.\\n- Multiple Edit calls may run in one response only when they do not target the same file.\\n- DO NOT issue consecutive Edit calls on the same file. A previous Edit can invalidate a later Edit's \`old_string\`, causing \`old_string not found\`. Read the file again before the next Edit.\\n- A write lock serializes same-file edits in response order, but serialization does not make stale \`old_string\` valid.\\n- For pure CRLF files, Read shows LF; use LF in \`old_string\` and \`new_string\`, and Edit writes CRLF back.\\n- For mixed endings or lone carriage returns, Read shows carriage returns as \\\\r; include actual \\\\r escapes in those positions.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to the text file to edit. Relative paths resolve against the working directory; a path outside the working directory must be absolute." }, "old_string": { "type": "string", "minLength": 1, "description": "Exact content to replace from the Read output view, without the line-number prefix. Use LF for pure CRLF files; use actual \\\\r escapes where Read shows \\\\r." }, "new_string": { "type": "string", "description": "Replacement text in the same Read output view. LF is written back as CRLF only for pure CRLF files." }, "replace_all": { "description": "Set true only when every occurrence of old_string should be replaced.", "type": "boolean" } }, "required": [ "path", "old_string", "new_string" ], "additionalProperties": false } }, { "name": "EnterPlanMode", "description": "Use this tool proactively when you're about to start a non-trivial implementation task.\\nGetting user sign-off on your approach via ExitPlanMode before writing code prevents wasted effort.\\n\\nUse it when ANY of these conditions apply:\\n\\n1. New Feature Implementation - e.g. \\"Add a caching layer to the API\\"\\n2. Multiple Valid Approaches - e.g. \\"Optimize database queries\\" (indexing vs rewrite vs caching)\\n3. Code Modifications - e.g. \\"Refactor auth module to support OAuth\\"\\n4. Architectural Decisions - e.g. \\"Add WebSocket support\\"\\n5. Multi-File Changes - involves more than 2-3 files\\n6. Unclear Requirements - need exploration to understand scope\\n7. User Preferences Matter - if user input would materially change the implementation approach, use EnterPlanMode to structure the decision\\n\\nPermission mode notes:\\n- EnterPlanMode enters plan mode automatically without an approval prompt in all permission modes.\\n- In yolo and manual modes, ExitPlanMode still presents the plan to the user for approval.\\n- In auto permission mode, do not use AskUserQuestion; make the best decision from available context.\\n- In auto permission mode, ExitPlanMode exits plan mode without asking the user.\\n- Use EnterPlanMode only when planning itself adds value.\\n\\nWhen NOT to use:\\n- Single-line or few-line fixes (typos, obvious bugs, small tweaks)\\n- User gave very specific, detailed instructions\\n- Pure research/exploration tasks\\n\\nOnce you are in plan mode, a reminder walks you through the workflow (explore → design → write the plan file → \`ExitPlanMode\`) and enforces read-only access. For non-trivial tasks where you are unsure of the codebase structure or relevant code paths, use \`Agent(subagent_type=\\"explore\\")\` to investigate first when the \`Agent\` tool is available.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "ExitPlanMode", "description": "Use this tool when you are in plan mode and have finished writing your plan to the plan file and are ready for user approval.\\n\\n## How This Tool Works\\n- You should have already written your plan to the plan file specified in the plan mode reminder.\\n- This tool does NOT take the plan content as a parameter - it reads the plan from the file you wrote.\\n- The user will see the contents of your plan file when they review it. In auto permission mode, the tool reads the file and exits plan mode without asking the user.\\n\\n## When to Use\\nOnly use this tool for tasks that require planning implementation steps. For research tasks (searching files, reading code, understanding the codebase), do NOT use this tool.\\n\\n## What a good plan contains\\nList specific, verifiable steps grounded in the actual codebase — real files, functions, and commands, in a sensible order. Each step should be concrete enough to act on and to check. Avoid vague filler like \\"improve performance\\" or \\"add tests\\"; say what to change and where.\\n\\n## Multiple Approaches\\nIf your plan offers multiple alternative approaches, pass them via the \`options\` parameter so the user can choose which one to execute — see the \`options\` parameter for the format, count, and reserved labels. In yolo and manual modes the user sees all options alongside the host's Reject and Revise controls.\\n\\n## Before Using\\n- In auto permission mode, do NOT use AskUserQuestion; make the best decision from available context.\\n- In auto permission mode, this tool exits plan mode without asking the user.\\n- In yolo and manual modes, this tool still presents the plan to the user for approval.\\n- If auto permission mode is not active and you have unresolved questions, use AskUserQuestion first.\\n- If auto permission mode is not active and you have multiple approaches and haven't narrowed down yet, consider using AskUserQuestion first to let the user choose, then write a plan for the chosen approach only.\\n- Once your plan is finalized, use THIS tool to request approval.\\n- Do NOT use AskUserQuestion to ask \\"Is this plan OK?\\" or \\"Should I proceed?\\" - that is exactly what ExitPlanMode does.\\n- If rejected, revise based on feedback and call ExitPlanMode again.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "options": { "description": "When the plan contains multiple alternative approaches, list them here so the user can choose which one to execute. Provide up to 3 options; 2-3 distinct approaches work best when the plan offers a real choice. Passing a single option is allowed and is equivalent to a plain plan approval. Each option represents a distinct approach from the plan. Do not use \\"Reject\\", \\"Revise\\", \\"Approve\\", or \\"Reject and Exit\\" as labels.", "minItems": 1, "maxItems": 3, "type": "array", "items": { "type": "object", "properties": { "label": { "type": "string", "minLength": 1, "maxLength": 80, "description": "Short name for this option (1-8 words). Append \\"(Recommended)\\" if you recommend this option." }, "description": { "default": "", "description": "Brief summary of this approach and its trade-offs.", "type": "string" } }, "required": [ "label" ], "additionalProperties": false } } }, "additionalProperties": false } }, { "name": "FetchURL", "description": "Fetch content from a URL. The content is returned either as the main text extracted from the page, or as the full response body verbatim; a note at the top of the result states which of the two you received, so you can judge how complete it is. Use this when you need to read a specific web page.\\n\\nOnly fully-formed public \`http\`/\`https\` URLs are supported; other schemes and private or loopback addresses are not fetched. Very large pages may be truncated or refused. The fetch carries no login or session for the target site, so pages behind authentication (private repositories, internal dashboards) return a login page or an error instead of the real content — if the text you get back looks like a generic landing or sign-in page, treat that as the login wall, not the answer, and reach the content through a credentialed route (an authenticated CLI or MCP tool) instead.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "url": { "type": "string", "description": "The URL to fetch content from." } }, "required": [ "url" ], "additionalProperties": false } }, { "name": "GetGoal", "description": "Read the current goal: its objective, completion criterion, status, and budgets (turns, tokens,\\ntime, and how much of each remains). When the goal has stopped, it also reports the terminal reason.\\n\\nUse \`GetGoal\` before deciding whether to continue working, report completion, report a blocker,\\nor respect a pause. It returns \`{ \\"goal\\": null }\` when there is no current goal.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "Glob", "description": "Find files by glob pattern, sorted by modification time (most recent first).\\n\\nPowered by ripgrep. Respects \`.gitignore\`, \`.ignore\`, and \`.rgignore\` by default — set \`include_ignored\` to also match ignored files (e.g. build outputs, \`node_modules\`). Sensitive files (such as \`.env\`) are always filtered out. Matches are files only — directories themselves are never listed; to find a directory, glob for a file inside it (e.g. \`**/fixtures/**\`).\\n\\nGood patterns:\\n- \`*.ts\` — all files matching an extension, at any depth below the search root (a bare pattern without \`/\` matches recursively)\\n- \`src/*.ts\` — files directly inside \`src/\` (one level, not recursive)\\n- \`src/**/*.ts\` — recursive walk with a subdirectory anchor and extension\\n- \`**/*.py\` — recursive walk from the search root for an extension\\n- \`*.{ts,tsx}\` — brace expansion is supported\\n- \`{src,test}/**/*.ts\` — cartesian brace expansion is supported too\\n\\nResults are capped at the first 100 matching paths. If a search would return more, a truncation marker is appended. Refine the pattern (extension, subdirectory) when 100 is not enough, or call again with a narrower anchor.\\n\\nLarge-directory caveat — avoid recursing into dependency / build output even with an anchor, especially when \`include_ignored\` is set:\\n- \`node_modules/**/*.js\`, \`.venv/**/*.py\`, \`__pycache__/**\`, \`target/**\` can produce thousands of results that truncate at the match cap and waste context. Prefer specific subpaths like \`node_modules/react/src/**/*.js\`.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "pattern": { "type": "string", "description": "Glob pattern to match files." }, "path": { "description": "Directory to search. Accepts an absolute path, or a path relative to the current working directory. Defaults to the current working directory.", "type": "string" }, "include_ignored": { "description": "Also match files excluded by ignore files such as \`.gitignore\`, \`.ignore\`, and \`.rgignore\` (for example \`node_modules\` or build outputs). Sensitive files (such as \`.env\`) remain filtered out for safety. VCS metadata directories (\`.git\` and similar) are always skipped, even when this is true. Defaults to false.", "type": "boolean" }, "include_dirs": { "description": "Deprecated and ignored. Results are always files-only — directories are never listed. Accepted only so older calls that still pass this flag are not rejected by parameter validation.", "type": "boolean" } }, "required": [ "pattern" ], "additionalProperties": false } }, { "name": "Grep", "description": "Search file contents using regular expressions (powered by ripgrep).\\n\\nUse Grep when the task is to find unknown content or unknown file locations. Do not use shell \`grep\` or \`rg\` directly; this tool applies workspace path policy, output limits, and sensitive-file filtering.\\nALWAYS use Grep tool instead of running \`grep\` or \`rg\` from a shell — direct shell calls bypass workspace policy, output limits, and sensitive-file filtering.\\nIf you already know a concrete file path and need to inspect its contents, use Read directly instead.\\n\\nWrite patterns in ripgrep regex syntax, which differs from POSIX \`grep\` syntax. For example, braces are special, so escape them as \`\\\\{\` to match a literal \`{\`.\\n\\nHidden files (dotfiles such as \`.gitlab-ci.yml\` or \`.eslintrc.json\`) are searched by default. To also search files excluded by \`.gitignore\` (such as \`node_modules\` or build outputs), set \`include_ignored\` to \`true\`. Sensitive files (such as \`.env\`) are always skipped for safety, even when \`include_ignored\` is \`true\`.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "pattern": { "type": "string", "description": "Regular expression to search for." }, "path": { "description": "File or directory to search. Accepts an absolute path, or a path relative to the current working directory. Omit to search the current working directory. Use Read instead when you already know a concrete file path and need its contents.", "type": "string" }, "glob": { "description": "Optional glob filter for which files to search, e.g. \`*.ts\`. Matched against each file's full absolute path, so a path-anchored pattern like \`src/**/*.ts\` silently matches nothing — use a basename pattern (\`*.ts\`), or anchor with \`**/\` (\`**/src/**/*.ts\`). To scope the search to a directory, use \`path\` instead.", "type": "string" }, "type": { "description": "Optional ripgrep file type filter, such as ts or py. Prefer this over \`glob\` when filtering by language or file kind: it is more efficient and less error-prone than an equivalent glob pattern.", "type": "string" }, "output_mode": { "description": "Shape of the result. \`content\` shows matching lines (honors \`-A\`, \`-B\`, \`-C\`, \`-n\`, and \`head_limit\`); \`files_with_matches\` shows only the paths of files that contain a match, most-recently-modified first (honors \`head_limit\`); \`count_matches\` shows per-file match counts as \`path:count\` lines, preceded by an aggregate total line. Defaults to \`files_with_matches\`.", "type": "string", "enum": [ "content", "files_with_matches", "count_matches" ] }, "-i": { "description": "Perform a case-insensitive search. Defaults to false.", "type": "boolean" }, "-n": { "description": "Prefix each matching line with its line number. Applies only when \`output_mode\` is \`content\`. Defaults to true.", "type": "boolean" }, "-A": { "description": "Number of lines to show after each match. Applies only when \`output_mode\` is \`content\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "-B": { "description": "Number of lines to show before each match. Applies only when \`output_mode\` is \`content\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "-C": { "description": "Number of lines to show before and after each match. Applies only when \`output_mode\` is \`content\`; takes precedence over \`-A\` and \`-B\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "head_limit": { "description": "Limit output to the first N lines/entries after offset. Defaults to 250. Pass 0 for unlimited.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "offset": { "description": "Number of leading lines/entries to skip before applying \`head_limit\`. Use it together with \`head_limit\` to page through large result sets. Defaults to 0.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "multiline": { "description": "Enable multiline matching, where the pattern can span line boundaries and \`.\` also matches newlines. Defaults to false.", "type": "boolean" }, "include_ignored": { "description": "Also search files excluded by ignore files such as \`.gitignore\`, \`.ignore\`, and \`.rgignore\` (for example \`node_modules\` or build outputs). Sensitive files (such as \`.env\`) remain filtered out for safety. VCS metadata directories (\`.git\` and similar) are always skipped, even when this is true. Defaults to false.", "type": "boolean" } }, "required": [ "pattern" ], "additionalProperties": false } }, { "name": "Read", "description": "Read a text file from the local filesystem.\\n\\nIf the user provides a concrete file path to a text file, call Read directly. Do not \`Glob\`, \`ls\`, or otherwise pre-check known text file paths; missing or invalid file paths return errors you can handle. Do not use Read for directories; use \`ls\` via Bash for a known directory, or Glob when you need files matching a name pattern (Glob lists files only, never directories). Use \`Grep\` only when the task is to search for unknown content or locations.\\n\\nWhen you need several files, prefer to read them in parallel: emit multiple \`Read\` calls in a single response instead of reading one file per turn.\\n\\n- Relative paths resolve against the working directory; a path outside the working directory must be absolute.\\n- Returns up to 1000 lines or 100 KB per call, whichever comes first; lines longer than 2000 chars are truncated mid-line.\\n- Page larger files with \`line_offset\` (1-based start line) and \`n_lines\`. Omit \`n_lines\` to read up to the 1000-line cap.\\n- Sensitive files (\`.env\` files, credential stores, SSH private keys, and similar secrets) are refused to protect secrets; do not attempt to read them. Templates and public keys are exempt: \`.env.example\` / \`.env.sample\` / \`.env.template\` and public SSH keys such as \`id_rsa.pub\` read normally.\\n- UTF-8 text files are read directly. UTF-16 LE/BE text files (with or without a BOM) are detected automatically and transcoded to UTF-8 for display; the status block notes the detected encoding, and Edit/Write on such a file still expect UTF-8 — convert its encoding first (e.g. with \`iconv\`). Other encodings (e.g. GBK), binary files, and files containing NUL bytes are refused.\\n- Negative line_offset reads from the end of the file (for example, -100 reads the last 100 lines); the absolute value cannot exceed 1000.\\n- Output format: \`<line-number>\\\\t<content>\` per line.\\n- A \`<system>...</system>\` status block is appended after the file content; it summarizes how much was read (line and byte counts, truncation, line-ending notes) and is not part of the file itself.\\n- Pure CRLF files are displayed with LF line endings; \`Edit\` matches this output and preserves CRLF when writing back.\\n- Mixed or lone carriage-return line endings are shown as \`\\\\r\` and require exact \`Edit.old_string\` escapes.\\n- After a successful \`Edit\`/\`Write\`, do not re-read solely to prove the write landed. When the task depends on an exact file, API, or output shape, inspect the final external contract before finishing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to a text file. Relative paths resolve against the working directory; a path outside the working directory must be absolute. Directories are not supported; use \`ls\` via Bash for a known directory, or Glob for pattern search." }, "line_offset": { "description": "The line number to start reading from. Omit to start at line 1. Negative values read from the end of the file; the absolute value cannot exceed 1000.", "anyOf": [ { "type": "integer", "minimum": 1, "maximum": 9007199254740991 }, { "type": "integer", "minimum": -1000, "maximum": -1 } ] }, "n_lines": { "description": "The number of lines to read; the tool also applies its internal cap. Omit to read up to the internal cap of 1000 lines.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 } }, "required": [ "path" ], "additionalProperties": false } }, { "name": "SetGoalBudget", "description": "Set a hard budget limit for the current goal.\\n\\nUse this only when the user clearly gives a runtime limit, such as:\\n\\n- \\"stop after 20 turns\\"\\n- \\"use no more than 500k tokens\\"\\n- \\"finish within 30 minutes\\"\\n\\nDo not invent limits. Do not call this for vague wording such as \\"spend some time\\" or\\n\\"try to be quick\\".\\n\\nIf the user gives a compound time, convert it to one supported unit before calling this tool.\\nFor example, \\"2 hours and 3 minutes\\" can be set as \`value: 123, unit: \\"minutes\\"\`.\\n\\nA time budget must be between 1 second and 24 hours — the tool rejects anything shorter or\\nlonger, telling the user it is not a reasonable goal budget. Turn and token budgets are not\\nbounded this way; they must be positive and are rounded to the nearest whole number (minimum 1).\\n\\nSupported units:\\n\\n- \`turns\`\\n- \`tokens\`\\n- \`milliseconds\`\\n- \`seconds\`\\n- \`minutes\`\\n- \`hours\`\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "value": { "type": "number", "exclusiveMinimum": 0, "description": "The positive numeric budget value." }, "unit": { "type": "string", "enum": [ "turns", "tokens", "milliseconds", "seconds", "minutes", "hours" ] } }, "required": [ "value", "unit" ], "additionalProperties": false } }, { "name": "Skill", "description": "Invoke a registered skill from the current skill listing. BLOCKING REQUIREMENT: when a skill from the listing matches the user's request, you MUST call this tool (not free-form text). Do not re-invoke a skill to repeat work already done: if a \`<skill-loaded>\` block for it with the same \`args\` is already present in the conversation, follow those instructions directly instead of calling the tool again. Do call the tool again when you need the skill with different arguments — the loaded block was expanded with the earlier \`args\` and will not reflect new inputs.", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "skill": { "type": "string", "description": "The exact name of the skill to invoke, spelled as it appears in the current skill listing (e.g. \\"commit\\", \\"pdf\\")." }, "args": { "description": "Optional argument string for the skill, written like a command line (e.g. \`-m \\"fix bug\\"\`, \`123\`, a file path). It is split on whitespace (quotes group a token) and expanded into the skill's placeholders ($NAME, $1, $ARGUMENTS); if the skill body has no placeholders, the whole string is still appended as a trailing \`ARGUMENTS:\` line. Omit it only when there is nothing to pass.", "type": "string" } }, "required": [ "skill" ], "additionalProperties": false } }, { "name": "TaskList", "description": "List background tasks and their current status.\\n\\nUse this tool to discover which background tasks exist and where each one\\nstands. It is the entry point for inspecting background work: it returns a\\ntask ID, status, and description for every task it reports, plus the command,\\nPID, and (once finished) exit code for shell tasks, and a stop reason for any\\ntask that ended early.\\n\\nGuidelines:\\n\\n- After a context compaction, or whenever you are unsure which background\\n  tasks are running or what their task IDs are, call this tool to\\n  re-enumerate them instead of guessing a task ID.\\n- Prefer the default \`active_only=true\`, which lists only non-terminal tasks.\\n  Pass \`active_only=false\` only when you specifically need to see tasks that\\n  have already finished. With \`active_only=false\` the result may also include\\n  \`lost\` tasks — tasks left over from a previous process that can no longer be\\n  inspected or controlled; treat them as already terminated.\\n- \`limit\` caps how many tasks are returned. It accepts a value between 1 and\\n  100 and defaults to 20 when omitted.\\n- This tool only lists tasks; it does not return their output. Use it first\\n  to locate the task ID you need, then call \`TaskOutput\` with that ID to read\\n  the task's output and details.\\n- This tool is read-only and does not change any state, so it is always safe\\n  to call, including in plan mode.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "active_only": { "default": true, "description": "Whether to list only non-terminal background tasks.", "type": "boolean" }, "limit": { "default": 20, "description": "Maximum number of tasks to return.", "type": "integer", "minimum": 1, "maximum": 100 } }, "additionalProperties": false } }, { "name": "TaskOutput", "description": "Retrieve a snapshot of a running or completed background task.\\n\\nUse this after \`Bash(run_in_background=true)\`, \`Agent(run_in_background=true)\`, or \`AskUserQuestion(background=true)\` to check progress, or to read the output of a task that has already completed.\\n\\nGuidelines:\\n- Prefer relying on automatic completion notifications. Use this tool only when you need task output before the automatic notification arrives.\\n- This tool is always non-blocking: it returns the current status/output snapshot immediately and never waits for the task to finish.\\n- Do not use TaskOutput to wait for a result you need before continuing — if your next step depends on the task's result, run that task in the foreground instead. TaskOutput is for a deliberate progress check you will act on without blocking, not a way to sit and wait for a background task you just launched.\\n- This tool returns structured task metadata, a fixed-size output preview, and an output_path for the full log.\\n- For a terminal task, the metadata also explains why it ended. A shell command that runs to completion reports \`status: completed\` on a zero exit, or \`status: failed\` with its non-zero \`exit_code\` — judge that failure from the \`exit_code\`, because a plain command failure carries no \`stop_reason\` and no \`terminal_reason\`. \`terminal_reason\` is a categorical label emitted only when the end is not an ordinary exit: \`timed_out\` when the deadline aborted it, \`stopped\` when it was explicitly stopped, or \`failed\` when it errored without producing an exit code; the \`stopped\` and \`failed\` cases also carry a human-readable \`stop_reason\`. A task that finished on its own with a clean exit carries neither \`stop_reason\` nor \`terminal_reason\`.\\n- The full, never-truncated log is always available at output_path; use the \`Read\` tool with that path to page through it, whether or not the preview was truncated.\\n- This tool works with the generic background task system and should remain the primary read path for future task types, not just bash.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "task_id": { "type": "string", "description": "The background task ID to inspect." } }, "required": [ "task_id" ], "additionalProperties": false } }, { "name": "TaskStop", "description": "Stop a running background task.\\n\\nOnly use this when a task must genuinely be cancelled — for a task that is\\nfinishing normally, wait for its completion notification or inspect it with\\n\`TaskOutput\` instead of stopping it.\\n\\nGuidelines:\\n- This is a general-purpose stop capability for any background task. It is not\\n  a bash-specific kill.\\n- Stopping a task is destructive: it may leave partial side effects behind.\\n  Use it with care.\\n- If the task has already finished, this tool simply returns its current\\n  status.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "task_id": { "type": "string", "description": "The background task ID to stop." }, "reason": { "default": "Stopped by TaskStop", "description": "Short reason recorded when the task is stopped.", "type": "string" } }, "required": [ "task_id" ], "additionalProperties": false } }, { "name": "TodoList", "description": "Use this tool to maintain a structured TODO list as you work through a multi-step task. Use it proactively and often when progress tracking helps the current work. This is especially useful in long-running investigations and implementation tasks with several tool calls; in plan mode, write the plan to the plan file rather than tracking it here.\\n\\n**When to use:**\\n- Multi-step tasks that span several tool calls\\n- Tracking investigation progress across a large codebase search\\n- Planning a sequence of edits before making them\\n- After receiving new multi-step instructions, capture the requirements as todos\\n- Before starting a tracked task, mark exactly one item as \`in_progress\`\\n- Immediately after finishing a tracked task, mark it \`done\`; do not batch completions at the end\\n\\n**When NOT to use:**\\n- Single-shot answers that complete in one or two tool calls\\n- Trivial requests where tracking adds no clarity\\n- Purely conversational or informational replies\\n\\n**Avoid churn:**\\n- Do not re-call this tool when nothing meaningful has changed since the last call — update the list only after real progress.\\n- When unsure of the current state, call query mode first (omit \`todos\`) to check the list before deciding what to update.\\n- If no available tool can move any task forward, tell the user where you are stuck instead of repeatedly re-ordering the same todos.\\n\\n**How to use:**\\n- Call with \`todos: [...]\` to replace the full list. Statuses: pending / in_progress / done.\\n- Call with no \`todos\` argument to retrieve the current list without changing it.\\n- Call with \`todos: []\` to clear the list.\\n- Keep titles short and actionable (e.g. \\"Read session-control.ts\\", \\"Add planMode flag to TurnManager\\").\\n- Update statuses as you make progress.\\n- When work is underway, keep exactly one task \`in_progress\`.\\n- Only mark a task \`done\` when it is fully accomplished.\\n- Never mark a task \`done\` if tests are failing, implementation is partial, unresolved errors remain, or required files/dependencies could not be found.\\n- If you encounter a blocker, keep the blocked task \`in_progress\` or add a new pending task describing what must be resolved.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "todos": { "description": "The updated todo list. Omit to read the current todo list without making changes. Pass an empty array to clear the list.", "type": "array", "items": { "type": "object", "properties": { "title": { "type": "string", "minLength": 1, "description": "Short, actionable title for the todo." }, "status": { "type": "string", "enum": [ "pending", "in_progress", "done" ], "description": "Current status of the todo." } }, "required": [ "title", "status" ], "additionalProperties": false } } }, "additionalProperties": false } }, { "name": "UpdateGoal", "description": "Set the status of the current goal. This is how you resume, complete, or block an autonomous goal.\\n\\n- \`active\` — resume a paused or blocked goal when the user explicitly asks you to work on that goal.\\n- \`complete\` — the objective is satisfied and any stated validation has passed. The goal ends and a completion summary is recorded. Before using this, verify the current state against the actual objective and every explicit requirement. Treat weak or indirect evidence as not complete. Do not use \`complete\` merely because a budget is nearly exhausted or you want to stop.\\n- \`blocked\` — a genuine impasse prevents useful progress: an external condition, required user input, missing credentials or permissions, a persistent technical failure, or an impossible, unsafe, or contradictory objective. For non-terminal blockers, do not use \`blocked\` the first time you hit the blocker. The same blocking condition must repeat for at least 3 consecutive goal turns before you call \`blocked\`, counting the original/user-triggered turn and automatic continuations. If a previously blocked goal is resumed, treat the resumed run as a fresh blocked audit. If the objective itself is impossible, unsafe, or contradictory, call \`blocked\` in the same turn instead of running more goal turns. Do not use \`blocked\` because the work is large, hard, slow, uncertain, incomplete, still needs validation, would benefit from clarification, or needs more goal turns. Once the 3-turn threshold is met and you cannot make meaningful progress without user input or an external-state change, call \`blocked\` instead of leaving the goal active.\\n\\nMost active goal turns should not call this tool. If you complete one useful slice of work and material work remains, end the turn normally without calling UpdateGoal; the runtime will prompt you to continue in the next goal turn. Call \`complete\` only when all required work is done, any stated validation has passed, and there is no useful next action. Do not call \`complete\` after only producing a plan, summary, first pass, or partial result. Call \`blocked\` only after the blocked audit threshold is met. If you call \`blocked\`, you will be prompted to explain the blocker in your next message. Setting the status is the machine-readable signal; the completion summary or blocker explanation is yours to write in the following message.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "status": { "type": "string", "enum": [ "active", "complete", "blocked" ], "description": "The lifecycle status to set for the current goal. Use \`blocked\` for impossible, unsafe, or contradictory objectives, or after the same non-terminal blocking condition repeats for at least 3 consecutive goal turns." } }, "required": [ "status" ], "additionalProperties": false } }, { "name": "WaitFor", "description": "Wait for background tasks to finish without ending the current turn.\\n\\nUse this when your next step depends on the result of a running background task (a sub-agent, a background bash command, or a background AskUserQuestion). The call suspends inside the current turn until the task finishes or the timeout elapses, then returns the outcome so you can keep working in the same turn. While waiting, no LLM requests are made.\\n\\nGuidelines:\\n\\n- Do not call WaitFor right after dispatching work whose result you do not need yet — finished background tasks notify you automatically. WaitFor is for the moment you genuinely cannot proceed without a result.\\n- \`timeout\` is required, in seconds, capped at 600. To wait longer, call WaitFor again; waking up periodically also lets you re-evaluate the situation.\\n- A timeout is not an error: the result lists the tasks that are still running, and you decide whether to wait again or do other work meanwhile.\\n- Without \`task_id\`, the wait ends as soon as any background task that was running at call time finishes. Tasks started during the wait are not covered by it; their completion arrives via the usual automatic notification.\\n- With \`task_id\`, the wait ends when that task finishes. An unknown \`task_id\` is an error; a task that has already finished returns immediately.\\n- When no background tasks are running, WaitFor returns immediately without waiting.\\n- When the wait ends because a task finished, the result also lists other tasks that finished during the wait window, so failures surface with context.\\n- Waiting has no side effects on the waited tasks: WaitFor never stops a task, and interrupting the wait (for example, a user interruption) leaves every task running.\\n- A finished task's result is delivered exactly once: tasks reported by WaitFor do not also produce an automatic completion notification.\\n- You can only wait for background tasks started by this agent; task IDs belonging to other agents are unknown here.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "timeout": { "type": "integer", "exclusiveMinimum": 0, "maximum": 600, "description": "Maximum time to wait, in seconds (1-600). A timeout is not an error: the tool returns the tasks that are still running, and you can call it again to keep waiting." }, "task_id": { "description": "The background task ID to wait for. When omitted, the wait ends as soon as any background task that was running at call time finishes.", "type": "string" } }, "required": [ "timeout" ], "additionalProperties": false } }, { "name": "Write", "description": "Create, append to, or replace a file entirely.\\n\\n- Missing parent directories are created automatically (like \`mkdir(parents=True, exist_ok=True)\`).\\n- Mode defaults to overwrite; append adds content at EOF without adding a newline.\\n- Write is NOT ALLOWED for incremental changes to existing files, including trivial, one-line, quick, or cosmetic edits. Use Edit instead.\\n- Use Write only when the file does not exist, you intend a complete replacement, or the new contents have little continuity with the old contents.\\n- Do not create unsolicited documentation files (\`*.md\` write-ups, \`README\`s, summaries) just because a task finished — write one only when the user asks for it, or when a task or project instruction requires it (e.g. the plan-mode plan file, created with Write when plan mode directs you to, or a changeset the repo mandates).\\n- Read before overwriting an existing file.\\n- Write ignores the Read/Edit line-number view. NEVER include line prefixes.\\n- Write outputs content literally, including supplied line endings: \\\\n stays LF, \\\\r\\\\n stays CRLF.\\n- For new content too large for one call, overwrite the first chunk, then append subsequent chunks. Never chunk Write to modify an existing file.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to the file to create, append to, or completely overwrite. Relative paths resolve against the working directory; a path outside the working directory must be absolute. Missing parent directories are created automatically." }, "content": { "type": "string", "description": "Raw full file content to write exactly as provided. This does not use the Read/Edit text view." }, "mode": { "description": "Write mode. Defaults to overwrite. append adds content to the end exactly as provided and does not add a newline.", "type": "string", "enum": [ "overwrite", "append" ] } }, "required": [ "path", "content" ], "additionalProperties": false } } ], "time": "<time>" }
-      [wire] llm.request                 { "agentId": "main", "kind": "loop", "provider": "openai", "model": "mock-model", "modelAlias": "mock-model", "thinkingEffort": "off", "maxTokens": 1000000, "toolSelect": false, "systemPromptHash": "ec9c34379c88babbc468ef2f3e0e08cd2f422c8c4a910664fb8bb394d703a575", "toolsHash": "d3e22e7861e40632ba8e85ac56a6f0adea6ca3acf2f00b249b2b656657b59473", "messageCount": 1, "turnStep": "0.1", "time": "<time>" }
+      [wire] llm.tools_snapshot          { "agentId": "main", "hash": "755845c4dac18464a7b45428e5f0f6c35626531988b85d7b6b70b5e03a45e3c2", "tools": [ { "name": "Agent", "description": "Launch a subagent to handle a task. The subagent runs as a same-process loop instance with its own context and wire file. Delegating also keeps the bulk of intermediate file contents out of your own context — you get a conclusion back instead of a pile of dumps.\\n\\nWriting the prompt:\\n- The subagent starts with zero context — it has not seen this conversation. Brief it like a colleague who just walked into the room: state the goal, list what you already know, hand over the specifics.\\n- Lookups (read this file, run that test): put the exact path or command in the prompt. The subagent should not have to search for things you already know.\\n- Investigations (figure out X, find why Y): give the question, not prescribed steps — fixed steps become dead weight when the premise is wrong.\\n- Do not delegate understanding. If the task hinges on a file path or line number, find it yourself first and write it into the prompt.\\n\\nUsage notes:\\n- When the task continues earlier work a subagent already did, prefer resuming that agent (pass its \`resume\` id) over spawning a fresh instance — the resumed agent keeps its prior context.\\n- A subagent's result is only visible to you, not to the user. When the user needs to see what a subagent produced, summarize the relevant parts yourself in your own reply.\\n- Subagents use a fixed 2-hour timeout. If one times out, resume the same agent instead of starting over.\\n\\nWhen NOT to use Agent: skip delegation for trivial work you can do directly — reading a file whose path you already know, searching a small known set of files, or any task that takes only a step or two. Delegation has a context-handoff cost; it pays off only when the task is substantial enough to outweigh it.\\n\\nOnce a subagent is running, leave that scope to it: do not redo its searches or reads in parallel, and do not abandon it midway and finish the job manually. Both undo the context savings the delegation was meant to buy.\\n\\n\\nWhen \`run_in_background=true\`, the subagent runs detached from this turn. The completion arrives in a later turn as a synthetic user-role message containing its result — you do not need to poll, sleep, or check on its progress. Continue with other work or respond to the user. Never fabricate or predict what the result will say.\\n\\nDefault to a foreground subagent (omit \`run_in_background\`) when your next step needs its result — foreground hands the result straight back. Reach for \`run_in_background=true\` only when you have other work to do while it runs and do not need its result to proceed. Never launch in the background and then immediately wait on it (by polling \`TaskOutput\`, sleeping, or otherwise): that just blocks the turn for no benefit — run it in the foreground instead.\\n\\n\\nAvailable agent types (pass via subagent_type):\\n- plan: Read-only implementation planning and architecture design. Use this agent when the parent agent needs a step-by-step implementation plan, key file identification, and architectural trade-off analysis before code changes are made.\\n  Tools: Read, ReadMediaFile, Glob, Grep, WebSearch, FetchURL\\n- coder: General software engineering agent — the only subagent type with file-editing tools; use it for any delegated task that must modify code. Use this agent for non-trivial software engineering work that may require reading files, editing code, running commands, and returning a compact but technically complete summary to the parent agent.\\n  Tools: Bash, CronCreate, CronDelete, CronList, Edit, EnterPlanMode, ExitPlanMode, Glob, Grep, Read, ReadMediaFile, Skill, TaskList, TaskOutput, TaskStop, TodoList, WaitFor, WebSearch, FetchURL, Write, mcp__*\\n- explore: Fast codebase exploration with prompt-enforced read-only behavior. Fast agent specialized for exploring codebases. Use this when you need to quickly find files by patterns (e.g. \\"src/**/*.yaml\\"), search code for keywords (e.g. \\"database connection\\"), or answer questions about the codebase (e.g. \\"how does the auth module work?\\"). When calling this agent, specify the desired thoroughness level: \\"quick\\" for basic searches, \\"medium\\" for moderate exploration, or \\"thorough\\" for comprehensive analysis across multiple locations and naming conventions. Use this agent for any read-only exploration that will clearly require more than 3 search queries. Prefer launching multiple explore agents concurrently when investigating independent questions.\\n  Tools: Bash, Read, ReadMediaFile, Glob, Grep, WebSearch, FetchURL", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "prompt": { "type": "string", "description": "Full task prompt for the subagent" }, "description": { "type": "string", "description": "Short task description (3-5 words) for UI display" }, "subagent_type": { "description": "One of the available agent types (see \\"Available agent types\\" in this tool description). Defaults to \\"coder\\" when omitted.", "type": "string" }, "resume": { "description": "Optional agent ID to resume instead of creating a new instance. When set, do not also pass subagent_type — the resumed agent keeps its own type, and supplying both is rejected.", "type": "string" }, "run_in_background": { "description": "If true, return immediately without waiting for completion. Prefer false unless the task can run independently and there is a clear benefit to not waiting.", "type": "boolean" } }, "required": [ "prompt", "description" ], "additionalProperties": false } }, { "name": "AgentDynamicWorkflow", "description": "Launch multiple subagents from one prompt template, existing agent resumes, or both.\\n\\nUse AgentDynamicWorkflow when many subagents should run the same kind of task over different inputs. The placeholder is exactly \`{{item}}\`. For example, with \`prompt_template\` set to \`Review {{item}} for likely regressions.\` and \`items\` set to \`[\\"src/a.ts\\", \\"src/b.ts\\"]\`, AgentDynamicWorkflow launches two new subagents with those two concrete prompts. For a few differently-shaped tasks, make separate \`Agent\` calls in one message instead.\\n\\nUse \`resume_agent_ids\` to continue subagents that already exist from earlier work, such as ones that failed or timed out: map each agent id to the prompt for that resumed subagent (usually \`continue\` if no extra information is needed). You may combine \`resume_agent_ids\` with \`items\` in the same call to resume existing subagents and launch new ones. Do not duplicate resumed work in \`items\`.\\n\\nEach of these is enforced — a violation is rejected before any subagent starts: provide at least 2 \`items\` unless you pass \`resume_agent_ids\`; whenever \`items\` are present, \`prompt_template\` is required and must contain \`{{item}}\`; and the filled-in prompts must be distinct (two items that expand to the same prompt are rejected).\\n\\nUse enough subagents to keep the work focused and parallel. AgentDynamicWorkflow supports up to 128 subagents, and launches are queued automatically, so it is safe to split large tasks into many clear, independent items.\\n\\nIf \`AgentDynamicWorkflow\` is called, that call must be the only tool call in the response.\\n\\nThis capability is called a \\"dynamic workflow\\" (or simply \\"workflow\\"). Never use the word \\"swarm\\" in the \`description\` field, in subagent prompts, or when talking to the user about this tool.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "description": { "type": "string", "minLength": 1, "description": "Short description for the whole dynamic workflow. It is shown to the user as the workflow and subagent title, so word it as a workflow or run and never use the word \\"swarm\\"." }, "subagent_type": { "description": "Subagent type used for every new subagent spawned from items; defaults to coder when omitted. Resumed subagents always keep their original type, so passing subagent_type together with resume_agent_ids is allowed — it only affects the item-based spawns.", "type": "string", "minLength": 1 }, "prompt_template": { "description": "Prompt template for each subagent. The {{item}} placeholder is replaced with each item value.", "type": "string", "minLength": 1 }, "items": { "description": "Values used to fill {{item}}. Each item launches one new subagent.", "maxItems": 128, "type": "array", "items": { "type": "string", "minLength": 1 } }, "resume_agent_ids": { "description": "Map of existing subagent agent_id to the prompt used to resume that subagent. These resumed subagents are launched before new item-based subagents.", "type": "object", "propertyNames": { "type": "string", "minLength": 1 }, "additionalProperties": { "type": "string", "minLength": 1 } } }, "required": [ "description" ], "additionalProperties": false } }, { "name": "AskUserQuestion", "description": "Use this tool when you need to ask the user questions with structured options during execution. This allows you to:\\n1. Collect user preferences or requirements before proceeding\\n2. Resolve ambiguous or underspecified instructions\\n3. Let the user decide between implementation approaches as you work\\n4. Present concrete options when multiple valid directions exist\\n\\n**When NOT to use:**\\n- When you can infer the answer from context — be decisive and proceed\\n- Trivial decisions that don't materially affect the outcome\\n\\nOverusing this tool interrupts the user's flow. Only use it when the user's input genuinely changes your next action.\\n\\n**Usage notes:**\\n- Users always have an \\"Other\\" option for custom input — don't create one yourself\\n- Use multi_select to allow multiple answers to be selected for a question\\n- Keep option labels concise (1-5 words), use descriptions for trade-offs and details\\n- Each question should have 2-4 meaningful, distinct options\\n- Question texts must be unique across the call, and option labels must be unique within each question\\n- You can ask 1-4 questions at a time; group related questions to minimize interruptions\\n- If you recommend a specific option, list it first and append \\"(Recommended)\\" to its label\\n- The result is JSON with an \`answers\` object keyed by question text; each value is the chosen option's label (comma-separated labels for multi_select, or the user's own words if they picked \\"Other\\"); if \`answers\` is empty and a \`note\` says the user dismissed it, they chose not to answer — do not treat this as selecting the recommended option; decide based on context and do not re-ask the same question\\n- Set background=true when you can keep working without the answer. This starts a background question task and returns a task_id immediately. The answer arrives automatically in a later turn — you do not need to poll, sleep, or check on it. Continue with other work; never fabricate or predict the answer.", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "questions": { "minItems": 1, "maxItems": 4, "type": "array", "items": { "type": "object", "properties": { "question": { "type": "string", "minLength": 1, "description": "A specific, actionable question. End with '?'." }, "header": { "default": "", "description": "Short category tag (max 12 chars, e.g. 'Auth', 'Style').", "type": "string" }, "options": { "minItems": 2, "maxItems": 4, "type": "array", "items": { "type": "object", "properties": { "label": { "type": "string", "minLength": 1, "description": "Concise display text (1-5 words). If recommended, append '(Recommended)'." }, "description": { "default": "", "description": "Brief explanation of trade-offs or implications.", "type": "string" } }, "required": [ "label" ], "additionalProperties": false }, "description": "2-4 meaningful, distinct options. Do NOT include an 'Other' option — the system adds one automatically." }, "multi_select": { "default": false, "description": "Whether the user can select multiple options.", "type": "boolean" } }, "required": [ "question", "options" ], "additionalProperties": false }, "description": "The questions to ask the user (1-4 questions)." }, "background": { "default": false, "description": "Set true to ask in the background and return immediately with a background task_id; you are notified automatically when the user answers — do not poll with TaskOutput while the question is pending.", "type": "boolean" } }, "required": [ "questions" ], "additionalProperties": false } }, { "name": "Bash", "description": "Execute a \`bash\` command. Use this for shell semantics — pipes, env, processes, git, package managers, build/test runners, anything genuinely interactive or multi-step.\\n\\n**Translate these to a dedicated tool instead:**\\n- \`cat\` / \`head\` / \`tail\` (known path) → \`Read\`\\n- \`sed\` / \`awk\` (in-place edit) → \`Edit\`\\n- \`echo > file\` / \`cat <<EOF\` → \`Write\`\\n- \`find\` / recursive \`ls\` to locate files by name pattern → \`Glob\` (plain \`ls <known-directory>\` is fine for listing a directory)\\n- \`grep\` / \`rg\` (search file contents) → \`Grep\`\\n- \`echo\` / \`printf\` (talk to the user) → just output text directly\\n\\nThe dedicated tools render in the per-tool permission UI and keep raw stdout out of the conversation; that is why they are worth reaching for whenever one fits.\\n\\n**Output:**\\nThe stdout and stderr will be combined and returned as a string. The output may be truncated if it is too long. If the command exits non-zero, the output ends with a \`Command failed with exit code: N\` line; a command killed by its timeout or interrupted by the user ends with its own message instead.\\n\\nIf \`run_in_background=true\`, the command will be started as a background task and this tool will return a task ID instead of waiting for command completion. When doing that, you must provide a short \`description\`. Background commands default to a 600s timeout and \`timeout\` is capped at 86400s; set \`disable_timeout=true\` only when the task should run without a timeout. You will be automatically notified when the task completes. After starting one, default to returning control to the user instead of immediately waiting on it. Use \`TaskOutput\` only for a non-blocking status/output snapshot — do not wait on a task you just launched, since its completion arrives automatically. Use \`TaskStop\` only if the task must be cancelled. If a human user wants to inspect background tasks themselves, point them to the background-task panel.\\n\\n**Guidelines for safety and security:**\\n- Each shell tool call will be executed in a fresh shell environment. The shell variables, current working directory changes, and the shell history is not preserved between calls. To run a command in a particular directory, pass the \`cwd\` argument (or use absolute paths) rather than relying on a \`cd\` from an earlier call.\\n- The tool call will return after the command is finished. You shall not use this tool to execute an interactive command or a command that may run forever. For possibly long-running foreground commands, set the \`timeout\` argument in seconds. Foreground commands default to 60s and allow up to 300s. When a foreground command hits its timeout it is moved to the background instead of being killed, and you will be automatically notified when it completes. The user can also move a running foreground command to the background at any time.\\n- Avoid using \`..\` to access files or directories outside of the working directory.\\n- Avoid modifying files outside of the working directory unless explicitly instructed to do so.\\n- Never run commands that require superuser privileges unless explicitly instructed to do so.\\n\\n**Guidelines for efficiency:**\\n- Use \`&&\` to chain commands that genuinely depend on each other, e.g. \`npm install && npm test\`. Independent read-only commands (separate \`git show\`, \`ls\`, or status checks) should be issued as separate parallel Bash calls in one response, not chained into a single call — chaining serializes their execution and mixes their output. Do not stitch outputs together with \`echo\` separators.\\n- Use \`;\` to run commands sequentially regardless of success/failure\\n- Use \`||\` for conditional execution (run second command only if first fails)\\n- Use pipe operations (\`|\`) and redirections (\`>\`, \`>>\`) to chain input and output between commands\\n- Always quote file paths containing spaces with double quotes (e.g., cd \\"/path with spaces/\\")\\n- Compose multi-step logic in a single call with \`if\` / \`case\` / \`for\` / \`while\` control flows.\\n- Prefer \`run_in_background=true\` for long-running builds, tests, watchers, or servers when you need the conversation to continue before the command finishes.\\n\\n**Commands available:**\\nThe following common command categories are usually available. Availability still depends on the host, so when in doubt run \`which <command>\` first to confirm a command exists before relying on it.\\n- Navigation and inspection: \`ls\`, \`pwd\`, \`cd\`, \`stat\`, \`file\`, \`du\`, \`df\`, \`tree\`\\n- File and directory management: \`cp\`, \`mv\`, \`rm\`, \`mkdir\`, \`touch\`, \`ln\`, \`chmod\`, \`chown\`\\n- Text and data processing: \`wc\`, \`sort\`, \`uniq\`, \`cut\`, \`tr\`, \`diff\`, \`xargs\`\\n- Archives and compression: \`tar\`, \`gzip\`, \`gunzip\`, \`zip\`, \`unzip\`\\n- Networking and transfer: \`curl\`, \`wget\`, \`ping\`, \`ssh\`, \`scp\`\\n- Version control: \`git\`; for GitHub-hosted work (PRs, issues, CI runs, API queries) prefer the \`gh\` CLI when installed — it carries the user's GitHub auth and can return structured JSON\\n- Process and system: \`ps\`, \`kill\`, \`top\`, \`env\`, \`date\`, \`uname\`, \`whoami\`\\n- Language and package toolchains: \`node\`, \`npm\`, \`pnpm\`, \`yarn\`, \`python\`, \`pip\` (use whichever the project actually relies on)\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "command": { "type": "string", "minLength": 1, "description": "The command to execute." }, "cwd": { "description": "The working directory in which to run the command. When omitted, the command runs in the session's working directory.", "type": "string" }, "timeout": { "default": 60, "description": "Optional timeout in seconds for the command to execute. Foreground default 60s, max 300s. Background default 600s, max 86400s. Ignored for background commands when disable_timeout=true.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 }, "description": { "description": "A short description for the background task. Required when run_in_background is true.", "type": "string" }, "run_in_background": { "description": "Whether to run the command as a background task.", "type": "boolean" }, "disable_timeout": { "description": "If true, do not apply a timeout to the command. Only applies when run_in_background is true.", "type": "boolean" } }, "required": [ "command" ], "additionalProperties": false } }, { "name": "CreateGoal", "description": "Create a durable, structured goal that the runtime will pursue across multiple turns.\\n\\nCall \`CreateGoal\` only when:\\n\\n- the user explicitly asks you to start a goal or work autonomously toward an outcome, or\\n- a host goal-intake prompt asks you to create one.\\n\\nDo NOT create a goal for greetings, ordinary questions, or vague requests that lack a\\nverifiable completion condition. A goal needs a checkable end state.\\n\\nWhen the request is vague, ask the user for the missing completion criterion before creating\\nthe goal. If the user clearly insists after you warn them that the wording is vague or risky,\\nrespect that and create the goal.\\n\\nInclude a \`completionCriterion\` when the user provides one, or when it can be stated without\\ninventing new requirements. Keep \`objective\` concise; reference long task descriptions by file\\npath rather than pasting them.\\n\\nCreating a goal fails if one already exists, so use \`replace: true\` only when the user explicitly\\nwants to abandon the current goal and start a new one.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "objective": { "type": "string", "minLength": 1, "description": "The objective to pursue. Must have a verifiable end state." }, "completionCriterion": { "description": "How to verify the goal is complete. Include when the user provides one.", "type": "string" }, "replace": { "description": "Replace an existing active, paused, or blocked goal instead of failing.", "type": "boolean" } }, "required": [ "objective" ], "additionalProperties": false } }, { "name": "CronCreate", "description": "Schedule a prompt to be enqueued at a future time. Use for both recurring schedules and one-shot reminders.\\n\\nUses standard 5-field cron in the user's local timezone: minute hour day-of-month month day-of-week. \`0 9 * * *\` means 9am local — no timezone conversion needed.\\n\\n## One-shot tasks (recurring: false)\\n\\nFor \\"remind me at X\\" or \\"at <time>, do Y\\" requests — fire once then auto-delete.\\nPin minute/hour/day-of-month/month to specific values:\\n  \\"remind me at 2:30pm today to check the deploy\\" → cron: \\"30 14 <today_dom> <today_month> *\\", recurring: false\\n  \\"tomorrow morning, run the smoke test\\" → cron: \\"57 8 <tomorrow_dom> <tomorrow_month> *\\", recurring: false\\n\\nOne-shots are best for near-term reminders. A task only fires while its session is still alive (see Session lifetime below), so favor near times — within hours or a few days — rather than scheduling weeks or months ahead.\\n\\n## Recurring jobs (recurring: true, the default)\\n\\nFor \\"every N minutes\\" / \\"every hour\\" / \\"weekdays at 9am\\" requests:\\n  \\"*/5 * * * *\\" (every 5 min), \\"0 * * * *\\" (hourly), \\"0 9 * * 1-5\\" (weekdays at 9am local)\\n\\n## Avoid the :00 and :30 minute marks when the task allows it\\n\\nEvery user who asks for \\"9am\\" gets \`0 9\`, and every user who asks for \\"hourly\\" gets \`0 *\` — which means requests from across the planet land on the API at the same instant. When the user's request is approximate, pick a minute that is NOT 0 or 30:\\n  \\"every morning around 9\\" → \\"57 8 * * *\\" or \\"3 9 * * *\\" (not \\"0 9 * * *\\")\\n  \\"hourly\\" → \\"7 * * * *\\" (not \\"0 * * * *\\")\\n  \\"in an hour or so, remind me to...\\" → pick whatever minute you land on, don't round\\n\\nOnly use minute 0 or 30 when the user names that exact time and clearly means it (\\"at 9:00 sharp\\", \\"at half past\\", coordinating with a meeting). When in doubt, nudge a few minutes early or late — the user will not notice, and the fleet will.\\n\\n## Coalesce semantics\\n\\nFires are delivered only while the session is idle: a fire that comes due during an active turn is held and delivered at the next idle moment, never injected mid-turn.\\n\\nIf the scheduler slept past multiple ideal fire times (laptop closed, long-running turn, etc.), only **one** fire is delivered when it wakes up. The origin carries \`coalescedCount\` showing how many ideal fires were collapsed into this single delivery. You should treat \`coalescedCount > 1\` as \\"I missed some checks; only the latest state matters\\" rather than running the prompt that many times.\\n\\n## Cron-fire envelope\\n\\nWhen a cron task fires, the prompt you scheduled is re-injected wrapped in an XML envelope that exposes the fire context:\\n\\n\`\`\`\\n<cron-fire jobId=\\"...\\" cron=\\"...\\" recurring=\\"true|false\\" coalescedCount=\\"N\\" stale=\\"true|false\\">\\n<prompt>\\nyour original prompt text, verbatim\\n</prompt>\\n</cron-fire>\\n\`\`\`\\n\\nThe envelope is parseable. Use \`coalescedCount > 1\` to know multiple ideal fires were collapsed into a single delivery (treat as \\"only the latest state matters\\"), and \`stale=\\"true\\"\` as a cue that the task is past its 7-day threshold.\\n\\n## 7-day stale behavior\\n\\nRecurring tasks that have been alive for more than 7 days fire one\\nfinal time with \`stale: true\` on the envelope, and the system then\\nauto-deletes the task. The flag is the model's notice that this is\\nthe last delivery. If the schedule is still wanted, call \`CronCreate\`\\nagain with the same \`cron\` and \`prompt\` — that resets \`createdAt\` and\\nstarts a fresh 7-day window. One-shot tasks are never marked stale.\\n\\n## Jitter behavior\\n\\nAnti-herd jitter is applied deterministically per task id:\\n  - Recurring: ideal fire time is shifted **forward** by an offset ≤ min(10% of the cron period, 15 minutes). A \`*/5 * * * *\` task can drift up to 30s; a \`0 9 * * *\` task can drift up to 15 minutes.\\n  - One-shot: only when the ideal fire lands on \`:00\` or \`:30\` of the hour, the fire is pulled **earlier** by ≤ 90 seconds. Other minutes pass through unchanged.\\n\\n## One-shot vs recurring — when to pick which\\n\\nUse \`recurring: false\` for \\"remind me at X\\" style requests, single deadlines, \\"in N minutes do Y\\", and any task that should not repeat. Use \`recurring: true\` for periodic polling (CI status, build watchers, scheduled reports), workday rituals, and anything the user explicitly described as recurring.\\n\\n## Session lifetime\\n\\nCron tasks live in the current session. When you exit, they\\nare persisted under the session homedir; resuming the same session\\nreloads them and the scheduler resumes from each task's \`createdAt\`. Fire times that fell during the offline window are\\ncollapsed into a single delivery via \`coalescedCount\` (and recurring\\ntasks past their 7-day window arrive with \`stale: true\` as their final\\ndelivery).\\n\\nTasks do **not** carry over into a brand-new session — they are scoped\\nto the resumed session id, not to the working directory.\\n\\n## Limits\\n\\nA session holds at most 50 live cron tasks; creating one beyond that is rejected. (The \`prompt\` body is also capped — see its parameter description.) Expressions that never fire within the next 5 years (e.g. \`0 0 31 2 *\`, an impossible date) are rejected at create time.\\n\\n## Returned fields\\n\\n\`id\` (ULID), \`cron\` (the normalized expression), \`humanSchedule\` (English summary), \`recurring\`,\\n\`nextFireAt\` (local ISO timestamp with numeric offset, or null). \`id\` is needed by \`CronDelete\`.\\n\\n## Tell the user how to cancel or modify\\n\\nAfter successfully creating a task, proactively tell the user how they can cancel or modify it later. Users have no direct \`/cron\` command or self-service UI to manage reminders themselves; they must ask the model to make changes (e.g. \\"cancel my 9am reminder\\" or \\"change my daily check to 10am\\"). Include the task \`id\` in your message so the user can reference it.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "cron": { "type": "string", "description": "5-field cron expression in local time: \\"M H DoM Mon DoW\\" (e.g. \\"*/5 * * * *\\" = every 5 minutes; \\"30 14 28 2 *\\" = Feb 28 at 2:30pm local — a pinned date like this repeats yearly unless you also pass recurring: false)." }, "prompt": { "type": "string", "minLength": 1, "maxLength": 8192, "description": "The prompt to enqueue at each fire time. Limited to 8 KiB (UTF-8)." }, "recurring": { "default": true, "description": "true (default) = fire on every cron match until deleted or auto-expired after 7 days. false = fire once at the next match, then auto-delete. Use false for \\"remind me at X\\" one-shot requests with pinned minute/hour/dom/month.", "type": "boolean" } }, "required": [ "cron", "prompt" ], "additionalProperties": false } }, { "name": "CronDelete", "description": "Cancel a scheduled cron job by id.\\n\\nUse this tool to remove a cron task previously scheduled with\\n\`CronCreate\`. The \`id\` is the ULID value returned by \`CronCreate\`, or\\nshown in the \`id:\` column of \`CronList\` — quote it verbatim, no\\nprefix.\\n\\nBehaviour by task kind:\\n\\n- **Recurring task** (\`recurring: true\`): stops all future fires\\n  immediately. The scheduler picks up the deletion on its next tick.\\n- **One-shot task** (\`recurring: false\`): cancels the pending fire if\\n  it has not happened yet. One-shots that have already fired\\n  auto-delete themselves, so calling \`CronDelete\` on a fired one-shot\\n  returns \\"no cron job with id ...\\".\\n\\nNot-found is reported as an error (not a silent no-op) so you can\\ncorrect yourself — typically by calling \`CronList\` to see which ids\\nare actually live, rather than re-trying with the same stale id.\\n\\nRefresh pattern (use when you want a stale recurring schedule to\\ncontinue):\\n\\nStale recurring tasks are auto-deleted by the system after their final\\nfire — there is nothing for \`CronDelete\` to remove at that point. To\\nkeep the schedule running, just call \`CronCreate\` with the same \`cron\`\\nand \`prompt\`. Use \`CronList\`'s \`prompt\` field to recall the original\\ntext after a context compaction.\\n\\n\`CronDelete\` remains the right call when you want to cancel a task\\nthat is still live (recurring not yet stale, or a one-shot still\\npending).\\n\\nGuidelines:\\n\\n- Users have no direct \`/cron\` command or self-service UI to delete\\n  tasks themselves; they must ask the model to cancel a reminder.\\n  When deleting on behalf of a user, confirm the action and report\\n  the result plainly.\\n- Cron deletion is irreversible — there is no undo. If you delete the\\n  wrong task, you must re-create it with \`CronCreate\`.\\n- If the model is unsure which id is current (e.g. after a context\\n  compaction), call \`CronList\` first rather than guessing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "id": { "type": "string", "description": "The cron job id (ULID) returned by CronCreate / CronList." } }, "required": [ "id" ], "additionalProperties": false } }, { "name": "CronList", "description": "List all cron jobs currently scheduled in this session.\\n\\nUse this tool to see every pending cron task — both recurring jobs and\\none-shot reminders — that you (or the user) have scheduled with\\n\`CronCreate\`. The output is the entry point for inspecting scheduled\\nwork: it returns a stable id, the original cron expression, a human\\nrendering, the next post-jitter fire time, the recurring flag, the\\ntask's age in days, and a stale indicator.\\n\\nEach record carries:\\n\\n- \`id\` — the task id (a ULID). Pass this to \`CronDelete\` to remove the\\n  task, or quote it in user-facing messages when asking for\\n  confirmation.\\n- \`cron\` — the verbatim 5-field cron expression as scheduled.\\n- \`humanSchedule\` — plain-English rendering (e.g. \`every 5 minutes\`).\\n- \`prompt\` — the scheduled prompt text, JSON-encoded so embedded\\n  newlines stay on one line. Truncated to 200 UTF-8 bytes with\\n  \`…(truncated)\` if longer. Use this to recall what a task is for\\n  after a context compaction, and as the source for the\\n  \`CronCreate\` refresh ritual.\\n- \`nextFireAt\` — local ISO timestamp with an explicit numeric offset\\n  for the next fire **after jitter has been applied**. The actual fire\\n  may land slightly before or after a round \`:00\` / \`:30\` minute mark\\n  due to herd-avoidance jitter; this is the value the scheduler will\\n  compare against, so it reflects what will really happen. \`null\` if\\n  the expression has no fire in the next 5 years (should not happen\\n  for tasks created through \`CronCreate\`, which validates).\\n- \`recurring\` — \`true\` for cadenced jobs, \`false\` for one-shots.\\n- \`ageDays\` — \`(now - createdAt) / day\`, two decimal places. Useful\\n  when deciding whether a long-running cron is still relevant.\\n- \`stale\` — \`true\` when a recurring task is older than 7 days. The\\n  system **auto-deletes the task after this fire** to bound session\\n  lifetime; the \`stale: true\` flag is the model's notice that this is\\n  the final delivery. To resume the same schedule, call \`CronCreate\`\\n  again with the original \`cron\` and \`prompt\` (the \`prompt\` row above\\n  carries it for exactly this purpose). One-shots are never marked\\n  stale — they fire at most once by construction.\\n\\nGuidelines:\\n\\n- This tool is read-only and never mutates state, so it is always\\n  safe to call (including in plan mode).\\n- Users cannot directly manage cron tasks themselves; if they want to\\n  cancel or modify a schedule, route the request through the model\\n  (i.e. call \`CronDelete\` or \`CronCreate\` on their behalf).\\n- The empty case returns \`cron_jobs: 0\\\\nNo cron jobs scheduled.\`. Cron\\n  tasks survive a resume of the same session but do not bleed into new\\n  sessions.\\n- After a context compaction, or whenever you are unsure which cron\\n  jobs are live, call this tool to re-enumerate them rather than\\n  guessing ids from earlier in the conversation.\\n- Records are separated by a line containing just \`---\`, in the\\n  insertion order they were scheduled.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "Edit", "description": "Perform exact replacements in existing files.\\n\\n- Edit is mandatory for every incremental change, especially small edits. DO NOT use Write or Bash \`sed\`.\\n- Read the target file before every Edit. DO NOT call Edit from memory, stale context, or a guessed \`old_string\`.\\n- Take \`old_string\` and \`new_string\` from the Read output view.\\n- Drop the line-number prefix and tab; match only file content.\\n- \`old_string\` must be unique unless \`replace_all\` is set.\\n- If \`old_string\` is ambiguous, add surrounding context. Use \`replace_all\` only when every occurrence should change — for example, renaming a symbol throughout the file.\\n- Multiple Edit calls may run in one response only when they do not target the same file.\\n- DO NOT issue consecutive Edit calls on the same file. A previous Edit can invalidate a later Edit's \`old_string\`, causing \`old_string not found\`. Read the file again before the next Edit.\\n- A write lock serializes same-file edits in response order, but serialization does not make stale \`old_string\` valid.\\n- For pure CRLF files, Read shows LF; use LF in \`old_string\` and \`new_string\`, and Edit writes CRLF back.\\n- For mixed endings or lone carriage returns, Read shows carriage returns as \\\\r; include actual \\\\r escapes in those positions.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to the text file to edit. Relative paths resolve against the working directory; a path outside the working directory must be absolute." }, "old_string": { "type": "string", "minLength": 1, "description": "Exact content to replace from the Read output view, without the line-number prefix. Use LF for pure CRLF files; use actual \\\\r escapes where Read shows \\\\r." }, "new_string": { "type": "string", "description": "Replacement text in the same Read output view. LF is written back as CRLF only for pure CRLF files." }, "replace_all": { "description": "Set true only when every occurrence of old_string should be replaced.", "type": "boolean" } }, "required": [ "path", "old_string", "new_string" ], "additionalProperties": false } }, { "name": "EnterPlanMode", "description": "Use this tool proactively when you're about to start a non-trivial implementation task.\\nGetting user sign-off on your approach via ExitPlanMode before writing code prevents wasted effort.\\n\\nUse it when ANY of these conditions apply:\\n\\n1. New Feature Implementation - e.g. \\"Add a caching layer to the API\\"\\n2. Multiple Valid Approaches - e.g. \\"Optimize database queries\\" (indexing vs rewrite vs caching)\\n3. Code Modifications - e.g. \\"Refactor auth module to support OAuth\\"\\n4. Architectural Decisions - e.g. \\"Add WebSocket support\\"\\n5. Multi-File Changes - involves more than 2-3 files\\n6. Unclear Requirements - need exploration to understand scope\\n7. User Preferences Matter - if user input would materially change the implementation approach, use EnterPlanMode to structure the decision\\n\\nPermission mode notes:\\n- EnterPlanMode enters plan mode automatically without an approval prompt in all permission modes.\\n- In yolo and manual modes, ExitPlanMode still presents the plan to the user for approval.\\n- In auto permission mode, do not use AskUserQuestion; make the best decision from available context.\\n- In auto permission mode, ExitPlanMode exits plan mode without asking the user.\\n- Use EnterPlanMode only when planning itself adds value.\\n\\nWhen NOT to use:\\n- Single-line or few-line fixes (typos, obvious bugs, small tweaks)\\n- User gave very specific, detailed instructions\\n- Pure research/exploration tasks\\n\\nOnce you are in plan mode, a reminder walks you through the workflow (explore → design → write the plan file → \`ExitPlanMode\`) and enforces read-only access. For non-trivial tasks where you are unsure of the codebase structure or relevant code paths, use \`Agent(subagent_type=\\"explore\\")\` to investigate first when the \`Agent\` tool is available.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "ExitPlanMode", "description": "Use this tool when you are in plan mode and have finished writing your plan to the plan file and are ready for user approval.\\n\\n## How This Tool Works\\n- You should have already written your plan to the plan file specified in the plan mode reminder.\\n- This tool does NOT take the plan content as a parameter - it reads the plan from the file you wrote.\\n- The user will see the contents of your plan file when they review it. In auto permission mode, the tool reads the file and exits plan mode without asking the user.\\n\\n## When to Use\\nOnly use this tool for tasks that require planning implementation steps. For research tasks (searching files, reading code, understanding the codebase), do NOT use this tool.\\n\\n## What a good plan contains\\nList specific, verifiable steps grounded in the actual codebase — real files, functions, and commands, in a sensible order. Each step should be concrete enough to act on and to check. Avoid vague filler like \\"improve performance\\" or \\"add tests\\"; say what to change and where.\\n\\n## Multiple Approaches\\nIf your plan offers multiple alternative approaches, pass them via the \`options\` parameter so the user can choose which one to execute — see the \`options\` parameter for the format, count, and reserved labels. In yolo and manual modes the user sees all options alongside the host's Reject and Revise controls.\\n\\n## Before Using\\n- In auto permission mode, do NOT use AskUserQuestion; make the best decision from available context.\\n- In auto permission mode, this tool exits plan mode without asking the user.\\n- In yolo and manual modes, this tool still presents the plan to the user for approval.\\n- If auto permission mode is not active and you have unresolved questions, use AskUserQuestion first.\\n- If auto permission mode is not active and you have multiple approaches and haven't narrowed down yet, consider using AskUserQuestion first to let the user choose, then write a plan for the chosen approach only.\\n- Once your plan is finalized, use THIS tool to request approval.\\n- Do NOT use AskUserQuestion to ask \\"Is this plan OK?\\" or \\"Should I proceed?\\" - that is exactly what ExitPlanMode does.\\n- If rejected, revise based on feedback and call ExitPlanMode again.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "options": { "description": "When the plan contains multiple alternative approaches, list them here so the user can choose which one to execute. Provide up to 3 options; 2-3 distinct approaches work best when the plan offers a real choice. Passing a single option is allowed and is equivalent to a plain plan approval. Each option represents a distinct approach from the plan. Do not use \\"Reject\\", \\"Revise\\", \\"Approve\\", or \\"Reject and Exit\\" as labels.", "minItems": 1, "maxItems": 3, "type": "array", "items": { "type": "object", "properties": { "label": { "type": "string", "minLength": 1, "maxLength": 80, "description": "Short name for this option (1-8 words). Append \\"(Recommended)\\" if you recommend this option." }, "description": { "default": "", "description": "Brief summary of this approach and its trade-offs.", "type": "string" } }, "required": [ "label" ], "additionalProperties": false } } }, "additionalProperties": false } }, { "name": "FetchURL", "description": "Fetch content from a URL. The content is returned either as the main text extracted from the page, or as the full response body verbatim; a note at the top of the result states which of the two you received, so you can judge how complete it is. Use this when you need to read a specific web page.\\n\\nOnly fully-formed public \`http\`/\`https\` URLs are supported; other schemes and private or loopback addresses are not fetched. Very large pages may be truncated or refused. The fetch carries no login or session for the target site, so pages behind authentication (private repositories, internal dashboards) return a login page or an error instead of the real content — if the text you get back looks like a generic landing or sign-in page, treat that as the login wall, not the answer, and reach the content through a credentialed route (an authenticated CLI or MCP tool) instead.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "url": { "type": "string", "description": "The URL to fetch content from." } }, "required": [ "url" ], "additionalProperties": false } }, { "name": "GetGoal", "description": "Read the current goal: its objective, completion criterion, status, and budgets (turns, tokens,\\ntime, and how much of each remains). When the goal has stopped, it also reports the terminal reason.\\n\\nUse \`GetGoal\` before deciding whether to continue working, report completion, report a blocker,\\nor respect a pause. It returns \`{ \\"goal\\": null }\` when there is no current goal.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "Glob", "description": "Find files by glob pattern, sorted by modification time (most recent first).\\n\\nPowered by ripgrep. Respects \`.gitignore\`, \`.ignore\`, and \`.rgignore\` by default — set \`include_ignored\` to also match ignored files (e.g. build outputs, \`node_modules\`). Sensitive files (such as \`.env\`) are always filtered out. Matches are files only — directories themselves are never listed; to find a directory, glob for a file inside it (e.g. \`**/fixtures/**\`).\\n\\nGood patterns:\\n- \`*.ts\` — all files matching an extension, at any depth below the search root (a bare pattern without \`/\` matches recursively)\\n- \`src/*.ts\` — files directly inside \`src/\` (one level, not recursive)\\n- \`src/**/*.ts\` — recursive walk with a subdirectory anchor and extension\\n- \`**/*.py\` — recursive walk from the search root for an extension\\n- \`*.{ts,tsx}\` — brace expansion is supported\\n- \`{src,test}/**/*.ts\` — cartesian brace expansion is supported too\\n\\nResults are capped at the first 100 matching paths. If a search would return more, a truncation marker is appended. Refine the pattern (extension, subdirectory) when 100 is not enough, or call again with a narrower anchor.\\n\\nLarge-directory caveat — avoid recursing into dependency / build output even with an anchor, especially when \`include_ignored\` is set:\\n- \`node_modules/**/*.js\`, \`.venv/**/*.py\`, \`__pycache__/**\`, \`target/**\` can produce thousands of results that truncate at the match cap and waste context. Prefer specific subpaths like \`node_modules/react/src/**/*.js\`.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "pattern": { "type": "string", "description": "Glob pattern to match files." }, "path": { "description": "Directory to search. Accepts an absolute path, or a path relative to the current working directory. Defaults to the current working directory.", "type": "string" }, "include_ignored": { "description": "Also match files excluded by ignore files such as \`.gitignore\`, \`.ignore\`, and \`.rgignore\` (for example \`node_modules\` or build outputs). Sensitive files (such as \`.env\`) remain filtered out for safety. VCS metadata directories (\`.git\` and similar) are always skipped, even when this is true. Defaults to false.", "type": "boolean" }, "include_dirs": { "description": "Deprecated and ignored. Results are always files-only — directories are never listed. Accepted only so older calls that still pass this flag are not rejected by parameter validation.", "type": "boolean" } }, "required": [ "pattern" ], "additionalProperties": false } }, { "name": "Grep", "description": "Search file contents using regular expressions (powered by ripgrep).\\n\\nUse Grep when the task is to find unknown content or unknown file locations. Do not use shell \`grep\` or \`rg\` directly; this tool applies workspace path policy, output limits, and sensitive-file filtering.\\nALWAYS use Grep tool instead of running \`grep\` or \`rg\` from a shell — direct shell calls bypass workspace policy, output limits, and sensitive-file filtering.\\nIf you already know a concrete file path and need to inspect its contents, use Read directly instead.\\n\\nWrite patterns in ripgrep regex syntax, which differs from POSIX \`grep\` syntax. For example, braces are special, so escape them as \`\\\\{\` to match a literal \`{\`.\\n\\nHidden files (dotfiles such as \`.gitlab-ci.yml\` or \`.eslintrc.json\`) are searched by default. To also search files excluded by \`.gitignore\` (such as \`node_modules\` or build outputs), set \`include_ignored\` to \`true\`. Sensitive files (such as \`.env\`) are always skipped for safety, even when \`include_ignored\` is \`true\`.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "pattern": { "type": "string", "description": "Regular expression to search for." }, "path": { "description": "File or directory to search. Accepts an absolute path, or a path relative to the current working directory. Omit to search the current working directory. Use Read instead when you already know a concrete file path and need its contents.", "type": "string" }, "glob": { "description": "Optional glob filter for which files to search, e.g. \`*.ts\`. Matched against each file's full absolute path, so a path-anchored pattern like \`src/**/*.ts\` silently matches nothing — use a basename pattern (\`*.ts\`), or anchor with \`**/\` (\`**/src/**/*.ts\`). To scope the search to a directory, use \`path\` instead.", "type": "string" }, "type": { "description": "Optional ripgrep file type filter, such as ts or py. Prefer this over \`glob\` when filtering by language or file kind: it is more efficient and less error-prone than an equivalent glob pattern.", "type": "string" }, "output_mode": { "description": "Shape of the result. \`content\` shows matching lines (honors \`-A\`, \`-B\`, \`-C\`, \`-n\`, and \`head_limit\`); \`files_with_matches\` shows only the paths of files that contain a match, most-recently-modified first (honors \`head_limit\`); \`count_matches\` shows per-file match counts as \`path:count\` lines, preceded by an aggregate total line. Defaults to \`files_with_matches\`.", "type": "string", "enum": [ "content", "files_with_matches", "count_matches" ] }, "-i": { "description": "Perform a case-insensitive search. Defaults to false.", "type": "boolean" }, "-n": { "description": "Prefix each matching line with its line number. Applies only when \`output_mode\` is \`content\`. Defaults to true.", "type": "boolean" }, "-A": { "description": "Number of lines to show after each match. Applies only when \`output_mode\` is \`content\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "-B": { "description": "Number of lines to show before each match. Applies only when \`output_mode\` is \`content\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "-C": { "description": "Number of lines to show before and after each match. Applies only when \`output_mode\` is \`content\`; takes precedence over \`-A\` and \`-B\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "head_limit": { "description": "Limit output to the first N lines/entries after offset. Defaults to 250. Pass 0 for unlimited.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "offset": { "description": "Number of leading lines/entries to skip before applying \`head_limit\`. Use it together with \`head_limit\` to page through large result sets. Defaults to 0.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "multiline": { "description": "Enable multiline matching, where the pattern can span line boundaries and \`.\` also matches newlines. Defaults to false.", "type": "boolean" }, "include_ignored": { "description": "Also search files excluded by ignore files such as \`.gitignore\`, \`.ignore\`, and \`.rgignore\` (for example \`node_modules\` or build outputs). Sensitive files (such as \`.env\`) remain filtered out for safety. VCS metadata directories (\`.git\` and similar) are always skipped, even when this is true. Defaults to false.", "type": "boolean" } }, "required": [ "pattern" ], "additionalProperties": false } }, { "name": "Read", "description": "Read a text file from the local filesystem.\\n\\nIf the user provides a concrete file path to a text file, call Read directly. Do not \`Glob\`, \`ls\`, or otherwise pre-check known text file paths; missing or invalid file paths return errors you can handle. Do not use Read for directories; use \`ls\` via Bash for a known directory, or Glob when you need files matching a name pattern (Glob lists files only, never directories). Use \`Grep\` only when the task is to search for unknown content or locations.\\n\\nWhen you need several files, prefer to read them in parallel: emit multiple \`Read\` calls in a single response instead of reading one file per turn.\\n\\n- Relative paths resolve against the working directory; a path outside the working directory must be absolute.\\n- Returns up to 1000 lines or 100 KB per call, whichever comes first; lines longer than 2000 chars are truncated mid-line.\\n- Page larger files with \`line_offset\` (1-based start line) and \`n_lines\`. Omit \`n_lines\` to read up to the 1000-line cap.\\n- Sensitive files (\`.env\` files, credential stores, SSH private keys, and similar secrets) are refused to protect secrets; do not attempt to read them. Templates and public keys are exempt: \`.env.example\` / \`.env.sample\` / \`.env.template\` and public SSH keys such as \`id_rsa.pub\` read normally.\\n- UTF-8 text files are read directly. UTF-16 LE/BE text files (with or without a BOM) are detected automatically and transcoded to UTF-8 for display; the status block notes the detected encoding, and Edit/Write on such a file still expect UTF-8 — convert its encoding first (e.g. with \`iconv\`). Other encodings (e.g. GBK), binary files, and files containing NUL bytes are refused.\\n- Negative line_offset reads from the end of the file (for example, -100 reads the last 100 lines); the absolute value cannot exceed 1000.\\n- Output format: \`<line-number>\\\\t<content>\` per line.\\n- A \`<system>...</system>\` status block is appended after the file content; it summarizes how much was read (line and byte counts, truncation, line-ending notes) and is not part of the file itself.\\n- Pure CRLF files are displayed with LF line endings; \`Edit\` matches this output and preserves CRLF when writing back.\\n- Mixed or lone carriage-return line endings are shown as \`\\\\r\` and require exact \`Edit.old_string\` escapes.\\n- After a successful \`Edit\`/\`Write\`, do not re-read solely to prove the write landed. When the task depends on an exact file, API, or output shape, inspect the final external contract before finishing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to a text file. Relative paths resolve against the working directory; a path outside the working directory must be absolute. Directories are not supported; use \`ls\` via Bash for a known directory, or Glob for pattern search." }, "line_offset": { "description": "The line number to start reading from. Omit to start at line 1. Negative values read from the end of the file; the absolute value cannot exceed 1000.", "anyOf": [ { "type": "integer", "minimum": 1, "maximum": 9007199254740991 }, { "type": "integer", "minimum": -1000, "maximum": -1 } ] }, "n_lines": { "description": "The number of lines to read; the tool also applies its internal cap. Omit to read up to the internal cap of 1000 lines.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 } }, "required": [ "path" ], "additionalProperties": false } }, { "name": "SetGoalBudget", "description": "Set a hard budget limit for the current goal.\\n\\nUse this only when the user clearly gives a runtime limit, such as:\\n\\n- \\"stop after 20 turns\\"\\n- \\"use no more than 500k tokens\\"\\n- \\"finish within 30 minutes\\"\\n\\nDo not invent limits. Do not call this for vague wording such as \\"spend some time\\" or\\n\\"try to be quick\\".\\n\\nIf the user gives a compound time, convert it to one supported unit before calling this tool.\\nFor example, \\"2 hours and 3 minutes\\" can be set as \`value: 123, unit: \\"minutes\\"\`.\\n\\nA time budget must be between 1 second and 24 hours — the tool rejects anything shorter or\\nlonger, telling the user it is not a reasonable goal budget. Turn and token budgets are not\\nbounded this way; they must be positive and are rounded to the nearest whole number (minimum 1).\\n\\nSupported units:\\n\\n- \`turns\`\\n- \`tokens\`\\n- \`milliseconds\`\\n- \`seconds\`\\n- \`minutes\`\\n- \`hours\`\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "value": { "type": "number", "exclusiveMinimum": 0, "description": "The positive numeric budget value." }, "unit": { "type": "string", "enum": [ "turns", "tokens", "milliseconds", "seconds", "minutes", "hours" ] } }, "required": [ "value", "unit" ], "additionalProperties": false } }, { "name": "Skill", "description": "Invoke a registered skill from the current skill listing. BLOCKING REQUIREMENT: when a skill from the listing matches the user's request, you MUST call this tool (not free-form text). Do not re-invoke a skill to repeat work already done: if a \`<skill-loaded>\` block for it with the same \`args\` is already present in the conversation, follow those instructions directly instead of calling the tool again. Do call the tool again when you need the skill with different arguments — the loaded block was expanded with the earlier \`args\` and will not reflect new inputs.", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "skill": { "type": "string", "description": "The exact name of the skill to invoke, spelled as it appears in the current skill listing (e.g. \\"commit\\", \\"pdf\\")." }, "args": { "description": "Optional argument string for the skill, written like a command line (e.g. \`-m \\"fix bug\\"\`, \`123\`, a file path). It is split on whitespace (quotes group a token) and expanded into the skill's placeholders ($NAME, $1, $ARGUMENTS); if the skill body has no placeholders, the whole string is still appended as a trailing \`ARGUMENTS:\` line. Omit it only when there is nothing to pass.", "type": "string" } }, "required": [ "skill" ], "additionalProperties": false } }, { "name": "TaskList", "description": "List background tasks and their current status.\\n\\nUse this tool to discover which background tasks exist and where each one\\nstands. It is the entry point for inspecting background work: it returns a\\ntask ID, status, and description for every task it reports, plus the command,\\nPID, and (once finished) exit code for shell tasks, and a stop reason for any\\ntask that ended early.\\n\\nGuidelines:\\n\\n- After a context compaction, or whenever you are unsure which background\\n  tasks are running or what their task IDs are, call this tool to\\n  re-enumerate them instead of guessing a task ID.\\n- Prefer the default \`active_only=true\`, which lists only non-terminal tasks.\\n  Pass \`active_only=false\` only when you specifically need to see tasks that\\n  have already finished. With \`active_only=false\` the result may also include\\n  \`lost\` tasks — tasks left over from a previous process that can no longer be\\n  inspected or controlled; treat them as already terminated.\\n- \`limit\` caps how many tasks are returned. It accepts a value between 1 and\\n  100 and defaults to 20 when omitted.\\n- This tool only lists tasks; it does not return their output. Use it first\\n  to locate the task ID you need, then call \`TaskOutput\` with that ID to read\\n  the task's output and details.\\n- This tool is read-only and does not change any state, so it is always safe\\n  to call, including in plan mode.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "active_only": { "default": true, "description": "Whether to list only non-terminal background tasks.", "type": "boolean" }, "limit": { "default": 20, "description": "Maximum number of tasks to return.", "type": "integer", "minimum": 1, "maximum": 100 } }, "additionalProperties": false } }, { "name": "TaskOutput", "description": "Retrieve a snapshot of a running or completed background task.\\n\\nUse this after \`Bash(run_in_background=true)\`, \`Agent(run_in_background=true)\`, or \`AskUserQuestion(background=true)\` to check progress, or to read the output of a task that has already completed.\\n\\nGuidelines:\\n- Prefer relying on automatic completion notifications. Use this tool only when you need task output before the automatic notification arrives.\\n- This tool is always non-blocking: it returns the current status/output snapshot immediately and never waits for the task to finish.\\n- Do not use TaskOutput to wait for a result you need before continuing — if your next step depends on the task's result, run that task in the foreground instead. TaskOutput is for a deliberate progress check you will act on without blocking, not a way to sit and wait for a background task you just launched.\\n- This tool returns structured task metadata, a fixed-size output preview, and an output_path for the full log.\\n- For a terminal task, the metadata also explains why it ended. A shell command that runs to completion reports \`status: completed\` on a zero exit, or \`status: failed\` with its non-zero \`exit_code\` — judge that failure from the \`exit_code\`, because a plain command failure carries no \`stop_reason\` and no \`terminal_reason\`. \`terminal_reason\` is a categorical label emitted only when the end is not an ordinary exit: \`timed_out\` when the deadline aborted it, \`stopped\` when it was explicitly stopped, or \`failed\` when it errored without producing an exit code; the \`stopped\` and \`failed\` cases also carry a human-readable \`stop_reason\`. A task that finished on its own with a clean exit carries neither \`stop_reason\` nor \`terminal_reason\`.\\n- The full, never-truncated log is always available at output_path; use the \`Read\` tool with that path to page through it, whether or not the preview was truncated.\\n- This tool works with the generic background task system and should remain the primary read path for future task types, not just bash.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "task_id": { "type": "string", "description": "The background task ID to inspect." } }, "required": [ "task_id" ], "additionalProperties": false } }, { "name": "TaskStop", "description": "Stop a running background task.\\n\\nOnly use this when a task must genuinely be cancelled — for a task that is\\nfinishing normally, wait for its completion notification or inspect it with\\n\`TaskOutput\` instead of stopping it.\\n\\nGuidelines:\\n- This is a general-purpose stop capability for any background task. It is not\\n  a bash-specific kill.\\n- Stopping a task is destructive: it may leave partial side effects behind.\\n  Use it with care.\\n- If the task has already finished, this tool simply returns its current\\n  status.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "task_id": { "type": "string", "description": "The background task ID to stop." }, "reason": { "default": "Stopped by TaskStop", "description": "Short reason recorded when the task is stopped.", "type": "string" } }, "required": [ "task_id" ], "additionalProperties": false } }, { "name": "TodoList", "description": "Use this tool to maintain a structured TODO list as you work through a multi-step task. Use it proactively and often when progress tracking helps the current work. This is especially useful in long-running investigations and implementation tasks with several tool calls; in plan mode, write the plan to the plan file rather than tracking it here.\\n\\n**When to use:**\\n- Multi-step tasks that span several tool calls\\n- Tracking investigation progress across a large codebase search\\n- Planning a sequence of edits before making them\\n- After receiving new multi-step instructions, capture the requirements as todos\\n- Before starting a tracked task, mark exactly one item as \`in_progress\`\\n- Immediately after finishing a tracked task, mark it \`done\`; do not batch completions at the end\\n\\n**When NOT to use:**\\n- Single-shot answers that complete in one or two tool calls\\n- Trivial requests where tracking adds no clarity\\n- Purely conversational or informational replies\\n\\n**Avoid churn:**\\n- Do not re-call this tool when nothing meaningful has changed since the last call — update the list only after real progress.\\n- When unsure of the current state, call query mode first (omit \`todos\`) to check the list before deciding what to update.\\n- If no available tool can move any task forward, tell the user where you are stuck instead of repeatedly re-ordering the same todos.\\n\\n**How to use:**\\n- Call with \`todos: [...]\` to replace the full list. Statuses: pending / in_progress / done.\\n- Call with no \`todos\` argument to retrieve the current list without changing it.\\n- Call with \`todos: []\` to clear the list.\\n- Keep titles short and actionable (e.g. \\"Read session-control.ts\\", \\"Add planMode flag to TurnManager\\").\\n- Update statuses as you make progress.\\n- When work is underway, keep exactly one task \`in_progress\`.\\n- Only mark a task \`done\` when it is fully accomplished.\\n- Never mark a task \`done\` if tests are failing, implementation is partial, unresolved errors remain, or required files/dependencies could not be found.\\n- If you encounter a blocker, keep the blocked task \`in_progress\` or add a new pending task describing what must be resolved.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "todos": { "description": "The updated todo list. Omit to read the current todo list without making changes. Pass an empty array to clear the list.", "type": "array", "items": { "type": "object", "properties": { "title": { "type": "string", "minLength": 1, "description": "Short, actionable title for the todo." }, "status": { "type": "string", "enum": [ "pending", "in_progress", "done" ], "description": "Current status of the todo." } }, "required": [ "title", "status" ], "additionalProperties": false } } }, "additionalProperties": false } }, { "name": "UpdateGoal", "description": "Set the status of the current goal. This is how you resume, complete, or block an autonomous goal.\\n\\n- \`active\` — resume a paused or blocked goal when the user explicitly asks you to work on that goal.\\n- \`complete\` — the objective is satisfied and any stated validation has passed. The goal ends and a completion summary is recorded. Before using this, verify the current state against the actual objective and every explicit requirement. Treat weak or indirect evidence as not complete. Do not use \`complete\` merely because a budget is nearly exhausted or you want to stop.\\n- \`blocked\` — a genuine impasse prevents useful progress: an external condition, required user input, missing credentials or permissions, a persistent technical failure, or an impossible, unsafe, or contradictory objective. For non-terminal blockers, do not use \`blocked\` the first time you hit the blocker. The same blocking condition must repeat for at least 3 consecutive goal turns before you call \`blocked\`, counting the original/user-triggered turn and automatic continuations. If a previously blocked goal is resumed, treat the resumed run as a fresh blocked audit. If the objective itself is impossible, unsafe, or contradictory, call \`blocked\` in the same turn instead of running more goal turns. Do not use \`blocked\` because the work is large, hard, slow, uncertain, incomplete, still needs validation, would benefit from clarification, or needs more goal turns. Once the 3-turn threshold is met and you cannot make meaningful progress without user input or an external-state change, call \`blocked\` instead of leaving the goal active.\\n\\nMost active goal turns should not call this tool. If you complete one useful slice of work and material work remains, end the turn normally without calling UpdateGoal; the runtime will prompt you to continue in the next goal turn. Call \`complete\` only when all required work is done, any stated validation has passed, and there is no useful next action. Do not call \`complete\` after only producing a plan, summary, first pass, or partial result. Call \`blocked\` only after the blocked audit threshold is met. If you call \`blocked\`, you will be prompted to explain the blocker in your next message. Setting the status is the machine-readable signal; the completion summary or blocker explanation is yours to write in the following message.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "status": { "type": "string", "enum": [ "active", "complete", "blocked" ], "description": "The lifecycle status to set for the current goal. Use \`blocked\` for impossible, unsafe, or contradictory objectives, or after the same non-terminal blocking condition repeats for at least 3 consecutive goal turns." } }, "required": [ "status" ], "additionalProperties": false } }, { "name": "WaitFor", "description": "Wait for background tasks to finish without ending the current turn.\\n\\nUse this when your next step depends on the result of a running background task (a sub-agent, a background bash command, or a background AskUserQuestion). The call suspends inside the current turn until the task finishes or the timeout elapses, then returns the outcome so you can keep working in the same turn. While waiting, no LLM requests are made.\\n\\nGuidelines:\\n\\n- Do not call WaitFor right after dispatching work whose result you do not need yet — finished background tasks notify you automatically. WaitFor is for the moment you genuinely cannot proceed without a result.\\n- \`timeout\` is required, in seconds, capped at 600. To wait longer, call WaitFor again; waking up periodically also lets you re-evaluate the situation.\\n- A timeout is not an error: the result lists the tasks that are still running, and you decide whether to wait again or do other work meanwhile.\\n- Without \`task_id\`, the wait ends as soon as any background task that was running at call time finishes. Tasks started during the wait are not covered by it; their completion arrives via the usual automatic notification.\\n- With \`task_id\`, the wait ends when that task finishes. An unknown \`task_id\` is an error; a task that has already finished returns immediately.\\n- When no background tasks are running, WaitFor returns immediately without waiting.\\n- When the wait ends because a task finished, the result also lists other tasks that finished during the wait window, so failures surface with context.\\n- Waiting has no side effects on the waited tasks: WaitFor never stops a task, and interrupting the wait (for example, a user interruption) leaves every task running.\\n- A finished task's result is delivered exactly once: tasks reported by WaitFor do not also produce an automatic completion notification.\\n- You can only wait for background tasks started by this agent; task IDs belonging to other agents are unknown here.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "timeout": { "type": "integer", "exclusiveMinimum": 0, "maximum": 600, "description": "Maximum time to wait, in seconds (1-600). A timeout is not an error: the tool returns the tasks that are still running, and you can call it again to keep waiting." }, "task_id": { "description": "The background task ID to wait for. When omitted, the wait ends as soon as any background task that was running at call time finishes.", "type": "string" } }, "required": [ "timeout" ], "additionalProperties": false } }, { "name": "Write", "description": "Create, append to, or replace a file entirely.\\n\\n- Missing parent directories are created automatically (like \`mkdir(parents=True, exist_ok=True)\`).\\n- Mode defaults to overwrite; append adds content at EOF without adding a newline.\\n- Write is NOT ALLOWED for incremental changes to existing files, including trivial, one-line, quick, or cosmetic edits. Use Edit instead.\\n- Use Write only when the file does not exist, you intend a complete replacement, or the new contents have little continuity with the old contents.\\n- Do not create unsolicited documentation files (\`*.md\` write-ups, \`README\`s, summaries) just because a task finished — write one only when the user asks for it, or when a task or project instruction requires it (e.g. the plan-mode plan file, created with Write when plan mode directs you to, or a changeset the repo mandates).\\n- Read before overwriting an existing file.\\n- Write ignores the Read/Edit line-number view. NEVER include line prefixes.\\n- Write outputs content literally, including supplied line endings: \\\\n stays LF, \\\\r\\\\n stays CRLF.\\n- For new content too large for one call, overwrite the first chunk, then append subsequent chunks. Never chunk Write to modify an existing file.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to the file to create, append to, or completely overwrite. Relative paths resolve against the working directory; a path outside the working directory must be absolute. Missing parent directories are created automatically." }, "content": { "type": "string", "description": "Raw full file content to write exactly as provided. This does not use the Read/Edit text view." }, "mode": { "description": "Write mode. Defaults to overwrite. append adds content to the end exactly as provided and does not add a newline.", "type": "string", "enum": [ "overwrite", "append" ] } }, "required": [ "path", "content" ], "additionalProperties": false } } ], "time": "<time>" }
+      [wire] llm.request                 { "agentId": "main", "kind": "loop", "provider": "openai", "model": "mock-model", "modelAlias": "mock-model", "thinkingEffort": "off", "maxTokens": 1000000, "toolSelect": false, "systemPromptHash": "ec9c34379c88babbc468ef2f3e0e08cd2f422c8c4a910664fb8bb394d703a575", "toolsHash": "755845c4dac18464a7b45428e5f0f6c35626531988b85d7b6b70b5e03a45e3c2", "messageCount": 1, "turnStep": "0.1", "time": "<time>" }
       [emit] assistant.delta             { "time": "<time>", "agentId": "main", "turnId": 0, "delta": "blocked" }
       [emit] agent.activity.updated      { "time": "<time>", "lifecycle": "ready", "turn": { "turnId": 0, "origin": { "kind": "user" }, "phase": "streaming", "stream": "assistant", "step": 1, "ending": false, "pendingApprovals": [], "activeToolCalls": [], "since": "<time>" }, "background": [], "agentId": "main" }
       [wire] usage.record                { "agentId": "main", "model": "mock-model", "usage": { "inputOther": 3, "output": 5, "inputCacheRead": 0, "inputCacheCreation": 0 }, "usageScope": "turn", "time": "<time>" }
@@ -814,6 +814,32 @@ describe('Agent loop', () => {
     expect(prompts).toEqual([undefined, 'hi']);
   });
 
+  it('carries the turn.started prompt for subagent system triggers', async () => {
+    const prompts: Array<string | undefined> = [];
+    const subscription = ctx.get(IEventBus).subscribe(TurnStarted, (event) => {
+      prompts.push(event.prompt);
+    });
+    ctx.mockNextResponse({ type: 'text', text: 'scanned' });
+
+    const subagent = (
+      await loop.enqueue(
+        new MessageStepRequest(
+          {
+            role: 'user',
+            content: [{ type: 'text', text: 'scan the repo' }],
+            toolCalls: [],
+            origin: { kind: 'system_trigger', name: 'subagent' },
+          },
+          { admission: 'newTurn' },
+        ),
+      ).assigned
+    ).turn;
+    await subagent.result;
+    subscription.dispose();
+
+    expect(prompts).toEqual(['scan the repo']);
+  });
+
   it('carries session file prompt attachments on turn started when the part id is omitted', async () => {
     const payloads: Array<readonly { kind: string; fileId: string }[] | undefined> = [];
     const subscription = ctx.get(IEventBus).subscribe(TurnStarted, (event) => {
diff --git a/packages/agent-core-v2/test/agent/questionTools/tools/ask-user.test.ts b/packages/agent-core-v2/test/agent/questionTools/tools/ask-user.test.ts
index fa8165b89..fea5a64f3 100644
--- a/packages/agent-core-v2/test/agent/questionTools/tools/ask-user.test.ts
+++ b/packages/agent-core-v2/test/agent/questionTools/tools/ask-user.test.ts
@@ -546,7 +546,7 @@ describe('AskUserQuestionTool', () => {
       expect(result.isError).toBe(false);
       expect(result.output).toContain('task_id: q_test_task_id');
       expect(result.output).toContain('automatic_notification: true');
-      expect(result.output).toContain('/tasks');
+      expect(result.output).toContain('human_shell_hint: The pending question is also visible in the client UI.');
       expect(registerTask).toHaveBeenCalledOnce();
       expect(registerTask.mock.calls[0]![1]).toMatchObject({ detached: true });
       expect(getTask).toHaveBeenCalledWith('q_test_task_id');
diff --git a/packages/agent-core-v2/test/agent/task/taskManager.test.ts b/packages/agent-core-v2/test/agent/task/taskManager.test.ts
index c56448b35..e96c9fe60 100644
--- a/packages/agent-core-v2/test/agent/task/taskManager.test.ts
+++ b/packages/agent-core-v2/test/agent/task/taskManager.test.ts
@@ -498,6 +498,30 @@ describe('AgentTaskService', () => {
     });
   });
 
+  it('keeps a detached process task running when the register-time signal aborts', async () => {
+    const { manager } = createAgentTaskService();
+    const { proc, killSpy } = pendingProcess();
+    const controller = new AbortController();
+    const taskId = manager.registerTask(
+      new ProcessTask(proc, 'sleep 10', 'foreground process'),
+      {
+        detached: false,
+        signal: controller.signal,
+      },
+    );
+
+    const waiting = manager.waitForForegroundRelease(taskId);
+    expect(manager.detach(taskId)).toMatchObject({ detached: true });
+    controller.abort();
+
+    await expect(waiting).resolves.toBe('detached');
+    expect(killSpy).not.toHaveBeenCalled();
+    expect(manager.getTask(taskId)).toMatchObject({
+      status: 'running',
+      detached: true,
+    });
+  });
+
   it('forwards foreground signal abort reasons to agent task controllers', async () => {
     const { manager } = createAgentTaskService();
     const foregroundController = new AbortController();
@@ -529,6 +553,30 @@ describe('AgentTaskService', () => {
     expect(isUserCancellation(subagentController.signal.reason)).toBe(true);
   });
 
+  it('does not forward register-time signal aborts to a detached agent task', async () => {
+    const { manager } = createAgentTaskService();
+    const foregroundController = new AbortController();
+    const subagentController = new AbortController();
+    const taskId = manager.registerTask(
+      agentTask(new Promise(() => {}), 'foreground agent', {
+        abortController: subagentController,
+      }),
+      {
+        detached: false,
+        signal: foregroundController.signal,
+      },
+    );
+
+    expect(manager.detach(taskId)).toMatchObject({ detached: true });
+    foregroundController.abort(userCancellationReason());
+
+    expect(subagentController.signal.aborted).toBe(false);
+    expect(manager.getTask(taskId)).toMatchObject({
+      status: 'running',
+      detached: true,
+    });
+  });
+
   it('does not count foreground tasks against the detached task limit', () => {
     const { manager } = createAgentTaskService({ maxRunningTasks: 1 });
     manager.registerTask(agentTask(new Promise(() => {}), 'foreground agent'), {
diff --git a/packages/agent-core-v2/test/app/codexLogin/codexLogin.test.ts b/packages/agent-core-v2/test/app/codexLogin/codexLogin.test.ts
index 5517a000c..35316d466 100644
--- a/packages/agent-core-v2/test/app/codexLogin/codexLogin.test.ts
+++ b/packages/agent-core-v2/test/app/codexLogin/codexLogin.test.ts
@@ -41,7 +41,7 @@ const MODELS: OpenAICodexModelInfo[] = [
 
 function callback(
   loopback = false,
-  waitForCode: () => Promise<{ code: string } | null> = async () => null,
+  waitForCode: OpenAICodexCallbackServer['waitForCode'] = async () => null,
 ): OpenAICodexCallbackServer {
   return {
     loopback,
@@ -190,6 +190,19 @@ describe('CodexLoginFlow', () => {
     flow.dispose();
   });
 
+  it('reports a cancelled state when the consent page denies authorization', async () => {
+    const flow = new CodexLoginFlow(
+      configStub().service,
+      deps({ startCallbackServer: async () => callback(true, async () => ({ denied: true })) }),
+    );
+
+    const start = await flow.start();
+    await vi.waitFor(() => expect(flow.status(start.login_id).state).toBe('cancelled'));
+
+    expect(flow.status(start.login_id).message).toBe('OpenAI Codex authorization was denied.');
+    flow.dispose();
+  });
+
   it('does not report cancellation after the atomic config commit starts', async () => {
     const config = configStub();
     const persisted = deferred<void>();
diff --git a/packages/agent-core-v2/test/os/backends/node-local/tools/bash.test.ts b/packages/agent-core-v2/test/os/backends/node-local/tools/bash.test.ts
index c84e7a3dc..0c4885a75 100644
--- a/packages/agent-core-v2/test/os/backends/node-local/tools/bash.test.ts
+++ b/packages/agent-core-v2/test/os/backends/node-local/tools/bash.test.ts
@@ -819,12 +819,12 @@ describe('BashTool', () => {
     expect(properties['timeout']?.default).toBe(60);
   });
 
-  it('renders the available commands section and the /tasks hint', () => {
+  it('renders the available commands section and the background-task panel hint', () => {
     const { runner } = createTestRunner(processWithOutput());
     const tool = bashTool(runner);
 
     expect(tool.description).toContain('Commands available');
-    expect(tool.description).toContain('/tasks');
+    expect(tool.description).toContain('background-task panel');
   });
 
   it('runs through runner.spawn, injects cwd, noninteractive env, and closes stdin', async () => {
@@ -1307,6 +1307,25 @@ describe('BashTool', () => {
     expect(description).toContain('run_in_background=true');
   });
 
+  it('strips every background sentence from the description when background execution is off', () => {
+    const { runner } = createTestRunner(processWithOutput());
+    const tool = bashTool(
+      runner,
+      createTestEnv(),
+      createTestCtx(),
+      createFakeTaskService().service,
+      stubToolPolicy((name) => name !== 'TaskList'),
+    );
+
+    const description = tool.description;
+    expect(description).toContain('Background execution is disabled for this agent.');
+    expect(description).not.toContain('background-task panel');
+    expect(description).not.toContain('run_in_background=true`,');
+    expect(description).not.toContain('move a running foreground command to the background');
+    expect(description).not.toContain('moved to the background instead of being killed');
+    expect(description).toContain('a foreground command that hits its timeout is killed');
+  });
+
   it('disables background execution when TaskList is inactive even if TaskOutput/TaskStop are active', async () => {
     const { runner, exec } = createTestRunner(processWithOutput());
     const tool = bashTool(
@@ -1393,8 +1412,11 @@ describe('BashTool background mode', () => {
     expect(result.output).not.toContain('after detach\n');
     expect(result.output).toContain(`task_id: ${task.taskId}`);
     expect(result.output).toContain('automatic_notification: true');
+    expect(result.output).toContain('The user moved this task to the background.');
+    expect(result.output).toContain('detached_by_user: true');
     expect(result.output).toContain('do NOT wait, poll, or call TaskOutput');
-    expect((result as { brief?: string }).brief).toBe(`Backgrounded ${task.taskId}`);
+    expect(result.output).toContain('human_shell_hint: The task is visible in the background-task panel.');
+    expect((result as { brief?: string }).brief).toBe(`Backgrounded ${task.taskId} by the user`);
     expect(service.getTask(task.taskId)).toMatchObject({ detached: true });
     await vi.waitFor(async () => {
       await expect(service.readOutput(task.taskId)).resolves.toContain('after detach\n');
@@ -1425,6 +1447,28 @@ describe('BashTool background mode', () => {
     await running;
   });
 
+  it('records the parent tool call id on the registered task', async () => {
+    const { proc, finish } = pendingProcess();
+    const { runner } = createTestRunner(proc);
+    const { service } = createFakeTaskService();
+    const tool = bashTool(runner, createTestEnv(), createTestCtx(), service);
+
+    const running = executeTool(tool, context({ command: 'sleep 10', timeout: 60 }));
+    await vi.waitFor(() => {
+      expect(service.list(false)).toHaveLength(1);
+    });
+    const task = service.list(false)[0]!;
+
+    expect(task).toMatchObject({
+      kind: 'process',
+      detached: false,
+      parentToolCallId: 'call_bash',
+    });
+
+    finish();
+    await running;
+  });
+
   it('applies the background timeout when a foreground command is detached', async () => {
     vi.useFakeTimers();
     try {
@@ -1469,6 +1513,10 @@ describe('BashTool background mode', () => {
         isError: false,
         brief: expect.stringContaining('after timeout'),
       });
+      expect(result.output).toContain('The task now runs in the background.');
+      expect(result.output).not.toContain('The user moved this task');
+      expect(result.output).not.toContain('detached_by_user');
+      expect(result.output).toContain('human_shell_hint: The task is visible in the background-task panel.');
       const taskId = /^task_id: (\S+)/m.exec(result.output as string)?.[1];
       expect(taskId).toBeDefined();
       expect(service.getTask(taskId!)).toMatchObject({ status: 'running', detached: true });
@@ -1803,8 +1851,8 @@ describe('BashTool background mode', () => {
     expect(output).toContain('automatic_notification: true');
     expect(output).toContain('do NOT wait, poll, or call TaskOutput on it');
     expect(output).not.toContain('block=false');
-    expect(output).toContain('human_shell_hint:');
-    expect(output).toContain('/tasks');
+    expect(output).toContain('human_shell_hint: The task is visible in the background-task panel.');
+    expect(output).not.toContain('/tasks');
   });
 
   it('rejects background command without description (description-required guard)', async () => {
diff --git a/packages/agent-core-v2/test/tool/tool.test.ts b/packages/agent-core-v2/test/tool/tool.test.ts
index aabf43885..631887c40 100644
--- a/packages/agent-core-v2/test/tool/tool.test.ts
+++ b/packages/agent-core-v2/test/tool/tool.test.ts
@@ -2573,6 +2573,7 @@ describe('Agent tool execution contract', () => {
     expect(result.output).toContain(`task_id: ${task.taskId}`);
     expect(result.output).toContain('agent_id: agent-child');
     expect(result.output).toContain('automatic_notification: true');
+    expect(result.output).toContain('note: The user moved this subagent to the background.');
 
     completion.resolve({ summary: 'finished later' });
     await expect(tasks.wait(task.taskId)).resolves.toMatchObject({
@@ -3885,8 +3886,8 @@ describe('Agent tools', () => {
         [emit] turn.step.started           { "time": "<time>", "agentId": "main", "turnId": 0, "step": 1, "stepId": "<uuid-1>" }
         [emit] agent.activity.updated      { "time": "<time>", "lifecycle": "ready", "turn": { "turnId": 0, "origin": { "kind": "user" }, "phase": "running", "step": 1, "ending": false, "pendingApprovals": [], "activeToolCalls": [], "since": "<time>" }, "background": [], "agentId": "main" }
         [wire] context.append_loop_event   { "agentId": "main", "event": { "type": "step.begin", "uuid": "<uuid-1>", "turnId": "0", "step": 1 }, "time": "<time>" }
-        [wire] llm.tools_snapshot          { "agentId": "main", "hash": "de9ec7eeaa58a8f75777f03e0ffe757f43975c59488deed982a7b2aca15a1d78", "tools": [ { "name": "Agent", "description": "Launch a subagent to handle a task. The subagent runs as a same-process loop instance with its own context and wire file. Delegating also keeps the bulk of intermediate file contents out of your own context — you get a conclusion back instead of a pile of dumps.\\n\\nWriting the prompt:\\n- The subagent starts with zero context — it has not seen this conversation. Brief it like a colleague who just walked into the room: state the goal, list what you already know, hand over the specifics.\\n- Lookups (read this file, run that test): put the exact path or command in the prompt. The subagent should not have to search for things you already know.\\n- Investigations (figure out X, find why Y): give the question, not prescribed steps — fixed steps become dead weight when the premise is wrong.\\n- Do not delegate understanding. If the task hinges on a file path or line number, find it yourself first and write it into the prompt.\\n\\nUsage notes:\\n- When the task continues earlier work a subagent already did, prefer resuming that agent (pass its \`resume\` id) over spawning a fresh instance — the resumed agent keeps its prior context.\\n- A subagent's result is only visible to you, not to the user. When the user needs to see what a subagent produced, summarize the relevant parts yourself in your own reply.\\n- Subagents use a fixed 2-hour timeout. If one times out, resume the same agent instead of starting over.\\n\\nWhen NOT to use Agent: skip delegation for trivial work you can do directly — reading a file whose path you already know, searching a small known set of files, or any task that takes only a step or two. Delegation has a context-handoff cost; it pays off only when the task is substantial enough to outweigh it.\\n\\nOnce a subagent is running, leave that scope to it: do not redo its searches or reads in parallel, and do not abandon it midway and finish the job manually. Both undo the context savings the delegation was meant to buy.\\n\\n\\nWhen \`run_in_background=true\`, the subagent runs detached from this turn. The completion arrives in a later turn as a synthetic user-role message containing its result — you do not need to poll, sleep, or check on its progress. Continue with other work or respond to the user. Never fabricate or predict what the result will say.\\n\\nDefault to a foreground subagent (omit \`run_in_background\`) when your next step needs its result — foreground hands the result straight back. Reach for \`run_in_background=true\` only when you have other work to do while it runs and do not need its result to proceed. Never launch in the background and then immediately wait on it (by polling \`TaskOutput\`, sleeping, or otherwise): that just blocks the turn for no benefit — run it in the foreground instead.\\n\\n\\nAvailable agent types (pass via subagent_type):\\n- plan: Read-only implementation planning and architecture design. Use this agent when the parent agent needs a step-by-step implementation plan, key file identification, and architectural trade-off analysis before code changes are made.\\n  Tools: Read, ReadMediaFile, Glob, Grep, WebSearch, FetchURL\\n- coder: General software engineering agent — the only subagent type with file-editing tools; use it for any delegated task that must modify code. Use this agent for non-trivial software engineering work that may require reading files, editing code, running commands, and returning a compact but technically complete summary to the parent agent.\\n  Tools: Bash, CronCreate, CronDelete, CronList, Edit, EnterPlanMode, ExitPlanMode, Glob, Grep, Read, ReadMediaFile, Skill, TaskList, TaskOutput, TaskStop, TodoList, WaitFor, WebSearch, FetchURL, Write, mcp__*\\n- explore: Fast codebase exploration with prompt-enforced read-only behavior. Fast agent specialized for exploring codebases. Use this when you need to quickly find files by patterns (e.g. \\"src/**/*.yaml\\"), search code for keywords (e.g. \\"database connection\\"), or answer questions about the codebase (e.g. \\"how does the auth module work?\\"). When calling this agent, specify the desired thoroughness level: \\"quick\\" for basic searches, \\"medium\\" for moderate exploration, or \\"thorough\\" for comprehensive analysis across multiple locations and naming conventions. Use this agent for any read-only exploration that will clearly require more than 3 search queries. Prefer launching multiple explore agents concurrently when investigating independent questions.\\n  Tools: Bash, Read, ReadMediaFile, Glob, Grep, WebSearch, FetchURL", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "prompt": { "type": "string", "description": "Full task prompt for the subagent" }, "description": { "type": "string", "description": "Short task description (3-5 words) for UI display" }, "subagent_type": { "description": "One of the available agent types (see \\"Available agent types\\" in this tool description). Defaults to \\"coder\\" when omitted.", "type": "string" }, "resume": { "description": "Optional agent ID to resume instead of creating a new instance. When set, do not also pass subagent_type — the resumed agent keeps its own type, and supplying both is rejected.", "type": "string" }, "run_in_background": { "description": "If true, return immediately without waiting for completion. Prefer false unless the task can run independently and there is a clear benefit to not waiting.", "type": "boolean" } }, "required": [ "prompt", "description" ], "additionalProperties": false } }, { "name": "AgentDynamicWorkflow", "description": "Launch multiple subagents from one prompt template, existing agent resumes, or both.\\n\\nUse AgentDynamicWorkflow when many subagents should run the same kind of task over different inputs. The placeholder is exactly \`{{item}}\`. For example, with \`prompt_template\` set to \`Review {{item}} for likely regressions.\` and \`items\` set to \`[\\"src/a.ts\\", \\"src/b.ts\\"]\`, AgentDynamicWorkflow launches two new subagents with those two concrete prompts. For a few differently-shaped tasks, make separate \`Agent\` calls in one message instead.\\n\\nUse \`resume_agent_ids\` to continue subagents that already exist from earlier work, such as ones that failed or timed out: map each agent id to the prompt for that resumed subagent (usually \`continue\` if no extra information is needed). You may combine \`resume_agent_ids\` with \`items\` in the same call to resume existing subagents and launch new ones. Do not duplicate resumed work in \`items\`.\\n\\nEach of these is enforced — a violation is rejected before any subagent starts: provide at least 2 \`items\` unless you pass \`resume_agent_ids\`; whenever \`items\` are present, \`prompt_template\` is required and must contain \`{{item}}\`; and the filled-in prompts must be distinct (two items that expand to the same prompt are rejected).\\n\\nUse enough subagents to keep the work focused and parallel. AgentDynamicWorkflow supports up to 128 subagents, and launches are queued automatically, so it is safe to split large tasks into many clear, independent items.\\n\\nIf \`AgentDynamicWorkflow\` is called, that call must be the only tool call in the response.\\n\\nThis capability is called a \\"dynamic workflow\\" (or simply \\"workflow\\"). Never use the word \\"swarm\\" in the \`description\` field, in subagent prompts, or when talking to the user about this tool.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "description": { "type": "string", "minLength": 1, "description": "Short description for the whole dynamic workflow. It is shown to the user as the workflow and subagent title, so word it as a workflow or run and never use the word \\"swarm\\"." }, "subagent_type": { "description": "Subagent type used for every new subagent spawned from items; defaults to coder when omitted. Resumed subagents always keep their original type, so passing subagent_type together with resume_agent_ids is allowed — it only affects the item-based spawns.", "type": "string", "minLength": 1 }, "prompt_template": { "description": "Prompt template for each subagent. The {{item}} placeholder is replaced with each item value.", "type": "string", "minLength": 1 }, "items": { "description": "Values used to fill {{item}}. Each item launches one new subagent.", "maxItems": 128, "type": "array", "items": { "type": "string", "minLength": 1 } }, "resume_agent_ids": { "description": "Map of existing subagent agent_id to the prompt used to resume that subagent. These resumed subagents are launched before new item-based subagents.", "type": "object", "propertyNames": { "type": "string", "minLength": 1 }, "additionalProperties": { "type": "string", "minLength": 1 } } }, "required": [ "description" ], "additionalProperties": false } }, { "name": "AskUserQuestion", "description": "Use this tool when you need to ask the user questions with structured options during execution. This allows you to:\\n1. Collect user preferences or requirements before proceeding\\n2. Resolve ambiguous or underspecified instructions\\n3. Let the user decide between implementation approaches as you work\\n4. Present concrete options when multiple valid directions exist\\n\\n**When NOT to use:**\\n- When you can infer the answer from context — be decisive and proceed\\n- Trivial decisions that don't materially affect the outcome\\n\\nOverusing this tool interrupts the user's flow. Only use it when the user's input genuinely changes your next action.\\n\\n**Usage notes:**\\n- Users always have an \\"Other\\" option for custom input — don't create one yourself\\n- Use multi_select to allow multiple answers to be selected for a question\\n- Keep option labels concise (1-5 words), use descriptions for trade-offs and details\\n- Each question should have 2-4 meaningful, distinct options\\n- Question texts must be unique across the call, and option labels must be unique within each question\\n- You can ask 1-4 questions at a time; group related questions to minimize interruptions\\n- If you recommend a specific option, list it first and append \\"(Recommended)\\" to its label\\n- The result is JSON with an \`answers\` object keyed by question text; each value is the chosen option's label (comma-separated labels for multi_select, or the user's own words if they picked \\"Other\\"); if \`answers\` is empty and a \`note\` says the user dismissed it, they chose not to answer — do not treat this as selecting the recommended option; decide based on context and do not re-ask the same question\\n- Set background=true when you can keep working without the answer. This starts a background question task and returns a task_id immediately. The answer arrives automatically in a later turn — you do not need to poll, sleep, or check on it. Continue with other work; never fabricate or predict the answer.", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "questions": { "minItems": 1, "maxItems": 4, "type": "array", "items": { "type": "object", "properties": { "question": { "type": "string", "minLength": 1, "description": "A specific, actionable question. End with '?'." }, "header": { "default": "", "description": "Short category tag (max 12 chars, e.g. 'Auth', 'Style').", "type": "string" }, "options": { "minItems": 2, "maxItems": 4, "type": "array", "items": { "type": "object", "properties": { "label": { "type": "string", "minLength": 1, "description": "Concise display text (1-5 words). If recommended, append '(Recommended)'." }, "description": { "default": "", "description": "Brief explanation of trade-offs or implications.", "type": "string" } }, "required": [ "label" ], "additionalProperties": false }, "description": "2-4 meaningful, distinct options. Do NOT include an 'Other' option — the system adds one automatically." }, "multi_select": { "default": false, "description": "Whether the user can select multiple options.", "type": "boolean" } }, "required": [ "question", "options" ], "additionalProperties": false }, "description": "The questions to ask the user (1-4 questions)." }, "background": { "default": false, "description": "Set true to ask in the background and return immediately with a background task_id; you are notified automatically when the user answers — do not poll with TaskOutput while the question is pending.", "type": "boolean" } }, "required": [ "questions" ], "additionalProperties": false } }, { "name": "Bash", "description": "Execute a \`bash\` command. Use this for shell semantics — pipes, env, processes, git, package managers, build/test runners, anything genuinely interactive or multi-step.\\n\\n**Translate these to a dedicated tool instead:**\\n- \`cat\` / \`head\` / \`tail\` (known path) → \`Read\`\\n- \`sed\` / \`awk\` (in-place edit) → \`Edit\`\\n- \`echo > file\` / \`cat <<EOF\` → \`Write\`\\n- \`find\` / recursive \`ls\` to locate files by name pattern → \`Glob\` (plain \`ls <known-directory>\` is fine for listing a directory)\\n- \`grep\` / \`rg\` (search file contents) → \`Grep\`\\n- \`echo\` / \`printf\` (talk to the user) → just output text directly\\n\\nThe dedicated tools render in the per-tool permission UI and keep raw stdout out of the conversation; that is why they are worth reaching for whenever one fits.\\n\\n**Output:**\\nThe stdout and stderr will be combined and returned as a string. The output may be truncated if it is too long. If the command exits non-zero, the output ends with a \`Command failed with exit code: N\` line; a command killed by its timeout or interrupted by the user ends with its own message instead.\\n\\nIf \`run_in_background=true\`, the command will be started as a background task and this tool will return a task ID instead of waiting for command completion. When doing that, you must provide a short \`description\`. Background commands default to a 600s timeout and \`timeout\` is capped at 86400s; set \`disable_timeout=true\` only when the task should run without a timeout. You will be automatically notified when the task completes. After starting one, default to returning control to the user instead of immediately waiting on it. Use \`TaskOutput\` only for a non-blocking status/output snapshot — do not wait on a task you just launched, since its completion arrives automatically. Use \`TaskStop\` only if the task must be cancelled. If a human user wants to inspect background tasks themselves, point them to the \`/tasks\` command, which opens an interactive panel; it has no subcommands.\\n\\n**Guidelines for safety and security:**\\n- Each shell tool call will be executed in a fresh shell environment. The shell variables, current working directory changes, and the shell history is not preserved between calls. To run a command in a particular directory, pass the \`cwd\` argument (or use absolute paths) rather than relying on a \`cd\` from an earlier call.\\n- The tool call will return after the command is finished. You shall not use this tool to execute an interactive command or a command that may run forever. For possibly long-running foreground commands, set the \`timeout\` argument in seconds. Foreground commands default to 60s and allow up to 300s. When a foreground command hits its timeout it is moved to the background instead of being killed, and you will be automatically notified when it completes.\\n- Avoid using \`..\` to access files or directories outside of the working directory.\\n- Avoid modifying files outside of the working directory unless explicitly instructed to do so.\\n- Never run commands that require superuser privileges unless explicitly instructed to do so.\\n\\n**Guidelines for efficiency:**\\n- Use \`&&\` to chain commands that genuinely depend on each other, e.g. \`npm install && npm test\`. Independent read-only commands (separate \`git show\`, \`ls\`, or status checks) should be issued as separate parallel Bash calls in one response, not chained into a single call — chaining serializes their execution and mixes their output. Do not stitch outputs together with \`echo\` separators.\\n- Use \`;\` to run commands sequentially regardless of success/failure\\n- Use \`||\` for conditional execution (run second command only if first fails)\\n- Use pipe operations (\`|\`) and redirections (\`>\`, \`>>\`) to chain input and output between commands\\n- Always quote file paths containing spaces with double quotes (e.g., cd \\"/path with spaces/\\")\\n- Compose multi-step logic in a single call with \`if\` / \`case\` / \`for\` / \`while\` control flows.\\n- Prefer \`run_in_background=true\` for long-running builds, tests, watchers, or servers when you need the conversation to continue before the command finishes.\\n\\n**Commands available:**\\nThe following common command categories are usually available. Availability still depends on the host, so when in doubt run \`which <command>\` first to confirm a command exists before relying on it.\\n- Navigation and inspection: \`ls\`, \`pwd\`, \`cd\`, \`stat\`, \`file\`, \`du\`, \`df\`, \`tree\`\\n- File and directory management: \`cp\`, \`mv\`, \`rm\`, \`mkdir\`, \`touch\`, \`ln\`, \`chmod\`, \`chown\`\\n- Text and data processing: \`wc\`, \`sort\`, \`uniq\`, \`cut\`, \`tr\`, \`diff\`, \`xargs\`\\n- Archives and compression: \`tar\`, \`gzip\`, \`gunzip\`, \`zip\`, \`unzip\`\\n- Networking and transfer: \`curl\`, \`wget\`, \`ping\`, \`ssh\`, \`scp\`\\n- Version control: \`git\`; for GitHub-hosted work (PRs, issues, CI runs, API queries) prefer the \`gh\` CLI when installed — it carries the user's GitHub auth and can return structured JSON\\n- Process and system: \`ps\`, \`kill\`, \`top\`, \`env\`, \`date\`, \`uname\`, \`whoami\`\\n- Language and package toolchains: \`node\`, \`npm\`, \`pnpm\`, \`yarn\`, \`python\`, \`pip\` (use whichever the project actually relies on)\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "command": { "type": "string", "minLength": 1, "description": "The command to execute." }, "cwd": { "description": "The working directory in which to run the command. When omitted, the command runs in the session's working directory.", "type": "string" }, "timeout": { "default": 60, "description": "Optional timeout in seconds for the command to execute. Foreground default 60s, max 300s. Background default 600s, max 86400s. Ignored for background commands when disable_timeout=true.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 }, "description": { "description": "A short description for the background task. Required when run_in_background is true.", "type": "string" }, "run_in_background": { "description": "Whether to run the command as a background task.", "type": "boolean" }, "disable_timeout": { "description": "If true, do not apply a timeout to the command. Only applies when run_in_background is true.", "type": "boolean" } }, "required": [ "command" ], "additionalProperties": false } }, { "name": "CreateGoal", "description": "Create a durable, structured goal that the runtime will pursue across multiple turns.\\n\\nCall \`CreateGoal\` only when:\\n\\n- the user explicitly asks you to start a goal or work autonomously toward an outcome, or\\n- a host goal-intake prompt asks you to create one.\\n\\nDo NOT create a goal for greetings, ordinary questions, or vague requests that lack a\\nverifiable completion condition. A goal needs a checkable end state.\\n\\nWhen the request is vague, ask the user for the missing completion criterion before creating\\nthe goal. If the user clearly insists after you warn them that the wording is vague or risky,\\nrespect that and create the goal.\\n\\nInclude a \`completionCriterion\` when the user provides one, or when it can be stated without\\ninventing new requirements. Keep \`objective\` concise; reference long task descriptions by file\\npath rather than pasting them.\\n\\nCreating a goal fails if one already exists, so use \`replace: true\` only when the user explicitly\\nwants to abandon the current goal and start a new one.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "objective": { "type": "string", "minLength": 1, "description": "The objective to pursue. Must have a verifiable end state." }, "completionCriterion": { "description": "How to verify the goal is complete. Include when the user provides one.", "type": "string" }, "replace": { "description": "Replace an existing active, paused, or blocked goal instead of failing.", "type": "boolean" } }, "required": [ "objective" ], "additionalProperties": false } }, { "name": "CronCreate", "description": "Schedule a prompt to be enqueued at a future time. Use for both recurring schedules and one-shot reminders.\\n\\nUses standard 5-field cron in the user's local timezone: minute hour day-of-month month day-of-week. \`0 9 * * *\` means 9am local — no timezone conversion needed.\\n\\n## One-shot tasks (recurring: false)\\n\\nFor \\"remind me at X\\" or \\"at <time>, do Y\\" requests — fire once then auto-delete.\\nPin minute/hour/day-of-month/month to specific values:\\n  \\"remind me at 2:30pm today to check the deploy\\" → cron: \\"30 14 <today_dom> <today_month> *\\", recurring: false\\n  \\"tomorrow morning, run the smoke test\\" → cron: \\"57 8 <tomorrow_dom> <tomorrow_month> *\\", recurring: false\\n\\nOne-shots are best for near-term reminders. A task only fires while its session is still alive (see Session lifetime below), so favor near times — within hours or a few days — rather than scheduling weeks or months ahead.\\n\\n## Recurring jobs (recurring: true, the default)\\n\\nFor \\"every N minutes\\" / \\"every hour\\" / \\"weekdays at 9am\\" requests:\\n  \\"*/5 * * * *\\" (every 5 min), \\"0 * * * *\\" (hourly), \\"0 9 * * 1-5\\" (weekdays at 9am local)\\n\\n## Avoid the :00 and :30 minute marks when the task allows it\\n\\nEvery user who asks for \\"9am\\" gets \`0 9\`, and every user who asks for \\"hourly\\" gets \`0 *\` — which means requests from across the planet land on the API at the same instant. When the user's request is approximate, pick a minute that is NOT 0 or 30:\\n  \\"every morning around 9\\" → \\"57 8 * * *\\" or \\"3 9 * * *\\" (not \\"0 9 * * *\\")\\n  \\"hourly\\" → \\"7 * * * *\\" (not \\"0 * * * *\\")\\n  \\"in an hour or so, remind me to...\\" → pick whatever minute you land on, don't round\\n\\nOnly use minute 0 or 30 when the user names that exact time and clearly means it (\\"at 9:00 sharp\\", \\"at half past\\", coordinating with a meeting). When in doubt, nudge a few minutes early or late — the user will not notice, and the fleet will.\\n\\n## Coalesce semantics\\n\\nFires are delivered only while the session is idle: a fire that comes due during an active turn is held and delivered at the next idle moment, never injected mid-turn.\\n\\nIf the scheduler slept past multiple ideal fire times (laptop closed, long-running turn, etc.), only **one** fire is delivered when it wakes up. The origin carries \`coalescedCount\` showing how many ideal fires were collapsed into this single delivery. You should treat \`coalescedCount > 1\` as \\"I missed some checks; only the latest state matters\\" rather than running the prompt that many times.\\n\\n## Cron-fire envelope\\n\\nWhen a cron task fires, the prompt you scheduled is re-injected wrapped in an XML envelope that exposes the fire context:\\n\\n\`\`\`\\n<cron-fire jobId=\\"...\\" cron=\\"...\\" recurring=\\"true|false\\" coalescedCount=\\"N\\" stale=\\"true|false\\">\\n<prompt>\\nyour original prompt text, verbatim\\n</prompt>\\n</cron-fire>\\n\`\`\`\\n\\nThe envelope is parseable. Use \`coalescedCount > 1\` to know multiple ideal fires were collapsed into a single delivery (treat as \\"only the latest state matters\\"), and \`stale=\\"true\\"\` as a cue that the task is past its 7-day threshold.\\n\\n## 7-day stale behavior\\n\\nRecurring tasks that have been alive for more than 7 days fire one\\nfinal time with \`stale: true\` on the envelope, and the system then\\nauto-deletes the task. The flag is the model's notice that this is\\nthe last delivery. If the schedule is still wanted, call \`CronCreate\`\\nagain with the same \`cron\` and \`prompt\` — that resets \`createdAt\` and\\nstarts a fresh 7-day window. One-shot tasks are never marked stale.\\n\\n## Jitter behavior\\n\\nAnti-herd jitter is applied deterministically per task id:\\n  - Recurring: ideal fire time is shifted **forward** by an offset ≤ min(10% of the cron period, 15 minutes). A \`*/5 * * * *\` task can drift up to 30s; a \`0 9 * * *\` task can drift up to 15 minutes.\\n  - One-shot: only when the ideal fire lands on \`:00\` or \`:30\` of the hour, the fire is pulled **earlier** by ≤ 90 seconds. Other minutes pass through unchanged.\\n\\n## One-shot vs recurring — when to pick which\\n\\nUse \`recurring: false\` for \\"remind me at X\\" style requests, single deadlines, \\"in N minutes do Y\\", and any task that should not repeat. Use \`recurring: true\` for periodic polling (CI status, build watchers, scheduled reports), workday rituals, and anything the user explicitly described as recurring.\\n\\n## Session lifetime\\n\\nCron tasks live in the current session. When you exit, they\\nare persisted under the session homedir; resuming the same session\\nreloads them and the scheduler resumes from each task's \`createdAt\`. Fire times that fell during the offline window are\\ncollapsed into a single delivery via \`coalescedCount\` (and recurring\\ntasks past their 7-day window arrive with \`stale: true\` as their final\\ndelivery).\\n\\nTasks do **not** carry over into a brand-new session — they are scoped\\nto the resumed session id, not to the working directory.\\n\\n## Limits\\n\\nA session holds at most 50 live cron tasks; creating one beyond that is rejected. (The \`prompt\` body is also capped — see its parameter description.) Expressions that never fire within the next 5 years (e.g. \`0 0 31 2 *\`, an impossible date) are rejected at create time.\\n\\n## Returned fields\\n\\n\`id\` (ULID), \`cron\` (the normalized expression), \`humanSchedule\` (English summary), \`recurring\`,\\n\`nextFireAt\` (local ISO timestamp with numeric offset, or null). \`id\` is needed by \`CronDelete\`.\\n\\n## Tell the user how to cancel or modify\\n\\nAfter successfully creating a task, proactively tell the user how they can cancel or modify it later. Users have no direct \`/cron\` command or self-service UI to manage reminders themselves; they must ask the model to make changes (e.g. \\"cancel my 9am reminder\\" or \\"change my daily check to 10am\\"). Include the task \`id\` in your message so the user can reference it.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "cron": { "type": "string", "description": "5-field cron expression in local time: \\"M H DoM Mon DoW\\" (e.g. \\"*/5 * * * *\\" = every 5 minutes; \\"30 14 28 2 *\\" = Feb 28 at 2:30pm local — a pinned date like this repeats yearly unless you also pass recurring: false)." }, "prompt": { "type": "string", "minLength": 1, "maxLength": 8192, "description": "The prompt to enqueue at each fire time. Limited to 8 KiB (UTF-8)." }, "recurring": { "default": true, "description": "true (default) = fire on every cron match until deleted or auto-expired after 7 days. false = fire once at the next match, then auto-delete. Use false for \\"remind me at X\\" one-shot requests with pinned minute/hour/dom/month.", "type": "boolean" } }, "required": [ "cron", "prompt" ], "additionalProperties": false } }, { "name": "CronDelete", "description": "Cancel a scheduled cron job by id.\\n\\nUse this tool to remove a cron task previously scheduled with\\n\`CronCreate\`. The \`id\` is the ULID value returned by \`CronCreate\`, or\\nshown in the \`id:\` column of \`CronList\` — quote it verbatim, no\\nprefix.\\n\\nBehaviour by task kind:\\n\\n- **Recurring task** (\`recurring: true\`): stops all future fires\\n  immediately. The scheduler picks up the deletion on its next tick.\\n- **One-shot task** (\`recurring: false\`): cancels the pending fire if\\n  it has not happened yet. One-shots that have already fired\\n  auto-delete themselves, so calling \`CronDelete\` on a fired one-shot\\n  returns \\"no cron job with id ...\\".\\n\\nNot-found is reported as an error (not a silent no-op) so you can\\ncorrect yourself — typically by calling \`CronList\` to see which ids\\nare actually live, rather than re-trying with the same stale id.\\n\\nRefresh pattern (use when you want a stale recurring schedule to\\ncontinue):\\n\\nStale recurring tasks are auto-deleted by the system after their final\\nfire — there is nothing for \`CronDelete\` to remove at that point. To\\nkeep the schedule running, just call \`CronCreate\` with the same \`cron\`\\nand \`prompt\`. Use \`CronList\`'s \`prompt\` field to recall the original\\ntext after a context compaction.\\n\\n\`CronDelete\` remains the right call when you want to cancel a task\\nthat is still live (recurring not yet stale, or a one-shot still\\npending).\\n\\nGuidelines:\\n\\n- Users have no direct \`/cron\` command or self-service UI to delete\\n  tasks themselves; they must ask the model to cancel a reminder.\\n  When deleting on behalf of a user, confirm the action and report\\n  the result plainly.\\n- Cron deletion is irreversible — there is no undo. If you delete the\\n  wrong task, you must re-create it with \`CronCreate\`.\\n- If the model is unsure which id is current (e.g. after a context\\n  compaction), call \`CronList\` first rather than guessing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "id": { "type": "string", "description": "The cron job id (ULID) returned by CronCreate / CronList." } }, "required": [ "id" ], "additionalProperties": false } }, { "name": "CronList", "description": "List all cron jobs currently scheduled in this session.\\n\\nUse this tool to see every pending cron task — both recurring jobs and\\none-shot reminders — that you (or the user) have scheduled with\\n\`CronCreate\`. The output is the entry point for inspecting scheduled\\nwork: it returns a stable id, the original cron expression, a human\\nrendering, the next post-jitter fire time, the recurring flag, the\\ntask's age in days, and a stale indicator.\\n\\nEach record carries:\\n\\n- \`id\` — the task id (a ULID). Pass this to \`CronDelete\` to remove the\\n  task, or quote it in user-facing messages when asking for\\n  confirmation.\\n- \`cron\` — the verbatim 5-field cron expression as scheduled.\\n- \`humanSchedule\` — plain-English rendering (e.g. \`every 5 minutes\`).\\n- \`prompt\` — the scheduled prompt text, JSON-encoded so embedded\\n  newlines stay on one line. Truncated to 200 UTF-8 bytes with\\n  \`…(truncated)\` if longer. Use this to recall what a task is for\\n  after a context compaction, and as the source for the\\n  \`CronCreate\` refresh ritual.\\n- \`nextFireAt\` — local ISO timestamp with an explicit numeric offset\\n  for the next fire **after jitter has been applied**. The actual fire\\n  may land slightly before or after a round \`:00\` / \`:30\` minute mark\\n  due to herd-avoidance jitter; this is the value the scheduler will\\n  compare against, so it reflects what will really happen. \`null\` if\\n  the expression has no fire in the next 5 years (should not happen\\n  for tasks created through \`CronCreate\`, which validates).\\n- \`recurring\` — \`true\` for cadenced jobs, \`false\` for one-shots.\\n- \`ageDays\` — \`(now - createdAt) / day\`, two decimal places. Useful\\n  when deciding whether a long-running cron is still relevant.\\n- \`stale\` — \`true\` when a recurring task is older than 7 days. The\\n  system **auto-deletes the task after this fire** to bound session\\n  lifetime; the \`stale: true\` flag is the model's notice that this is\\n  the final delivery. To resume the same schedule, call \`CronCreate\`\\n  again with the original \`cron\` and \`prompt\` (the \`prompt\` row above\\n  carries it for exactly this purpose). One-shots are never marked\\n  stale — they fire at most once by construction.\\n\\nGuidelines:\\n\\n- This tool is read-only and never mutates state, so it is always\\n  safe to call (including in plan mode).\\n- Users cannot directly manage cron tasks themselves; if they want to\\n  cancel or modify a schedule, route the request through the model\\n  (i.e. call \`CronDelete\` or \`CronCreate\` on their behalf).\\n- The empty case returns \`cron_jobs: 0\\\\nNo cron jobs scheduled.\`. Cron\\n  tasks survive a resume of the same session but do not bleed into new\\n  sessions.\\n- After a context compaction, or whenever you are unsure which cron\\n  jobs are live, call this tool to re-enumerate them rather than\\n  guessing ids from earlier in the conversation.\\n- Records are separated by a line containing just \`---\`, in the\\n  insertion order they were scheduled.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "Edit", "description": "Perform exact replacements in existing files.\\n\\n- Edit is mandatory for every incremental change, especially small edits. DO NOT use Write or Bash \`sed\`.\\n- Read the target file before every Edit. DO NOT call Edit from memory, stale context, or a guessed \`old_string\`.\\n- Take \`old_string\` and \`new_string\` from the Read output view.\\n- Drop the line-number prefix and tab; match only file content.\\n- \`old_string\` must be unique unless \`replace_all\` is set.\\n- If \`old_string\` is ambiguous, add surrounding context. Use \`replace_all\` only when every occurrence should change — for example, renaming a symbol throughout the file.\\n- Multiple Edit calls may run in one response only when they do not target the same file.\\n- DO NOT issue consecutive Edit calls on the same file. A previous Edit can invalidate a later Edit's \`old_string\`, causing \`old_string not found\`. Read the file again before the next Edit.\\n- A write lock serializes same-file edits in response order, but serialization does not make stale \`old_string\` valid.\\n- For pure CRLF files, Read shows LF; use LF in \`old_string\` and \`new_string\`, and Edit writes CRLF back.\\n- For mixed endings or lone carriage returns, Read shows carriage returns as \\\\r; include actual \\\\r escapes in those positions.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to the text file to edit. Relative paths resolve against the working directory; a path outside the working directory must be absolute." }, "old_string": { "type": "string", "minLength": 1, "description": "Exact content to replace from the Read output view, without the line-number prefix. Use LF for pure CRLF files; use actual \\\\r escapes where Read shows \\\\r." }, "new_string": { "type": "string", "description": "Replacement text in the same Read output view. LF is written back as CRLF only for pure CRLF files." }, "replace_all": { "description": "Set true only when every occurrence of old_string should be replaced.", "type": "boolean" } }, "required": [ "path", "old_string", "new_string" ], "additionalProperties": false } }, { "name": "EnterPlanMode", "description": "Use this tool proactively when you're about to start a non-trivial implementation task.\\nGetting user sign-off on your approach via ExitPlanMode before writing code prevents wasted effort.\\n\\nUse it when ANY of these conditions apply:\\n\\n1. New Feature Implementation - e.g. \\"Add a caching layer to the API\\"\\n2. Multiple Valid Approaches - e.g. \\"Optimize database queries\\" (indexing vs rewrite vs caching)\\n3. Code Modifications - e.g. \\"Refactor auth module to support OAuth\\"\\n4. Architectural Decisions - e.g. \\"Add WebSocket support\\"\\n5. Multi-File Changes - involves more than 2-3 files\\n6. Unclear Requirements - need exploration to understand scope\\n7. User Preferences Matter - if user input would materially change the implementation approach, use EnterPlanMode to structure the decision\\n\\nPermission mode notes:\\n- EnterPlanMode enters plan mode automatically without an approval prompt in all permission modes.\\n- In yolo and manual modes, ExitPlanMode still presents the plan to the user for approval.\\n- In auto permission mode, do not use AskUserQuestion; make the best decision from available context.\\n- In auto permission mode, ExitPlanMode exits plan mode without asking the user.\\n- Use EnterPlanMode only when planning itself adds value.\\n\\nWhen NOT to use:\\n- Single-line or few-line fixes (typos, obvious bugs, small tweaks)\\n- User gave very specific, detailed instructions\\n- Pure research/exploration tasks\\n\\nOnce you are in plan mode, a reminder walks you through the workflow (explore → design → write the plan file → \`ExitPlanMode\`) and enforces read-only access. For non-trivial tasks where you are unsure of the codebase structure or relevant code paths, use \`Agent(subagent_type=\\"explore\\")\` to investigate first when the \`Agent\` tool is available.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "ExitPlanMode", "description": "Use this tool when you are in plan mode and have finished writing your plan to the plan file and are ready for user approval.\\n\\n## How This Tool Works\\n- You should have already written your plan to the plan file specified in the plan mode reminder.\\n- This tool does NOT take the plan content as a parameter - it reads the plan from the file you wrote.\\n- The user will see the contents of your plan file when they review it. In auto permission mode, the tool reads the file and exits plan mode without asking the user.\\n\\n## When to Use\\nOnly use this tool for tasks that require planning implementation steps. For research tasks (searching files, reading code, understanding the codebase), do NOT use this tool.\\n\\n## What a good plan contains\\nList specific, verifiable steps grounded in the actual codebase — real files, functions, and commands, in a sensible order. Each step should be concrete enough to act on and to check. Avoid vague filler like \\"improve performance\\" or \\"add tests\\"; say what to change and where.\\n\\n## Multiple Approaches\\nIf your plan offers multiple alternative approaches, pass them via the \`options\` parameter so the user can choose which one to execute — see the \`options\` parameter for the format, count, and reserved labels. In yolo and manual modes the user sees all options alongside the host's Reject and Revise controls.\\n\\n## Before Using\\n- In auto permission mode, do NOT use AskUserQuestion; make the best decision from available context.\\n- In auto permission mode, this tool exits plan mode without asking the user.\\n- In yolo and manual modes, this tool still presents the plan to the user for approval.\\n- If auto permission mode is not active and you have unresolved questions, use AskUserQuestion first.\\n- If auto permission mode is not active and you have multiple approaches and haven't narrowed down yet, consider using AskUserQuestion first to let the user choose, then write a plan for the chosen approach only.\\n- Once your plan is finalized, use THIS tool to request approval.\\n- Do NOT use AskUserQuestion to ask \\"Is this plan OK?\\" or \\"Should I proceed?\\" - that is exactly what ExitPlanMode does.\\n- If rejected, revise based on feedback and call ExitPlanMode again.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "options": { "description": "When the plan contains multiple alternative approaches, list them here so the user can choose which one to execute. Provide up to 3 options; 2-3 distinct approaches work best when the plan offers a real choice. Passing a single option is allowed and is equivalent to a plain plan approval. Each option represents a distinct approach from the plan. Do not use \\"Reject\\", \\"Revise\\", \\"Approve\\", or \\"Reject and Exit\\" as labels.", "minItems": 1, "maxItems": 3, "type": "array", "items": { "type": "object", "properties": { "label": { "type": "string", "minLength": 1, "maxLength": 80, "description": "Short name for this option (1-8 words). Append \\"(Recommended)\\" if you recommend this option." }, "description": { "default": "", "description": "Brief summary of this approach and its trade-offs.", "type": "string" } }, "required": [ "label" ], "additionalProperties": false } } }, "additionalProperties": false } }, { "name": "FetchURL", "description": "Fetch content from a URL. The content is returned either as the main text extracted from the page, or as the full response body verbatim; a note at the top of the result states which of the two you received, so you can judge how complete it is. Use this when you need to read a specific web page.\\n\\nOnly fully-formed public \`http\`/\`https\` URLs are supported; other schemes and private or loopback addresses are not fetched. Very large pages may be truncated or refused. The fetch carries no login or session for the target site, so pages behind authentication (private repositories, internal dashboards) return a login page or an error instead of the real content — if the text you get back looks like a generic landing or sign-in page, treat that as the login wall, not the answer, and reach the content through a credentialed route (an authenticated CLI or MCP tool) instead.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "url": { "type": "string", "description": "The URL to fetch content from." } }, "required": [ "url" ], "additionalProperties": false } }, { "name": "GetGoal", "description": "Read the current goal: its objective, completion criterion, status, and budgets (turns, tokens,\\ntime, and how much of each remains). When the goal has stopped, it also reports the terminal reason.\\n\\nUse \`GetGoal\` before deciding whether to continue working, report completion, report a blocker,\\nor respect a pause. It returns \`{ \\"goal\\": null }\` when there is no current goal.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "Glob", "description": "Find files by glob pattern, sorted by modification time (most recent first).\\n\\nPowered by ripgrep. Respects \`.gitignore\`, \`.ignore\`, and \`.rgignore\` by default — set \`include_ignored\` to also match ignored files (e.g. build outputs, \`node_modules\`). Sensitive files (such as \`.env\`) are always filtered out. Matches are files only — directories themselves are never listed; to find a directory, glob for a file inside it (e.g. \`**/fixtures/**\`).\\n\\nGood patterns:\\n- \`*.ts\` — all files matching an extension, at any depth below the search root (a bare pattern without \`/\` matches recursively)\\n- \`src/*.ts\` — files directly inside \`src/\` (one level, not recursive)\\n- \`src/**/*.ts\` — recursive walk with a subdirectory anchor and extension\\n- \`**/*.py\` — recursive walk from the search root for an extension\\n- \`*.{ts,tsx}\` — brace expansion is supported\\n- \`{src,test}/**/*.ts\` — cartesian brace expansion is supported too\\n\\nResults are capped at the first 100 matching paths. If a search would return more, a truncation marker is appended. Refine the pattern (extension, subdirectory) when 100 is not enough, or call again with a narrower anchor.\\n\\nLarge-directory caveat — avoid recursing into dependency / build output even with an anchor, especially when \`include_ignored\` is set:\\n- \`node_modules/**/*.js\`, \`.venv/**/*.py\`, \`__pycache__/**\`, \`target/**\` can produce thousands of results that truncate at the match cap and waste context. Prefer specific subpaths like \`node_modules/react/src/**/*.js\`.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "pattern": { "type": "string", "description": "Glob pattern to match files." }, "path": { "description": "Directory to search. Accepts an absolute path, or a path relative to the current working directory. Defaults to the current working directory.", "type": "string" }, "include_ignored": { "description": "Also match files excluded by ignore files such as \`.gitignore\`, \`.ignore\`, and \`.rgignore\` (for example \`node_modules\` or build outputs). Sensitive files (such as \`.env\`) remain filtered out for safety. VCS metadata directories (\`.git\` and similar) are always skipped, even when this is true. Defaults to false.", "type": "boolean" }, "include_dirs": { "description": "Deprecated and ignored. Results are always files-only — directories are never listed. Accepted only so older calls that still pass this flag are not rejected by parameter validation.", "type": "boolean" } }, "required": [ "pattern" ], "additionalProperties": false } }, { "name": "Grep", "description": "Search file contents using regular expressions (powered by ripgrep).\\n\\nUse Grep when the task is to find unknown content or unknown file locations. Do not use shell \`grep\` or \`rg\` directly; this tool applies workspace path policy, output limits, and sensitive-file filtering.\\nALWAYS use Grep tool instead of running \`grep\` or \`rg\` from a shell — direct shell calls bypass workspace policy, output limits, and sensitive-file filtering.\\nIf you already know a concrete file path and need to inspect its contents, use Read directly instead.\\n\\nWrite patterns in ripgrep regex syntax, which differs from POSIX \`grep\` syntax. For example, braces are special, so escape them as \`\\\\{\` to match a literal \`{\`.\\n\\nHidden files (dotfiles such as \`.gitlab-ci.yml\` or \`.eslintrc.json\`) are searched by default. To also search files excluded by \`.gitignore\` (such as \`node_modules\` or build outputs), set \`include_ignored\` to \`true\`. Sensitive files (such as \`.env\`) are always skipped for safety, even when \`include_ignored\` is \`true\`.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "pattern": { "type": "string", "description": "Regular expression to search for." }, "path": { "description": "File or directory to search. Accepts an absolute path, or a path relative to the current working directory. Omit to search the current working directory. Use Read instead when you already know a concrete file path and need its contents.", "type": "string" }, "glob": { "description": "Optional glob filter for which files to search, e.g. \`*.ts\`. Matched against each file's full absolute path, so a path-anchored pattern like \`src/**/*.ts\` silently matches nothing — use a basename pattern (\`*.ts\`), or anchor with \`**/\` (\`**/src/**/*.ts\`). To scope the search to a directory, use \`path\` instead.", "type": "string" }, "type": { "description": "Optional ripgrep file type filter, such as ts or py. Prefer this over \`glob\` when filtering by language or file kind: it is more efficient and less error-prone than an equivalent glob pattern.", "type": "string" }, "output_mode": { "description": "Shape of the result. \`content\` shows matching lines (honors \`-A\`, \`-B\`, \`-C\`, \`-n\`, and \`head_limit\`); \`files_with_matches\` shows only the paths of files that contain a match, most-recently-modified first (honors \`head_limit\`); \`count_matches\` shows per-file match counts as \`path:count\` lines, preceded by an aggregate total line. Defaults to \`files_with_matches\`.", "type": "string", "enum": [ "content", "files_with_matches", "count_matches" ] }, "-i": { "description": "Perform a case-insensitive search. Defaults to false.", "type": "boolean" }, "-n": { "description": "Prefix each matching line with its line number. Applies only when \`output_mode\` is \`content\`. Defaults to true.", "type": "boolean" }, "-A": { "description": "Number of lines to show after each match. Applies only when \`output_mode\` is \`content\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "-B": { "description": "Number of lines to show before each match. Applies only when \`output_mode\` is \`content\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "-C": { "description": "Number of lines to show before and after each match. Applies only when \`output_mode\` is \`content\`; takes precedence over \`-A\` and \`-B\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "head_limit": { "description": "Limit output to the first N lines/entries after offset. Defaults to 250. Pass 0 for unlimited.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "offset": { "description": "Number of leading lines/entries to skip before applying \`head_limit\`. Use it together with \`head_limit\` to page through large result sets. Defaults to 0.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "multiline": { "description": "Enable multiline matching, where the pattern can span line boundaries and \`.\` also matches newlines. Defaults to false.", "type": "boolean" }, "include_ignored": { "description": "Also search files excluded by ignore files such as \`.gitignore\`, \`.ignore\`, and \`.rgignore\` (for example \`node_modules\` or build outputs). Sensitive files (such as \`.env\`) remain filtered out for safety. VCS metadata directories (\`.git\` and similar) are always skipped, even when this is true. Defaults to false.", "type": "boolean" } }, "required": [ "pattern" ], "additionalProperties": false } }, { "name": "Lookup", "description": "Look up a short test value.", "parameters": { "type": "object", "properties": { "query": { "type": "string" } }, "required": [ "query" ], "additionalProperties": false } }, { "name": "Read", "description": "Read a text file from the local filesystem.\\n\\nIf the user provides a concrete file path to a text file, call Read directly. Do not \`Glob\`, \`ls\`, or otherwise pre-check known text file paths; missing or invalid file paths return errors you can handle. Do not use Read for directories; use \`ls\` via Bash for a known directory, or Glob when you need files matching a name pattern (Glob lists files only, never directories). Use \`Grep\` only when the task is to search for unknown content or locations.\\n\\nWhen you need several files, prefer to read them in parallel: emit multiple \`Read\` calls in a single response instead of reading one file per turn.\\n\\n- Relative paths resolve against the working directory; a path outside the working directory must be absolute.\\n- Returns up to 1000 lines or 100 KB per call, whichever comes first; lines longer than 2000 chars are truncated mid-line.\\n- Page larger files with \`line_offset\` (1-based start line) and \`n_lines\`. Omit \`n_lines\` to read up to the 1000-line cap.\\n- Sensitive files (\`.env\` files, credential stores, SSH private keys, and similar secrets) are refused to protect secrets; do not attempt to read them. Templates and public keys are exempt: \`.env.example\` / \`.env.sample\` / \`.env.template\` and public SSH keys such as \`id_rsa.pub\` read normally.\\n- UTF-8 text files are read directly. UTF-16 LE/BE text files (with or without a BOM) are detected automatically and transcoded to UTF-8 for display; the status block notes the detected encoding, and Edit/Write on such a file still expect UTF-8 — convert its encoding first (e.g. with \`iconv\`). Other encodings (e.g. GBK), binary files, and files containing NUL bytes are refused.\\n- Negative line_offset reads from the end of the file (for example, -100 reads the last 100 lines); the absolute value cannot exceed 1000.\\n- Output format: \`<line-number>\\\\t<content>\` per line.\\n- A \`<system>...</system>\` status block is appended after the file content; it summarizes how much was read (line and byte counts, truncation, line-ending notes) and is not part of the file itself.\\n- Pure CRLF files are displayed with LF line endings; \`Edit\` matches this output and preserves CRLF when writing back.\\n- Mixed or lone carriage-return line endings are shown as \`\\\\r\` and require exact \`Edit.old_string\` escapes.\\n- After a successful \`Edit\`/\`Write\`, do not re-read solely to prove the write landed. When the task depends on an exact file, API, or output shape, inspect the final external contract before finishing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to a text file. Relative paths resolve against the working directory; a path outside the working directory must be absolute. Directories are not supported; use \`ls\` via Bash for a known directory, or Glob for pattern search." }, "line_offset": { "description": "The line number to start reading from. Omit to start at line 1. Negative values read from the end of the file; the absolute value cannot exceed 1000.", "anyOf": [ { "type": "integer", "minimum": 1, "maximum": 9007199254740991 }, { "type": "integer", "minimum": -1000, "maximum": -1 } ] }, "n_lines": { "description": "The number of lines to read; the tool also applies its internal cap. Omit to read up to the internal cap of 1000 lines.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 } }, "required": [ "path" ], "additionalProperties": false } }, { "name": "SetGoalBudget", "description": "Set a hard budget limit for the current goal.\\n\\nUse this only when the user clearly gives a runtime limit, such as:\\n\\n- \\"stop after 20 turns\\"\\n- \\"use no more than 500k tokens\\"\\n- \\"finish within 30 minutes\\"\\n\\nDo not invent limits. Do not call this for vague wording such as \\"spend some time\\" or\\n\\"try to be quick\\".\\n\\nIf the user gives a compound time, convert it to one supported unit before calling this tool.\\nFor example, \\"2 hours and 3 minutes\\" can be set as \`value: 123, unit: \\"minutes\\"\`.\\n\\nA time budget must be between 1 second and 24 hours — the tool rejects anything shorter or\\nlonger, telling the user it is not a reasonable goal budget. Turn and token budgets are not\\nbounded this way; they must be positive and are rounded to the nearest whole number (minimum 1).\\n\\nSupported units:\\n\\n- \`turns\`\\n- \`tokens\`\\n- \`milliseconds\`\\n- \`seconds\`\\n- \`minutes\`\\n- \`hours\`\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "value": { "type": "number", "exclusiveMinimum": 0, "description": "The positive numeric budget value." }, "unit": { "type": "string", "enum": [ "turns", "tokens", "milliseconds", "seconds", "minutes", "hours" ] } }, "required": [ "value", "unit" ], "additionalProperties": false } }, { "name": "Skill", "description": "Invoke a registered skill from the current skill listing. BLOCKING REQUIREMENT: when a skill from the listing matches the user's request, you MUST call this tool (not free-form text). Do not re-invoke a skill to repeat work already done: if a \`<skill-loaded>\` block for it with the same \`args\` is already present in the conversation, follow those instructions directly instead of calling the tool again. Do call the tool again when you need the skill with different arguments — the loaded block was expanded with the earlier \`args\` and will not reflect new inputs.", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "skill": { "type": "string", "description": "The exact name of the skill to invoke, spelled as it appears in the current skill listing (e.g. \\"commit\\", \\"pdf\\")." }, "args": { "description": "Optional argument string for the skill, written like a command line (e.g. \`-m \\"fix bug\\"\`, \`123\`, a file path). It is split on whitespace (quotes group a token) and expanded into the skill's placeholders ($NAME, $1, $ARGUMENTS); if the skill body has no placeholders, the whole string is still appended as a trailing \`ARGUMENTS:\` line. Omit it only when there is nothing to pass.", "type": "string" } }, "required": [ "skill" ], "additionalProperties": false } }, { "name": "TaskList", "description": "List background tasks and their current status.\\n\\nUse this tool to discover which background tasks exist and where each one\\nstands. It is the entry point for inspecting background work: it returns a\\ntask ID, status, and description for every task it reports, plus the command,\\nPID, and (once finished) exit code for shell tasks, and a stop reason for any\\ntask that ended early.\\n\\nGuidelines:\\n\\n- After a context compaction, or whenever you are unsure which background\\n  tasks are running or what their task IDs are, call this tool to\\n  re-enumerate them instead of guessing a task ID.\\n- Prefer the default \`active_only=true\`, which lists only non-terminal tasks.\\n  Pass \`active_only=false\` only when you specifically need to see tasks that\\n  have already finished. With \`active_only=false\` the result may also include\\n  \`lost\` tasks — tasks left over from a previous process that can no longer be\\n  inspected or controlled; treat them as already terminated.\\n- \`limit\` caps how many tasks are returned. It accepts a value between 1 and\\n  100 and defaults to 20 when omitted.\\n- This tool only lists tasks; it does not return their output. Use it first\\n  to locate the task ID you need, then call \`TaskOutput\` with that ID to read\\n  the task's output and details.\\n- This tool is read-only and does not change any state, so it is always safe\\n  to call, including in plan mode.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "active_only": { "default": true, "description": "Whether to list only non-terminal background tasks.", "type": "boolean" }, "limit": { "default": 20, "description": "Maximum number of tasks to return.", "type": "integer", "minimum": 1, "maximum": 100 } }, "additionalProperties": false } }, { "name": "TaskOutput", "description": "Retrieve a snapshot of a running or completed background task.\\n\\nUse this after \`Bash(run_in_background=true)\`, \`Agent(run_in_background=true)\`, or \`AskUserQuestion(background=true)\` to check progress, or to read the output of a task that has already completed.\\n\\nGuidelines:\\n- Prefer relying on automatic completion notifications. Use this tool only when you need task output before the automatic notification arrives.\\n- This tool is always non-blocking: it returns the current status/output snapshot immediately and never waits for the task to finish.\\n- Do not use TaskOutput to wait for a result you need before continuing — if your next step depends on the task's result, run that task in the foreground instead. TaskOutput is for a deliberate progress check you will act on without blocking, not a way to sit and wait for a background task you just launched.\\n- This tool returns structured task metadata, a fixed-size output preview, and an output_path for the full log.\\n- For a terminal task, the metadata also explains why it ended. A shell command that runs to completion reports \`status: completed\` on a zero exit, or \`status: failed\` with its non-zero \`exit_code\` — judge that failure from the \`exit_code\`, because a plain command failure carries no \`stop_reason\` and no \`terminal_reason\`. \`terminal_reason\` is a categorical label emitted only when the end is not an ordinary exit: \`timed_out\` when the deadline aborted it, \`stopped\` when it was explicitly stopped, or \`failed\` when it errored without producing an exit code; the \`stopped\` and \`failed\` cases also carry a human-readable \`stop_reason\`. A task that finished on its own with a clean exit carries neither \`stop_reason\` nor \`terminal_reason\`.\\n- The full, never-truncated log is always available at output_path; use the \`Read\` tool with that path to page through it, whether or not the preview was truncated.\\n- This tool works with the generic background task system and should remain the primary read path for future task types, not just bash.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "task_id": { "type": "string", "description": "The background task ID to inspect." } }, "required": [ "task_id" ], "additionalProperties": false } }, { "name": "TaskStop", "description": "Stop a running background task.\\n\\nOnly use this when a task must genuinely be cancelled — for a task that is\\nfinishing normally, wait for its completion notification or inspect it with\\n\`TaskOutput\` instead of stopping it.\\n\\nGuidelines:\\n- This is a general-purpose stop capability for any background task. It is not\\n  a bash-specific kill.\\n- Stopping a task is destructive: it may leave partial side effects behind.\\n  Use it with care.\\n- If the task has already finished, this tool simply returns its current\\n  status.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "task_id": { "type": "string", "description": "The background task ID to stop." }, "reason": { "default": "Stopped by TaskStop", "description": "Short reason recorded when the task is stopped.", "type": "string" } }, "required": [ "task_id" ], "additionalProperties": false } }, { "name": "TodoList", "description": "Use this tool to maintain a structured TODO list as you work through a multi-step task. Use it proactively and often when progress tracking helps the current work. This is especially useful in long-running investigations and implementation tasks with several tool calls; in plan mode, write the plan to the plan file rather than tracking it here.\\n\\n**When to use:**\\n- Multi-step tasks that span several tool calls\\n- Tracking investigation progress across a large codebase search\\n- Planning a sequence of edits before making them\\n- After receiving new multi-step instructions, capture the requirements as todos\\n- Before starting a tracked task, mark exactly one item as \`in_progress\`\\n- Immediately after finishing a tracked task, mark it \`done\`; do not batch completions at the end\\n\\n**When NOT to use:**\\n- Single-shot answers that complete in one or two tool calls\\n- Trivial requests where tracking adds no clarity\\n- Purely conversational or informational replies\\n\\n**Avoid churn:**\\n- Do not re-call this tool when nothing meaningful has changed since the last call — update the list only after real progress.\\n- When unsure of the current state, call query mode first (omit \`todos\`) to check the list before deciding what to update.\\n- If no available tool can move any task forward, tell the user where you are stuck instead of repeatedly re-ordering the same todos.\\n\\n**How to use:**\\n- Call with \`todos: [...]\` to replace the full list. Statuses: pending / in_progress / done.\\n- Call with no \`todos\` argument to retrieve the current list without changing it.\\n- Call with \`todos: []\` to clear the list.\\n- Keep titles short and actionable (e.g. \\"Read session-control.ts\\", \\"Add planMode flag to TurnManager\\").\\n- Update statuses as you make progress.\\n- When work is underway, keep exactly one task \`in_progress\`.\\n- Only mark a task \`done\` when it is fully accomplished.\\n- Never mark a task \`done\` if tests are failing, implementation is partial, unresolved errors remain, or required files/dependencies could not be found.\\n- If you encounter a blocker, keep the blocked task \`in_progress\` or add a new pending task describing what must be resolved.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "todos": { "description": "The updated todo list. Omit to read the current todo list without making changes. Pass an empty array to clear the list.", "type": "array", "items": { "type": "object", "properties": { "title": { "type": "string", "minLength": 1, "description": "Short, actionable title for the todo." }, "status": { "type": "string", "enum": [ "pending", "in_progress", "done" ], "description": "Current status of the todo." } }, "required": [ "title", "status" ], "additionalProperties": false } } }, "additionalProperties": false } }, { "name": "UpdateGoal", "description": "Set the status of the current goal. This is how you resume, complete, or block an autonomous goal.\\n\\n- \`active\` — resume a paused or blocked goal when the user explicitly asks you to work on that goal.\\n- \`complete\` — the objective is satisfied and any stated validation has passed. The goal ends and a completion summary is recorded. Before using this, verify the current state against the actual objective and every explicit requirement. Treat weak or indirect evidence as not complete. Do not use \`complete\` merely because a budget is nearly exhausted or you want to stop.\\n- \`blocked\` — a genuine impasse prevents useful progress: an external condition, required user input, missing credentials or permissions, a persistent technical failure, or an impossible, unsafe, or contradictory objective. For non-terminal blockers, do not use \`blocked\` the first time you hit the blocker. The same blocking condition must repeat for at least 3 consecutive goal turns before you call \`blocked\`, counting the original/user-triggered turn and automatic continuations. If a previously blocked goal is resumed, treat the resumed run as a fresh blocked audit. If the objective itself is impossible, unsafe, or contradictory, call \`blocked\` in the same turn instead of running more goal turns. Do not use \`blocked\` because the work is large, hard, slow, uncertain, incomplete, still needs validation, would benefit from clarification, or needs more goal turns. Once the 3-turn threshold is met and you cannot make meaningful progress without user input or an external-state change, call \`blocked\` instead of leaving the goal active.\\n\\nMost active goal turns should not call this tool. If you complete one useful slice of work and material work remains, end the turn normally without calling UpdateGoal; the runtime will prompt you to continue in the next goal turn. Call \`complete\` only when all required work is done, any stated validation has passed, and there is no useful next action. Do not call \`complete\` after only producing a plan, summary, first pass, or partial result. Call \`blocked\` only after the blocked audit threshold is met. If you call \`blocked\`, you will be prompted to explain the blocker in your next message. Setting the status is the machine-readable signal; the completion summary or blocker explanation is yours to write in the following message.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "status": { "type": "string", "enum": [ "active", "complete", "blocked" ], "description": "The lifecycle status to set for the current goal. Use \`blocked\` for impossible, unsafe, or contradictory objectives, or after the same non-terminal blocking condition repeats for at least 3 consecutive goal turns." } }, "required": [ "status" ], "additionalProperties": false } }, { "name": "WaitFor", "description": "Wait for background tasks to finish without ending the current turn.\\n\\nUse this when your next step depends on the result of a running background task (a sub-agent, a background bash command, or a background AskUserQuestion). The call suspends inside the current turn until the task finishes or the timeout elapses, then returns the outcome so you can keep working in the same turn. While waiting, no LLM requests are made.\\n\\nGuidelines:\\n\\n- Do not call WaitFor right after dispatching work whose result you do not need yet — finished background tasks notify you automatically. WaitFor is for the moment you genuinely cannot proceed without a result.\\n- \`timeout\` is required, in seconds, capped at 600. To wait longer, call WaitFor again; waking up periodically also lets you re-evaluate the situation.\\n- A timeout is not an error: the result lists the tasks that are still running, and you decide whether to wait again or do other work meanwhile.\\n- Without \`task_id\`, the wait ends as soon as any background task that was running at call time finishes. Tasks started during the wait are not covered by it; their completion arrives via the usual automatic notification.\\n- With \`task_id\`, the wait ends when that task finishes. An unknown \`task_id\` is an error; a task that has already finished returns immediately.\\n- When no background tasks are running, WaitFor returns immediately without waiting.\\n- When the wait ends because a task finished, the result also lists other tasks that finished during the wait window, so failures surface with context.\\n- Waiting has no side effects on the waited tasks: WaitFor never stops a task, and interrupting the wait (for example, a user interruption) leaves every task running.\\n- A finished task's result is delivered exactly once: tasks reported by WaitFor do not also produce an automatic completion notification.\\n- You can only wait for background tasks started by this agent; task IDs belonging to other agents are unknown here.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "timeout": { "type": "integer", "exclusiveMinimum": 0, "maximum": 600, "description": "Maximum time to wait, in seconds (1-600). A timeout is not an error: the tool returns the tasks that are still running, and you can call it again to keep waiting." }, "task_id": { "description": "The background task ID to wait for. When omitted, the wait ends as soon as any background task that was running at call time finishes.", "type": "string" } }, "required": [ "timeout" ], "additionalProperties": false } }, { "name": "Write", "description": "Create, append to, or replace a file entirely.\\n\\n- Missing parent directories are created automatically (like \`mkdir(parents=True, exist_ok=True)\`).\\n- Mode defaults to overwrite; append adds content at EOF without adding a newline.\\n- Write is NOT ALLOWED for incremental changes to existing files, including trivial, one-line, quick, or cosmetic edits. Use Edit instead.\\n- Use Write only when the file does not exist, you intend a complete replacement, or the new contents have little continuity with the old contents.\\n- Do not create unsolicited documentation files (\`*.md\` write-ups, \`README\`s, summaries) just because a task finished — write one only when the user asks for it, or when a task or project instruction requires it (e.g. the plan-mode plan file, created with Write when plan mode directs you to, or a changeset the repo mandates).\\n- Read before overwriting an existing file.\\n- Write ignores the Read/Edit line-number view. NEVER include line prefixes.\\n- Write outputs content literally, including supplied line endings: \\\\n stays LF, \\\\r\\\\n stays CRLF.\\n- For new content too large for one call, overwrite the first chunk, then append subsequent chunks. Never chunk Write to modify an existing file.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to the file to create, append to, or completely overwrite. Relative paths resolve against the working directory; a path outside the working directory must be absolute. Missing parent directories are created automatically." }, "content": { "type": "string", "description": "Raw full file content to write exactly as provided. This does not use the Read/Edit text view." }, "mode": { "description": "Write mode. Defaults to overwrite. append adds content to the end exactly as provided and does not add a newline.", "type": "string", "enum": [ "overwrite", "append" ] } }, "required": [ "path", "content" ], "additionalProperties": false } } ], "time": "<time>" }
-        [wire] llm.request                 { "agentId": "main", "kind": "loop", "provider": "openai", "model": "mock-model", "modelAlias": "mock-model", "thinkingEffort": "off", "maxTokens": 1000000, "toolSelect": false, "systemPromptHash": "ec9c34379c88babbc468ef2f3e0e08cd2f422c8c4a910664fb8bb394d703a575", "toolsHash": "de9ec7eeaa58a8f75777f03e0ffe757f43975c59488deed982a7b2aca15a1d78", "messageCount": 2, "turnStep": "0.1", "time": "<time>" }
+        [wire] llm.tools_snapshot          { "agentId": "main", "hash": "592fbd4c44cf7bb79fd0b574697181fd58f78ff5a7864fd92f782f3e5427522f", "tools": [ { "name": "Agent", "description": "Launch a subagent to handle a task. The subagent runs as a same-process loop instance with its own context and wire file. Delegating also keeps the bulk of intermediate file contents out of your own context — you get a conclusion back instead of a pile of dumps.\\n\\nWriting the prompt:\\n- The subagent starts with zero context — it has not seen this conversation. Brief it like a colleague who just walked into the room: state the goal, list what you already know, hand over the specifics.\\n- Lookups (read this file, run that test): put the exact path or command in the prompt. The subagent should not have to search for things you already know.\\n- Investigations (figure out X, find why Y): give the question, not prescribed steps — fixed steps become dead weight when the premise is wrong.\\n- Do not delegate understanding. If the task hinges on a file path or line number, find it yourself first and write it into the prompt.\\n\\nUsage notes:\\n- When the task continues earlier work a subagent already did, prefer resuming that agent (pass its \`resume\` id) over spawning a fresh instance — the resumed agent keeps its prior context.\\n- A subagent's result is only visible to you, not to the user. When the user needs to see what a subagent produced, summarize the relevant parts yourself in your own reply.\\n- Subagents use a fixed 2-hour timeout. If one times out, resume the same agent instead of starting over.\\n\\nWhen NOT to use Agent: skip delegation for trivial work you can do directly — reading a file whose path you already know, searching a small known set of files, or any task that takes only a step or two. Delegation has a context-handoff cost; it pays off only when the task is substantial enough to outweigh it.\\n\\nOnce a subagent is running, leave that scope to it: do not redo its searches or reads in parallel, and do not abandon it midway and finish the job manually. Both undo the context savings the delegation was meant to buy.\\n\\n\\nWhen \`run_in_background=true\`, the subagent runs detached from this turn. The completion arrives in a later turn as a synthetic user-role message containing its result — you do not need to poll, sleep, or check on its progress. Continue with other work or respond to the user. Never fabricate or predict what the result will say.\\n\\nDefault to a foreground subagent (omit \`run_in_background\`) when your next step needs its result — foreground hands the result straight back. Reach for \`run_in_background=true\` only when you have other work to do while it runs and do not need its result to proceed. Never launch in the background and then immediately wait on it (by polling \`TaskOutput\`, sleeping, or otherwise): that just blocks the turn for no benefit — run it in the foreground instead.\\n\\n\\nAvailable agent types (pass via subagent_type):\\n- plan: Read-only implementation planning and architecture design. Use this agent when the parent agent needs a step-by-step implementation plan, key file identification, and architectural trade-off analysis before code changes are made.\\n  Tools: Read, ReadMediaFile, Glob, Grep, WebSearch, FetchURL\\n- coder: General software engineering agent — the only subagent type with file-editing tools; use it for any delegated task that must modify code. Use this agent for non-trivial software engineering work that may require reading files, editing code, running commands, and returning a compact but technically complete summary to the parent agent.\\n  Tools: Bash, CronCreate, CronDelete, CronList, Edit, EnterPlanMode, ExitPlanMode, Glob, Grep, Read, ReadMediaFile, Skill, TaskList, TaskOutput, TaskStop, TodoList, WaitFor, WebSearch, FetchURL, Write, mcp__*\\n- explore: Fast codebase exploration with prompt-enforced read-only behavior. Fast agent specialized for exploring codebases. Use this when you need to quickly find files by patterns (e.g. \\"src/**/*.yaml\\"), search code for keywords (e.g. \\"database connection\\"), or answer questions about the codebase (e.g. \\"how does the auth module work?\\"). When calling this agent, specify the desired thoroughness level: \\"quick\\" for basic searches, \\"medium\\" for moderate exploration, or \\"thorough\\" for comprehensive analysis across multiple locations and naming conventions. Use this agent for any read-only exploration that will clearly require more than 3 search queries. Prefer launching multiple explore agents concurrently when investigating independent questions.\\n  Tools: Bash, Read, ReadMediaFile, Glob, Grep, WebSearch, FetchURL", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "prompt": { "type": "string", "description": "Full task prompt for the subagent" }, "description": { "type": "string", "description": "Short task description (3-5 words) for UI display" }, "subagent_type": { "description": "One of the available agent types (see \\"Available agent types\\" in this tool description). Defaults to \\"coder\\" when omitted.", "type": "string" }, "resume": { "description": "Optional agent ID to resume instead of creating a new instance. When set, do not also pass subagent_type — the resumed agent keeps its own type, and supplying both is rejected.", "type": "string" }, "run_in_background": { "description": "If true, return immediately without waiting for completion. Prefer false unless the task can run independently and there is a clear benefit to not waiting.", "type": "boolean" } }, "required": [ "prompt", "description" ], "additionalProperties": false } }, { "name": "AgentDynamicWorkflow", "description": "Launch multiple subagents from one prompt template, existing agent resumes, or both.\\n\\nUse AgentDynamicWorkflow when many subagents should run the same kind of task over different inputs. The placeholder is exactly \`{{item}}\`. For example, with \`prompt_template\` set to \`Review {{item}} for likely regressions.\` and \`items\` set to \`[\\"src/a.ts\\", \\"src/b.ts\\"]\`, AgentDynamicWorkflow launches two new subagents with those two concrete prompts. For a few differently-shaped tasks, make separate \`Agent\` calls in one message instead.\\n\\nUse \`resume_agent_ids\` to continue subagents that already exist from earlier work, such as ones that failed or timed out: map each agent id to the prompt for that resumed subagent (usually \`continue\` if no extra information is needed). You may combine \`resume_agent_ids\` with \`items\` in the same call to resume existing subagents and launch new ones. Do not duplicate resumed work in \`items\`.\\n\\nEach of these is enforced — a violation is rejected before any subagent starts: provide at least 2 \`items\` unless you pass \`resume_agent_ids\`; whenever \`items\` are present, \`prompt_template\` is required and must contain \`{{item}}\`; and the filled-in prompts must be distinct (two items that expand to the same prompt are rejected).\\n\\nUse enough subagents to keep the work focused and parallel. AgentDynamicWorkflow supports up to 128 subagents, and launches are queued automatically, so it is safe to split large tasks into many clear, independent items.\\n\\nIf \`AgentDynamicWorkflow\` is called, that call must be the only tool call in the response.\\n\\nThis capability is called a \\"dynamic workflow\\" (or simply \\"workflow\\"). Never use the word \\"swarm\\" in the \`description\` field, in subagent prompts, or when talking to the user about this tool.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "description": { "type": "string", "minLength": 1, "description": "Short description for the whole dynamic workflow. It is shown to the user as the workflow and subagent title, so word it as a workflow or run and never use the word \\"swarm\\"." }, "subagent_type": { "description": "Subagent type used for every new subagent spawned from items; defaults to coder when omitted. Resumed subagents always keep their original type, so passing subagent_type together with resume_agent_ids is allowed — it only affects the item-based spawns.", "type": "string", "minLength": 1 }, "prompt_template": { "description": "Prompt template for each subagent. The {{item}} placeholder is replaced with each item value.", "type": "string", "minLength": 1 }, "items": { "description": "Values used to fill {{item}}. Each item launches one new subagent.", "maxItems": 128, "type": "array", "items": { "type": "string", "minLength": 1 } }, "resume_agent_ids": { "description": "Map of existing subagent agent_id to the prompt used to resume that subagent. These resumed subagents are launched before new item-based subagents.", "type": "object", "propertyNames": { "type": "string", "minLength": 1 }, "additionalProperties": { "type": "string", "minLength": 1 } } }, "required": [ "description" ], "additionalProperties": false } }, { "name": "AskUserQuestion", "description": "Use this tool when you need to ask the user questions with structured options during execution. This allows you to:\\n1. Collect user preferences or requirements before proceeding\\n2. Resolve ambiguous or underspecified instructions\\n3. Let the user decide between implementation approaches as you work\\n4. Present concrete options when multiple valid directions exist\\n\\n**When NOT to use:**\\n- When you can infer the answer from context — be decisive and proceed\\n- Trivial decisions that don't materially affect the outcome\\n\\nOverusing this tool interrupts the user's flow. Only use it when the user's input genuinely changes your next action.\\n\\n**Usage notes:**\\n- Users always have an \\"Other\\" option for custom input — don't create one yourself\\n- Use multi_select to allow multiple answers to be selected for a question\\n- Keep option labels concise (1-5 words), use descriptions for trade-offs and details\\n- Each question should have 2-4 meaningful, distinct options\\n- Question texts must be unique across the call, and option labels must be unique within each question\\n- You can ask 1-4 questions at a time; group related questions to minimize interruptions\\n- If you recommend a specific option, list it first and append \\"(Recommended)\\" to its label\\n- The result is JSON with an \`answers\` object keyed by question text; each value is the chosen option's label (comma-separated labels for multi_select, or the user's own words if they picked \\"Other\\"); if \`answers\` is empty and a \`note\` says the user dismissed it, they chose not to answer — do not treat this as selecting the recommended option; decide based on context and do not re-ask the same question\\n- Set background=true when you can keep working without the answer. This starts a background question task and returns a task_id immediately. The answer arrives automatically in a later turn — you do not need to poll, sleep, or check on it. Continue with other work; never fabricate or predict the answer.", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "questions": { "minItems": 1, "maxItems": 4, "type": "array", "items": { "type": "object", "properties": { "question": { "type": "string", "minLength": 1, "description": "A specific, actionable question. End with '?'." }, "header": { "default": "", "description": "Short category tag (max 12 chars, e.g. 'Auth', 'Style').", "type": "string" }, "options": { "minItems": 2, "maxItems": 4, "type": "array", "items": { "type": "object", "properties": { "label": { "type": "string", "minLength": 1, "description": "Concise display text (1-5 words). If recommended, append '(Recommended)'." }, "description": { "default": "", "description": "Brief explanation of trade-offs or implications.", "type": "string" } }, "required": [ "label" ], "additionalProperties": false }, "description": "2-4 meaningful, distinct options. Do NOT include an 'Other' option — the system adds one automatically." }, "multi_select": { "default": false, "description": "Whether the user can select multiple options.", "type": "boolean" } }, "required": [ "question", "options" ], "additionalProperties": false }, "description": "The questions to ask the user (1-4 questions)." }, "background": { "default": false, "description": "Set true to ask in the background and return immediately with a background task_id; you are notified automatically when the user answers — do not poll with TaskOutput while the question is pending.", "type": "boolean" } }, "required": [ "questions" ], "additionalProperties": false } }, { "name": "Bash", "description": "Execute a \`bash\` command. Use this for shell semantics — pipes, env, processes, git, package managers, build/test runners, anything genuinely interactive or multi-step.\\n\\n**Translate these to a dedicated tool instead:**\\n- \`cat\` / \`head\` / \`tail\` (known path) → \`Read\`\\n- \`sed\` / \`awk\` (in-place edit) → \`Edit\`\\n- \`echo > file\` / \`cat <<EOF\` → \`Write\`\\n- \`find\` / recursive \`ls\` to locate files by name pattern → \`Glob\` (plain \`ls <known-directory>\` is fine for listing a directory)\\n- \`grep\` / \`rg\` (search file contents) → \`Grep\`\\n- \`echo\` / \`printf\` (talk to the user) → just output text directly\\n\\nThe dedicated tools render in the per-tool permission UI and keep raw stdout out of the conversation; that is why they are worth reaching for whenever one fits.\\n\\n**Output:**\\nThe stdout and stderr will be combined and returned as a string. The output may be truncated if it is too long. If the command exits non-zero, the output ends with a \`Command failed with exit code: N\` line; a command killed by its timeout or interrupted by the user ends with its own message instead.\\n\\nIf \`run_in_background=true\`, the command will be started as a background task and this tool will return a task ID instead of waiting for command completion. When doing that, you must provide a short \`description\`. Background commands default to a 600s timeout and \`timeout\` is capped at 86400s; set \`disable_timeout=true\` only when the task should run without a timeout. You will be automatically notified when the task completes. After starting one, default to returning control to the user instead of immediately waiting on it. Use \`TaskOutput\` only for a non-blocking status/output snapshot — do not wait on a task you just launched, since its completion arrives automatically. Use \`TaskStop\` only if the task must be cancelled. If a human user wants to inspect background tasks themselves, point them to the background-task panel.\\n\\n**Guidelines for safety and security:**\\n- Each shell tool call will be executed in a fresh shell environment. The shell variables, current working directory changes, and the shell history is not preserved between calls. To run a command in a particular directory, pass the \`cwd\` argument (or use absolute paths) rather than relying on a \`cd\` from an earlier call.\\n- The tool call will return after the command is finished. You shall not use this tool to execute an interactive command or a command that may run forever. For possibly long-running foreground commands, set the \`timeout\` argument in seconds. Foreground commands default to 60s and allow up to 300s. When a foreground command hits its timeout it is moved to the background instead of being killed, and you will be automatically notified when it completes. The user can also move a running foreground command to the background at any time.\\n- Avoid using \`..\` to access files or directories outside of the working directory.\\n- Avoid modifying files outside of the working directory unless explicitly instructed to do so.\\n- Never run commands that require superuser privileges unless explicitly instructed to do so.\\n\\n**Guidelines for efficiency:**\\n- Use \`&&\` to chain commands that genuinely depend on each other, e.g. \`npm install && npm test\`. Independent read-only commands (separate \`git show\`, \`ls\`, or status checks) should be issued as separate parallel Bash calls in one response, not chained into a single call — chaining serializes their execution and mixes their output. Do not stitch outputs together with \`echo\` separators.\\n- Use \`;\` to run commands sequentially regardless of success/failure\\n- Use \`||\` for conditional execution (run second command only if first fails)\\n- Use pipe operations (\`|\`) and redirections (\`>\`, \`>>\`) to chain input and output between commands\\n- Always quote file paths containing spaces with double quotes (e.g., cd \\"/path with spaces/\\")\\n- Compose multi-step logic in a single call with \`if\` / \`case\` / \`for\` / \`while\` control flows.\\n- Prefer \`run_in_background=true\` for long-running builds, tests, watchers, or servers when you need the conversation to continue before the command finishes.\\n\\n**Commands available:**\\nThe following common command categories are usually available. Availability still depends on the host, so when in doubt run \`which <command>\` first to confirm a command exists before relying on it.\\n- Navigation and inspection: \`ls\`, \`pwd\`, \`cd\`, \`stat\`, \`file\`, \`du\`, \`df\`, \`tree\`\\n- File and directory management: \`cp\`, \`mv\`, \`rm\`, \`mkdir\`, \`touch\`, \`ln\`, \`chmod\`, \`chown\`\\n- Text and data processing: \`wc\`, \`sort\`, \`uniq\`, \`cut\`, \`tr\`, \`diff\`, \`xargs\`\\n- Archives and compression: \`tar\`, \`gzip\`, \`gunzip\`, \`zip\`, \`unzip\`\\n- Networking and transfer: \`curl\`, \`wget\`, \`ping\`, \`ssh\`, \`scp\`\\n- Version control: \`git\`; for GitHub-hosted work (PRs, issues, CI runs, API queries) prefer the \`gh\` CLI when installed — it carries the user's GitHub auth and can return structured JSON\\n- Process and system: \`ps\`, \`kill\`, \`top\`, \`env\`, \`date\`, \`uname\`, \`whoami\`\\n- Language and package toolchains: \`node\`, \`npm\`, \`pnpm\`, \`yarn\`, \`python\`, \`pip\` (use whichever the project actually relies on)\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "command": { "type": "string", "minLength": 1, "description": "The command to execute." }, "cwd": { "description": "The working directory in which to run the command. When omitted, the command runs in the session's working directory.", "type": "string" }, "timeout": { "default": 60, "description": "Optional timeout in seconds for the command to execute. Foreground default 60s, max 300s. Background default 600s, max 86400s. Ignored for background commands when disable_timeout=true.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 }, "description": { "description": "A short description for the background task. Required when run_in_background is true.", "type": "string" }, "run_in_background": { "description": "Whether to run the command as a background task.", "type": "boolean" }, "disable_timeout": { "description": "If true, do not apply a timeout to the command. Only applies when run_in_background is true.", "type": "boolean" } }, "required": [ "command" ], "additionalProperties": false } }, { "name": "CreateGoal", "description": "Create a durable, structured goal that the runtime will pursue across multiple turns.\\n\\nCall \`CreateGoal\` only when:\\n\\n- the user explicitly asks you to start a goal or work autonomously toward an outcome, or\\n- a host goal-intake prompt asks you to create one.\\n\\nDo NOT create a goal for greetings, ordinary questions, or vague requests that lack a\\nverifiable completion condition. A goal needs a checkable end state.\\n\\nWhen the request is vague, ask the user for the missing completion criterion before creating\\nthe goal. If the user clearly insists after you warn them that the wording is vague or risky,\\nrespect that and create the goal.\\n\\nInclude a \`completionCriterion\` when the user provides one, or when it can be stated without\\ninventing new requirements. Keep \`objective\` concise; reference long task descriptions by file\\npath rather than pasting them.\\n\\nCreating a goal fails if one already exists, so use \`replace: true\` only when the user explicitly\\nwants to abandon the current goal and start a new one.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "objective": { "type": "string", "minLength": 1, "description": "The objective to pursue. Must have a verifiable end state." }, "completionCriterion": { "description": "How to verify the goal is complete. Include when the user provides one.", "type": "string" }, "replace": { "description": "Replace an existing active, paused, or blocked goal instead of failing.", "type": "boolean" } }, "required": [ "objective" ], "additionalProperties": false } }, { "name": "CronCreate", "description": "Schedule a prompt to be enqueued at a future time. Use for both recurring schedules and one-shot reminders.\\n\\nUses standard 5-field cron in the user's local timezone: minute hour day-of-month month day-of-week. \`0 9 * * *\` means 9am local — no timezone conversion needed.\\n\\n## One-shot tasks (recurring: false)\\n\\nFor \\"remind me at X\\" or \\"at <time>, do Y\\" requests — fire once then auto-delete.\\nPin minute/hour/day-of-month/month to specific values:\\n  \\"remind me at 2:30pm today to check the deploy\\" → cron: \\"30 14 <today_dom> <today_month> *\\", recurring: false\\n  \\"tomorrow morning, run the smoke test\\" → cron: \\"57 8 <tomorrow_dom> <tomorrow_month> *\\", recurring: false\\n\\nOne-shots are best for near-term reminders. A task only fires while its session is still alive (see Session lifetime below), so favor near times — within hours or a few days — rather than scheduling weeks or months ahead.\\n\\n## Recurring jobs (recurring: true, the default)\\n\\nFor \\"every N minutes\\" / \\"every hour\\" / \\"weekdays at 9am\\" requests:\\n  \\"*/5 * * * *\\" (every 5 min), \\"0 * * * *\\" (hourly), \\"0 9 * * 1-5\\" (weekdays at 9am local)\\n\\n## Avoid the :00 and :30 minute marks when the task allows it\\n\\nEvery user who asks for \\"9am\\" gets \`0 9\`, and every user who asks for \\"hourly\\" gets \`0 *\` — which means requests from across the planet land on the API at the same instant. When the user's request is approximate, pick a minute that is NOT 0 or 30:\\n  \\"every morning around 9\\" → \\"57 8 * * *\\" or \\"3 9 * * *\\" (not \\"0 9 * * *\\")\\n  \\"hourly\\" → \\"7 * * * *\\" (not \\"0 * * * *\\")\\n  \\"in an hour or so, remind me to...\\" → pick whatever minute you land on, don't round\\n\\nOnly use minute 0 or 30 when the user names that exact time and clearly means it (\\"at 9:00 sharp\\", \\"at half past\\", coordinating with a meeting). When in doubt, nudge a few minutes early or late — the user will not notice, and the fleet will.\\n\\n## Coalesce semantics\\n\\nFires are delivered only while the session is idle: a fire that comes due during an active turn is held and delivered at the next idle moment, never injected mid-turn.\\n\\nIf the scheduler slept past multiple ideal fire times (laptop closed, long-running turn, etc.), only **one** fire is delivered when it wakes up. The origin carries \`coalescedCount\` showing how many ideal fires were collapsed into this single delivery. You should treat \`coalescedCount > 1\` as \\"I missed some checks; only the latest state matters\\" rather than running the prompt that many times.\\n\\n## Cron-fire envelope\\n\\nWhen a cron task fires, the prompt you scheduled is re-injected wrapped in an XML envelope that exposes the fire context:\\n\\n\`\`\`\\n<cron-fire jobId=\\"...\\" cron=\\"...\\" recurring=\\"true|false\\" coalescedCount=\\"N\\" stale=\\"true|false\\">\\n<prompt>\\nyour original prompt text, verbatim\\n</prompt>\\n</cron-fire>\\n\`\`\`\\n\\nThe envelope is parseable. Use \`coalescedCount > 1\` to know multiple ideal fires were collapsed into a single delivery (treat as \\"only the latest state matters\\"), and \`stale=\\"true\\"\` as a cue that the task is past its 7-day threshold.\\n\\n## 7-day stale behavior\\n\\nRecurring tasks that have been alive for more than 7 days fire one\\nfinal time with \`stale: true\` on the envelope, and the system then\\nauto-deletes the task. The flag is the model's notice that this is\\nthe last delivery. If the schedule is still wanted, call \`CronCreate\`\\nagain with the same \`cron\` and \`prompt\` — that resets \`createdAt\` and\\nstarts a fresh 7-day window. One-shot tasks are never marked stale.\\n\\n## Jitter behavior\\n\\nAnti-herd jitter is applied deterministically per task id:\\n  - Recurring: ideal fire time is shifted **forward** by an offset ≤ min(10% of the cron period, 15 minutes). A \`*/5 * * * *\` task can drift up to 30s; a \`0 9 * * *\` task can drift up to 15 minutes.\\n  - One-shot: only when the ideal fire lands on \`:00\` or \`:30\` of the hour, the fire is pulled **earlier** by ≤ 90 seconds. Other minutes pass through unchanged.\\n\\n## One-shot vs recurring — when to pick which\\n\\nUse \`recurring: false\` for \\"remind me at X\\" style requests, single deadlines, \\"in N minutes do Y\\", and any task that should not repeat. Use \`recurring: true\` for periodic polling (CI status, build watchers, scheduled reports), workday rituals, and anything the user explicitly described as recurring.\\n\\n## Session lifetime\\n\\nCron tasks live in the current session. When you exit, they\\nare persisted under the session homedir; resuming the same session\\nreloads them and the scheduler resumes from each task's \`createdAt\`. Fire times that fell during the offline window are\\ncollapsed into a single delivery via \`coalescedCount\` (and recurring\\ntasks past their 7-day window arrive with \`stale: true\` as their final\\ndelivery).\\n\\nTasks do **not** carry over into a brand-new session — they are scoped\\nto the resumed session id, not to the working directory.\\n\\n## Limits\\n\\nA session holds at most 50 live cron tasks; creating one beyond that is rejected. (The \`prompt\` body is also capped — see its parameter description.) Expressions that never fire within the next 5 years (e.g. \`0 0 31 2 *\`, an impossible date) are rejected at create time.\\n\\n## Returned fields\\n\\n\`id\` (ULID), \`cron\` (the normalized expression), \`humanSchedule\` (English summary), \`recurring\`,\\n\`nextFireAt\` (local ISO timestamp with numeric offset, or null). \`id\` is needed by \`CronDelete\`.\\n\\n## Tell the user how to cancel or modify\\n\\nAfter successfully creating a task, proactively tell the user how they can cancel or modify it later. Users have no direct \`/cron\` command or self-service UI to manage reminders themselves; they must ask the model to make changes (e.g. \\"cancel my 9am reminder\\" or \\"change my daily check to 10am\\"). Include the task \`id\` in your message so the user can reference it.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "cron": { "type": "string", "description": "5-field cron expression in local time: \\"M H DoM Mon DoW\\" (e.g. \\"*/5 * * * *\\" = every 5 minutes; \\"30 14 28 2 *\\" = Feb 28 at 2:30pm local — a pinned date like this repeats yearly unless you also pass recurring: false)." }, "prompt": { "type": "string", "minLength": 1, "maxLength": 8192, "description": "The prompt to enqueue at each fire time. Limited to 8 KiB (UTF-8)." }, "recurring": { "default": true, "description": "true (default) = fire on every cron match until deleted or auto-expired after 7 days. false = fire once at the next match, then auto-delete. Use false for \\"remind me at X\\" one-shot requests with pinned minute/hour/dom/month.", "type": "boolean" } }, "required": [ "cron", "prompt" ], "additionalProperties": false } }, { "name": "CronDelete", "description": "Cancel a scheduled cron job by id.\\n\\nUse this tool to remove a cron task previously scheduled with\\n\`CronCreate\`. The \`id\` is the ULID value returned by \`CronCreate\`, or\\nshown in the \`id:\` column of \`CronList\` — quote it verbatim, no\\nprefix.\\n\\nBehaviour by task kind:\\n\\n- **Recurring task** (\`recurring: true\`): stops all future fires\\n  immediately. The scheduler picks up the deletion on its next tick.\\n- **One-shot task** (\`recurring: false\`): cancels the pending fire if\\n  it has not happened yet. One-shots that have already fired\\n  auto-delete themselves, so calling \`CronDelete\` on a fired one-shot\\n  returns \\"no cron job with id ...\\".\\n\\nNot-found is reported as an error (not a silent no-op) so you can\\ncorrect yourself — typically by calling \`CronList\` to see which ids\\nare actually live, rather than re-trying with the same stale id.\\n\\nRefresh pattern (use when you want a stale recurring schedule to\\ncontinue):\\n\\nStale recurring tasks are auto-deleted by the system after their final\\nfire — there is nothing for \`CronDelete\` to remove at that point. To\\nkeep the schedule running, just call \`CronCreate\` with the same \`cron\`\\nand \`prompt\`. Use \`CronList\`'s \`prompt\` field to recall the original\\ntext after a context compaction.\\n\\n\`CronDelete\` remains the right call when you want to cancel a task\\nthat is still live (recurring not yet stale, or a one-shot still\\npending).\\n\\nGuidelines:\\n\\n- Users have no direct \`/cron\` command or self-service UI to delete\\n  tasks themselves; they must ask the model to cancel a reminder.\\n  When deleting on behalf of a user, confirm the action and report\\n  the result plainly.\\n- Cron deletion is irreversible — there is no undo. If you delete the\\n  wrong task, you must re-create it with \`CronCreate\`.\\n- If the model is unsure which id is current (e.g. after a context\\n  compaction), call \`CronList\` first rather than guessing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "id": { "type": "string", "description": "The cron job id (ULID) returned by CronCreate / CronList." } }, "required": [ "id" ], "additionalProperties": false } }, { "name": "CronList", "description": "List all cron jobs currently scheduled in this session.\\n\\nUse this tool to see every pending cron task — both recurring jobs and\\none-shot reminders — that you (or the user) have scheduled with\\n\`CronCreate\`. The output is the entry point for inspecting scheduled\\nwork: it returns a stable id, the original cron expression, a human\\nrendering, the next post-jitter fire time, the recurring flag, the\\ntask's age in days, and a stale indicator.\\n\\nEach record carries:\\n\\n- \`id\` — the task id (a ULID). Pass this to \`CronDelete\` to remove the\\n  task, or quote it in user-facing messages when asking for\\n  confirmation.\\n- \`cron\` — the verbatim 5-field cron expression as scheduled.\\n- \`humanSchedule\` — plain-English rendering (e.g. \`every 5 minutes\`).\\n- \`prompt\` — the scheduled prompt text, JSON-encoded so embedded\\n  newlines stay on one line. Truncated to 200 UTF-8 bytes with\\n  \`…(truncated)\` if longer. Use this to recall what a task is for\\n  after a context compaction, and as the source for the\\n  \`CronCreate\` refresh ritual.\\n- \`nextFireAt\` — local ISO timestamp with an explicit numeric offset\\n  for the next fire **after jitter has been applied**. The actual fire\\n  may land slightly before or after a round \`:00\` / \`:30\` minute mark\\n  due to herd-avoidance jitter; this is the value the scheduler will\\n  compare against, so it reflects what will really happen. \`null\` if\\n  the expression has no fire in the next 5 years (should not happen\\n  for tasks created through \`CronCreate\`, which validates).\\n- \`recurring\` — \`true\` for cadenced jobs, \`false\` for one-shots.\\n- \`ageDays\` — \`(now - createdAt) / day\`, two decimal places. Useful\\n  when deciding whether a long-running cron is still relevant.\\n- \`stale\` — \`true\` when a recurring task is older than 7 days. The\\n  system **auto-deletes the task after this fire** to bound session\\n  lifetime; the \`stale: true\` flag is the model's notice that this is\\n  the final delivery. To resume the same schedule, call \`CronCreate\`\\n  again with the original \`cron\` and \`prompt\` (the \`prompt\` row above\\n  carries it for exactly this purpose). One-shots are never marked\\n  stale — they fire at most once by construction.\\n\\nGuidelines:\\n\\n- This tool is read-only and never mutates state, so it is always\\n  safe to call (including in plan mode).\\n- Users cannot directly manage cron tasks themselves; if they want to\\n  cancel or modify a schedule, route the request through the model\\n  (i.e. call \`CronDelete\` or \`CronCreate\` on their behalf).\\n- The empty case returns \`cron_jobs: 0\\\\nNo cron jobs scheduled.\`. Cron\\n  tasks survive a resume of the same session but do not bleed into new\\n  sessions.\\n- After a context compaction, or whenever you are unsure which cron\\n  jobs are live, call this tool to re-enumerate them rather than\\n  guessing ids from earlier in the conversation.\\n- Records are separated by a line containing just \`---\`, in the\\n  insertion order they were scheduled.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "Edit", "description": "Perform exact replacements in existing files.\\n\\n- Edit is mandatory for every incremental change, especially small edits. DO NOT use Write or Bash \`sed\`.\\n- Read the target file before every Edit. DO NOT call Edit from memory, stale context, or a guessed \`old_string\`.\\n- Take \`old_string\` and \`new_string\` from the Read output view.\\n- Drop the line-number prefix and tab; match only file content.\\n- \`old_string\` must be unique unless \`replace_all\` is set.\\n- If \`old_string\` is ambiguous, add surrounding context. Use \`replace_all\` only when every occurrence should change — for example, renaming a symbol throughout the file.\\n- Multiple Edit calls may run in one response only when they do not target the same file.\\n- DO NOT issue consecutive Edit calls on the same file. A previous Edit can invalidate a later Edit's \`old_string\`, causing \`old_string not found\`. Read the file again before the next Edit.\\n- A write lock serializes same-file edits in response order, but serialization does not make stale \`old_string\` valid.\\n- For pure CRLF files, Read shows LF; use LF in \`old_string\` and \`new_string\`, and Edit writes CRLF back.\\n- For mixed endings or lone carriage returns, Read shows carriage returns as \\\\r; include actual \\\\r escapes in those positions.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to the text file to edit. Relative paths resolve against the working directory; a path outside the working directory must be absolute." }, "old_string": { "type": "string", "minLength": 1, "description": "Exact content to replace from the Read output view, without the line-number prefix. Use LF for pure CRLF files; use actual \\\\r escapes where Read shows \\\\r." }, "new_string": { "type": "string", "description": "Replacement text in the same Read output view. LF is written back as CRLF only for pure CRLF files." }, "replace_all": { "description": "Set true only when every occurrence of old_string should be replaced.", "type": "boolean" } }, "required": [ "path", "old_string", "new_string" ], "additionalProperties": false } }, { "name": "EnterPlanMode", "description": "Use this tool proactively when you're about to start a non-trivial implementation task.\\nGetting user sign-off on your approach via ExitPlanMode before writing code prevents wasted effort.\\n\\nUse it when ANY of these conditions apply:\\n\\n1. New Feature Implementation - e.g. \\"Add a caching layer to the API\\"\\n2. Multiple Valid Approaches - e.g. \\"Optimize database queries\\" (indexing vs rewrite vs caching)\\n3. Code Modifications - e.g. \\"Refactor auth module to support OAuth\\"\\n4. Architectural Decisions - e.g. \\"Add WebSocket support\\"\\n5. Multi-File Changes - involves more than 2-3 files\\n6. Unclear Requirements - need exploration to understand scope\\n7. User Preferences Matter - if user input would materially change the implementation approach, use EnterPlanMode to structure the decision\\n\\nPermission mode notes:\\n- EnterPlanMode enters plan mode automatically without an approval prompt in all permission modes.\\n- In yolo and manual modes, ExitPlanMode still presents the plan to the user for approval.\\n- In auto permission mode, do not use AskUserQuestion; make the best decision from available context.\\n- In auto permission mode, ExitPlanMode exits plan mode without asking the user.\\n- Use EnterPlanMode only when planning itself adds value.\\n\\nWhen NOT to use:\\n- Single-line or few-line fixes (typos, obvious bugs, small tweaks)\\n- User gave very specific, detailed instructions\\n- Pure research/exploration tasks\\n\\nOnce you are in plan mode, a reminder walks you through the workflow (explore → design → write the plan file → \`ExitPlanMode\`) and enforces read-only access. For non-trivial tasks where you are unsure of the codebase structure or relevant code paths, use \`Agent(subagent_type=\\"explore\\")\` to investigate first when the \`Agent\` tool is available.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "ExitPlanMode", "description": "Use this tool when you are in plan mode and have finished writing your plan to the plan file and are ready for user approval.\\n\\n## How This Tool Works\\n- You should have already written your plan to the plan file specified in the plan mode reminder.\\n- This tool does NOT take the plan content as a parameter - it reads the plan from the file you wrote.\\n- The user will see the contents of your plan file when they review it. In auto permission mode, the tool reads the file and exits plan mode without asking the user.\\n\\n## When to Use\\nOnly use this tool for tasks that require planning implementation steps. For research tasks (searching files, reading code, understanding the codebase), do NOT use this tool.\\n\\n## What a good plan contains\\nList specific, verifiable steps grounded in the actual codebase — real files, functions, and commands, in a sensible order. Each step should be concrete enough to act on and to check. Avoid vague filler like \\"improve performance\\" or \\"add tests\\"; say what to change and where.\\n\\n## Multiple Approaches\\nIf your plan offers multiple alternative approaches, pass them via the \`options\` parameter so the user can choose which one to execute — see the \`options\` parameter for the format, count, and reserved labels. In yolo and manual modes the user sees all options alongside the host's Reject and Revise controls.\\n\\n## Before Using\\n- In auto permission mode, do NOT use AskUserQuestion; make the best decision from available context.\\n- In auto permission mode, this tool exits plan mode without asking the user.\\n- In yolo and manual modes, this tool still presents the plan to the user for approval.\\n- If auto permission mode is not active and you have unresolved questions, use AskUserQuestion first.\\n- If auto permission mode is not active and you have multiple approaches and haven't narrowed down yet, consider using AskUserQuestion first to let the user choose, then write a plan for the chosen approach only.\\n- Once your plan is finalized, use THIS tool to request approval.\\n- Do NOT use AskUserQuestion to ask \\"Is this plan OK?\\" or \\"Should I proceed?\\" - that is exactly what ExitPlanMode does.\\n- If rejected, revise based on feedback and call ExitPlanMode again.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "options": { "description": "When the plan contains multiple alternative approaches, list them here so the user can choose which one to execute. Provide up to 3 options; 2-3 distinct approaches work best when the plan offers a real choice. Passing a single option is allowed and is equivalent to a plain plan approval. Each option represents a distinct approach from the plan. Do not use \\"Reject\\", \\"Revise\\", \\"Approve\\", or \\"Reject and Exit\\" as labels.", "minItems": 1, "maxItems": 3, "type": "array", "items": { "type": "object", "properties": { "label": { "type": "string", "minLength": 1, "maxLength": 80, "description": "Short name for this option (1-8 words). Append \\"(Recommended)\\" if you recommend this option." }, "description": { "default": "", "description": "Brief summary of this approach and its trade-offs.", "type": "string" } }, "required": [ "label" ], "additionalProperties": false } } }, "additionalProperties": false } }, { "name": "FetchURL", "description": "Fetch content from a URL. The content is returned either as the main text extracted from the page, or as the full response body verbatim; a note at the top of the result states which of the two you received, so you can judge how complete it is. Use this when you need to read a specific web page.\\n\\nOnly fully-formed public \`http\`/\`https\` URLs are supported; other schemes and private or loopback addresses are not fetched. Very large pages may be truncated or refused. The fetch carries no login or session for the target site, so pages behind authentication (private repositories, internal dashboards) return a login page or an error instead of the real content — if the text you get back looks like a generic landing or sign-in page, treat that as the login wall, not the answer, and reach the content through a credentialed route (an authenticated CLI or MCP tool) instead.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "url": { "type": "string", "description": "The URL to fetch content from." } }, "required": [ "url" ], "additionalProperties": false } }, { "name": "GetGoal", "description": "Read the current goal: its objective, completion criterion, status, and budgets (turns, tokens,\\ntime, and how much of each remains). When the goal has stopped, it also reports the terminal reason.\\n\\nUse \`GetGoal\` before deciding whether to continue working, report completion, report a blocker,\\nor respect a pause. It returns \`{ \\"goal\\": null }\` when there is no current goal.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "Glob", "description": "Find files by glob pattern, sorted by modification time (most recent first).\\n\\nPowered by ripgrep. Respects \`.gitignore\`, \`.ignore\`, and \`.rgignore\` by default — set \`include_ignored\` to also match ignored files (e.g. build outputs, \`node_modules\`). Sensitive files (such as \`.env\`) are always filtered out. Matches are files only — directories themselves are never listed; to find a directory, glob for a file inside it (e.g. \`**/fixtures/**\`).\\n\\nGood patterns:\\n- \`*.ts\` — all files matching an extension, at any depth below the search root (a bare pattern without \`/\` matches recursively)\\n- \`src/*.ts\` — files directly inside \`src/\` (one level, not recursive)\\n- \`src/**/*.ts\` — recursive walk with a subdirectory anchor and extension\\n- \`**/*.py\` — recursive walk from the search root for an extension\\n- \`*.{ts,tsx}\` — brace expansion is supported\\n- \`{src,test}/**/*.ts\` — cartesian brace expansion is supported too\\n\\nResults are capped at the first 100 matching paths. If a search would return more, a truncation marker is appended. Refine the pattern (extension, subdirectory) when 100 is not enough, or call again with a narrower anchor.\\n\\nLarge-directory caveat — avoid recursing into dependency / build output even with an anchor, especially when \`include_ignored\` is set:\\n- \`node_modules/**/*.js\`, \`.venv/**/*.py\`, \`__pycache__/**\`, \`target/**\` can produce thousands of results that truncate at the match cap and waste context. Prefer specific subpaths like \`node_modules/react/src/**/*.js\`.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "pattern": { "type": "string", "description": "Glob pattern to match files." }, "path": { "description": "Directory to search. Accepts an absolute path, or a path relative to the current working directory. Defaults to the current working directory.", "type": "string" }, "include_ignored": { "description": "Also match files excluded by ignore files such as \`.gitignore\`, \`.ignore\`, and \`.rgignore\` (for example \`node_modules\` or build outputs). Sensitive files (such as \`.env\`) remain filtered out for safety. VCS metadata directories (\`.git\` and similar) are always skipped, even when this is true. Defaults to false.", "type": "boolean" }, "include_dirs": { "description": "Deprecated and ignored. Results are always files-only — directories are never listed. Accepted only so older calls that still pass this flag are not rejected by parameter validation.", "type": "boolean" } }, "required": [ "pattern" ], "additionalProperties": false } }, { "name": "Grep", "description": "Search file contents using regular expressions (powered by ripgrep).\\n\\nUse Grep when the task is to find unknown content or unknown file locations. Do not use shell \`grep\` or \`rg\` directly; this tool applies workspace path policy, output limits, and sensitive-file filtering.\\nALWAYS use Grep tool instead of running \`grep\` or \`rg\` from a shell — direct shell calls bypass workspace policy, output limits, and sensitive-file filtering.\\nIf you already know a concrete file path and need to inspect its contents, use Read directly instead.\\n\\nWrite patterns in ripgrep regex syntax, which differs from POSIX \`grep\` syntax. For example, braces are special, so escape them as \`\\\\{\` to match a literal \`{\`.\\n\\nHidden files (dotfiles such as \`.gitlab-ci.yml\` or \`.eslintrc.json\`) are searched by default. To also search files excluded by \`.gitignore\` (such as \`node_modules\` or build outputs), set \`include_ignored\` to \`true\`. Sensitive files (such as \`.env\`) are always skipped for safety, even when \`include_ignored\` is \`true\`.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "pattern": { "type": "string", "description": "Regular expression to search for." }, "path": { "description": "File or directory to search. Accepts an absolute path, or a path relative to the current working directory. Omit to search the current working directory. Use Read instead when you already know a concrete file path and need its contents.", "type": "string" }, "glob": { "description": "Optional glob filter for which files to search, e.g. \`*.ts\`. Matched against each file's full absolute path, so a path-anchored pattern like \`src/**/*.ts\` silently matches nothing — use a basename pattern (\`*.ts\`), or anchor with \`**/\` (\`**/src/**/*.ts\`). To scope the search to a directory, use \`path\` instead.", "type": "string" }, "type": { "description": "Optional ripgrep file type filter, such as ts or py. Prefer this over \`glob\` when filtering by language or file kind: it is more efficient and less error-prone than an equivalent glob pattern.", "type": "string" }, "output_mode": { "description": "Shape of the result. \`content\` shows matching lines (honors \`-A\`, \`-B\`, \`-C\`, \`-n\`, and \`head_limit\`); \`files_with_matches\` shows only the paths of files that contain a match, most-recently-modified first (honors \`head_limit\`); \`count_matches\` shows per-file match counts as \`path:count\` lines, preceded by an aggregate total line. Defaults to \`files_with_matches\`.", "type": "string", "enum": [ "content", "files_with_matches", "count_matches" ] }, "-i": { "description": "Perform a case-insensitive search. Defaults to false.", "type": "boolean" }, "-n": { "description": "Prefix each matching line with its line number. Applies only when \`output_mode\` is \`content\`. Defaults to true.", "type": "boolean" }, "-A": { "description": "Number of lines to show after each match. Applies only when \`output_mode\` is \`content\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "-B": { "description": "Number of lines to show before each match. Applies only when \`output_mode\` is \`content\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "-C": { "description": "Number of lines to show before and after each match. Applies only when \`output_mode\` is \`content\`; takes precedence over \`-A\` and \`-B\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "head_limit": { "description": "Limit output to the first N lines/entries after offset. Defaults to 250. Pass 0 for unlimited.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "offset": { "description": "Number of leading lines/entries to skip before applying \`head_limit\`. Use it together with \`head_limit\` to page through large result sets. Defaults to 0.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "multiline": { "description": "Enable multiline matching, where the pattern can span line boundaries and \`.\` also matches newlines. Defaults to false.", "type": "boolean" }, "include_ignored": { "description": "Also search files excluded by ignore files such as \`.gitignore\`, \`.ignore\`, and \`.rgignore\` (for example \`node_modules\` or build outputs). Sensitive files (such as \`.env\`) remain filtered out for safety. VCS metadata directories (\`.git\` and similar) are always skipped, even when this is true. Defaults to false.", "type": "boolean" } }, "required": [ "pattern" ], "additionalProperties": false } }, { "name": "Lookup", "description": "Look up a short test value.", "parameters": { "type": "object", "properties": { "query": { "type": "string" } }, "required": [ "query" ], "additionalProperties": false } }, { "name": "Read", "description": "Read a text file from the local filesystem.\\n\\nIf the user provides a concrete file path to a text file, call Read directly. Do not \`Glob\`, \`ls\`, or otherwise pre-check known text file paths; missing or invalid file paths return errors you can handle. Do not use Read for directories; use \`ls\` via Bash for a known directory, or Glob when you need files matching a name pattern (Glob lists files only, never directories). Use \`Grep\` only when the task is to search for unknown content or locations.\\n\\nWhen you need several files, prefer to read them in parallel: emit multiple \`Read\` calls in a single response instead of reading one file per turn.\\n\\n- Relative paths resolve against the working directory; a path outside the working directory must be absolute.\\n- Returns up to 1000 lines or 100 KB per call, whichever comes first; lines longer than 2000 chars are truncated mid-line.\\n- Page larger files with \`line_offset\` (1-based start line) and \`n_lines\`. Omit \`n_lines\` to read up to the 1000-line cap.\\n- Sensitive files (\`.env\` files, credential stores, SSH private keys, and similar secrets) are refused to protect secrets; do not attempt to read them. Templates and public keys are exempt: \`.env.example\` / \`.env.sample\` / \`.env.template\` and public SSH keys such as \`id_rsa.pub\` read normally.\\n- UTF-8 text files are read directly. UTF-16 LE/BE text files (with or without a BOM) are detected automatically and transcoded to UTF-8 for display; the status block notes the detected encoding, and Edit/Write on such a file still expect UTF-8 — convert its encoding first (e.g. with \`iconv\`). Other encodings (e.g. GBK), binary files, and files containing NUL bytes are refused.\\n- Negative line_offset reads from the end of the file (for example, -100 reads the last 100 lines); the absolute value cannot exceed 1000.\\n- Output format: \`<line-number>\\\\t<content>\` per line.\\n- A \`<system>...</system>\` status block is appended after the file content; it summarizes how much was read (line and byte counts, truncation, line-ending notes) and is not part of the file itself.\\n- Pure CRLF files are displayed with LF line endings; \`Edit\` matches this output and preserves CRLF when writing back.\\n- Mixed or lone carriage-return line endings are shown as \`\\\\r\` and require exact \`Edit.old_string\` escapes.\\n- After a successful \`Edit\`/\`Write\`, do not re-read solely to prove the write landed. When the task depends on an exact file, API, or output shape, inspect the final external contract before finishing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to a text file. Relative paths resolve against the working directory; a path outside the working directory must be absolute. Directories are not supported; use \`ls\` via Bash for a known directory, or Glob for pattern search." }, "line_offset": { "description": "The line number to start reading from. Omit to start at line 1. Negative values read from the end of the file; the absolute value cannot exceed 1000.", "anyOf": [ { "type": "integer", "minimum": 1, "maximum": 9007199254740991 }, { "type": "integer", "minimum": -1000, "maximum": -1 } ] }, "n_lines": { "description": "The number of lines to read; the tool also applies its internal cap. Omit to read up to the internal cap of 1000 lines.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 } }, "required": [ "path" ], "additionalProperties": false } }, { "name": "SetGoalBudget", "description": "Set a hard budget limit for the current goal.\\n\\nUse this only when the user clearly gives a runtime limit, such as:\\n\\n- \\"stop after 20 turns\\"\\n- \\"use no more than 500k tokens\\"\\n- \\"finish within 30 minutes\\"\\n\\nDo not invent limits. Do not call this for vague wording such as \\"spend some time\\" or\\n\\"try to be quick\\".\\n\\nIf the user gives a compound time, convert it to one supported unit before calling this tool.\\nFor example, \\"2 hours and 3 minutes\\" can be set as \`value: 123, unit: \\"minutes\\"\`.\\n\\nA time budget must be between 1 second and 24 hours — the tool rejects anything shorter or\\nlonger, telling the user it is not a reasonable goal budget. Turn and token budgets are not\\nbounded this way; they must be positive and are rounded to the nearest whole number (minimum 1).\\n\\nSupported units:\\n\\n- \`turns\`\\n- \`tokens\`\\n- \`milliseconds\`\\n- \`seconds\`\\n- \`minutes\`\\n- \`hours\`\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "value": { "type": "number", "exclusiveMinimum": 0, "description": "The positive numeric budget value." }, "unit": { "type": "string", "enum": [ "turns", "tokens", "milliseconds", "seconds", "minutes", "hours" ] } }, "required": [ "value", "unit" ], "additionalProperties": false } }, { "name": "Skill", "description": "Invoke a registered skill from the current skill listing. BLOCKING REQUIREMENT: when a skill from the listing matches the user's request, you MUST call this tool (not free-form text). Do not re-invoke a skill to repeat work already done: if a \`<skill-loaded>\` block for it with the same \`args\` is already present in the conversation, follow those instructions directly instead of calling the tool again. Do call the tool again when you need the skill with different arguments — the loaded block was expanded with the earlier \`args\` and will not reflect new inputs.", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "skill": { "type": "string", "description": "The exact name of the skill to invoke, spelled as it appears in the current skill listing (e.g. \\"commit\\", \\"pdf\\")." }, "args": { "description": "Optional argument string for the skill, written like a command line (e.g. \`-m \\"fix bug\\"\`, \`123\`, a file path). It is split on whitespace (quotes group a token) and expanded into the skill's placeholders ($NAME, $1, $ARGUMENTS); if the skill body has no placeholders, the whole string is still appended as a trailing \`ARGUMENTS:\` line. Omit it only when there is nothing to pass.", "type": "string" } }, "required": [ "skill" ], "additionalProperties": false } }, { "name": "TaskList", "description": "List background tasks and their current status.\\n\\nUse this tool to discover which background tasks exist and where each one\\nstands. It is the entry point for inspecting background work: it returns a\\ntask ID, status, and description for every task it reports, plus the command,\\nPID, and (once finished) exit code for shell tasks, and a stop reason for any\\ntask that ended early.\\n\\nGuidelines:\\n\\n- After a context compaction, or whenever you are unsure which background\\n  tasks are running or what their task IDs are, call this tool to\\n  re-enumerate them instead of guessing a task ID.\\n- Prefer the default \`active_only=true\`, which lists only non-terminal tasks.\\n  Pass \`active_only=false\` only when you specifically need to see tasks that\\n  have already finished. With \`active_only=false\` the result may also include\\n  \`lost\` tasks — tasks left over from a previous process that can no longer be\\n  inspected or controlled; treat them as already terminated.\\n- \`limit\` caps how many tasks are returned. It accepts a value between 1 and\\n  100 and defaults to 20 when omitted.\\n- This tool only lists tasks; it does not return their output. Use it first\\n  to locate the task ID you need, then call \`TaskOutput\` with that ID to read\\n  the task's output and details.\\n- This tool is read-only and does not change any state, so it is always safe\\n  to call, including in plan mode.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "active_only": { "default": true, "description": "Whether to list only non-terminal background tasks.", "type": "boolean" }, "limit": { "default": 20, "description": "Maximum number of tasks to return.", "type": "integer", "minimum": 1, "maximum": 100 } }, "additionalProperties": false } }, { "name": "TaskOutput", "description": "Retrieve a snapshot of a running or completed background task.\\n\\nUse this after \`Bash(run_in_background=true)\`, \`Agent(run_in_background=true)\`, or \`AskUserQuestion(background=true)\` to check progress, or to read the output of a task that has already completed.\\n\\nGuidelines:\\n- Prefer relying on automatic completion notifications. Use this tool only when you need task output before the automatic notification arrives.\\n- This tool is always non-blocking: it returns the current status/output snapshot immediately and never waits for the task to finish.\\n- Do not use TaskOutput to wait for a result you need before continuing — if your next step depends on the task's result, run that task in the foreground instead. TaskOutput is for a deliberate progress check you will act on without blocking, not a way to sit and wait for a background task you just launched.\\n- This tool returns structured task metadata, a fixed-size output preview, and an output_path for the full log.\\n- For a terminal task, the metadata also explains why it ended. A shell command that runs to completion reports \`status: completed\` on a zero exit, or \`status: failed\` with its non-zero \`exit_code\` — judge that failure from the \`exit_code\`, because a plain command failure carries no \`stop_reason\` and no \`terminal_reason\`. \`terminal_reason\` is a categorical label emitted only when the end is not an ordinary exit: \`timed_out\` when the deadline aborted it, \`stopped\` when it was explicitly stopped, or \`failed\` when it errored without producing an exit code; the \`stopped\` and \`failed\` cases also carry a human-readable \`stop_reason\`. A task that finished on its own with a clean exit carries neither \`stop_reason\` nor \`terminal_reason\`.\\n- The full, never-truncated log is always available at output_path; use the \`Read\` tool with that path to page through it, whether or not the preview was truncated.\\n- This tool works with the generic background task system and should remain the primary read path for future task types, not just bash.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "task_id": { "type": "string", "description": "The background task ID to inspect." } }, "required": [ "task_id" ], "additionalProperties": false } }, { "name": "TaskStop", "description": "Stop a running background task.\\n\\nOnly use this when a task must genuinely be cancelled — for a task that is\\nfinishing normally, wait for its completion notification or inspect it with\\n\`TaskOutput\` instead of stopping it.\\n\\nGuidelines:\\n- This is a general-purpose stop capability for any background task. It is not\\n  a bash-specific kill.\\n- Stopping a task is destructive: it may leave partial side effects behind.\\n  Use it with care.\\n- If the task has already finished, this tool simply returns its current\\n  status.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "task_id": { "type": "string", "description": "The background task ID to stop." }, "reason": { "default": "Stopped by TaskStop", "description": "Short reason recorded when the task is stopped.", "type": "string" } }, "required": [ "task_id" ], "additionalProperties": false } }, { "name": "TodoList", "description": "Use this tool to maintain a structured TODO list as you work through a multi-step task. Use it proactively and often when progress tracking helps the current work. This is especially useful in long-running investigations and implementation tasks with several tool calls; in plan mode, write the plan to the plan file rather than tracking it here.\\n\\n**When to use:**\\n- Multi-step tasks that span several tool calls\\n- Tracking investigation progress across a large codebase search\\n- Planning a sequence of edits before making them\\n- After receiving new multi-step instructions, capture the requirements as todos\\n- Before starting a tracked task, mark exactly one item as \`in_progress\`\\n- Immediately after finishing a tracked task, mark it \`done\`; do not batch completions at the end\\n\\n**When NOT to use:**\\n- Single-shot answers that complete in one or two tool calls\\n- Trivial requests where tracking adds no clarity\\n- Purely conversational or informational replies\\n\\n**Avoid churn:**\\n- Do not re-call this tool when nothing meaningful has changed since the last call — update the list only after real progress.\\n- When unsure of the current state, call query mode first (omit \`todos\`) to check the list before deciding what to update.\\n- If no available tool can move any task forward, tell the user where you are stuck instead of repeatedly re-ordering the same todos.\\n\\n**How to use:**\\n- Call with \`todos: [...]\` to replace the full list. Statuses: pending / in_progress / done.\\n- Call with no \`todos\` argument to retrieve the current list without changing it.\\n- Call with \`todos: []\` to clear the list.\\n- Keep titles short and actionable (e.g. \\"Read session-control.ts\\", \\"Add planMode flag to TurnManager\\").\\n- Update statuses as you make progress.\\n- When work is underway, keep exactly one task \`in_progress\`.\\n- Only mark a task \`done\` when it is fully accomplished.\\n- Never mark a task \`done\` if tests are failing, implementation is partial, unresolved errors remain, or required files/dependencies could not be found.\\n- If you encounter a blocker, keep the blocked task \`in_progress\` or add a new pending task describing what must be resolved.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "todos": { "description": "The updated todo list. Omit to read the current todo list without making changes. Pass an empty array to clear the list.", "type": "array", "items": { "type": "object", "properties": { "title": { "type": "string", "minLength": 1, "description": "Short, actionable title for the todo." }, "status": { "type": "string", "enum": [ "pending", "in_progress", "done" ], "description": "Current status of the todo." } }, "required": [ "title", "status" ], "additionalProperties": false } } }, "additionalProperties": false } }, { "name": "UpdateGoal", "description": "Set the status of the current goal. This is how you resume, complete, or block an autonomous goal.\\n\\n- \`active\` — resume a paused or blocked goal when the user explicitly asks you to work on that goal.\\n- \`complete\` — the objective is satisfied and any stated validation has passed. The goal ends and a completion summary is recorded. Before using this, verify the current state against the actual objective and every explicit requirement. Treat weak or indirect evidence as not complete. Do not use \`complete\` merely because a budget is nearly exhausted or you want to stop.\\n- \`blocked\` — a genuine impasse prevents useful progress: an external condition, required user input, missing credentials or permissions, a persistent technical failure, or an impossible, unsafe, or contradictory objective. For non-terminal blockers, do not use \`blocked\` the first time you hit the blocker. The same blocking condition must repeat for at least 3 consecutive goal turns before you call \`blocked\`, counting the original/user-triggered turn and automatic continuations. If a previously blocked goal is resumed, treat the resumed run as a fresh blocked audit. If the objective itself is impossible, unsafe, or contradictory, call \`blocked\` in the same turn instead of running more goal turns. Do not use \`blocked\` because the work is large, hard, slow, uncertain, incomplete, still needs validation, would benefit from clarification, or needs more goal turns. Once the 3-turn threshold is met and you cannot make meaningful progress without user input or an external-state change, call \`blocked\` instead of leaving the goal active.\\n\\nMost active goal turns should not call this tool. If you complete one useful slice of work and material work remains, end the turn normally without calling UpdateGoal; the runtime will prompt you to continue in the next goal turn. Call \`complete\` only when all required work is done, any stated validation has passed, and there is no useful next action. Do not call \`complete\` after only producing a plan, summary, first pass, or partial result. Call \`blocked\` only after the blocked audit threshold is met. If you call \`blocked\`, you will be prompted to explain the blocker in your next message. Setting the status is the machine-readable signal; the completion summary or blocker explanation is yours to write in the following message.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "status": { "type": "string", "enum": [ "active", "complete", "blocked" ], "description": "The lifecycle status to set for the current goal. Use \`blocked\` for impossible, unsafe, or contradictory objectives, or after the same non-terminal blocking condition repeats for at least 3 consecutive goal turns." } }, "required": [ "status" ], "additionalProperties": false } }, { "name": "WaitFor", "description": "Wait for background tasks to finish without ending the current turn.\\n\\nUse this when your next step depends on the result of a running background task (a sub-agent, a background bash command, or a background AskUserQuestion). The call suspends inside the current turn until the task finishes or the timeout elapses, then returns the outcome so you can keep working in the same turn. While waiting, no LLM requests are made.\\n\\nGuidelines:\\n\\n- Do not call WaitFor right after dispatching work whose result you do not need yet — finished background tasks notify you automatically. WaitFor is for the moment you genuinely cannot proceed without a result.\\n- \`timeout\` is required, in seconds, capped at 600. To wait longer, call WaitFor again; waking up periodically also lets you re-evaluate the situation.\\n- A timeout is not an error: the result lists the tasks that are still running, and you decide whether to wait again or do other work meanwhile.\\n- Without \`task_id\`, the wait ends as soon as any background task that was running at call time finishes. Tasks started during the wait are not covered by it; their completion arrives via the usual automatic notification.\\n- With \`task_id\`, the wait ends when that task finishes. An unknown \`task_id\` is an error; a task that has already finished returns immediately.\\n- When no background tasks are running, WaitFor returns immediately without waiting.\\n- When the wait ends because a task finished, the result also lists other tasks that finished during the wait window, so failures surface with context.\\n- Waiting has no side effects on the waited tasks: WaitFor never stops a task, and interrupting the wait (for example, a user interruption) leaves every task running.\\n- A finished task's result is delivered exactly once: tasks reported by WaitFor do not also produce an automatic completion notification.\\n- You can only wait for background tasks started by this agent; task IDs belonging to other agents are unknown here.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "timeout": { "type": "integer", "exclusiveMinimum": 0, "maximum": 600, "description": "Maximum time to wait, in seconds (1-600). A timeout is not an error: the tool returns the tasks that are still running, and you can call it again to keep waiting." }, "task_id": { "description": "The background task ID to wait for. When omitted, the wait ends as soon as any background task that was running at call time finishes.", "type": "string" } }, "required": [ "timeout" ], "additionalProperties": false } }, { "name": "Write", "description": "Create, append to, or replace a file entirely.\\n\\n- Missing parent directories are created automatically (like \`mkdir(parents=True, exist_ok=True)\`).\\n- Mode defaults to overwrite; append adds content at EOF without adding a newline.\\n- Write is NOT ALLOWED for incremental changes to existing files, including trivial, one-line, quick, or cosmetic edits. Use Edit instead.\\n- Use Write only when the file does not exist, you intend a complete replacement, or the new contents have little continuity with the old contents.\\n- Do not create unsolicited documentation files (\`*.md\` write-ups, \`README\`s, summaries) just because a task finished — write one only when the user asks for it, or when a task or project instruction requires it (e.g. the plan-mode plan file, created with Write when plan mode directs you to, or a changeset the repo mandates).\\n- Read before overwriting an existing file.\\n- Write ignores the Read/Edit line-number view. NEVER include line prefixes.\\n- Write outputs content literally, including supplied line endings: \\\\n stays LF, \\\\r\\\\n stays CRLF.\\n- For new content too large for one call, overwrite the first chunk, then append subsequent chunks. Never chunk Write to modify an existing file.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to the file to create, append to, or completely overwrite. Relative paths resolve against the working directory; a path outside the working directory must be absolute. Missing parent directories are created automatically." }, "content": { "type": "string", "description": "Raw full file content to write exactly as provided. This does not use the Read/Edit text view." }, "mode": { "description": "Write mode. Defaults to overwrite. append adds content to the end exactly as provided and does not add a newline.", "type": "string", "enum": [ "overwrite", "append" ] } }, "required": [ "path", "content" ], "additionalProperties": false } } ], "time": "<time>" }
+        [wire] llm.request                 { "agentId": "main", "kind": "loop", "provider": "openai", "model": "mock-model", "modelAlias": "mock-model", "thinkingEffort": "off", "maxTokens": 1000000, "toolSelect": false, "systemPromptHash": "ec9c34379c88babbc468ef2f3e0e08cd2f422c8c4a910664fb8bb394d703a575", "toolsHash": "592fbd4c44cf7bb79fd0b574697181fd58f78ff5a7864fd92f782f3e5427522f", "messageCount": 2, "turnStep": "0.1", "time": "<time>" }
         [emit] assistant.delta             { "time": "<time>", "agentId": "main", "turnId": 0, "delta": "I will look it up." }
         [emit] agent.activity.updated      { "time": "<time>", "lifecycle": "ready", "turn": { "turnId": 0, "origin": { "kind": "user" }, "phase": "streaming", "stream": "assistant", "step": 1, "ending": false, "pendingApprovals": [], "activeToolCalls": [], "since": "<time>" }, "background": [], "agentId": "main" }
         [emit] tool.call.delta             { "time": "<time>", "agentId": "main", "turnId": 0, "toolCallId": "call_lookup", "name": "Lookup", "argumentsPart": "{\\"query\\":\\"moon\\"}" }
@@ -3921,7 +3922,7 @@ describe('Agent tools', () => {
         [emit] turn.step.started           { "time": "<time>", "agentId": "main", "turnId": 0, "step": 2, "stepId": "<uuid-4>" }
         [emit] agent.activity.updated      { "time": "<time>", "lifecycle": "ready", "turn": { "turnId": 0, "origin": { "kind": "user" }, "phase": "running", "step": 2, "ending": false, "pendingApprovals": [], "activeToolCalls": [], "since": "<time>" }, "background": [], "agentId": "main" }
         [wire] context.append_loop_event   { "agentId": "main", "event": { "type": "step.begin", "uuid": "<uuid-4>", "turnId": "0", "step": 2 }, "time": "<time>" }
-        [wire] llm.request                 { "agentId": "main", "kind": "loop", "provider": "openai", "model": "mock-model", "modelAlias": "mock-model", "thinkingEffort": "off", "maxTokens": 1000000, "toolSelect": false, "systemPromptHash": "ec9c34379c88babbc468ef2f3e0e08cd2f422c8c4a910664fb8bb394d703a575", "toolsHash": "de9ec7eeaa58a8f75777f03e0ffe757f43975c59488deed982a7b2aca15a1d78", "messageCount": 4, "turnStep": "0.2", "time": "<time>" }
+        [wire] llm.request                 { "agentId": "main", "kind": "loop", "provider": "openai", "model": "mock-model", "modelAlias": "mock-model", "thinkingEffort": "off", "maxTokens": 1000000, "toolSelect": false, "systemPromptHash": "ec9c34379c88babbc468ef2f3e0e08cd2f422c8c4a910664fb8bb394d703a575", "toolsHash": "592fbd4c44cf7bb79fd0b574697181fd58f78ff5a7864fd92f782f3e5427522f", "messageCount": 4, "turnStep": "0.2", "time": "<time>" }
         [emit] assistant.delta             { "time": "<time>", "agentId": "main", "turnId": 0, "delta": "The lookup result is moon-result." }
         [emit] agent.activity.updated      { "time": "<time>", "lifecycle": "ready", "turn": { "turnId": 0, "origin": { "kind": "user" }, "phase": "streaming", "stream": "assistant", "step": 2, "ending": false, "pendingApprovals": [], "activeToolCalls": [], "since": "<time>" }, "background": [], "agentId": "main" }
         [wire] usage.record                { "agentId": "main", "model": "mock-model", "usage": { "inputOther": 164, "output": 12, "inputCacheRead": 0, "inputCacheCreation": 0 }, "usageScope": "turn", "time": "<time>" }
@@ -3963,8 +3964,8 @@ describe('Agent tools', () => {
         [emit] turn.step.started              { "time": "<time>", "agentId": "main", "turnId": 1, "step": 1, "stepId": "<uuid-6>" }
         [emit] agent.activity.updated         { "time": "<time>", "lifecycle": "ready", "turn": { "turnId": 1, "origin": { "kind": "user" }, "phase": "running", "step": 1, "ending": false, "pendingApprovals": [], "activeToolCalls": [], "since": "<time>" }, "background": [], "agentId": "main" }
         [wire] context.append_loop_event      { "agentId": "main", "event": { "type": "step.begin", "uuid": "<uuid-6>", "turnId": "1", "step": 1 }, "time": "<time>" }
-        [wire] llm.tools_snapshot             { "agentId": "main", "hash": "d3e22e7861e40632ba8e85ac56a6f0adea6ca3acf2f00b249b2b656657b59473", "tools": [ { "name": "Agent", "description": "Launch a subagent to handle a task. The subagent runs as a same-process loop instance with its own context and wire file. Delegating also keeps the bulk of intermediate file contents out of your own context — you get a conclusion back instead of a pile of dumps.\\n\\nWriting the prompt:\\n- The subagent starts with zero context — it has not seen this conversation. Brief it like a colleague who just walked into the room: state the goal, list what you already know, hand over the specifics.\\n- Lookups (read this file, run that test): put the exact path or command in the prompt. The subagent should not have to search for things you already know.\\n- Investigations (figure out X, find why Y): give the question, not prescribed steps — fixed steps become dead weight when the premise is wrong.\\n- Do not delegate understanding. If the task hinges on a file path or line number, find it yourself first and write it into the prompt.\\n\\nUsage notes:\\n- When the task continues earlier work a subagent already did, prefer resuming that agent (pass its \`resume\` id) over spawning a fresh instance — the resumed agent keeps its prior context.\\n- A subagent's result is only visible to you, not to the user. When the user needs to see what a subagent produced, summarize the relevant parts yourself in your own reply.\\n- Subagents use a fixed 2-hour timeout. If one times out, resume the same agent instead of starting over.\\n\\nWhen NOT to use Agent: skip delegation for trivial work you can do directly — reading a file whose path you already know, searching a small known set of files, or any task that takes only a step or two. Delegation has a context-handoff cost; it pays off only when the task is substantial enough to outweigh it.\\n\\nOnce a subagent is running, leave that scope to it: do not redo its searches or reads in parallel, and do not abandon it midway and finish the job manually. Both undo the context savings the delegation was meant to buy.\\n\\n\\nWhen \`run_in_background=true\`, the subagent runs detached from this turn. The completion arrives in a later turn as a synthetic user-role message containing its result — you do not need to poll, sleep, or check on its progress. Continue with other work or respond to the user. Never fabricate or predict what the result will say.\\n\\nDefault to a foreground subagent (omit \`run_in_background\`) when your next step needs its result — foreground hands the result straight back. Reach for \`run_in_background=true\` only when you have other work to do while it runs and do not need its result to proceed. Never launch in the background and then immediately wait on it (by polling \`TaskOutput\`, sleeping, or otherwise): that just blocks the turn for no benefit — run it in the foreground instead.\\n\\n\\nAvailable agent types (pass via subagent_type):\\n- plan: Read-only implementation planning and architecture design. Use this agent when the parent agent needs a step-by-step implementation plan, key file identification, and architectural trade-off analysis before code changes are made.\\n  Tools: Read, ReadMediaFile, Glob, Grep, WebSearch, FetchURL\\n- coder: General software engineering agent — the only subagent type with file-editing tools; use it for any delegated task that must modify code. Use this agent for non-trivial software engineering work that may require reading files, editing code, running commands, and returning a compact but technically complete summary to the parent agent.\\n  Tools: Bash, CronCreate, CronDelete, CronList, Edit, EnterPlanMode, ExitPlanMode, Glob, Grep, Read, ReadMediaFile, Skill, TaskList, TaskOutput, TaskStop, TodoList, WaitFor, WebSearch, FetchURL, Write, mcp__*\\n- explore: Fast codebase exploration with prompt-enforced read-only behavior. Fast agent specialized for exploring codebases. Use this when you need to quickly find files by patterns (e.g. \\"src/**/*.yaml\\"), search code for keywords (e.g. \\"database connection\\"), or answer questions about the codebase (e.g. \\"how does the auth module work?\\"). When calling this agent, specify the desired thoroughness level: \\"quick\\" for basic searches, \\"medium\\" for moderate exploration, or \\"thorough\\" for comprehensive analysis across multiple locations and naming conventions. Use this agent for any read-only exploration that will clearly require more than 3 search queries. Prefer launching multiple explore agents concurrently when investigating independent questions.\\n  Tools: Bash, Read, ReadMediaFile, Glob, Grep, WebSearch, FetchURL", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "prompt": { "type": "string", "description": "Full task prompt for the subagent" }, "description": { "type": "string", "description": "Short task description (3-5 words) for UI display" }, "subagent_type": { "description": "One of the available agent types (see \\"Available agent types\\" in this tool description). Defaults to \\"coder\\" when omitted.", "type": "string" }, "resume": { "description": "Optional agent ID to resume instead of creating a new instance. When set, do not also pass subagent_type — the resumed agent keeps its own type, and supplying both is rejected.", "type": "string" }, "run_in_background": { "description": "If true, return immediately without waiting for completion. Prefer false unless the task can run independently and there is a clear benefit to not waiting.", "type": "boolean" } }, "required": [ "prompt", "description" ], "additionalProperties": false } }, { "name": "AgentDynamicWorkflow", "description": "Launch multiple subagents from one prompt template, existing agent resumes, or both.\\n\\nUse AgentDynamicWorkflow when many subagents should run the same kind of task over different inputs. The placeholder is exactly \`{{item}}\`. For example, with \`prompt_template\` set to \`Review {{item}} for likely regressions.\` and \`items\` set to \`[\\"src/a.ts\\", \\"src/b.ts\\"]\`, AgentDynamicWorkflow launches two new subagents with those two concrete prompts. For a few differently-shaped tasks, make separate \`Agent\` calls in one message instead.\\n\\nUse \`resume_agent_ids\` to continue subagents that already exist from earlier work, such as ones that failed or timed out: map each agent id to the prompt for that resumed subagent (usually \`continue\` if no extra information is needed). You may combine \`resume_agent_ids\` with \`items\` in the same call to resume existing subagents and launch new ones. Do not duplicate resumed work in \`items\`.\\n\\nEach of these is enforced — a violation is rejected before any subagent starts: provide at least 2 \`items\` unless you pass \`resume_agent_ids\`; whenever \`items\` are present, \`prompt_template\` is required and must contain \`{{item}}\`; and the filled-in prompts must be distinct (two items that expand to the same prompt are rejected).\\n\\nUse enough subagents to keep the work focused and parallel. AgentDynamicWorkflow supports up to 128 subagents, and launches are queued automatically, so it is safe to split large tasks into many clear, independent items.\\n\\nIf \`AgentDynamicWorkflow\` is called, that call must be the only tool call in the response.\\n\\nThis capability is called a \\"dynamic workflow\\" (or simply \\"workflow\\"). Never use the word \\"swarm\\" in the \`description\` field, in subagent prompts, or when talking to the user about this tool.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "description": { "type": "string", "minLength": 1, "description": "Short description for the whole dynamic workflow. It is shown to the user as the workflow and subagent title, so word it as a workflow or run and never use the word \\"swarm\\"." }, "subagent_type": { "description": "Subagent type used for every new subagent spawned from items; defaults to coder when omitted. Resumed subagents always keep their original type, so passing subagent_type together with resume_agent_ids is allowed — it only affects the item-based spawns.", "type": "string", "minLength": 1 }, "prompt_template": { "description": "Prompt template for each subagent. The {{item}} placeholder is replaced with each item value.", "type": "string", "minLength": 1 }, "items": { "description": "Values used to fill {{item}}. Each item launches one new subagent.", "maxItems": 128, "type": "array", "items": { "type": "string", "minLength": 1 } }, "resume_agent_ids": { "description": "Map of existing subagent agent_id to the prompt used to resume that subagent. These resumed subagents are launched before new item-based subagents.", "type": "object", "propertyNames": { "type": "string", "minLength": 1 }, "additionalProperties": { "type": "string", "minLength": 1 } } }, "required": [ "description" ], "additionalProperties": false } }, { "name": "AskUserQuestion", "description": "Use this tool when you need to ask the user questions with structured options during execution. This allows you to:\\n1. Collect user preferences or requirements before proceeding\\n2. Resolve ambiguous or underspecified instructions\\n3. Let the user decide between implementation approaches as you work\\n4. Present concrete options when multiple valid directions exist\\n\\n**When NOT to use:**\\n- When you can infer the answer from context — be decisive and proceed\\n- Trivial decisions that don't materially affect the outcome\\n\\nOverusing this tool interrupts the user's flow. Only use it when the user's input genuinely changes your next action.\\n\\n**Usage notes:**\\n- Users always have an \\"Other\\" option for custom input — don't create one yourself\\n- Use multi_select to allow multiple answers to be selected for a question\\n- Keep option labels concise (1-5 words), use descriptions for trade-offs and details\\n- Each question should have 2-4 meaningful, distinct options\\n- Question texts must be unique across the call, and option labels must be unique within each question\\n- You can ask 1-4 questions at a time; group related questions to minimize interruptions\\n- If you recommend a specific option, list it first and append \\"(Recommended)\\" to its label\\n- The result is JSON with an \`answers\` object keyed by question text; each value is the chosen option's label (comma-separated labels for multi_select, or the user's own words if they picked \\"Other\\"); if \`answers\` is empty and a \`note\` says the user dismissed it, they chose not to answer — do not treat this as selecting the recommended option; decide based on context and do not re-ask the same question\\n- Set background=true when you can keep working without the answer. This starts a background question task and returns a task_id immediately. The answer arrives automatically in a later turn — you do not need to poll, sleep, or check on it. Continue with other work; never fabricate or predict the answer.", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "questions": { "minItems": 1, "maxItems": 4, "type": "array", "items": { "type": "object", "properties": { "question": { "type": "string", "minLength": 1, "description": "A specific, actionable question. End with '?'." }, "header": { "default": "", "description": "Short category tag (max 12 chars, e.g. 'Auth', 'Style').", "type": "string" }, "options": { "minItems": 2, "maxItems": 4, "type": "array", "items": { "type": "object", "properties": { "label": { "type": "string", "minLength": 1, "description": "Concise display text (1-5 words). If recommended, append '(Recommended)'." }, "description": { "default": "", "description": "Brief explanation of trade-offs or implications.", "type": "string" } }, "required": [ "label" ], "additionalProperties": false }, "description": "2-4 meaningful, distinct options. Do NOT include an 'Other' option — the system adds one automatically." }, "multi_select": { "default": false, "description": "Whether the user can select multiple options.", "type": "boolean" } }, "required": [ "question", "options" ], "additionalProperties": false }, "description": "The questions to ask the user (1-4 questions)." }, "background": { "default": false, "description": "Set true to ask in the background and return immediately with a background task_id; you are notified automatically when the user answers — do not poll with TaskOutput while the question is pending.", "type": "boolean" } }, "required": [ "questions" ], "additionalProperties": false } }, { "name": "Bash", "description": "Execute a \`bash\` command. Use this for shell semantics — pipes, env, processes, git, package managers, build/test runners, anything genuinely interactive or multi-step.\\n\\n**Translate these to a dedicated tool instead:**\\n- \`cat\` / \`head\` / \`tail\` (known path) → \`Read\`\\n- \`sed\` / \`awk\` (in-place edit) → \`Edit\`\\n- \`echo > file\` / \`cat <<EOF\` → \`Write\`\\n- \`find\` / recursive \`ls\` to locate files by name pattern → \`Glob\` (plain \`ls <known-directory>\` is fine for listing a directory)\\n- \`grep\` / \`rg\` (search file contents) → \`Grep\`\\n- \`echo\` / \`printf\` (talk to the user) → just output text directly\\n\\nThe dedicated tools render in the per-tool permission UI and keep raw stdout out of the conversation; that is why they are worth reaching for whenever one fits.\\n\\n**Output:**\\nThe stdout and stderr will be combined and returned as a string. The output may be truncated if it is too long. If the command exits non-zero, the output ends with a \`Command failed with exit code: N\` line; a command killed by its timeout or interrupted by the user ends with its own message instead.\\n\\nIf \`run_in_background=true\`, the command will be started as a background task and this tool will return a task ID instead of waiting for command completion. When doing that, you must provide a short \`description\`. Background commands default to a 600s timeout and \`timeout\` is capped at 86400s; set \`disable_timeout=true\` only when the task should run without a timeout. You will be automatically notified when the task completes. After starting one, default to returning control to the user instead of immediately waiting on it. Use \`TaskOutput\` only for a non-blocking status/output snapshot — do not wait on a task you just launched, since its completion arrives automatically. Use \`TaskStop\` only if the task must be cancelled. If a human user wants to inspect background tasks themselves, point them to the \`/tasks\` command, which opens an interactive panel; it has no subcommands.\\n\\n**Guidelines for safety and security:**\\n- Each shell tool call will be executed in a fresh shell environment. The shell variables, current working directory changes, and the shell history is not preserved between calls. To run a command in a particular directory, pass the \`cwd\` argument (or use absolute paths) rather than relying on a \`cd\` from an earlier call.\\n- The tool call will return after the command is finished. You shall not use this tool to execute an interactive command or a command that may run forever. For possibly long-running foreground commands, set the \`timeout\` argument in seconds. Foreground commands default to 60s and allow up to 300s. When a foreground command hits its timeout it is moved to the background instead of being killed, and you will be automatically notified when it completes.\\n- Avoid using \`..\` to access files or directories outside of the working directory.\\n- Avoid modifying files outside of the working directory unless explicitly instructed to do so.\\n- Never run commands that require superuser privileges unless explicitly instructed to do so.\\n\\n**Guidelines for efficiency:**\\n- Use \`&&\` to chain commands that genuinely depend on each other, e.g. \`npm install && npm test\`. Independent read-only commands (separate \`git show\`, \`ls\`, or status checks) should be issued as separate parallel Bash calls in one response, not chained into a single call — chaining serializes their execution and mixes their output. Do not stitch outputs together with \`echo\` separators.\\n- Use \`;\` to run commands sequentially regardless of success/failure\\n- Use \`||\` for conditional execution (run second command only if first fails)\\n- Use pipe operations (\`|\`) and redirections (\`>\`, \`>>\`) to chain input and output between commands\\n- Always quote file paths containing spaces with double quotes (e.g., cd \\"/path with spaces/\\")\\n- Compose multi-step logic in a single call with \`if\` / \`case\` / \`for\` / \`while\` control flows.\\n- Prefer \`run_in_background=true\` for long-running builds, tests, watchers, or servers when you need the conversation to continue before the command finishes.\\n\\n**Commands available:**\\nThe following common command categories are usually available. Availability still depends on the host, so when in doubt run \`which <command>\` first to confirm a command exists before relying on it.\\n- Navigation and inspection: \`ls\`, \`pwd\`, \`cd\`, \`stat\`, \`file\`, \`du\`, \`df\`, \`tree\`\\n- File and directory management: \`cp\`, \`mv\`, \`rm\`, \`mkdir\`, \`touch\`, \`ln\`, \`chmod\`, \`chown\`\\n- Text and data processing: \`wc\`, \`sort\`, \`uniq\`, \`cut\`, \`tr\`, \`diff\`, \`xargs\`\\n- Archives and compression: \`tar\`, \`gzip\`, \`gunzip\`, \`zip\`, \`unzip\`\\n- Networking and transfer: \`curl\`, \`wget\`, \`ping\`, \`ssh\`, \`scp\`\\n- Version control: \`git\`; for GitHub-hosted work (PRs, issues, CI runs, API queries) prefer the \`gh\` CLI when installed — it carries the user's GitHub auth and can return structured JSON\\n- Process and system: \`ps\`, \`kill\`, \`top\`, \`env\`, \`date\`, \`uname\`, \`whoami\`\\n- Language and package toolchains: \`node\`, \`npm\`, \`pnpm\`, \`yarn\`, \`python\`, \`pip\` (use whichever the project actually relies on)\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "command": { "type": "string", "minLength": 1, "description": "The command to execute." }, "cwd": { "description": "The working directory in which to run the command. When omitted, the command runs in the session's working directory.", "type": "string" }, "timeout": { "default": 60, "description": "Optional timeout in seconds for the command to execute. Foreground default 60s, max 300s. Background default 600s, max 86400s. Ignored for background commands when disable_timeout=true.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 }, "description": { "description": "A short description for the background task. Required when run_in_background is true.", "type": "string" }, "run_in_background": { "description": "Whether to run the command as a background task.", "type": "boolean" }, "disable_timeout": { "description": "If true, do not apply a timeout to the command. Only applies when run_in_background is true.", "type": "boolean" } }, "required": [ "command" ], "additionalProperties": false } }, { "name": "CreateGoal", "description": "Create a durable, structured goal that the runtime will pursue across multiple turns.\\n\\nCall \`CreateGoal\` only when:\\n\\n- the user explicitly asks you to start a goal or work autonomously toward an outcome, or\\n- a host goal-intake prompt asks you to create one.\\n\\nDo NOT create a goal for greetings, ordinary questions, or vague requests that lack a\\nverifiable completion condition. A goal needs a checkable end state.\\n\\nWhen the request is vague, ask the user for the missing completion criterion before creating\\nthe goal. If the user clearly insists after you warn them that the wording is vague or risky,\\nrespect that and create the goal.\\n\\nInclude a \`completionCriterion\` when the user provides one, or when it can be stated without\\ninventing new requirements. Keep \`objective\` concise; reference long task descriptions by file\\npath rather than pasting them.\\n\\nCreating a goal fails if one already exists, so use \`replace: true\` only when the user explicitly\\nwants to abandon the current goal and start a new one.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "objective": { "type": "string", "minLength": 1, "description": "The objective to pursue. Must have a verifiable end state." }, "completionCriterion": { "description": "How to verify the goal is complete. Include when the user provides one.", "type": "string" }, "replace": { "description": "Replace an existing active, paused, or blocked goal instead of failing.", "type": "boolean" } }, "required": [ "objective" ], "additionalProperties": false } }, { "name": "CronCreate", "description": "Schedule a prompt to be enqueued at a future time. Use for both recurring schedules and one-shot reminders.\\n\\nUses standard 5-field cron in the user's local timezone: minute hour day-of-month month day-of-week. \`0 9 * * *\` means 9am local — no timezone conversion needed.\\n\\n## One-shot tasks (recurring: false)\\n\\nFor \\"remind me at X\\" or \\"at <time>, do Y\\" requests — fire once then auto-delete.\\nPin minute/hour/day-of-month/month to specific values:\\n  \\"remind me at 2:30pm today to check the deploy\\" → cron: \\"30 14 <today_dom> <today_month> *\\", recurring: false\\n  \\"tomorrow morning, run the smoke test\\" → cron: \\"57 8 <tomorrow_dom> <tomorrow_month> *\\", recurring: false\\n\\nOne-shots are best for near-term reminders. A task only fires while its session is still alive (see Session lifetime below), so favor near times — within hours or a few days — rather than scheduling weeks or months ahead.\\n\\n## Recurring jobs (recurring: true, the default)\\n\\nFor \\"every N minutes\\" / \\"every hour\\" / \\"weekdays at 9am\\" requests:\\n  \\"*/5 * * * *\\" (every 5 min), \\"0 * * * *\\" (hourly), \\"0 9 * * 1-5\\" (weekdays at 9am local)\\n\\n## Avoid the :00 and :30 minute marks when the task allows it\\n\\nEvery user who asks for \\"9am\\" gets \`0 9\`, and every user who asks for \\"hourly\\" gets \`0 *\` — which means requests from across the planet land on the API at the same instant. When the user's request is approximate, pick a minute that is NOT 0 or 30:\\n  \\"every morning around 9\\" → \\"57 8 * * *\\" or \\"3 9 * * *\\" (not \\"0 9 * * *\\")\\n  \\"hourly\\" → \\"7 * * * *\\" (not \\"0 * * * *\\")\\n  \\"in an hour or so, remind me to...\\" → pick whatever minute you land on, don't round\\n\\nOnly use minute 0 or 30 when the user names that exact time and clearly means it (\\"at 9:00 sharp\\", \\"at half past\\", coordinating with a meeting). When in doubt, nudge a few minutes early or late — the user will not notice, and the fleet will.\\n\\n## Coalesce semantics\\n\\nFires are delivered only while the session is idle: a fire that comes due during an active turn is held and delivered at the next idle moment, never injected mid-turn.\\n\\nIf the scheduler slept past multiple ideal fire times (laptop closed, long-running turn, etc.), only **one** fire is delivered when it wakes up. The origin carries \`coalescedCount\` showing how many ideal fires were collapsed into this single delivery. You should treat \`coalescedCount > 1\` as \\"I missed some checks; only the latest state matters\\" rather than running the prompt that many times.\\n\\n## Cron-fire envelope\\n\\nWhen a cron task fires, the prompt you scheduled is re-injected wrapped in an XML envelope that exposes the fire context:\\n\\n\`\`\`\\n<cron-fire jobId=\\"...\\" cron=\\"...\\" recurring=\\"true|false\\" coalescedCount=\\"N\\" stale=\\"true|false\\">\\n<prompt>\\nyour original prompt text, verbatim\\n</prompt>\\n</cron-fire>\\n\`\`\`\\n\\nThe envelope is parseable. Use \`coalescedCount > 1\` to know multiple ideal fires were collapsed into a single delivery (treat as \\"only the latest state matters\\"), and \`stale=\\"true\\"\` as a cue that the task is past its 7-day threshold.\\n\\n## 7-day stale behavior\\n\\nRecurring tasks that have been alive for more than 7 days fire one\\nfinal time with \`stale: true\` on the envelope, and the system then\\nauto-deletes the task. The flag is the model's notice that this is\\nthe last delivery. If the schedule is still wanted, call \`CronCreate\`\\nagain with the same \`cron\` and \`prompt\` — that resets \`createdAt\` and\\nstarts a fresh 7-day window. One-shot tasks are never marked stale.\\n\\n## Jitter behavior\\n\\nAnti-herd jitter is applied deterministically per task id:\\n  - Recurring: ideal fire time is shifted **forward** by an offset ≤ min(10% of the cron period, 15 minutes). A \`*/5 * * * *\` task can drift up to 30s; a \`0 9 * * *\` task can drift up to 15 minutes.\\n  - One-shot: only when the ideal fire lands on \`:00\` or \`:30\` of the hour, the fire is pulled **earlier** by ≤ 90 seconds. Other minutes pass through unchanged.\\n\\n## One-shot vs recurring — when to pick which\\n\\nUse \`recurring: false\` for \\"remind me at X\\" style requests, single deadlines, \\"in N minutes do Y\\", and any task that should not repeat. Use \`recurring: true\` for periodic polling (CI status, build watchers, scheduled reports), workday rituals, and anything the user explicitly described as recurring.\\n\\n## Session lifetime\\n\\nCron tasks live in the current session. When you exit, they\\nare persisted under the session homedir; resuming the same session\\nreloads them and the scheduler resumes from each task's \`createdAt\`. Fire times that fell during the offline window are\\ncollapsed into a single delivery via \`coalescedCount\` (and recurring\\ntasks past their 7-day window arrive with \`stale: true\` as their final\\ndelivery).\\n\\nTasks do **not** carry over into a brand-new session — they are scoped\\nto the resumed session id, not to the working directory.\\n\\n## Limits\\n\\nA session holds at most 50 live cron tasks; creating one beyond that is rejected. (The \`prompt\` body is also capped — see its parameter description.) Expressions that never fire within the next 5 years (e.g. \`0 0 31 2 *\`, an impossible date) are rejected at create time.\\n\\n## Returned fields\\n\\n\`id\` (ULID), \`cron\` (the normalized expression), \`humanSchedule\` (English summary), \`recurring\`,\\n\`nextFireAt\` (local ISO timestamp with numeric offset, or null). \`id\` is needed by \`CronDelete\`.\\n\\n## Tell the user how to cancel or modify\\n\\nAfter successfully creating a task, proactively tell the user how they can cancel or modify it later. Users have no direct \`/cron\` command or self-service UI to manage reminders themselves; they must ask the model to make changes (e.g. \\"cancel my 9am reminder\\" or \\"change my daily check to 10am\\"). Include the task \`id\` in your message so the user can reference it.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "cron": { "type": "string", "description": "5-field cron expression in local time: \\"M H DoM Mon DoW\\" (e.g. \\"*/5 * * * *\\" = every 5 minutes; \\"30 14 28 2 *\\" = Feb 28 at 2:30pm local — a pinned date like this repeats yearly unless you also pass recurring: false)." }, "prompt": { "type": "string", "minLength": 1, "maxLength": 8192, "description": "The prompt to enqueue at each fire time. Limited to 8 KiB (UTF-8)." }, "recurring": { "default": true, "description": "true (default) = fire on every cron match until deleted or auto-expired after 7 days. false = fire once at the next match, then auto-delete. Use false for \\"remind me at X\\" one-shot requests with pinned minute/hour/dom/month.", "type": "boolean" } }, "required": [ "cron", "prompt" ], "additionalProperties": false } }, { "name": "CronDelete", "description": "Cancel a scheduled cron job by id.\\n\\nUse this tool to remove a cron task previously scheduled with\\n\`CronCreate\`. The \`id\` is the ULID value returned by \`CronCreate\`, or\\nshown in the \`id:\` column of \`CronList\` — quote it verbatim, no\\nprefix.\\n\\nBehaviour by task kind:\\n\\n- **Recurring task** (\`recurring: true\`): stops all future fires\\n  immediately. The scheduler picks up the deletion on its next tick.\\n- **One-shot task** (\`recurring: false\`): cancels the pending fire if\\n  it has not happened yet. One-shots that have already fired\\n  auto-delete themselves, so calling \`CronDelete\` on a fired one-shot\\n  returns \\"no cron job with id ...\\".\\n\\nNot-found is reported as an error (not a silent no-op) so you can\\ncorrect yourself — typically by calling \`CronList\` to see which ids\\nare actually live, rather than re-trying with the same stale id.\\n\\nRefresh pattern (use when you want a stale recurring schedule to\\ncontinue):\\n\\nStale recurring tasks are auto-deleted by the system after their final\\nfire — there is nothing for \`CronDelete\` to remove at that point. To\\nkeep the schedule running, just call \`CronCreate\` with the same \`cron\`\\nand \`prompt\`. Use \`CronList\`'s \`prompt\` field to recall the original\\ntext after a context compaction.\\n\\n\`CronDelete\` remains the right call when you want to cancel a task\\nthat is still live (recurring not yet stale, or a one-shot still\\npending).\\n\\nGuidelines:\\n\\n- Users have no direct \`/cron\` command or self-service UI to delete\\n  tasks themselves; they must ask the model to cancel a reminder.\\n  When deleting on behalf of a user, confirm the action and report\\n  the result plainly.\\n- Cron deletion is irreversible — there is no undo. If you delete the\\n  wrong task, you must re-create it with \`CronCreate\`.\\n- If the model is unsure which id is current (e.g. after a context\\n  compaction), call \`CronList\` first rather than guessing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "id": { "type": "string", "description": "The cron job id (ULID) returned by CronCreate / CronList." } }, "required": [ "id" ], "additionalProperties": false } }, { "name": "CronList", "description": "List all cron jobs currently scheduled in this session.\\n\\nUse this tool to see every pending cron task — both recurring jobs and\\none-shot reminders — that you (or the user) have scheduled with\\n\`CronCreate\`. The output is the entry point for inspecting scheduled\\nwork: it returns a stable id, the original cron expression, a human\\nrendering, the next post-jitter fire time, the recurring flag, the\\ntask's age in days, and a stale indicator.\\n\\nEach record carries:\\n\\n- \`id\` — the task id (a ULID). Pass this to \`CronDelete\` to remove the\\n  task, or quote it in user-facing messages when asking for\\n  confirmation.\\n- \`cron\` — the verbatim 5-field cron expression as scheduled.\\n- \`humanSchedule\` — plain-English rendering (e.g. \`every 5 minutes\`).\\n- \`prompt\` — the scheduled prompt text, JSON-encoded so embedded\\n  newlines stay on one line. Truncated to 200 UTF-8 bytes with\\n  \`…(truncated)\` if longer. Use this to recall what a task is for\\n  after a context compaction, and as the source for the\\n  \`CronCreate\` refresh ritual.\\n- \`nextFireAt\` — local ISO timestamp with an explicit numeric offset\\n  for the next fire **after jitter has been applied**. The actual fire\\n  may land slightly before or after a round \`:00\` / \`:30\` minute mark\\n  due to herd-avoidance jitter; this is the value the scheduler will\\n  compare against, so it reflects what will really happen. \`null\` if\\n  the expression has no fire in the next 5 years (should not happen\\n  for tasks created through \`CronCreate\`, which validates).\\n- \`recurring\` — \`true\` for cadenced jobs, \`false\` for one-shots.\\n- \`ageDays\` — \`(now - createdAt) / day\`, two decimal places. Useful\\n  when deciding whether a long-running cron is still relevant.\\n- \`stale\` — \`true\` when a recurring task is older than 7 days. The\\n  system **auto-deletes the task after this fire** to bound session\\n  lifetime; the \`stale: true\` flag is the model's notice that this is\\n  the final delivery. To resume the same schedule, call \`CronCreate\`\\n  again with the original \`cron\` and \`prompt\` (the \`prompt\` row above\\n  carries it for exactly this purpose). One-shots are never marked\\n  stale — they fire at most once by construction.\\n\\nGuidelines:\\n\\n- This tool is read-only and never mutates state, so it is always\\n  safe to call (including in plan mode).\\n- Users cannot directly manage cron tasks themselves; if they want to\\n  cancel or modify a schedule, route the request through the model\\n  (i.e. call \`CronDelete\` or \`CronCreate\` on their behalf).\\n- The empty case returns \`cron_jobs: 0\\\\nNo cron jobs scheduled.\`. Cron\\n  tasks survive a resume of the same session but do not bleed into new\\n  sessions.\\n- After a context compaction, or whenever you are unsure which cron\\n  jobs are live, call this tool to re-enumerate them rather than\\n  guessing ids from earlier in the conversation.\\n- Records are separated by a line containing just \`---\`, in the\\n  insertion order they were scheduled.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "Edit", "description": "Perform exact replacements in existing files.\\n\\n- Edit is mandatory for every incremental change, especially small edits. DO NOT use Write or Bash \`sed\`.\\n- Read the target file before every Edit. DO NOT call Edit from memory, stale context, or a guessed \`old_string\`.\\n- Take \`old_string\` and \`new_string\` from the Read output view.\\n- Drop the line-number prefix and tab; match only file content.\\n- \`old_string\` must be unique unless \`replace_all\` is set.\\n- If \`old_string\` is ambiguous, add surrounding context. Use \`replace_all\` only when every occurrence should change — for example, renaming a symbol throughout the file.\\n- Multiple Edit calls may run in one response only when they do not target the same file.\\n- DO NOT issue consecutive Edit calls on the same file. A previous Edit can invalidate a later Edit's \`old_string\`, causing \`old_string not found\`. Read the file again before the next Edit.\\n- A write lock serializes same-file edits in response order, but serialization does not make stale \`old_string\` valid.\\n- For pure CRLF files, Read shows LF; use LF in \`old_string\` and \`new_string\`, and Edit writes CRLF back.\\n- For mixed endings or lone carriage returns, Read shows carriage returns as \\\\r; include actual \\\\r escapes in those positions.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to the text file to edit. Relative paths resolve against the working directory; a path outside the working directory must be absolute." }, "old_string": { "type": "string", "minLength": 1, "description": "Exact content to replace from the Read output view, without the line-number prefix. Use LF for pure CRLF files; use actual \\\\r escapes where Read shows \\\\r." }, "new_string": { "type": "string", "description": "Replacement text in the same Read output view. LF is written back as CRLF only for pure CRLF files." }, "replace_all": { "description": "Set true only when every occurrence of old_string should be replaced.", "type": "boolean" } }, "required": [ "path", "old_string", "new_string" ], "additionalProperties": false } }, { "name": "EnterPlanMode", "description": "Use this tool proactively when you're about to start a non-trivial implementation task.\\nGetting user sign-off on your approach via ExitPlanMode before writing code prevents wasted effort.\\n\\nUse it when ANY of these conditions apply:\\n\\n1. New Feature Implementation - e.g. \\"Add a caching layer to the API\\"\\n2. Multiple Valid Approaches - e.g. \\"Optimize database queries\\" (indexing vs rewrite vs caching)\\n3. Code Modifications - e.g. \\"Refactor auth module to support OAuth\\"\\n4. Architectural Decisions - e.g. \\"Add WebSocket support\\"\\n5. Multi-File Changes - involves more than 2-3 files\\n6. Unclear Requirements - need exploration to understand scope\\n7. User Preferences Matter - if user input would materially change the implementation approach, use EnterPlanMode to structure the decision\\n\\nPermission mode notes:\\n- EnterPlanMode enters plan mode automatically without an approval prompt in all permission modes.\\n- In yolo and manual modes, ExitPlanMode still presents the plan to the user for approval.\\n- In auto permission mode, do not use AskUserQuestion; make the best decision from available context.\\n- In auto permission mode, ExitPlanMode exits plan mode without asking the user.\\n- Use EnterPlanMode only when planning itself adds value.\\n\\nWhen NOT to use:\\n- Single-line or few-line fixes (typos, obvious bugs, small tweaks)\\n- User gave very specific, detailed instructions\\n- Pure research/exploration tasks\\n\\nOnce you are in plan mode, a reminder walks you through the workflow (explore → design → write the plan file → \`ExitPlanMode\`) and enforces read-only access. For non-trivial tasks where you are unsure of the codebase structure or relevant code paths, use \`Agent(subagent_type=\\"explore\\")\` to investigate first when the \`Agent\` tool is available.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "ExitPlanMode", "description": "Use this tool when you are in plan mode and have finished writing your plan to the plan file and are ready for user approval.\\n\\n## How This Tool Works\\n- You should have already written your plan to the plan file specified in the plan mode reminder.\\n- This tool does NOT take the plan content as a parameter - it reads the plan from the file you wrote.\\n- The user will see the contents of your plan file when they review it. In auto permission mode, the tool reads the file and exits plan mode without asking the user.\\n\\n## When to Use\\nOnly use this tool for tasks that require planning implementation steps. For research tasks (searching files, reading code, understanding the codebase), do NOT use this tool.\\n\\n## What a good plan contains\\nList specific, verifiable steps grounded in the actual codebase — real files, functions, and commands, in a sensible order. Each step should be concrete enough to act on and to check. Avoid vague filler like \\"improve performance\\" or \\"add tests\\"; say what to change and where.\\n\\n## Multiple Approaches\\nIf your plan offers multiple alternative approaches, pass them via the \`options\` parameter so the user can choose which one to execute — see the \`options\` parameter for the format, count, and reserved labels. In yolo and manual modes the user sees all options alongside the host's Reject and Revise controls.\\n\\n## Before Using\\n- In auto permission mode, do NOT use AskUserQuestion; make the best decision from available context.\\n- In auto permission mode, this tool exits plan mode without asking the user.\\n- In yolo and manual modes, this tool still presents the plan to the user for approval.\\n- If auto permission mode is not active and you have unresolved questions, use AskUserQuestion first.\\n- If auto permission mode is not active and you have multiple approaches and haven't narrowed down yet, consider using AskUserQuestion first to let the user choose, then write a plan for the chosen approach only.\\n- Once your plan is finalized, use THIS tool to request approval.\\n- Do NOT use AskUserQuestion to ask \\"Is this plan OK?\\" or \\"Should I proceed?\\" - that is exactly what ExitPlanMode does.\\n- If rejected, revise based on feedback and call ExitPlanMode again.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "options": { "description": "When the plan contains multiple alternative approaches, list them here so the user can choose which one to execute. Provide up to 3 options; 2-3 distinct approaches work best when the plan offers a real choice. Passing a single option is allowed and is equivalent to a plain plan approval. Each option represents a distinct approach from the plan. Do not use \\"Reject\\", \\"Revise\\", \\"Approve\\", or \\"Reject and Exit\\" as labels.", "minItems": 1, "maxItems": 3, "type": "array", "items": { "type": "object", "properties": { "label": { "type": "string", "minLength": 1, "maxLength": 80, "description": "Short name for this option (1-8 words). Append \\"(Recommended)\\" if you recommend this option." }, "description": { "default": "", "description": "Brief summary of this approach and its trade-offs.", "type": "string" } }, "required": [ "label" ], "additionalProperties": false } } }, "additionalProperties": false } }, { "name": "FetchURL", "description": "Fetch content from a URL. The content is returned either as the main text extracted from the page, or as the full response body verbatim; a note at the top of the result states which of the two you received, so you can judge how complete it is. Use this when you need to read a specific web page.\\n\\nOnly fully-formed public \`http\`/\`https\` URLs are supported; other schemes and private or loopback addresses are not fetched. Very large pages may be truncated or refused. The fetch carries no login or session for the target site, so pages behind authentication (private repositories, internal dashboards) return a login page or an error instead of the real content — if the text you get back looks like a generic landing or sign-in page, treat that as the login wall, not the answer, and reach the content through a credentialed route (an authenticated CLI or MCP tool) instead.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "url": { "type": "string", "description": "The URL to fetch content from." } }, "required": [ "url" ], "additionalProperties": false } }, { "name": "GetGoal", "description": "Read the current goal: its objective, completion criterion, status, and budgets (turns, tokens,\\ntime, and how much of each remains). When the goal has stopped, it also reports the terminal reason.\\n\\nUse \`GetGoal\` before deciding whether to continue working, report completion, report a blocker,\\nor respect a pause. It returns \`{ \\"goal\\": null }\` when there is no current goal.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "Glob", "description": "Find files by glob pattern, sorted by modification time (most recent first).\\n\\nPowered by ripgrep. Respects \`.gitignore\`, \`.ignore\`, and \`.rgignore\` by default — set \`include_ignored\` to also match ignored files (e.g. build outputs, \`node_modules\`). Sensitive files (such as \`.env\`) are always filtered out. Matches are files only — directories themselves are never listed; to find a directory, glob for a file inside it (e.g. \`**/fixtures/**\`).\\n\\nGood patterns:\\n- \`*.ts\` — all files matching an extension, at any depth below the search root (a bare pattern without \`/\` matches recursively)\\n- \`src/*.ts\` — files directly inside \`src/\` (one level, not recursive)\\n- \`src/**/*.ts\` — recursive walk with a subdirectory anchor and extension\\n- \`**/*.py\` — recursive walk from the search root for an extension\\n- \`*.{ts,tsx}\` — brace expansion is supported\\n- \`{src,test}/**/*.ts\` — cartesian brace expansion is supported too\\n\\nResults are capped at the first 100 matching paths. If a search would return more, a truncation marker is appended. Refine the pattern (extension, subdirectory) when 100 is not enough, or call again with a narrower anchor.\\n\\nLarge-directory caveat — avoid recursing into dependency / build output even with an anchor, especially when \`include_ignored\` is set:\\n- \`node_modules/**/*.js\`, \`.venv/**/*.py\`, \`__pycache__/**\`, \`target/**\` can produce thousands of results that truncate at the match cap and waste context. Prefer specific subpaths like \`node_modules/react/src/**/*.js\`.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "pattern": { "type": "string", "description": "Glob pattern to match files." }, "path": { "description": "Directory to search. Accepts an absolute path, or a path relative to the current working directory. Defaults to the current working directory.", "type": "string" }, "include_ignored": { "description": "Also match files excluded by ignore files such as \`.gitignore\`, \`.ignore\`, and \`.rgignore\` (for example \`node_modules\` or build outputs). Sensitive files (such as \`.env\`) remain filtered out for safety. VCS metadata directories (\`.git\` and similar) are always skipped, even when this is true. Defaults to false.", "type": "boolean" }, "include_dirs": { "description": "Deprecated and ignored. Results are always files-only — directories are never listed. Accepted only so older calls that still pass this flag are not rejected by parameter validation.", "type": "boolean" } }, "required": [ "pattern" ], "additionalProperties": false } }, { "name": "Grep", "description": "Search file contents using regular expressions (powered by ripgrep).\\n\\nUse Grep when the task is to find unknown content or unknown file locations. Do not use shell \`grep\` or \`rg\` directly; this tool applies workspace path policy, output limits, and sensitive-file filtering.\\nALWAYS use Grep tool instead of running \`grep\` or \`rg\` from a shell — direct shell calls bypass workspace policy, output limits, and sensitive-file filtering.\\nIf you already know a concrete file path and need to inspect its contents, use Read directly instead.\\n\\nWrite patterns in ripgrep regex syntax, which differs from POSIX \`grep\` syntax. For example, braces are special, so escape them as \`\\\\{\` to match a literal \`{\`.\\n\\nHidden files (dotfiles such as \`.gitlab-ci.yml\` or \`.eslintrc.json\`) are searched by default. To also search files excluded by \`.gitignore\` (such as \`node_modules\` or build outputs), set \`include_ignored\` to \`true\`. Sensitive files (such as \`.env\`) are always skipped for safety, even when \`include_ignored\` is \`true\`.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "pattern": { "type": "string", "description": "Regular expression to search for." }, "path": { "description": "File or directory to search. Accepts an absolute path, or a path relative to the current working directory. Omit to search the current working directory. Use Read instead when you already know a concrete file path and need its contents.", "type": "string" }, "glob": { "description": "Optional glob filter for which files to search, e.g. \`*.ts\`. Matched against each file's full absolute path, so a path-anchored pattern like \`src/**/*.ts\` silently matches nothing — use a basename pattern (\`*.ts\`), or anchor with \`**/\` (\`**/src/**/*.ts\`). To scope the search to a directory, use \`path\` instead.", "type": "string" }, "type": { "description": "Optional ripgrep file type filter, such as ts or py. Prefer this over \`glob\` when filtering by language or file kind: it is more efficient and less error-prone than an equivalent glob pattern.", "type": "string" }, "output_mode": { "description": "Shape of the result. \`content\` shows matching lines (honors \`-A\`, \`-B\`, \`-C\`, \`-n\`, and \`head_limit\`); \`files_with_matches\` shows only the paths of files that contain a match, most-recently-modified first (honors \`head_limit\`); \`count_matches\` shows per-file match counts as \`path:count\` lines, preceded by an aggregate total line. Defaults to \`files_with_matches\`.", "type": "string", "enum": [ "content", "files_with_matches", "count_matches" ] }, "-i": { "description": "Perform a case-insensitive search. Defaults to false.", "type": "boolean" }, "-n": { "description": "Prefix each matching line with its line number. Applies only when \`output_mode\` is \`content\`. Defaults to true.", "type": "boolean" }, "-A": { "description": "Number of lines to show after each match. Applies only when \`output_mode\` is \`content\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "-B": { "description": "Number of lines to show before each match. Applies only when \`output_mode\` is \`content\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "-C": { "description": "Number of lines to show before and after each match. Applies only when \`output_mode\` is \`content\`; takes precedence over \`-A\` and \`-B\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "head_limit": { "description": "Limit output to the first N lines/entries after offset. Defaults to 250. Pass 0 for unlimited.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "offset": { "description": "Number of leading lines/entries to skip before applying \`head_limit\`. Use it together with \`head_limit\` to page through large result sets. Defaults to 0.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "multiline": { "description": "Enable multiline matching, where the pattern can span line boundaries and \`.\` also matches newlines. Defaults to false.", "type": "boolean" }, "include_ignored": { "description": "Also search files excluded by ignore files such as \`.gitignore\`, \`.ignore\`, and \`.rgignore\` (for example \`node_modules\` or build outputs). Sensitive files (such as \`.env\`) remain filtered out for safety. VCS metadata directories (\`.git\` and similar) are always skipped, even when this is true. Defaults to false.", "type": "boolean" } }, "required": [ "pattern" ], "additionalProperties": false } }, { "name": "Read", "description": "Read a text file from the local filesystem.\\n\\nIf the user provides a concrete file path to a text file, call Read directly. Do not \`Glob\`, \`ls\`, or otherwise pre-check known text file paths; missing or invalid file paths return errors you can handle. Do not use Read for directories; use \`ls\` via Bash for a known directory, or Glob when you need files matching a name pattern (Glob lists files only, never directories). Use \`Grep\` only when the task is to search for unknown content or locations.\\n\\nWhen you need several files, prefer to read them in parallel: emit multiple \`Read\` calls in a single response instead of reading one file per turn.\\n\\n- Relative paths resolve against the working directory; a path outside the working directory must be absolute.\\n- Returns up to 1000 lines or 100 KB per call, whichever comes first; lines longer than 2000 chars are truncated mid-line.\\n- Page larger files with \`line_offset\` (1-based start line) and \`n_lines\`. Omit \`n_lines\` to read up to the 1000-line cap.\\n- Sensitive files (\`.env\` files, credential stores, SSH private keys, and similar secrets) are refused to protect secrets; do not attempt to read them. Templates and public keys are exempt: \`.env.example\` / \`.env.sample\` / \`.env.template\` and public SSH keys such as \`id_rsa.pub\` read normally.\\n- UTF-8 text files are read directly. UTF-16 LE/BE text files (with or without a BOM) are detected automatically and transcoded to UTF-8 for display; the status block notes the detected encoding, and Edit/Write on such a file still expect UTF-8 — convert its encoding first (e.g. with \`iconv\`). Other encodings (e.g. GBK), binary files, and files containing NUL bytes are refused.\\n- Negative line_offset reads from the end of the file (for example, -100 reads the last 100 lines); the absolute value cannot exceed 1000.\\n- Output format: \`<line-number>\\\\t<content>\` per line.\\n- A \`<system>...</system>\` status block is appended after the file content; it summarizes how much was read (line and byte counts, truncation, line-ending notes) and is not part of the file itself.\\n- Pure CRLF files are displayed with LF line endings; \`Edit\` matches this output and preserves CRLF when writing back.\\n- Mixed or lone carriage-return line endings are shown as \`\\\\r\` and require exact \`Edit.old_string\` escapes.\\n- After a successful \`Edit\`/\`Write\`, do not re-read solely to prove the write landed. When the task depends on an exact file, API, or output shape, inspect the final external contract before finishing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to a text file. Relative paths resolve against the working directory; a path outside the working directory must be absolute. Directories are not supported; use \`ls\` via Bash for a known directory, or Glob for pattern search." }, "line_offset": { "description": "The line number to start reading from. Omit to start at line 1. Negative values read from the end of the file; the absolute value cannot exceed 1000.", "anyOf": [ { "type": "integer", "minimum": 1, "maximum": 9007199254740991 }, { "type": "integer", "minimum": -1000, "maximum": -1 } ] }, "n_lines": { "description": "The number of lines to read; the tool also applies its internal cap. Omit to read up to the internal cap of 1000 lines.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 } }, "required": [ "path" ], "additionalProperties": false } }, { "name": "SetGoalBudget", "description": "Set a hard budget limit for the current goal.\\n\\nUse this only when the user clearly gives a runtime limit, such as:\\n\\n- \\"stop after 20 turns\\"\\n- \\"use no more than 500k tokens\\"\\n- \\"finish within 30 minutes\\"\\n\\nDo not invent limits. Do not call this for vague wording such as \\"spend some time\\" or\\n\\"try to be quick\\".\\n\\nIf the user gives a compound time, convert it to one supported unit before calling this tool.\\nFor example, \\"2 hours and 3 minutes\\" can be set as \`value: 123, unit: \\"minutes\\"\`.\\n\\nA time budget must be between 1 second and 24 hours — the tool rejects anything shorter or\\nlonger, telling the user it is not a reasonable goal budget. Turn and token budgets are not\\nbounded this way; they must be positive and are rounded to the nearest whole number (minimum 1).\\n\\nSupported units:\\n\\n- \`turns\`\\n- \`tokens\`\\n- \`milliseconds\`\\n- \`seconds\`\\n- \`minutes\`\\n- \`hours\`\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "value": { "type": "number", "exclusiveMinimum": 0, "description": "The positive numeric budget value." }, "unit": { "type": "string", "enum": [ "turns", "tokens", "milliseconds", "seconds", "minutes", "hours" ] } }, "required": [ "value", "unit" ], "additionalProperties": false } }, { "name": "Skill", "description": "Invoke a registered skill from the current skill listing. BLOCKING REQUIREMENT: when a skill from the listing matches the user's request, you MUST call this tool (not free-form text). Do not re-invoke a skill to repeat work already done: if a \`<skill-loaded>\` block for it with the same \`args\` is already present in the conversation, follow those instructions directly instead of calling the tool again. Do call the tool again when you need the skill with different arguments — the loaded block was expanded with the earlier \`args\` and will not reflect new inputs.", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "skill": { "type": "string", "description": "The exact name of the skill to invoke, spelled as it appears in the current skill listing (e.g. \\"commit\\", \\"pdf\\")." }, "args": { "description": "Optional argument string for the skill, written like a command line (e.g. \`-m \\"fix bug\\"\`, \`123\`, a file path). It is split on whitespace (quotes group a token) and expanded into the skill's placeholders ($NAME, $1, $ARGUMENTS); if the skill body has no placeholders, the whole string is still appended as a trailing \`ARGUMENTS:\` line. Omit it only when there is nothing to pass.", "type": "string" } }, "required": [ "skill" ], "additionalProperties": false } }, { "name": "TaskList", "description": "List background tasks and their current status.\\n\\nUse this tool to discover which background tasks exist and where each one\\nstands. It is the entry point for inspecting background work: it returns a\\ntask ID, status, and description for every task it reports, plus the command,\\nPID, and (once finished) exit code for shell tasks, and a stop reason for any\\ntask that ended early.\\n\\nGuidelines:\\n\\n- After a context compaction, or whenever you are unsure which background\\n  tasks are running or what their task IDs are, call this tool to\\n  re-enumerate them instead of guessing a task ID.\\n- Prefer the default \`active_only=true\`, which lists only non-terminal tasks.\\n  Pass \`active_only=false\` only when you specifically need to see tasks that\\n  have already finished. With \`active_only=false\` the result may also include\\n  \`lost\` tasks — tasks left over from a previous process that can no longer be\\n  inspected or controlled; treat them as already terminated.\\n- \`limit\` caps how many tasks are returned. It accepts a value between 1 and\\n  100 and defaults to 20 when omitted.\\n- This tool only lists tasks; it does not return their output. Use it first\\n  to locate the task ID you need, then call \`TaskOutput\` with that ID to read\\n  the task's output and details.\\n- This tool is read-only and does not change any state, so it is always safe\\n  to call, including in plan mode.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "active_only": { "default": true, "description": "Whether to list only non-terminal background tasks.", "type": "boolean" }, "limit": { "default": 20, "description": "Maximum number of tasks to return.", "type": "integer", "minimum": 1, "maximum": 100 } }, "additionalProperties": false } }, { "name": "TaskOutput", "description": "Retrieve a snapshot of a running or completed background task.\\n\\nUse this after \`Bash(run_in_background=true)\`, \`Agent(run_in_background=true)\`, or \`AskUserQuestion(background=true)\` to check progress, or to read the output of a task that has already completed.\\n\\nGuidelines:\\n- Prefer relying on automatic completion notifications. Use this tool only when you need task output before the automatic notification arrives.\\n- This tool is always non-blocking: it returns the current status/output snapshot immediately and never waits for the task to finish.\\n- Do not use TaskOutput to wait for a result you need before continuing — if your next step depends on the task's result, run that task in the foreground instead. TaskOutput is for a deliberate progress check you will act on without blocking, not a way to sit and wait for a background task you just launched.\\n- This tool returns structured task metadata, a fixed-size output preview, and an output_path for the full log.\\n- For a terminal task, the metadata also explains why it ended. A shell command that runs to completion reports \`status: completed\` on a zero exit, or \`status: failed\` with its non-zero \`exit_code\` — judge that failure from the \`exit_code\`, because a plain command failure carries no \`stop_reason\` and no \`terminal_reason\`. \`terminal_reason\` is a categorical label emitted only when the end is not an ordinary exit: \`timed_out\` when the deadline aborted it, \`stopped\` when it was explicitly stopped, or \`failed\` when it errored without producing an exit code; the \`stopped\` and \`failed\` cases also carry a human-readable \`stop_reason\`. A task that finished on its own with a clean exit carries neither \`stop_reason\` nor \`terminal_reason\`.\\n- The full, never-truncated log is always available at output_path; use the \`Read\` tool with that path to page through it, whether or not the preview was truncated.\\n- This tool works with the generic background task system and should remain the primary read path for future task types, not just bash.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "task_id": { "type": "string", "description": "The background task ID to inspect." } }, "required": [ "task_id" ], "additionalProperties": false } }, { "name": "TaskStop", "description": "Stop a running background task.\\n\\nOnly use this when a task must genuinely be cancelled — for a task that is\\nfinishing normally, wait for its completion notification or inspect it with\\n\`TaskOutput\` instead of stopping it.\\n\\nGuidelines:\\n- This is a general-purpose stop capability for any background task. It is not\\n  a bash-specific kill.\\n- Stopping a task is destructive: it may leave partial side effects behind.\\n  Use it with care.\\n- If the task has already finished, this tool simply returns its current\\n  status.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "task_id": { "type": "string", "description": "The background task ID to stop." }, "reason": { "default": "Stopped by TaskStop", "description": "Short reason recorded when the task is stopped.", "type": "string" } }, "required": [ "task_id" ], "additionalProperties": false } }, { "name": "TodoList", "description": "Use this tool to maintain a structured TODO list as you work through a multi-step task. Use it proactively and often when progress tracking helps the current work. This is especially useful in long-running investigations and implementation tasks with several tool calls; in plan mode, write the plan to the plan file rather than tracking it here.\\n\\n**When to use:**\\n- Multi-step tasks that span several tool calls\\n- Tracking investigation progress across a large codebase search\\n- Planning a sequence of edits before making them\\n- After receiving new multi-step instructions, capture the requirements as todos\\n- Before starting a tracked task, mark exactly one item as \`in_progress\`\\n- Immediately after finishing a tracked task, mark it \`done\`; do not batch completions at the end\\n\\n**When NOT to use:**\\n- Single-shot answers that complete in one or two tool calls\\n- Trivial requests where tracking adds no clarity\\n- Purely conversational or informational replies\\n\\n**Avoid churn:**\\n- Do not re-call this tool when nothing meaningful has changed since the last call — update the list only after real progress.\\n- When unsure of the current state, call query mode first (omit \`todos\`) to check the list before deciding what to update.\\n- If no available tool can move any task forward, tell the user where you are stuck instead of repeatedly re-ordering the same todos.\\n\\n**How to use:**\\n- Call with \`todos: [...]\` to replace the full list. Statuses: pending / in_progress / done.\\n- Call with no \`todos\` argument to retrieve the current list without changing it.\\n- Call with \`todos: []\` to clear the list.\\n- Keep titles short and actionable (e.g. \\"Read session-control.ts\\", \\"Add planMode flag to TurnManager\\").\\n- Update statuses as you make progress.\\n- When work is underway, keep exactly one task \`in_progress\`.\\n- Only mark a task \`done\` when it is fully accomplished.\\n- Never mark a task \`done\` if tests are failing, implementation is partial, unresolved errors remain, or required files/dependencies could not be found.\\n- If you encounter a blocker, keep the blocked task \`in_progress\` or add a new pending task describing what must be resolved.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "todos": { "description": "The updated todo list. Omit to read the current todo list without making changes. Pass an empty array to clear the list.", "type": "array", "items": { "type": "object", "properties": { "title": { "type": "string", "minLength": 1, "description": "Short, actionable title for the todo." }, "status": { "type": "string", "enum": [ "pending", "in_progress", "done" ], "description": "Current status of the todo." } }, "required": [ "title", "status" ], "additionalProperties": false } } }, "additionalProperties": false } }, { "name": "UpdateGoal", "description": "Set the status of the current goal. This is how you resume, complete, or block an autonomous goal.\\n\\n- \`active\` — resume a paused or blocked goal when the user explicitly asks you to work on that goal.\\n- \`complete\` — the objective is satisfied and any stated validation has passed. The goal ends and a completion summary is recorded. Before using this, verify the current state against the actual objective and every explicit requirement. Treat weak or indirect evidence as not complete. Do not use \`complete\` merely because a budget is nearly exhausted or you want to stop.\\n- \`blocked\` — a genuine impasse prevents useful progress: an external condition, required user input, missing credentials or permissions, a persistent technical failure, or an impossible, unsafe, or contradictory objective. For non-terminal blockers, do not use \`blocked\` the first time you hit the blocker. The same blocking condition must repeat for at least 3 consecutive goal turns before you call \`blocked\`, counting the original/user-triggered turn and automatic continuations. If a previously blocked goal is resumed, treat the resumed run as a fresh blocked audit. If the objective itself is impossible, unsafe, or contradictory, call \`blocked\` in the same turn instead of running more goal turns. Do not use \`blocked\` because the work is large, hard, slow, uncertain, incomplete, still needs validation, would benefit from clarification, or needs more goal turns. Once the 3-turn threshold is met and you cannot make meaningful progress without user input or an external-state change, call \`blocked\` instead of leaving the goal active.\\n\\nMost active goal turns should not call this tool. If you complete one useful slice of work and material work remains, end the turn normally without calling UpdateGoal; the runtime will prompt you to continue in the next goal turn. Call \`complete\` only when all required work is done, any stated validation has passed, and there is no useful next action. Do not call \`complete\` after only producing a plan, summary, first pass, or partial result. Call \`blocked\` only after the blocked audit threshold is met. If you call \`blocked\`, you will be prompted to explain the blocker in your next message. Setting the status is the machine-readable signal; the completion summary or blocker explanation is yours to write in the following message.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "status": { "type": "string", "enum": [ "active", "complete", "blocked" ], "description": "The lifecycle status to set for the current goal. Use \`blocked\` for impossible, unsafe, or contradictory objectives, or after the same non-terminal blocking condition repeats for at least 3 consecutive goal turns." } }, "required": [ "status" ], "additionalProperties": false } }, { "name": "WaitFor", "description": "Wait for background tasks to finish without ending the current turn.\\n\\nUse this when your next step depends on the result of a running background task (a sub-agent, a background bash command, or a background AskUserQuestion). The call suspends inside the current turn until the task finishes or the timeout elapses, then returns the outcome so you can keep working in the same turn. While waiting, no LLM requests are made.\\n\\nGuidelines:\\n\\n- Do not call WaitFor right after dispatching work whose result you do not need yet — finished background tasks notify you automatically. WaitFor is for the moment you genuinely cannot proceed without a result.\\n- \`timeout\` is required, in seconds, capped at 600. To wait longer, call WaitFor again; waking up periodically also lets you re-evaluate the situation.\\n- A timeout is not an error: the result lists the tasks that are still running, and you decide whether to wait again or do other work meanwhile.\\n- Without \`task_id\`, the wait ends as soon as any background task that was running at call time finishes. Tasks started during the wait are not covered by it; their completion arrives via the usual automatic notification.\\n- With \`task_id\`, the wait ends when that task finishes. An unknown \`task_id\` is an error; a task that has already finished returns immediately.\\n- When no background tasks are running, WaitFor returns immediately without waiting.\\n- When the wait ends because a task finished, the result also lists other tasks that finished during the wait window, so failures surface with context.\\n- Waiting has no side effects on the waited tasks: WaitFor never stops a task, and interrupting the wait (for example, a user interruption) leaves every task running.\\n- A finished task's result is delivered exactly once: tasks reported by WaitFor do not also produce an automatic completion notification.\\n- You can only wait for background tasks started by this agent; task IDs belonging to other agents are unknown here.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "timeout": { "type": "integer", "exclusiveMinimum": 0, "maximum": 600, "description": "Maximum time to wait, in seconds (1-600). A timeout is not an error: the tool returns the tasks that are still running, and you can call it again to keep waiting." }, "task_id": { "description": "The background task ID to wait for. When omitted, the wait ends as soon as any background task that was running at call time finishes.", "type": "string" } }, "required": [ "timeout" ], "additionalProperties": false } }, { "name": "Write", "description": "Create, append to, or replace a file entirely.\\n\\n- Missing parent directories are created automatically (like \`mkdir(parents=True, exist_ok=True)\`).\\n- Mode defaults to overwrite; append adds content at EOF without adding a newline.\\n- Write is NOT ALLOWED for incremental changes to existing files, including trivial, one-line, quick, or cosmetic edits. Use Edit instead.\\n- Use Write only when the file does not exist, you intend a complete replacement, or the new contents have little continuity with the old contents.\\n- Do not create unsolicited documentation files (\`*.md\` write-ups, \`README\`s, summaries) just because a task finished — write one only when the user asks for it, or when a task or project instruction requires it (e.g. the plan-mode plan file, created with Write when plan mode directs you to, or a changeset the repo mandates).\\n- Read before overwriting an existing file.\\n- Write ignores the Read/Edit line-number view. NEVER include line prefixes.\\n- Write outputs content literally, including supplied line endings: \\\\n stays LF, \\\\r\\\\n stays CRLF.\\n- For new content too large for one call, overwrite the first chunk, then append subsequent chunks. Never chunk Write to modify an existing file.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to the file to create, append to, or completely overwrite. Relative paths resolve against the working directory; a path outside the working directory must be absolute. Missing parent directories are created automatically." }, "content": { "type": "string", "description": "Raw full file content to write exactly as provided. This does not use the Read/Edit text view." }, "mode": { "description": "Write mode. Defaults to overwrite. append adds content to the end exactly as provided and does not add a newline.", "type": "string", "enum": [ "overwrite", "append" ] } }, "required": [ "path", "content" ], "additionalProperties": false } } ], "time": "<time>" }
-        [wire] llm.request                    { "agentId": "main", "kind": "loop", "provider": "openai", "model": "mock-model", "modelAlias": "mock-model", "thinkingEffort": "off", "maxTokens": 1000000, "toolSelect": false, "systemPromptHash": "ec9c34379c88babbc468ef2f3e0e08cd2f422c8c4a910664fb8bb394d703a575", "toolsHash": "d3e22e7861e40632ba8e85ac56a6f0adea6ca3acf2f00b249b2b656657b59473", "messageCount": 6, "turnStep": "1.1", "time": "<time>" }
+        [wire] llm.tools_snapshot             { "agentId": "main", "hash": "755845c4dac18464a7b45428e5f0f6c35626531988b85d7b6b70b5e03a45e3c2", "tools": [ { "name": "Agent", "description": "Launch a subagent to handle a task. The subagent runs as a same-process loop instance with its own context and wire file. Delegating also keeps the bulk of intermediate file contents out of your own context — you get a conclusion back instead of a pile of dumps.\\n\\nWriting the prompt:\\n- The subagent starts with zero context — it has not seen this conversation. Brief it like a colleague who just walked into the room: state the goal, list what you already know, hand over the specifics.\\n- Lookups (read this file, run that test): put the exact path or command in the prompt. The subagent should not have to search for things you already know.\\n- Investigations (figure out X, find why Y): give the question, not prescribed steps — fixed steps become dead weight when the premise is wrong.\\n- Do not delegate understanding. If the task hinges on a file path or line number, find it yourself first and write it into the prompt.\\n\\nUsage notes:\\n- When the task continues earlier work a subagent already did, prefer resuming that agent (pass its \`resume\` id) over spawning a fresh instance — the resumed agent keeps its prior context.\\n- A subagent's result is only visible to you, not to the user. When the user needs to see what a subagent produced, summarize the relevant parts yourself in your own reply.\\n- Subagents use a fixed 2-hour timeout. If one times out, resume the same agent instead of starting over.\\n\\nWhen NOT to use Agent: skip delegation for trivial work you can do directly — reading a file whose path you already know, searching a small known set of files, or any task that takes only a step or two. Delegation has a context-handoff cost; it pays off only when the task is substantial enough to outweigh it.\\n\\nOnce a subagent is running, leave that scope to it: do not redo its searches or reads in parallel, and do not abandon it midway and finish the job manually. Both undo the context savings the delegation was meant to buy.\\n\\n\\nWhen \`run_in_background=true\`, the subagent runs detached from this turn. The completion arrives in a later turn as a synthetic user-role message containing its result — you do not need to poll, sleep, or check on its progress. Continue with other work or respond to the user. Never fabricate or predict what the result will say.\\n\\nDefault to a foreground subagent (omit \`run_in_background\`) when your next step needs its result — foreground hands the result straight back. Reach for \`run_in_background=true\` only when you have other work to do while it runs and do not need its result to proceed. Never launch in the background and then immediately wait on it (by polling \`TaskOutput\`, sleeping, or otherwise): that just blocks the turn for no benefit — run it in the foreground instead.\\n\\n\\nAvailable agent types (pass via subagent_type):\\n- plan: Read-only implementation planning and architecture design. Use this agent when the parent agent needs a step-by-step implementation plan, key file identification, and architectural trade-off analysis before code changes are made.\\n  Tools: Read, ReadMediaFile, Glob, Grep, WebSearch, FetchURL\\n- coder: General software engineering agent — the only subagent type with file-editing tools; use it for any delegated task that must modify code. Use this agent for non-trivial software engineering work that may require reading files, editing code, running commands, and returning a compact but technically complete summary to the parent agent.\\n  Tools: Bash, CronCreate, CronDelete, CronList, Edit, EnterPlanMode, ExitPlanMode, Glob, Grep, Read, ReadMediaFile, Skill, TaskList, TaskOutput, TaskStop, TodoList, WaitFor, WebSearch, FetchURL, Write, mcp__*\\n- explore: Fast codebase exploration with prompt-enforced read-only behavior. Fast agent specialized for exploring codebases. Use this when you need to quickly find files by patterns (e.g. \\"src/**/*.yaml\\"), search code for keywords (e.g. \\"database connection\\"), or answer questions about the codebase (e.g. \\"how does the auth module work?\\"). When calling this agent, specify the desired thoroughness level: \\"quick\\" for basic searches, \\"medium\\" for moderate exploration, or \\"thorough\\" for comprehensive analysis across multiple locations and naming conventions. Use this agent for any read-only exploration that will clearly require more than 3 search queries. Prefer launching multiple explore agents concurrently when investigating independent questions.\\n  Tools: Bash, Read, ReadMediaFile, Glob, Grep, WebSearch, FetchURL", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "prompt": { "type": "string", "description": "Full task prompt for the subagent" }, "description": { "type": "string", "description": "Short task description (3-5 words) for UI display" }, "subagent_type": { "description": "One of the available agent types (see \\"Available agent types\\" in this tool description). Defaults to \\"coder\\" when omitted.", "type": "string" }, "resume": { "description": "Optional agent ID to resume instead of creating a new instance. When set, do not also pass subagent_type — the resumed agent keeps its own type, and supplying both is rejected.", "type": "string" }, "run_in_background": { "description": "If true, return immediately without waiting for completion. Prefer false unless the task can run independently and there is a clear benefit to not waiting.", "type": "boolean" } }, "required": [ "prompt", "description" ], "additionalProperties": false } }, { "name": "AgentDynamicWorkflow", "description": "Launch multiple subagents from one prompt template, existing agent resumes, or both.\\n\\nUse AgentDynamicWorkflow when many subagents should run the same kind of task over different inputs. The placeholder is exactly \`{{item}}\`. For example, with \`prompt_template\` set to \`Review {{item}} for likely regressions.\` and \`items\` set to \`[\\"src/a.ts\\", \\"src/b.ts\\"]\`, AgentDynamicWorkflow launches two new subagents with those two concrete prompts. For a few differently-shaped tasks, make separate \`Agent\` calls in one message instead.\\n\\nUse \`resume_agent_ids\` to continue subagents that already exist from earlier work, such as ones that failed or timed out: map each agent id to the prompt for that resumed subagent (usually \`continue\` if no extra information is needed). You may combine \`resume_agent_ids\` with \`items\` in the same call to resume existing subagents and launch new ones. Do not duplicate resumed work in \`items\`.\\n\\nEach of these is enforced — a violation is rejected before any subagent starts: provide at least 2 \`items\` unless you pass \`resume_agent_ids\`; whenever \`items\` are present, \`prompt_template\` is required and must contain \`{{item}}\`; and the filled-in prompts must be distinct (two items that expand to the same prompt are rejected).\\n\\nUse enough subagents to keep the work focused and parallel. AgentDynamicWorkflow supports up to 128 subagents, and launches are queued automatically, so it is safe to split large tasks into many clear, independent items.\\n\\nIf \`AgentDynamicWorkflow\` is called, that call must be the only tool call in the response.\\n\\nThis capability is called a \\"dynamic workflow\\" (or simply \\"workflow\\"). Never use the word \\"swarm\\" in the \`description\` field, in subagent prompts, or when talking to the user about this tool.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "description": { "type": "string", "minLength": 1, "description": "Short description for the whole dynamic workflow. It is shown to the user as the workflow and subagent title, so word it as a workflow or run and never use the word \\"swarm\\"." }, "subagent_type": { "description": "Subagent type used for every new subagent spawned from items; defaults to coder when omitted. Resumed subagents always keep their original type, so passing subagent_type together with resume_agent_ids is allowed — it only affects the item-based spawns.", "type": "string", "minLength": 1 }, "prompt_template": { "description": "Prompt template for each subagent. The {{item}} placeholder is replaced with each item value.", "type": "string", "minLength": 1 }, "items": { "description": "Values used to fill {{item}}. Each item launches one new subagent.", "maxItems": 128, "type": "array", "items": { "type": "string", "minLength": 1 } }, "resume_agent_ids": { "description": "Map of existing subagent agent_id to the prompt used to resume that subagent. These resumed subagents are launched before new item-based subagents.", "type": "object", "propertyNames": { "type": "string", "minLength": 1 }, "additionalProperties": { "type": "string", "minLength": 1 } } }, "required": [ "description" ], "additionalProperties": false } }, { "name": "AskUserQuestion", "description": "Use this tool when you need to ask the user questions with structured options during execution. This allows you to:\\n1. Collect user preferences or requirements before proceeding\\n2. Resolve ambiguous or underspecified instructions\\n3. Let the user decide between implementation approaches as you work\\n4. Present concrete options when multiple valid directions exist\\n\\n**When NOT to use:**\\n- When you can infer the answer from context — be decisive and proceed\\n- Trivial decisions that don't materially affect the outcome\\n\\nOverusing this tool interrupts the user's flow. Only use it when the user's input genuinely changes your next action.\\n\\n**Usage notes:**\\n- Users always have an \\"Other\\" option for custom input — don't create one yourself\\n- Use multi_select to allow multiple answers to be selected for a question\\n- Keep option labels concise (1-5 words), use descriptions for trade-offs and details\\n- Each question should have 2-4 meaningful, distinct options\\n- Question texts must be unique across the call, and option labels must be unique within each question\\n- You can ask 1-4 questions at a time; group related questions to minimize interruptions\\n- If you recommend a specific option, list it first and append \\"(Recommended)\\" to its label\\n- The result is JSON with an \`answers\` object keyed by question text; each value is the chosen option's label (comma-separated labels for multi_select, or the user's own words if they picked \\"Other\\"); if \`answers\` is empty and a \`note\` says the user dismissed it, they chose not to answer — do not treat this as selecting the recommended option; decide based on context and do not re-ask the same question\\n- Set background=true when you can keep working without the answer. This starts a background question task and returns a task_id immediately. The answer arrives automatically in a later turn — you do not need to poll, sleep, or check on it. Continue with other work; never fabricate or predict the answer.", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "questions": { "minItems": 1, "maxItems": 4, "type": "array", "items": { "type": "object", "properties": { "question": { "type": "string", "minLength": 1, "description": "A specific, actionable question. End with '?'." }, "header": { "default": "", "description": "Short category tag (max 12 chars, e.g. 'Auth', 'Style').", "type": "string" }, "options": { "minItems": 2, "maxItems": 4, "type": "array", "items": { "type": "object", "properties": { "label": { "type": "string", "minLength": 1, "description": "Concise display text (1-5 words). If recommended, append '(Recommended)'." }, "description": { "default": "", "description": "Brief explanation of trade-offs or implications.", "type": "string" } }, "required": [ "label" ], "additionalProperties": false }, "description": "2-4 meaningful, distinct options. Do NOT include an 'Other' option — the system adds one automatically." }, "multi_select": { "default": false, "description": "Whether the user can select multiple options.", "type": "boolean" } }, "required": [ "question", "options" ], "additionalProperties": false }, "description": "The questions to ask the user (1-4 questions)." }, "background": { "default": false, "description": "Set true to ask in the background and return immediately with a background task_id; you are notified automatically when the user answers — do not poll with TaskOutput while the question is pending.", "type": "boolean" } }, "required": [ "questions" ], "additionalProperties": false } }, { "name": "Bash", "description": "Execute a \`bash\` command. Use this for shell semantics — pipes, env, processes, git, package managers, build/test runners, anything genuinely interactive or multi-step.\\n\\n**Translate these to a dedicated tool instead:**\\n- \`cat\` / \`head\` / \`tail\` (known path) → \`Read\`\\n- \`sed\` / \`awk\` (in-place edit) → \`Edit\`\\n- \`echo > file\` / \`cat <<EOF\` → \`Write\`\\n- \`find\` / recursive \`ls\` to locate files by name pattern → \`Glob\` (plain \`ls <known-directory>\` is fine for listing a directory)\\n- \`grep\` / \`rg\` (search file contents) → \`Grep\`\\n- \`echo\` / \`printf\` (talk to the user) → just output text directly\\n\\nThe dedicated tools render in the per-tool permission UI and keep raw stdout out of the conversation; that is why they are worth reaching for whenever one fits.\\n\\n**Output:**\\nThe stdout and stderr will be combined and returned as a string. The output may be truncated if it is too long. If the command exits non-zero, the output ends with a \`Command failed with exit code: N\` line; a command killed by its timeout or interrupted by the user ends with its own message instead.\\n\\nIf \`run_in_background=true\`, the command will be started as a background task and this tool will return a task ID instead of waiting for command completion. When doing that, you must provide a short \`description\`. Background commands default to a 600s timeout and \`timeout\` is capped at 86400s; set \`disable_timeout=true\` only when the task should run without a timeout. You will be automatically notified when the task completes. After starting one, default to returning control to the user instead of immediately waiting on it. Use \`TaskOutput\` only for a non-blocking status/output snapshot — do not wait on a task you just launched, since its completion arrives automatically. Use \`TaskStop\` only if the task must be cancelled. If a human user wants to inspect background tasks themselves, point them to the background-task panel.\\n\\n**Guidelines for safety and security:**\\n- Each shell tool call will be executed in a fresh shell environment. The shell variables, current working directory changes, and the shell history is not preserved between calls. To run a command in a particular directory, pass the \`cwd\` argument (or use absolute paths) rather than relying on a \`cd\` from an earlier call.\\n- The tool call will return after the command is finished. You shall not use this tool to execute an interactive command or a command that may run forever. For possibly long-running foreground commands, set the \`timeout\` argument in seconds. Foreground commands default to 60s and allow up to 300s. When a foreground command hits its timeout it is moved to the background instead of being killed, and you will be automatically notified when it completes. The user can also move a running foreground command to the background at any time.\\n- Avoid using \`..\` to access files or directories outside of the working directory.\\n- Avoid modifying files outside of the working directory unless explicitly instructed to do so.\\n- Never run commands that require superuser privileges unless explicitly instructed to do so.\\n\\n**Guidelines for efficiency:**\\n- Use \`&&\` to chain commands that genuinely depend on each other, e.g. \`npm install && npm test\`. Independent read-only commands (separate \`git show\`, \`ls\`, or status checks) should be issued as separate parallel Bash calls in one response, not chained into a single call — chaining serializes their execution and mixes their output. Do not stitch outputs together with \`echo\` separators.\\n- Use \`;\` to run commands sequentially regardless of success/failure\\n- Use \`||\` for conditional execution (run second command only if first fails)\\n- Use pipe operations (\`|\`) and redirections (\`>\`, \`>>\`) to chain input and output between commands\\n- Always quote file paths containing spaces with double quotes (e.g., cd \\"/path with spaces/\\")\\n- Compose multi-step logic in a single call with \`if\` / \`case\` / \`for\` / \`while\` control flows.\\n- Prefer \`run_in_background=true\` for long-running builds, tests, watchers, or servers when you need the conversation to continue before the command finishes.\\n\\n**Commands available:**\\nThe following common command categories are usually available. Availability still depends on the host, so when in doubt run \`which <command>\` first to confirm a command exists before relying on it.\\n- Navigation and inspection: \`ls\`, \`pwd\`, \`cd\`, \`stat\`, \`file\`, \`du\`, \`df\`, \`tree\`\\n- File and directory management: \`cp\`, \`mv\`, \`rm\`, \`mkdir\`, \`touch\`, \`ln\`, \`chmod\`, \`chown\`\\n- Text and data processing: \`wc\`, \`sort\`, \`uniq\`, \`cut\`, \`tr\`, \`diff\`, \`xargs\`\\n- Archives and compression: \`tar\`, \`gzip\`, \`gunzip\`, \`zip\`, \`unzip\`\\n- Networking and transfer: \`curl\`, \`wget\`, \`ping\`, \`ssh\`, \`scp\`\\n- Version control: \`git\`; for GitHub-hosted work (PRs, issues, CI runs, API queries) prefer the \`gh\` CLI when installed — it carries the user's GitHub auth and can return structured JSON\\n- Process and system: \`ps\`, \`kill\`, \`top\`, \`env\`, \`date\`, \`uname\`, \`whoami\`\\n- Language and package toolchains: \`node\`, \`npm\`, \`pnpm\`, \`yarn\`, \`python\`, \`pip\` (use whichever the project actually relies on)\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "command": { "type": "string", "minLength": 1, "description": "The command to execute." }, "cwd": { "description": "The working directory in which to run the command. When omitted, the command runs in the session's working directory.", "type": "string" }, "timeout": { "default": 60, "description": "Optional timeout in seconds for the command to execute. Foreground default 60s, max 300s. Background default 600s, max 86400s. Ignored for background commands when disable_timeout=true.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 }, "description": { "description": "A short description for the background task. Required when run_in_background is true.", "type": "string" }, "run_in_background": { "description": "Whether to run the command as a background task.", "type": "boolean" }, "disable_timeout": { "description": "If true, do not apply a timeout to the command. Only applies when run_in_background is true.", "type": "boolean" } }, "required": [ "command" ], "additionalProperties": false } }, { "name": "CreateGoal", "description": "Create a durable, structured goal that the runtime will pursue across multiple turns.\\n\\nCall \`CreateGoal\` only when:\\n\\n- the user explicitly asks you to start a goal or work autonomously toward an outcome, or\\n- a host goal-intake prompt asks you to create one.\\n\\nDo NOT create a goal for greetings, ordinary questions, or vague requests that lack a\\nverifiable completion condition. A goal needs a checkable end state.\\n\\nWhen the request is vague, ask the user for the missing completion criterion before creating\\nthe goal. If the user clearly insists after you warn them that the wording is vague or risky,\\nrespect that and create the goal.\\n\\nInclude a \`completionCriterion\` when the user provides one, or when it can be stated without\\ninventing new requirements. Keep \`objective\` concise; reference long task descriptions by file\\npath rather than pasting them.\\n\\nCreating a goal fails if one already exists, so use \`replace: true\` only when the user explicitly\\nwants to abandon the current goal and start a new one.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "objective": { "type": "string", "minLength": 1, "description": "The objective to pursue. Must have a verifiable end state." }, "completionCriterion": { "description": "How to verify the goal is complete. Include when the user provides one.", "type": "string" }, "replace": { "description": "Replace an existing active, paused, or blocked goal instead of failing.", "type": "boolean" } }, "required": [ "objective" ], "additionalProperties": false } }, { "name": "CronCreate", "description": "Schedule a prompt to be enqueued at a future time. Use for both recurring schedules and one-shot reminders.\\n\\nUses standard 5-field cron in the user's local timezone: minute hour day-of-month month day-of-week. \`0 9 * * *\` means 9am local — no timezone conversion needed.\\n\\n## One-shot tasks (recurring: false)\\n\\nFor \\"remind me at X\\" or \\"at <time>, do Y\\" requests — fire once then auto-delete.\\nPin minute/hour/day-of-month/month to specific values:\\n  \\"remind me at 2:30pm today to check the deploy\\" → cron: \\"30 14 <today_dom> <today_month> *\\", recurring: false\\n  \\"tomorrow morning, run the smoke test\\" → cron: \\"57 8 <tomorrow_dom> <tomorrow_month> *\\", recurring: false\\n\\nOne-shots are best for near-term reminders. A task only fires while its session is still alive (see Session lifetime below), so favor near times — within hours or a few days — rather than scheduling weeks or months ahead.\\n\\n## Recurring jobs (recurring: true, the default)\\n\\nFor \\"every N minutes\\" / \\"every hour\\" / \\"weekdays at 9am\\" requests:\\n  \\"*/5 * * * *\\" (every 5 min), \\"0 * * * *\\" (hourly), \\"0 9 * * 1-5\\" (weekdays at 9am local)\\n\\n## Avoid the :00 and :30 minute marks when the task allows it\\n\\nEvery user who asks for \\"9am\\" gets \`0 9\`, and every user who asks for \\"hourly\\" gets \`0 *\` — which means requests from across the planet land on the API at the same instant. When the user's request is approximate, pick a minute that is NOT 0 or 30:\\n  \\"every morning around 9\\" → \\"57 8 * * *\\" or \\"3 9 * * *\\" (not \\"0 9 * * *\\")\\n  \\"hourly\\" → \\"7 * * * *\\" (not \\"0 * * * *\\")\\n  \\"in an hour or so, remind me to...\\" → pick whatever minute you land on, don't round\\n\\nOnly use minute 0 or 30 when the user names that exact time and clearly means it (\\"at 9:00 sharp\\", \\"at half past\\", coordinating with a meeting). When in doubt, nudge a few minutes early or late — the user will not notice, and the fleet will.\\n\\n## Coalesce semantics\\n\\nFires are delivered only while the session is idle: a fire that comes due during an active turn is held and delivered at the next idle moment, never injected mid-turn.\\n\\nIf the scheduler slept past multiple ideal fire times (laptop closed, long-running turn, etc.), only **one** fire is delivered when it wakes up. The origin carries \`coalescedCount\` showing how many ideal fires were collapsed into this single delivery. You should treat \`coalescedCount > 1\` as \\"I missed some checks; only the latest state matters\\" rather than running the prompt that many times.\\n\\n## Cron-fire envelope\\n\\nWhen a cron task fires, the prompt you scheduled is re-injected wrapped in an XML envelope that exposes the fire context:\\n\\n\`\`\`\\n<cron-fire jobId=\\"...\\" cron=\\"...\\" recurring=\\"true|false\\" coalescedCount=\\"N\\" stale=\\"true|false\\">\\n<prompt>\\nyour original prompt text, verbatim\\n</prompt>\\n</cron-fire>\\n\`\`\`\\n\\nThe envelope is parseable. Use \`coalescedCount > 1\` to know multiple ideal fires were collapsed into a single delivery (treat as \\"only the latest state matters\\"), and \`stale=\\"true\\"\` as a cue that the task is past its 7-day threshold.\\n\\n## 7-day stale behavior\\n\\nRecurring tasks that have been alive for more than 7 days fire one\\nfinal time with \`stale: true\` on the envelope, and the system then\\nauto-deletes the task. The flag is the model's notice that this is\\nthe last delivery. If the schedule is still wanted, call \`CronCreate\`\\nagain with the same \`cron\` and \`prompt\` — that resets \`createdAt\` and\\nstarts a fresh 7-day window. One-shot tasks are never marked stale.\\n\\n## Jitter behavior\\n\\nAnti-herd jitter is applied deterministically per task id:\\n  - Recurring: ideal fire time is shifted **forward** by an offset ≤ min(10% of the cron period, 15 minutes). A \`*/5 * * * *\` task can drift up to 30s; a \`0 9 * * *\` task can drift up to 15 minutes.\\n  - One-shot: only when the ideal fire lands on \`:00\` or \`:30\` of the hour, the fire is pulled **earlier** by ≤ 90 seconds. Other minutes pass through unchanged.\\n\\n## One-shot vs recurring — when to pick which\\n\\nUse \`recurring: false\` for \\"remind me at X\\" style requests, single deadlines, \\"in N minutes do Y\\", and any task that should not repeat. Use \`recurring: true\` for periodic polling (CI status, build watchers, scheduled reports), workday rituals, and anything the user explicitly described as recurring.\\n\\n## Session lifetime\\n\\nCron tasks live in the current session. When you exit, they\\nare persisted under the session homedir; resuming the same session\\nreloads them and the scheduler resumes from each task's \`createdAt\`. Fire times that fell during the offline window are\\ncollapsed into a single delivery via \`coalescedCount\` (and recurring\\ntasks past their 7-day window arrive with \`stale: true\` as their final\\ndelivery).\\n\\nTasks do **not** carry over into a brand-new session — they are scoped\\nto the resumed session id, not to the working directory.\\n\\n## Limits\\n\\nA session holds at most 50 live cron tasks; creating one beyond that is rejected. (The \`prompt\` body is also capped — see its parameter description.) Expressions that never fire within the next 5 years (e.g. \`0 0 31 2 *\`, an impossible date) are rejected at create time.\\n\\n## Returned fields\\n\\n\`id\` (ULID), \`cron\` (the normalized expression), \`humanSchedule\` (English summary), \`recurring\`,\\n\`nextFireAt\` (local ISO timestamp with numeric offset, or null). \`id\` is needed by \`CronDelete\`.\\n\\n## Tell the user how to cancel or modify\\n\\nAfter successfully creating a task, proactively tell the user how they can cancel or modify it later. Users have no direct \`/cron\` command or self-service UI to manage reminders themselves; they must ask the model to make changes (e.g. \\"cancel my 9am reminder\\" or \\"change my daily check to 10am\\"). Include the task \`id\` in your message so the user can reference it.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "cron": { "type": "string", "description": "5-field cron expression in local time: \\"M H DoM Mon DoW\\" (e.g. \\"*/5 * * * *\\" = every 5 minutes; \\"30 14 28 2 *\\" = Feb 28 at 2:30pm local — a pinned date like this repeats yearly unless you also pass recurring: false)." }, "prompt": { "type": "string", "minLength": 1, "maxLength": 8192, "description": "The prompt to enqueue at each fire time. Limited to 8 KiB (UTF-8)." }, "recurring": { "default": true, "description": "true (default) = fire on every cron match until deleted or auto-expired after 7 days. false = fire once at the next match, then auto-delete. Use false for \\"remind me at X\\" one-shot requests with pinned minute/hour/dom/month.", "type": "boolean" } }, "required": [ "cron", "prompt" ], "additionalProperties": false } }, { "name": "CronDelete", "description": "Cancel a scheduled cron job by id.\\n\\nUse this tool to remove a cron task previously scheduled with\\n\`CronCreate\`. The \`id\` is the ULID value returned by \`CronCreate\`, or\\nshown in the \`id:\` column of \`CronList\` — quote it verbatim, no\\nprefix.\\n\\nBehaviour by task kind:\\n\\n- **Recurring task** (\`recurring: true\`): stops all future fires\\n  immediately. The scheduler picks up the deletion on its next tick.\\n- **One-shot task** (\`recurring: false\`): cancels the pending fire if\\n  it has not happened yet. One-shots that have already fired\\n  auto-delete themselves, so calling \`CronDelete\` on a fired one-shot\\n  returns \\"no cron job with id ...\\".\\n\\nNot-found is reported as an error (not a silent no-op) so you can\\ncorrect yourself — typically by calling \`CronList\` to see which ids\\nare actually live, rather than re-trying with the same stale id.\\n\\nRefresh pattern (use when you want a stale recurring schedule to\\ncontinue):\\n\\nStale recurring tasks are auto-deleted by the system after their final\\nfire — there is nothing for \`CronDelete\` to remove at that point. To\\nkeep the schedule running, just call \`CronCreate\` with the same \`cron\`\\nand \`prompt\`. Use \`CronList\`'s \`prompt\` field to recall the original\\ntext after a context compaction.\\n\\n\`CronDelete\` remains the right call when you want to cancel a task\\nthat is still live (recurring not yet stale, or a one-shot still\\npending).\\n\\nGuidelines:\\n\\n- Users have no direct \`/cron\` command or self-service UI to delete\\n  tasks themselves; they must ask the model to cancel a reminder.\\n  When deleting on behalf of a user, confirm the action and report\\n  the result plainly.\\n- Cron deletion is irreversible — there is no undo. If you delete the\\n  wrong task, you must re-create it with \`CronCreate\`.\\n- If the model is unsure which id is current (e.g. after a context\\n  compaction), call \`CronList\` first rather than guessing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "id": { "type": "string", "description": "The cron job id (ULID) returned by CronCreate / CronList." } }, "required": [ "id" ], "additionalProperties": false } }, { "name": "CronList", "description": "List all cron jobs currently scheduled in this session.\\n\\nUse this tool to see every pending cron task — both recurring jobs and\\none-shot reminders — that you (or the user) have scheduled with\\n\`CronCreate\`. The output is the entry point for inspecting scheduled\\nwork: it returns a stable id, the original cron expression, a human\\nrendering, the next post-jitter fire time, the recurring flag, the\\ntask's age in days, and a stale indicator.\\n\\nEach record carries:\\n\\n- \`id\` — the task id (a ULID). Pass this to \`CronDelete\` to remove the\\n  task, or quote it in user-facing messages when asking for\\n  confirmation.\\n- \`cron\` — the verbatim 5-field cron expression as scheduled.\\n- \`humanSchedule\` — plain-English rendering (e.g. \`every 5 minutes\`).\\n- \`prompt\` — the scheduled prompt text, JSON-encoded so embedded\\n  newlines stay on one line. Truncated to 200 UTF-8 bytes with\\n  \`…(truncated)\` if longer. Use this to recall what a task is for\\n  after a context compaction, and as the source for the\\n  \`CronCreate\` refresh ritual.\\n- \`nextFireAt\` — local ISO timestamp with an explicit numeric offset\\n  for the next fire **after jitter has been applied**. The actual fire\\n  may land slightly before or after a round \`:00\` / \`:30\` minute mark\\n  due to herd-avoidance jitter; this is the value the scheduler will\\n  compare against, so it reflects what will really happen. \`null\` if\\n  the expression has no fire in the next 5 years (should not happen\\n  for tasks created through \`CronCreate\`, which validates).\\n- \`recurring\` — \`true\` for cadenced jobs, \`false\` for one-shots.\\n- \`ageDays\` — \`(now - createdAt) / day\`, two decimal places. Useful\\n  when deciding whether a long-running cron is still relevant.\\n- \`stale\` — \`true\` when a recurring task is older than 7 days. The\\n  system **auto-deletes the task after this fire** to bound session\\n  lifetime; the \`stale: true\` flag is the model's notice that this is\\n  the final delivery. To resume the same schedule, call \`CronCreate\`\\n  again with the original \`cron\` and \`prompt\` (the \`prompt\` row above\\n  carries it for exactly this purpose). One-shots are never marked\\n  stale — they fire at most once by construction.\\n\\nGuidelines:\\n\\n- This tool is read-only and never mutates state, so it is always\\n  safe to call (including in plan mode).\\n- Users cannot directly manage cron tasks themselves; if they want to\\n  cancel or modify a schedule, route the request through the model\\n  (i.e. call \`CronDelete\` or \`CronCreate\` on their behalf).\\n- The empty case returns \`cron_jobs: 0\\\\nNo cron jobs scheduled.\`. Cron\\n  tasks survive a resume of the same session but do not bleed into new\\n  sessions.\\n- After a context compaction, or whenever you are unsure which cron\\n  jobs are live, call this tool to re-enumerate them rather than\\n  guessing ids from earlier in the conversation.\\n- Records are separated by a line containing just \`---\`, in the\\n  insertion order they were scheduled.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "Edit", "description": "Perform exact replacements in existing files.\\n\\n- Edit is mandatory for every incremental change, especially small edits. DO NOT use Write or Bash \`sed\`.\\n- Read the target file before every Edit. DO NOT call Edit from memory, stale context, or a guessed \`old_string\`.\\n- Take \`old_string\` and \`new_string\` from the Read output view.\\n- Drop the line-number prefix and tab; match only file content.\\n- \`old_string\` must be unique unless \`replace_all\` is set.\\n- If \`old_string\` is ambiguous, add surrounding context. Use \`replace_all\` only when every occurrence should change — for example, renaming a symbol throughout the file.\\n- Multiple Edit calls may run in one response only when they do not target the same file.\\n- DO NOT issue consecutive Edit calls on the same file. A previous Edit can invalidate a later Edit's \`old_string\`, causing \`old_string not found\`. Read the file again before the next Edit.\\n- A write lock serializes same-file edits in response order, but serialization does not make stale \`old_string\` valid.\\n- For pure CRLF files, Read shows LF; use LF in \`old_string\` and \`new_string\`, and Edit writes CRLF back.\\n- For mixed endings or lone carriage returns, Read shows carriage returns as \\\\r; include actual \\\\r escapes in those positions.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to the text file to edit. Relative paths resolve against the working directory; a path outside the working directory must be absolute." }, "old_string": { "type": "string", "minLength": 1, "description": "Exact content to replace from the Read output view, without the line-number prefix. Use LF for pure CRLF files; use actual \\\\r escapes where Read shows \\\\r." }, "new_string": { "type": "string", "description": "Replacement text in the same Read output view. LF is written back as CRLF only for pure CRLF files." }, "replace_all": { "description": "Set true only when every occurrence of old_string should be replaced.", "type": "boolean" } }, "required": [ "path", "old_string", "new_string" ], "additionalProperties": false } }, { "name": "EnterPlanMode", "description": "Use this tool proactively when you're about to start a non-trivial implementation task.\\nGetting user sign-off on your approach via ExitPlanMode before writing code prevents wasted effort.\\n\\nUse it when ANY of these conditions apply:\\n\\n1. New Feature Implementation - e.g. \\"Add a caching layer to the API\\"\\n2. Multiple Valid Approaches - e.g. \\"Optimize database queries\\" (indexing vs rewrite vs caching)\\n3. Code Modifications - e.g. \\"Refactor auth module to support OAuth\\"\\n4. Architectural Decisions - e.g. \\"Add WebSocket support\\"\\n5. Multi-File Changes - involves more than 2-3 files\\n6. Unclear Requirements - need exploration to understand scope\\n7. User Preferences Matter - if user input would materially change the implementation approach, use EnterPlanMode to structure the decision\\n\\nPermission mode notes:\\n- EnterPlanMode enters plan mode automatically without an approval prompt in all permission modes.\\n- In yolo and manual modes, ExitPlanMode still presents the plan to the user for approval.\\n- In auto permission mode, do not use AskUserQuestion; make the best decision from available context.\\n- In auto permission mode, ExitPlanMode exits plan mode without asking the user.\\n- Use EnterPlanMode only when planning itself adds value.\\n\\nWhen NOT to use:\\n- Single-line or few-line fixes (typos, obvious bugs, small tweaks)\\n- User gave very specific, detailed instructions\\n- Pure research/exploration tasks\\n\\nOnce you are in plan mode, a reminder walks you through the workflow (explore → design → write the plan file → \`ExitPlanMode\`) and enforces read-only access. For non-trivial tasks where you are unsure of the codebase structure or relevant code paths, use \`Agent(subagent_type=\\"explore\\")\` to investigate first when the \`Agent\` tool is available.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "ExitPlanMode", "description": "Use this tool when you are in plan mode and have finished writing your plan to the plan file and are ready for user approval.\\n\\n## How This Tool Works\\n- You should have already written your plan to the plan file specified in the plan mode reminder.\\n- This tool does NOT take the plan content as a parameter - it reads the plan from the file you wrote.\\n- The user will see the contents of your plan file when they review it. In auto permission mode, the tool reads the file and exits plan mode without asking the user.\\n\\n## When to Use\\nOnly use this tool for tasks that require planning implementation steps. For research tasks (searching files, reading code, understanding the codebase), do NOT use this tool.\\n\\n## What a good plan contains\\nList specific, verifiable steps grounded in the actual codebase — real files, functions, and commands, in a sensible order. Each step should be concrete enough to act on and to check. Avoid vague filler like \\"improve performance\\" or \\"add tests\\"; say what to change and where.\\n\\n## Multiple Approaches\\nIf your plan offers multiple alternative approaches, pass them via the \`options\` parameter so the user can choose which one to execute — see the \`options\` parameter for the format, count, and reserved labels. In yolo and manual modes the user sees all options alongside the host's Reject and Revise controls.\\n\\n## Before Using\\n- In auto permission mode, do NOT use AskUserQuestion; make the best decision from available context.\\n- In auto permission mode, this tool exits plan mode without asking the user.\\n- In yolo and manual modes, this tool still presents the plan to the user for approval.\\n- If auto permission mode is not active and you have unresolved questions, use AskUserQuestion first.\\n- If auto permission mode is not active and you have multiple approaches and haven't narrowed down yet, consider using AskUserQuestion first to let the user choose, then write a plan for the chosen approach only.\\n- Once your plan is finalized, use THIS tool to request approval.\\n- Do NOT use AskUserQuestion to ask \\"Is this plan OK?\\" or \\"Should I proceed?\\" - that is exactly what ExitPlanMode does.\\n- If rejected, revise based on feedback and call ExitPlanMode again.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "options": { "description": "When the plan contains multiple alternative approaches, list them here so the user can choose which one to execute. Provide up to 3 options; 2-3 distinct approaches work best when the plan offers a real choice. Passing a single option is allowed and is equivalent to a plain plan approval. Each option represents a distinct approach from the plan. Do not use \\"Reject\\", \\"Revise\\", \\"Approve\\", or \\"Reject and Exit\\" as labels.", "minItems": 1, "maxItems": 3, "type": "array", "items": { "type": "object", "properties": { "label": { "type": "string", "minLength": 1, "maxLength": 80, "description": "Short name for this option (1-8 words). Append \\"(Recommended)\\" if you recommend this option." }, "description": { "default": "", "description": "Brief summary of this approach and its trade-offs.", "type": "string" } }, "required": [ "label" ], "additionalProperties": false } } }, "additionalProperties": false } }, { "name": "FetchURL", "description": "Fetch content from a URL. The content is returned either as the main text extracted from the page, or as the full response body verbatim; a note at the top of the result states which of the two you received, so you can judge how complete it is. Use this when you need to read a specific web page.\\n\\nOnly fully-formed public \`http\`/\`https\` URLs are supported; other schemes and private or loopback addresses are not fetched. Very large pages may be truncated or refused. The fetch carries no login or session for the target site, so pages behind authentication (private repositories, internal dashboards) return a login page or an error instead of the real content — if the text you get back looks like a generic landing or sign-in page, treat that as the login wall, not the answer, and reach the content through a credentialed route (an authenticated CLI or MCP tool) instead.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "url": { "type": "string", "description": "The URL to fetch content from." } }, "required": [ "url" ], "additionalProperties": false } }, { "name": "GetGoal", "description": "Read the current goal: its objective, completion criterion, status, and budgets (turns, tokens,\\ntime, and how much of each remains). When the goal has stopped, it also reports the terminal reason.\\n\\nUse \`GetGoal\` before deciding whether to continue working, report completion, report a blocker,\\nor respect a pause. It returns \`{ \\"goal\\": null }\` when there is no current goal.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": {}, "additionalProperties": false } }, { "name": "Glob", "description": "Find files by glob pattern, sorted by modification time (most recent first).\\n\\nPowered by ripgrep. Respects \`.gitignore\`, \`.ignore\`, and \`.rgignore\` by default — set \`include_ignored\` to also match ignored files (e.g. build outputs, \`node_modules\`). Sensitive files (such as \`.env\`) are always filtered out. Matches are files only — directories themselves are never listed; to find a directory, glob for a file inside it (e.g. \`**/fixtures/**\`).\\n\\nGood patterns:\\n- \`*.ts\` — all files matching an extension, at any depth below the search root (a bare pattern without \`/\` matches recursively)\\n- \`src/*.ts\` — files directly inside \`src/\` (one level, not recursive)\\n- \`src/**/*.ts\` — recursive walk with a subdirectory anchor and extension\\n- \`**/*.py\` — recursive walk from the search root for an extension\\n- \`*.{ts,tsx}\` — brace expansion is supported\\n- \`{src,test}/**/*.ts\` — cartesian brace expansion is supported too\\n\\nResults are capped at the first 100 matching paths. If a search would return more, a truncation marker is appended. Refine the pattern (extension, subdirectory) when 100 is not enough, or call again with a narrower anchor.\\n\\nLarge-directory caveat — avoid recursing into dependency / build output even with an anchor, especially when \`include_ignored\` is set:\\n- \`node_modules/**/*.js\`, \`.venv/**/*.py\`, \`__pycache__/**\`, \`target/**\` can produce thousands of results that truncate at the match cap and waste context. Prefer specific subpaths like \`node_modules/react/src/**/*.js\`.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "pattern": { "type": "string", "description": "Glob pattern to match files." }, "path": { "description": "Directory to search. Accepts an absolute path, or a path relative to the current working directory. Defaults to the current working directory.", "type": "string" }, "include_ignored": { "description": "Also match files excluded by ignore files such as \`.gitignore\`, \`.ignore\`, and \`.rgignore\` (for example \`node_modules\` or build outputs). Sensitive files (such as \`.env\`) remain filtered out for safety. VCS metadata directories (\`.git\` and similar) are always skipped, even when this is true. Defaults to false.", "type": "boolean" }, "include_dirs": { "description": "Deprecated and ignored. Results are always files-only — directories are never listed. Accepted only so older calls that still pass this flag are not rejected by parameter validation.", "type": "boolean" } }, "required": [ "pattern" ], "additionalProperties": false } }, { "name": "Grep", "description": "Search file contents using regular expressions (powered by ripgrep).\\n\\nUse Grep when the task is to find unknown content or unknown file locations. Do not use shell \`grep\` or \`rg\` directly; this tool applies workspace path policy, output limits, and sensitive-file filtering.\\nALWAYS use Grep tool instead of running \`grep\` or \`rg\` from a shell — direct shell calls bypass workspace policy, output limits, and sensitive-file filtering.\\nIf you already know a concrete file path and need to inspect its contents, use Read directly instead.\\n\\nWrite patterns in ripgrep regex syntax, which differs from POSIX \`grep\` syntax. For example, braces are special, so escape them as \`\\\\{\` to match a literal \`{\`.\\n\\nHidden files (dotfiles such as \`.gitlab-ci.yml\` or \`.eslintrc.json\`) are searched by default. To also search files excluded by \`.gitignore\` (such as \`node_modules\` or build outputs), set \`include_ignored\` to \`true\`. Sensitive files (such as \`.env\`) are always skipped for safety, even when \`include_ignored\` is \`true\`.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "pattern": { "type": "string", "description": "Regular expression to search for." }, "path": { "description": "File or directory to search. Accepts an absolute path, or a path relative to the current working directory. Omit to search the current working directory. Use Read instead when you already know a concrete file path and need its contents.", "type": "string" }, "glob": { "description": "Optional glob filter for which files to search, e.g. \`*.ts\`. Matched against each file's full absolute path, so a path-anchored pattern like \`src/**/*.ts\` silently matches nothing — use a basename pattern (\`*.ts\`), or anchor with \`**/\` (\`**/src/**/*.ts\`). To scope the search to a directory, use \`path\` instead.", "type": "string" }, "type": { "description": "Optional ripgrep file type filter, such as ts or py. Prefer this over \`glob\` when filtering by language or file kind: it is more efficient and less error-prone than an equivalent glob pattern.", "type": "string" }, "output_mode": { "description": "Shape of the result. \`content\` shows matching lines (honors \`-A\`, \`-B\`, \`-C\`, \`-n\`, and \`head_limit\`); \`files_with_matches\` shows only the paths of files that contain a match, most-recently-modified first (honors \`head_limit\`); \`count_matches\` shows per-file match counts as \`path:count\` lines, preceded by an aggregate total line. Defaults to \`files_with_matches\`.", "type": "string", "enum": [ "content", "files_with_matches", "count_matches" ] }, "-i": { "description": "Perform a case-insensitive search. Defaults to false.", "type": "boolean" }, "-n": { "description": "Prefix each matching line with its line number. Applies only when \`output_mode\` is \`content\`. Defaults to true.", "type": "boolean" }, "-A": { "description": "Number of lines to show after each match. Applies only when \`output_mode\` is \`content\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "-B": { "description": "Number of lines to show before each match. Applies only when \`output_mode\` is \`content\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "-C": { "description": "Number of lines to show before and after each match. Applies only when \`output_mode\` is \`content\`; takes precedence over \`-A\` and \`-B\`.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "head_limit": { "description": "Limit output to the first N lines/entries after offset. Defaults to 250. Pass 0 for unlimited.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "offset": { "description": "Number of leading lines/entries to skip before applying \`head_limit\`. Use it together with \`head_limit\` to page through large result sets. Defaults to 0.", "type": "integer", "minimum": 0, "maximum": 9007199254740991 }, "multiline": { "description": "Enable multiline matching, where the pattern can span line boundaries and \`.\` also matches newlines. Defaults to false.", "type": "boolean" }, "include_ignored": { "description": "Also search files excluded by ignore files such as \`.gitignore\`, \`.ignore\`, and \`.rgignore\` (for example \`node_modules\` or build outputs). Sensitive files (such as \`.env\`) remain filtered out for safety. VCS metadata directories (\`.git\` and similar) are always skipped, even when this is true. Defaults to false.", "type": "boolean" } }, "required": [ "pattern" ], "additionalProperties": false } }, { "name": "Read", "description": "Read a text file from the local filesystem.\\n\\nIf the user provides a concrete file path to a text file, call Read directly. Do not \`Glob\`, \`ls\`, or otherwise pre-check known text file paths; missing or invalid file paths return errors you can handle. Do not use Read for directories; use \`ls\` via Bash for a known directory, or Glob when you need files matching a name pattern (Glob lists files only, never directories). Use \`Grep\` only when the task is to search for unknown content or locations.\\n\\nWhen you need several files, prefer to read them in parallel: emit multiple \`Read\` calls in a single response instead of reading one file per turn.\\n\\n- Relative paths resolve against the working directory; a path outside the working directory must be absolute.\\n- Returns up to 1000 lines or 100 KB per call, whichever comes first; lines longer than 2000 chars are truncated mid-line.\\n- Page larger files with \`line_offset\` (1-based start line) and \`n_lines\`. Omit \`n_lines\` to read up to the 1000-line cap.\\n- Sensitive files (\`.env\` files, credential stores, SSH private keys, and similar secrets) are refused to protect secrets; do not attempt to read them. Templates and public keys are exempt: \`.env.example\` / \`.env.sample\` / \`.env.template\` and public SSH keys such as \`id_rsa.pub\` read normally.\\n- UTF-8 text files are read directly. UTF-16 LE/BE text files (with or without a BOM) are detected automatically and transcoded to UTF-8 for display; the status block notes the detected encoding, and Edit/Write on such a file still expect UTF-8 — convert its encoding first (e.g. with \`iconv\`). Other encodings (e.g. GBK), binary files, and files containing NUL bytes are refused.\\n- Negative line_offset reads from the end of the file (for example, -100 reads the last 100 lines); the absolute value cannot exceed 1000.\\n- Output format: \`<line-number>\\\\t<content>\` per line.\\n- A \`<system>...</system>\` status block is appended after the file content; it summarizes how much was read (line and byte counts, truncation, line-ending notes) and is not part of the file itself.\\n- Pure CRLF files are displayed with LF line endings; \`Edit\` matches this output and preserves CRLF when writing back.\\n- Mixed or lone carriage-return line endings are shown as \`\\\\r\` and require exact \`Edit.old_string\` escapes.\\n- After a successful \`Edit\`/\`Write\`, do not re-read solely to prove the write landed. When the task depends on an exact file, API, or output shape, inspect the final external contract before finishing.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to a text file. Relative paths resolve against the working directory; a path outside the working directory must be absolute. Directories are not supported; use \`ls\` via Bash for a known directory, or Glob for pattern search." }, "line_offset": { "description": "The line number to start reading from. Omit to start at line 1. Negative values read from the end of the file; the absolute value cannot exceed 1000.", "anyOf": [ { "type": "integer", "minimum": 1, "maximum": 9007199254740991 }, { "type": "integer", "minimum": -1000, "maximum": -1 } ] }, "n_lines": { "description": "The number of lines to read; the tool also applies its internal cap. Omit to read up to the internal cap of 1000 lines.", "type": "integer", "exclusiveMinimum": 0, "maximum": 9007199254740991 } }, "required": [ "path" ], "additionalProperties": false } }, { "name": "SetGoalBudget", "description": "Set a hard budget limit for the current goal.\\n\\nUse this only when the user clearly gives a runtime limit, such as:\\n\\n- \\"stop after 20 turns\\"\\n- \\"use no more than 500k tokens\\"\\n- \\"finish within 30 minutes\\"\\n\\nDo not invent limits. Do not call this for vague wording such as \\"spend some time\\" or\\n\\"try to be quick\\".\\n\\nIf the user gives a compound time, convert it to one supported unit before calling this tool.\\nFor example, \\"2 hours and 3 minutes\\" can be set as \`value: 123, unit: \\"minutes\\"\`.\\n\\nA time budget must be between 1 second and 24 hours — the tool rejects anything shorter or\\nlonger, telling the user it is not a reasonable goal budget. Turn and token budgets are not\\nbounded this way; they must be positive and are rounded to the nearest whole number (minimum 1).\\n\\nSupported units:\\n\\n- \`turns\`\\n- \`tokens\`\\n- \`milliseconds\`\\n- \`seconds\`\\n- \`minutes\`\\n- \`hours\`\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "value": { "type": "number", "exclusiveMinimum": 0, "description": "The positive numeric budget value." }, "unit": { "type": "string", "enum": [ "turns", "tokens", "milliseconds", "seconds", "minutes", "hours" ] } }, "required": [ "value", "unit" ], "additionalProperties": false } }, { "name": "Skill", "description": "Invoke a registered skill from the current skill listing. BLOCKING REQUIREMENT: when a skill from the listing matches the user's request, you MUST call this tool (not free-form text). Do not re-invoke a skill to repeat work already done: if a \`<skill-loaded>\` block for it with the same \`args\` is already present in the conversation, follow those instructions directly instead of calling the tool again. Do call the tool again when you need the skill with different arguments — the loaded block was expanded with the earlier \`args\` and will not reflect new inputs.", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "skill": { "type": "string", "description": "The exact name of the skill to invoke, spelled as it appears in the current skill listing (e.g. \\"commit\\", \\"pdf\\")." }, "args": { "description": "Optional argument string for the skill, written like a command line (e.g. \`-m \\"fix bug\\"\`, \`123\`, a file path). It is split on whitespace (quotes group a token) and expanded into the skill's placeholders ($NAME, $1, $ARGUMENTS); if the skill body has no placeholders, the whole string is still appended as a trailing \`ARGUMENTS:\` line. Omit it only when there is nothing to pass.", "type": "string" } }, "required": [ "skill" ], "additionalProperties": false } }, { "name": "TaskList", "description": "List background tasks and their current status.\\n\\nUse this tool to discover which background tasks exist and where each one\\nstands. It is the entry point for inspecting background work: it returns a\\ntask ID, status, and description for every task it reports, plus the command,\\nPID, and (once finished) exit code for shell tasks, and a stop reason for any\\ntask that ended early.\\n\\nGuidelines:\\n\\n- After a context compaction, or whenever you are unsure which background\\n  tasks are running or what their task IDs are, call this tool to\\n  re-enumerate them instead of guessing a task ID.\\n- Prefer the default \`active_only=true\`, which lists only non-terminal tasks.\\n  Pass \`active_only=false\` only when you specifically need to see tasks that\\n  have already finished. With \`active_only=false\` the result may also include\\n  \`lost\` tasks — tasks left over from a previous process that can no longer be\\n  inspected or controlled; treat them as already terminated.\\n- \`limit\` caps how many tasks are returned. It accepts a value between 1 and\\n  100 and defaults to 20 when omitted.\\n- This tool only lists tasks; it does not return their output. Use it first\\n  to locate the task ID you need, then call \`TaskOutput\` with that ID to read\\n  the task's output and details.\\n- This tool is read-only and does not change any state, so it is always safe\\n  to call, including in plan mode.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "active_only": { "default": true, "description": "Whether to list only non-terminal background tasks.", "type": "boolean" }, "limit": { "default": 20, "description": "Maximum number of tasks to return.", "type": "integer", "minimum": 1, "maximum": 100 } }, "additionalProperties": false } }, { "name": "TaskOutput", "description": "Retrieve a snapshot of a running or completed background task.\\n\\nUse this after \`Bash(run_in_background=true)\`, \`Agent(run_in_background=true)\`, or \`AskUserQuestion(background=true)\` to check progress, or to read the output of a task that has already completed.\\n\\nGuidelines:\\n- Prefer relying on automatic completion notifications. Use this tool only when you need task output before the automatic notification arrives.\\n- This tool is always non-blocking: it returns the current status/output snapshot immediately and never waits for the task to finish.\\n- Do not use TaskOutput to wait for a result you need before continuing — if your next step depends on the task's result, run that task in the foreground instead. TaskOutput is for a deliberate progress check you will act on without blocking, not a way to sit and wait for a background task you just launched.\\n- This tool returns structured task metadata, a fixed-size output preview, and an output_path for the full log.\\n- For a terminal task, the metadata also explains why it ended. A shell command that runs to completion reports \`status: completed\` on a zero exit, or \`status: failed\` with its non-zero \`exit_code\` — judge that failure from the \`exit_code\`, because a plain command failure carries no \`stop_reason\` and no \`terminal_reason\`. \`terminal_reason\` is a categorical label emitted only when the end is not an ordinary exit: \`timed_out\` when the deadline aborted it, \`stopped\` when it was explicitly stopped, or \`failed\` when it errored without producing an exit code; the \`stopped\` and \`failed\` cases also carry a human-readable \`stop_reason\`. A task that finished on its own with a clean exit carries neither \`stop_reason\` nor \`terminal_reason\`.\\n- The full, never-truncated log is always available at output_path; use the \`Read\` tool with that path to page through it, whether or not the preview was truncated.\\n- This tool works with the generic background task system and should remain the primary read path for future task types, not just bash.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "task_id": { "type": "string", "description": "The background task ID to inspect." } }, "required": [ "task_id" ], "additionalProperties": false } }, { "name": "TaskStop", "description": "Stop a running background task.\\n\\nOnly use this when a task must genuinely be cancelled — for a task that is\\nfinishing normally, wait for its completion notification or inspect it with\\n\`TaskOutput\` instead of stopping it.\\n\\nGuidelines:\\n- This is a general-purpose stop capability for any background task. It is not\\n  a bash-specific kill.\\n- Stopping a task is destructive: it may leave partial side effects behind.\\n  Use it with care.\\n- If the task has already finished, this tool simply returns its current\\n  status.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "task_id": { "type": "string", "description": "The background task ID to stop." }, "reason": { "default": "Stopped by TaskStop", "description": "Short reason recorded when the task is stopped.", "type": "string" } }, "required": [ "task_id" ], "additionalProperties": false } }, { "name": "TodoList", "description": "Use this tool to maintain a structured TODO list as you work through a multi-step task. Use it proactively and often when progress tracking helps the current work. This is especially useful in long-running investigations and implementation tasks with several tool calls; in plan mode, write the plan to the plan file rather than tracking it here.\\n\\n**When to use:**\\n- Multi-step tasks that span several tool calls\\n- Tracking investigation progress across a large codebase search\\n- Planning a sequence of edits before making them\\n- After receiving new multi-step instructions, capture the requirements as todos\\n- Before starting a tracked task, mark exactly one item as \`in_progress\`\\n- Immediately after finishing a tracked task, mark it \`done\`; do not batch completions at the end\\n\\n**When NOT to use:**\\n- Single-shot answers that complete in one or two tool calls\\n- Trivial requests where tracking adds no clarity\\n- Purely conversational or informational replies\\n\\n**Avoid churn:**\\n- Do not re-call this tool when nothing meaningful has changed since the last call — update the list only after real progress.\\n- When unsure of the current state, call query mode first (omit \`todos\`) to check the list before deciding what to update.\\n- If no available tool can move any task forward, tell the user where you are stuck instead of repeatedly re-ordering the same todos.\\n\\n**How to use:**\\n- Call with \`todos: [...]\` to replace the full list. Statuses: pending / in_progress / done.\\n- Call with no \`todos\` argument to retrieve the current list without changing it.\\n- Call with \`todos: []\` to clear the list.\\n- Keep titles short and actionable (e.g. \\"Read session-control.ts\\", \\"Add planMode flag to TurnManager\\").\\n- Update statuses as you make progress.\\n- When work is underway, keep exactly one task \`in_progress\`.\\n- Only mark a task \`done\` when it is fully accomplished.\\n- Never mark a task \`done\` if tests are failing, implementation is partial, unresolved errors remain, or required files/dependencies could not be found.\\n- If you encounter a blocker, keep the blocked task \`in_progress\` or add a new pending task describing what must be resolved.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "todos": { "description": "The updated todo list. Omit to read the current todo list without making changes. Pass an empty array to clear the list.", "type": "array", "items": { "type": "object", "properties": { "title": { "type": "string", "minLength": 1, "description": "Short, actionable title for the todo." }, "status": { "type": "string", "enum": [ "pending", "in_progress", "done" ], "description": "Current status of the todo." } }, "required": [ "title", "status" ], "additionalProperties": false } } }, "additionalProperties": false } }, { "name": "UpdateGoal", "description": "Set the status of the current goal. This is how you resume, complete, or block an autonomous goal.\\n\\n- \`active\` — resume a paused or blocked goal when the user explicitly asks you to work on that goal.\\n- \`complete\` — the objective is satisfied and any stated validation has passed. The goal ends and a completion summary is recorded. Before using this, verify the current state against the actual objective and every explicit requirement. Treat weak or indirect evidence as not complete. Do not use \`complete\` merely because a budget is nearly exhausted or you want to stop.\\n- \`blocked\` — a genuine impasse prevents useful progress: an external condition, required user input, missing credentials or permissions, a persistent technical failure, or an impossible, unsafe, or contradictory objective. For non-terminal blockers, do not use \`blocked\` the first time you hit the blocker. The same blocking condition must repeat for at least 3 consecutive goal turns before you call \`blocked\`, counting the original/user-triggered turn and automatic continuations. If a previously blocked goal is resumed, treat the resumed run as a fresh blocked audit. If the objective itself is impossible, unsafe, or contradictory, call \`blocked\` in the same turn instead of running more goal turns. Do not use \`blocked\` because the work is large, hard, slow, uncertain, incomplete, still needs validation, would benefit from clarification, or needs more goal turns. Once the 3-turn threshold is met and you cannot make meaningful progress without user input or an external-state change, call \`blocked\` instead of leaving the goal active.\\n\\nMost active goal turns should not call this tool. If you complete one useful slice of work and material work remains, end the turn normally without calling UpdateGoal; the runtime will prompt you to continue in the next goal turn. Call \`complete\` only when all required work is done, any stated validation has passed, and there is no useful next action. Do not call \`complete\` after only producing a plan, summary, first pass, or partial result. Call \`blocked\` only after the blocked audit threshold is met. If you call \`blocked\`, you will be prompted to explain the blocker in your next message. Setting the status is the machine-readable signal; the completion summary or blocker explanation is yours to write in the following message.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "status": { "type": "string", "enum": [ "active", "complete", "blocked" ], "description": "The lifecycle status to set for the current goal. Use \`blocked\` for impossible, unsafe, or contradictory objectives, or after the same non-terminal blocking condition repeats for at least 3 consecutive goal turns." } }, "required": [ "status" ], "additionalProperties": false } }, { "name": "WaitFor", "description": "Wait for background tasks to finish without ending the current turn.\\n\\nUse this when your next step depends on the result of a running background task (a sub-agent, a background bash command, or a background AskUserQuestion). The call suspends inside the current turn until the task finishes or the timeout elapses, then returns the outcome so you can keep working in the same turn. While waiting, no LLM requests are made.\\n\\nGuidelines:\\n\\n- Do not call WaitFor right after dispatching work whose result you do not need yet — finished background tasks notify you automatically. WaitFor is for the moment you genuinely cannot proceed without a result.\\n- \`timeout\` is required, in seconds, capped at 600. To wait longer, call WaitFor again; waking up periodically also lets you re-evaluate the situation.\\n- A timeout is not an error: the result lists the tasks that are still running, and you decide whether to wait again or do other work meanwhile.\\n- Without \`task_id\`, the wait ends as soon as any background task that was running at call time finishes. Tasks started during the wait are not covered by it; their completion arrives via the usual automatic notification.\\n- With \`task_id\`, the wait ends when that task finishes. An unknown \`task_id\` is an error; a task that has already finished returns immediately.\\n- When no background tasks are running, WaitFor returns immediately without waiting.\\n- When the wait ends because a task finished, the result also lists other tasks that finished during the wait window, so failures surface with context.\\n- Waiting has no side effects on the waited tasks: WaitFor never stops a task, and interrupting the wait (for example, a user interruption) leaves every task running.\\n- A finished task's result is delivered exactly once: tasks reported by WaitFor do not also produce an automatic completion notification.\\n- You can only wait for background tasks started by this agent; task IDs belonging to other agents are unknown here.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "timeout": { "type": "integer", "exclusiveMinimum": 0, "maximum": 600, "description": "Maximum time to wait, in seconds (1-600). A timeout is not an error: the tool returns the tasks that are still running, and you can call it again to keep waiting." }, "task_id": { "description": "The background task ID to wait for. When omitted, the wait ends as soon as any background task that was running at call time finishes.", "type": "string" } }, "required": [ "timeout" ], "additionalProperties": false } }, { "name": "Write", "description": "Create, append to, or replace a file entirely.\\n\\n- Missing parent directories are created automatically (like \`mkdir(parents=True, exist_ok=True)\`).\\n- Mode defaults to overwrite; append adds content at EOF without adding a newline.\\n- Write is NOT ALLOWED for incremental changes to existing files, including trivial, one-line, quick, or cosmetic edits. Use Edit instead.\\n- Use Write only when the file does not exist, you intend a complete replacement, or the new contents have little continuity with the old contents.\\n- Do not create unsolicited documentation files (\`*.md\` write-ups, \`README\`s, summaries) just because a task finished — write one only when the user asks for it, or when a task or project instruction requires it (e.g. the plan-mode plan file, created with Write when plan mode directs you to, or a changeset the repo mandates).\\n- Read before overwriting an existing file.\\n- Write ignores the Read/Edit line-number view. NEVER include line prefixes.\\n- Write outputs content literally, including supplied line endings: \\\\n stays LF, \\\\r\\\\n stays CRLF.\\n- For new content too large for one call, overwrite the first chunk, then append subsequent chunks. Never chunk Write to modify an existing file.\\n", "parameters": { "$schema": "http://json-schema.org/draft-07/schema#", "type": "object", "properties": { "path": { "type": "string", "description": "Path to the file to create, append to, or completely overwrite. Relative paths resolve against the working directory; a path outside the working directory must be absolute. Missing parent directories are created automatically." }, "content": { "type": "string", "description": "Raw full file content to write exactly as provided. This does not use the Read/Edit text view." }, "mode": { "description": "Write mode. Defaults to overwrite. append adds content to the end exactly as provided and does not add a newline.", "type": "string", "enum": [ "overwrite", "append" ] } }, "required": [ "path", "content" ], "additionalProperties": false } } ], "time": "<time>" }
+        [wire] llm.request                    { "agentId": "main", "kind": "loop", "provider": "openai", "model": "mock-model", "modelAlias": "mock-model", "thinkingEffort": "off", "maxTokens": 1000000, "toolSelect": false, "systemPromptHash": "ec9c34379c88babbc468ef2f3e0e08cd2f422c8c4a910664fb8bb394d703a575", "toolsHash": "755845c4dac18464a7b45428e5f0f6c35626531988b85d7b6b70b5e03a45e3c2", "messageCount": 6, "turnStep": "1.1", "time": "<time>" }
         [emit] assistant.delta                { "time": "<time>", "agentId": "main", "turnId": 1, "delta": "No lookup tool is available." }
         [emit] agent.activity.updated         { "time": "<time>", "lifecycle": "ready", "turn": { "turnId": 1, "origin": { "kind": "user" }, "phase": "streaming", "stream": "assistant", "step": 1, "ending": false, "pendingApprovals": [], "activeToolCalls": [], "since": "<time>" }, "background": [], "agentId": "main" }
         [wire] usage.record                   { "agentId": "main", "model": "mock-model", "usage": { "inputOther": 184, "output": 10, "inputCacheRead": 0, "inputCacheCreation": 0 }, "usageScope": "turn", "time": "<time>" }
diff --git a/packages/agent-core-v2/test/wire/wireService.test.ts b/packages/agent-core-v2/test/wire/wireService.test.ts
index 012b5a0c4..ea367086d 100644
--- a/packages/agent-core-v2/test/wire/wireService.test.ts
+++ b/packages/agent-core-v2/test/wire/wireService.test.ts
@@ -14,6 +14,7 @@ import { IAppendLogStore } from '#/persistence/interface/appendLogStore';
 import { IFileSystemStorageService, StorageError, StorageErrors } from '#/persistence/interface/storage';
 import { WIRE_PROTOCOL_VERSION } from '#/wire/migration/migration';
 import { wireJournalBackupKey } from '#/wire/repair';
+import { WireError, WireErrors } from '#/wire/errors';
 import { IWireService } from '#/wire/wire';
 import { AGENT_WIRE_RECORD_KEY, type WireRecord } from '#/wire/record';
 
@@ -612,6 +613,188 @@ describe('WireService corruption repair', () => {
       },
     ]);
   });
+
+  it('retries a failed repair before the next append and heals the journal first', async () => {
+    const capture: RepairCapture = { warnings: [], events: [] };
+    const svc = wireWithCapture(KEY, capture);
+    const prefix = `${currentMetadata()}\n`;
+    const raw = `${prefix}GARBAGE\n`;
+    await seedCorrupt(raw);
+    const originalWrite = storage.write.bind(storage);
+    storage.write = async (scope, key, data, options) => {
+      if (key === AGENT_WIRE_RECORD_KEY) throw new Error('disk full');
+      return originalWrite(scope, key, data, options);
+    };
+    await collect(svc.readJournal());
+    storage.write = originalWrite;
+
+    svc.appendRecord({ type: 'wire.test.new', time: 7 });
+    await svc.flush();
+
+    expect(await rawBytes()).toBe(`${prefix}${JSON.stringify({ type: 'wire.test.new', time: 7 })}\n`);
+    expect(await rawBytes(BACKUP_KEY)).toBe(raw);
+    expect(await collect(svc.readJournal())).toEqual([
+      { type: 'metadata', protocol_version: WIRE_PROTOCOL_VERSION, created_at: 1 },
+      { type: 'wire.test.new', time: 7 },
+    ]);
+    expect(capture.events).toEqual([
+      {
+        name: 'wire_repair',
+        payload: {
+          kind: 'corrupted',
+          outcome: 'failed',
+          dropped_count: 1,
+          backup_created: true,
+        },
+      },
+      {
+        name: 'wire_repair',
+        payload: {
+          kind: 'corrupted',
+          outcome: 'repaired',
+          dropped_count: 1,
+          backup_created: false,
+        },
+      },
+    ]);
+  });
+
+  it('retries a pending repair on flush even when nothing was appended', async () => {
+    const capture: RepairCapture = { warnings: [], events: [] };
+    const svc = wireWithCapture(KEY, capture);
+    const prefix = `${currentMetadata()}\n`;
+    const raw = `${prefix}GARBAGE\n`;
+    await seedCorrupt(raw);
+    const originalWrite = storage.write.bind(storage);
+    storage.write = async (scope, key, data, options) => {
+      if (key === AGENT_WIRE_RECORD_KEY) throw new Error('disk full');
+      return originalWrite(scope, key, data, options);
+    };
+    await collect(svc.readJournal());
+    storage.write = originalWrite;
+
+    await svc.flush();
+
+    expect(await rawBytes()).toBe(prefix);
+    expect(await rawBytes(BACKUP_KEY)).toBe(raw);
+    expect(capture.events).toMatchObject([
+      { name: 'wire_repair', payload: { outcome: 'failed' } },
+      { name: 'wire_repair', payload: { outcome: 'repaired' } },
+    ]);
+  });
+
+  it('reports the still-broken journal from flush when nothing was appended', async () => {
+    const capture: RepairCapture = { warnings: [], events: [] };
+    const svc = wireWithCapture(KEY, capture);
+    const prefix = `${currentMetadata()}\n`;
+    const raw = `${prefix}GARBAGE\n`;
+    await seedCorrupt(raw);
+    const originalWrite = storage.write.bind(storage);
+    storage.write = async (scope, key, data, options) => {
+      if (key === AGENT_WIRE_RECORD_KEY) throw new Error('disk full');
+      return originalWrite(scope, key, data, options);
+    };
+    await collect(svc.readJournal());
+
+    await expect(svc.flush()).rejects.toThrow('Wire journal repair did not complete');
+    expect(await rawBytes()).toBe(raw);
+  });
+
+  it('refuses to append behind the corrupted tail while a failed repair keeps failing', async () => {
+    const capture: RepairCapture = { warnings: [], events: [] };
+    const svc = wireWithCapture(KEY, capture);
+    const prefix = `${currentMetadata()}\n`;
+    const raw = `${prefix}GARBAGE\n`;
+    await seedCorrupt(raw);
+    const originalWrite = storage.write.bind(storage);
+    storage.write = async (scope, key, data, options) => {
+      if (key === AGENT_WIRE_RECORD_KEY) throw new Error('disk full');
+      return originalWrite(scope, key, data, options);
+    };
+    await collect(svc.readJournal());
+
+    const unexpected: unknown[] = [];
+    setUnexpectedErrorHandler((error) => unexpected.push(error));
+    try {
+      svc.appendRecord({ type: 'wire.test.doomed', time: 8 });
+      await expect(svc.flush()).rejects.toThrow('Wire journal repair did not complete');
+
+      expect(await rawBytes()).toBe(raw);
+      expect(unexpected).toHaveLength(1);
+      expect(unexpected[0]).toBeInstanceOf(WireError);
+      expect(unexpected[0]).toMatchObject({ code: WireErrors.codes.RECORDS_WRITE_FAILED });
+      expect(capture.events).toEqual([
+        {
+          name: 'wire_repair',
+          payload: {
+            kind: 'corrupted',
+            outcome: 'failed',
+            dropped_count: 1,
+            backup_created: true,
+          },
+        },
+        {
+          name: 'wire_repair',
+          payload: {
+            kind: 'corrupted',
+            outcome: 'failed',
+            dropped_count: 1,
+            backup_created: false,
+          },
+        },
+      ]);
+    } finally {
+      resetUnexpectedErrorHandler();
+    }
+  });
+
+  it('surfaces the discarded record to flush callers when the repair never reaches the rewrite', async () => {
+    const capture: RepairCapture = { warnings: [], events: [] };
+    const svc = wireWithCapture(KEY, capture);
+    const prefix = `${currentMetadata()}\n`;
+    const raw = `${prefix}GARBAGE\n`;
+    await seedCorrupt(raw);
+    const originalWrite = storage.write.bind(storage);
+    storage.write = async (scope, key, data, options) => {
+      if (key === BACKUP_KEY) throw new Error('disk full');
+      return originalWrite(scope, key, data, options);
+    };
+    await collect(svc.readJournal());
+
+    const unexpected: unknown[] = [];
+    setUnexpectedErrorHandler((error) => unexpected.push(error));
+    try {
+      svc.appendRecord({ type: 'wire.test.doomed', time: 9 });
+      await expect(svc.flush()).rejects.toThrow('Wire journal repair did not complete');
+
+      expect(await rawBytes()).toBe(raw);
+      expect(unexpected).toHaveLength(1);
+      expect(unexpected[0]).toBeInstanceOf(WireError);
+      expect(unexpected[0]).toMatchObject({ code: WireErrors.codes.RECORDS_WRITE_FAILED });
+      expect(capture.events).toEqual([
+        {
+          name: 'wire_repair',
+          payload: {
+            kind: 'corrupted',
+            outcome: 'failed',
+            dropped_count: 1,
+            backup_created: false,
+          },
+        },
+        {
+          name: 'wire_repair',
+          payload: {
+            kind: 'corrupted',
+            outcome: 'failed',
+            dropped_count: 1,
+            backup_created: false,
+          },
+        },
+      ]);
+    } finally {
+      resetUnexpectedErrorHandler();
+    }
+  });
 });
 
 describe('WireService flush', () => {
diff --git a/packages/agent-gateway/src/protocol/rest-task.ts b/packages/agent-gateway/src/protocol/rest-task.ts
index fb0f4e2d6..459475383 100644
--- a/packages/agent-gateway/src/protocol/rest-task.ts
+++ b/packages/agent-gateway/src/protocol/rest-task.ts
@@ -26,6 +26,12 @@ export const cancelTaskResultSchema = z.object({
 });
 export type CancelTaskResult = z.infer<typeof cancelTaskResultSchema>;
 
+export const detachTaskResultSchema = z.object({
+  detached: z.boolean(),
+  status: taskStatusSchema,
+});
+export type DetachTaskResult = z.infer<typeof detachTaskResultSchema>;
+
 export const taskAlreadyFinishedDataSchema = z.object({
   cancelled: z.literal(false),
 });
diff --git a/packages/agent-gateway/src/routes/tasks.ts b/packages/agent-gateway/src/routes/tasks.ts
index cbe4fec91..ead8df209 100644
--- a/packages/agent-gateway/src/routes/tasks.ts
+++ b/packages/agent-gateway/src/routes/tasks.ts
@@ -8,6 +8,7 @@ import {
 import { ErrorCode } from '../protocol/error-codes';
 import {
   cancelTaskResultSchema,
+  detachTaskResultSchema,
   getTaskQuerySchema,
   getTaskResponseSchema,
   listTasksQuerySchema,
@@ -135,11 +136,11 @@ export function registerTasksRoutes(app: TasksRouteHost, core: Scope): void {
   );
   app.get(getRoute.path, getRoute.options, getRoute.handler as Parameters<TasksRouteHost['get']>[2]);
 
-  const cancelRoute = defineRoute(
+  const taskActionRoute = defineRoute(
     {
       method: 'POST',
       path: '/sessions/{session_id}/tasks/{tail}',
-      success: { data: cancelTaskResultSchema },
+      success: { data: z.union([cancelTaskResultSchema, detachTaskResultSchema]) },
       errors: {
         [ErrorCode.VALIDATION_FAILED]: { detailsSchema },
         [ErrorCode.SESSION_NOT_FOUND]: {},
@@ -149,9 +150,9 @@ export function registerTasksRoutes(app: TasksRouteHost, core: Scope): void {
           detailsSchema: z.object({ current_status: z.string() }),
         },
       },
-      description: 'Cancel a task',
+      description: 'Run a task action',
       tags: ['tasks'],
-      operationId: 'cancelTask',
+      operationId: 'runTaskAction',
     },
     async (req, reply) => {
       const { session_id, tail } = req.params as {
@@ -160,7 +161,7 @@ export function registerTasksRoutes(app: TasksRouteHost, core: Scope): void {
       };
       const parsed = parseActionSuffix({
         tail,
-        allowedActions: ['cancel'] as const,
+        allowedActions: ['cancel', 'detach'] as const,
         resourceLabel: 'task',
       });
       if (parsed.kind === 'invalid') {
@@ -190,18 +191,29 @@ export function registerTasksRoutes(app: TasksRouteHost, core: Scope): void {
         reply.send(taskNotFound(session_id, task_id, req.id));
         return;
       }
-      const wireStatus = toWireTask(session_id, found).status;
-      if (isTerminalStatus(wireStatus)) {
-        reply.send(taskAlreadyFinished(session_id, task_id, wireStatus, req.id));
+
+      if (parsed.action === 'cancel') {
+        const wireStatus = toWireTask(session_id, found).status;
+        if (isTerminalStatus(wireStatus)) {
+          reply.send(taskAlreadyFinished(session_id, task_id, wireStatus, req.id));
+          return;
+        }
+
+        await resolved.tasks?.stopByUser(task_id);
+        requestLog(req)?.info({ session_id, task_id }, 'task cancelled');
+        reply.send(okEnvelope({ cancelled: true as const }, req.id));
         return;
       }
 
-      await resolved.tasks?.stopByUser(task_id);
-      requestLog(req)?.info({ session_id, task_id }, 'task cancelled');
-      reply.send(okEnvelope({ cancelled: true as const }, req.id));
+      const detached = found.status === 'running' && found.detached === false;
+      const info = resolved.tasks?.detach(task_id) ?? found;
+      if (detached) {
+        requestLog(req)?.info({ session_id, task_id }, 'task detached');
+      }
+      reply.send(okEnvelope({ detached, status: mapStatus(info.status) }, req.id));
     },
   );
-  app.post(cancelRoute.path, cancelRoute.options, cancelRoute.handler as Parameters<TasksRouteHost['post']>[2]);
+  app.post(taskActionRoute.path, taskActionRoute.options, taskActionRoute.handler as Parameters<TasksRouteHost['post']>[2]);
 }
 
 type ResolvedTasks =
@@ -292,7 +304,10 @@ function toWireTask(
   if (info.kind === 'agent' && info.subagentType !== undefined) {
     base.subagent_type = info.subagentType;
   }
-  if (info.kind === 'agent' && info.parentToolCallId !== undefined) {
+  if (
+    (info.kind === 'agent' || info.kind === 'process') &&
+    info.parentToolCallId !== undefined
+  ) {
     base.parent_tool_call_id = info.parentToolCallId;
   }
   if (output !== undefined) {
diff --git a/packages/agent-gateway/test/search/wireExtract.test.ts b/packages/agent-gateway/test/search/wireExtract.test.ts
index 06f32ec46..1da4dfd60 100644
--- a/packages/agent-gateway/test/search/wireExtract.test.ts
+++ b/packages/agent-gateway/test/search/wireExtract.test.ts
@@ -56,6 +56,14 @@ describe('extractFromWireLine', () => {
     },
   );
 
+  it('filters out subagent system triggers even though they open transcript turns', () => {
+    expect(
+      extractFromWireLine(
+        userRecord('scan the repo', 1_700_000_000_000, { kind: 'system_trigger', name: 'subagent' }),
+      ),
+    ).toEqual([]);
+  });
+
   it.each(['skill_activation', 'plugin_command'])(
     'keeps %s messages the user typed as a slash command',
     (kind) => {
diff --git a/packages/agent-gateway/test/services/transcript.test.ts b/packages/agent-gateway/test/services/transcript.test.ts
index 5febc15c9..b5c5396fb 100644
--- a/packages/agent-gateway/test/services/transcript.test.ts
+++ b/packages/agent-gateway/test/services/transcript.test.ts
@@ -172,6 +172,31 @@ describe('AgentTranscriptProjector', () => {
     expect(turn.state).toBe('completed');
   });
 
+  it('projects the live prompt for subagent system triggers and keeps it through turn.ended', () => {
+    const projector = new AgentTranscriptProjector('main', TEST_SESSION_ID);
+    const tx = new AgentTranscript('main');
+    const feed = (event: ProjectorBusEvent): void => {
+      tx.apply(projector.map(event));
+    };
+
+    feed(
+      ev({
+        type: 'turn.started',
+        turnId: 0,
+        origin: { kind: 'system_trigger', name: 'subagent' },
+        prompt: 'scan the repo',
+        promptAttachments: [{ kind: 'image', fileId: 'file_1' }],
+      }),
+    );
+    feed(ev({ type: 'assistant.delta', turnId: 0, delta: 'scanning' }));
+    feed(ev({ type: 'turn.ended', turnId: 0, reason: 'completed' }));
+
+    const turn = turnOps('t0', tx.getItems());
+    expect(turn.prompt).toBe('scan the repo');
+    expect(turn.attachmentIds).toEqual(['t0.att1']);
+    expect(turn.state).toBe('completed');
+  });
+
   it('projects turn.started promptAttachments into attachment entities and turn.attachmentIds', () => {
     const projector = new AgentTranscriptProjector('main', TEST_SESSION_ID);
     const tx = new AgentTranscript('main');
diff --git a/packages/agent-gateway/test/tasks.test.ts b/packages/agent-gateway/test/tasks.test.ts
index 214360cbe..aa9eac560 100644
--- a/packages/agent-gateway/test/tasks.test.ts
+++ b/packages/agent-gateway/test/tasks.test.ts
@@ -352,6 +352,84 @@ describe('server-v2 /api/v1/sessions/{sid}/tasks', () => {
     expect(body.code).toBe(40001);
   });
 
+  it('detaches a running foreground task and flips run_in_background', async () => {
+    const id = await createSession();
+    const tasks = await mainAgentTasks(id);
+    const taskId = tasks.registerTask(fakeTask('process'), { detached: false });
+    await flush();
+
+    const detached = await postJson<{ detached: boolean; status: string }>(
+      `/api/v1/sessions/${id}/tasks/${taskId}:detach`,
+    );
+    expect(detached.body.code).toBe(0);
+    expect(detached.body.data).toEqual({ detached: true, status: 'running' });
+    expect(tasks.getTask(taskId)?.detached).toBe(true);
+
+    const got = await getJson<TaskWire>(`/api/v1/sessions/${id}/tasks/${taskId}`);
+    expect(got.body.data.run_in_background).toBe(true);
+  });
+
+  it('answers concurrent detach requests idempotently', async () => {
+    const id = await createSession();
+    const tasks = await mainAgentTasks(id);
+    const taskId = tasks.registerTask(fakeTask('process'), { detached: false });
+    await flush();
+
+    const [first, second] = await Promise.all([
+      postJson<{ detached: boolean; status: string }>(
+        `/api/v1/sessions/${id}/tasks/${taskId}:detach`,
+      ),
+      postJson<{ detached: boolean; status: string }>(
+        `/api/v1/sessions/${id}/tasks/${taskId}:detach`,
+      ),
+    ]);
+    expect(first.body.code).toBe(0);
+    expect(second.body.code).toBe(0);
+    const detachedFlags = [first.body.data.detached, second.body.data.detached].toSorted();
+    expect(detachedFlags).toEqual([false, true]);
+    expect(first.body.data.status).toBe('running');
+    expect(second.body.data.status).toBe('running');
+  });
+
+  it('reports detached: false for an already-background task', async () => {
+    const id = await createSession();
+    const tasks = await mainAgentTasks(id);
+    const taskId = tasks.registerTask(fakeTask('process'));
+    await flush();
+
+    const { body } = await postJson<{ detached: boolean; status: string }>(
+      `/api/v1/sessions/${id}/tasks/${taskId}:detach`,
+    );
+    expect(body.code).toBe(0);
+    expect(body.data).toEqual({ detached: false, status: 'running' });
+  });
+
+  it('reports detached: false with the current status for a finished task', async () => {
+    const id = await createSession();
+    const tasks = await mainAgentTasks(id);
+    const taskId = tasks.registerTask({
+      ...fakeTask('process'),
+      start: (sink) => {
+        void sink.settle({ status: 'completed' });
+      },
+    });
+    await flush();
+    expect(tasks.getTask(taskId)?.status).toBe('completed');
+
+    const { body } = await postJson<{ detached: boolean; status: string }>(
+      `/api/v1/sessions/${id}/tasks/${taskId}:detach`,
+    );
+    expect(body.code).toBe(0);
+    expect(body.data).toEqual({ detached: false, status: 'completed' });
+  });
+
+  it('detaching an unknown task returns 40406', async () => {
+    const id = await createSession();
+    await mainAgentTasks(id);
+    const { body } = await postJson<null>(`/api/v1/sessions/${id}/tasks/nope:detach`);
+    expect(body.code).toBe(40406);
+  });
+
   it('returns 40401 for an unknown session on all three endpoints', async () => {
     const list = await getJson<null>('/api/v1/sessions/nope/tasks');
     expect(list.body.code).toBe(40401);
@@ -361,5 +439,8 @@ describe('server-v2 /api/v1/sessions/{sid}/tasks', () => {
 
     const cancelled = await postJson<null>('/api/v1/sessions/nope/tasks/tid:cancel');
     expect(cancelled.body.code).toBe(40401);
+
+    const detached = await postJson<null>('/api/v1/sessions/nope/tasks/tid:detach');
+    expect(detached.body.code).toBe(40401);
   });
 });
diff --git a/packages/agent-gateway/test/transcript.test.ts b/packages/agent-gateway/test/transcript.test.ts
index 88fa67674..206904188 100644
--- a/packages/agent-gateway/test/transcript.test.ts
+++ b/packages/agent-gateway/test/transcript.test.ts
@@ -1108,9 +1108,10 @@ describe('server-v2 /api/v1/sessions/{sid}/transcript', () => {
     const main = byAgent.get('main')!;
     expect(main.messages.map((m) => [m.turn_id, m.prompt])).toEqual([
       ['t0', 'hi'],
+      ['t1', 'subagent run prompt'],
       ['t2', 'second question'],
     ]);
-    expect(main.messages[1]!.attachment_ids).toEqual(['att_1']);
+    expect(main.messages[2]!.attachment_ids).toEqual(['att_1']);
     expect(main.attachments).toEqual([
       expect.objectContaining({
         attachmentId: 'att_1',
diff --git a/packages/node-sdk/src/login/flows.ts b/packages/node-sdk/src/login/flows.ts
index 50dfdb10a..f84ae20c3 100644
--- a/packages/node-sdk/src/login/flows.ts
+++ b/packages/node-sdk/src/login/flows.ts
@@ -9,6 +9,7 @@ import {
   OPENAI_CODEX_PROVIDER_ID,
   OpenAICodexApiError,
   OpenPlatformApiError,
+  OAuthAccessDeniedError,
   runOpenAICodexOAuthFlow,
   type ProviderModelInfo,
   type PythinkerConfigShape,
@@ -196,6 +197,10 @@ async function handleOpenAICodexOAuthLogin(ui: LoginUi): Promise<boolean> {
       });
     } catch (error) {
       if (controller.signal.aborted) return false;
+      if (error instanceof OAuthAccessDeniedError) {
+        ui.showError(`OpenAI Codex login cancelled: ${formatErrorMessage(error)}`);
+        return false;
+      }
       ui.showError(`OpenAI Codex login failed: ${formatErrorMessage(error)}`);
       return false;
     }
diff --git a/packages/node-sdk/test/session-prompt-events.test.ts b/packages/node-sdk/test/session-prompt-events.test.ts
index 059df1c45..a76a165aa 100644
--- a/packages/node-sdk/test/session-prompt-events.test.ts
+++ b/packages/node-sdk/test/session-prompt-events.test.ts
@@ -14,7 +14,12 @@ import { PYTHINKER_CODE_PLATFORM } from '@pymodel/pythinker-code-oauth';
 import type * as KosongModule from '@pymodel/kosong';
 import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
 
-import { createPythinkerHarness, type Event, type PythinkerHarness } from '#/index';
+import {
+  createPythinkerHarness,
+  createPythinkerHarnessV2,
+  type Event,
+  type PythinkerHarness,
+} from '#/index';
 
 import { TEST_IDENTITY } from './test-identity';
 
@@ -346,6 +351,70 @@ describe('Session.prompt events', () => {
     }
   });
 
+  it('carries the prompt on the public turn.started event for subagent system triggers (v2 engine)', async () => {
+    const homeDir = await makeTempDir();
+    const workDir = await makeTempDir();
+    const harness = createPythinkerHarnessV2({
+      identity: TEST_IDENTITY,
+      homeDir,
+    });
+    const sseChunk = (delta: Record<string, unknown>, finishReason: string | null = null): string =>
+      `data: ${JSON.stringify({
+        id: 'chatcmpl-stub',
+        object: 'chat.completion.chunk',
+        created: 0,
+        model: 'fake-model',
+        choices: [{ index: 0, delta, finish_reason: finishReason }],
+      })}`;
+    const fetchStub = vi.spyOn(globalThis, 'fetch').mockImplementation(async (input) => {
+      const url = typeof input === 'string' ? input : input instanceof URL ? input.href : input.url;
+      if (url === 'https://model.example.test/v1/chat/completions') {
+        const body = [sseChunk({ role: 'assistant', content: 'init done' }), sseChunk({}, 'stop'), 'data: [DONE]']
+          .map((line) => line + '\n\n')
+          .join('');
+        return new Response(body, {
+          status: 200,
+          headers: { 'Content-Type': 'text/event-stream' },
+        });
+      }
+      throw new Error(`Unexpected fetch: ${url}`);
+    });
+
+    try {
+      await harness.setConfig({
+        providers: {
+          local: { type: 'openai', baseUrl: 'https://model.example.test/v1', apiKey: 'sk-test' },
+        },
+        models: {
+          'fake-model': { provider: 'local', model: 'fake-model', maxContextSize: 262144 },
+        },
+        defaultModel: 'fake-model',
+      });
+      const session = await harness.createSession({ id: 'ses_init_rpc_v2', workDir });
+      const events: Event[] = [];
+      const unsubscribe = session.onEvent((event) => {
+        events.push(event);
+      });
+
+      await session.init();
+      unsubscribe();
+
+      const spawned = events.find((event) => event.type === 'subagent.spawned');
+      expect(events).toContainEqual(
+        expect.objectContaining({
+          type: 'turn.started',
+          sessionId: session.id,
+          agentId: spawned?.type === 'subagent.spawned' ? spawned.subagentId : undefined,
+          origin: { kind: 'system_trigger', name: 'subagent' },
+          prompt: expect.stringContaining('Task requirements:'),
+        }),
+      );
+    } finally {
+      fetchStub.mockRestore();
+      await harness.close();
+    }
+  });
+
   it('includes persisted subagent replay only when resume explicitly requests it', async () => {
     const homeDir = await makeTempDir();
     const workDir = await makeTempDir();
diff --git a/packages/node-sdk/test/v1-v2-parity.test.ts b/packages/node-sdk/test/v1-v2-parity.test.ts
index 84deb8ea1..3a4850439 100644
--- a/packages/node-sdk/test/v1-v2-parity.test.ts
+++ b/packages/node-sdk/test/v1-v2-parity.test.ts
@@ -371,6 +371,7 @@ function projectBackgroundTask(info: BackgroundTaskInfo): unknown {
   delete projected['endedAt'];
   delete projected['timeoutMs'];
   delete projected['terminalNotificationSuppressed'];
+  delete projected['parentToolCallId'];
   return projected;
 }
 
diff --git a/packages/oauth/src/index.ts b/packages/oauth/src/index.ts
index bf01a9c9c..ff2f768ae 100644
--- a/packages/oauth/src/index.ts
+++ b/packages/oauth/src/index.ts
@@ -10,6 +10,7 @@ export {
   OPENAI_CODEX_CLI_CLIENT_VERSION,
   OPENAI_CODEX_OAUTH_PLATFORM_ID,
   OPENAI_CODEX_PROVIDER_ID,
+  OAuthAccessDeniedError,
   OPENAI_CODEX_REDIRECT_URI,
   OpenAICodexApiError,
   parseOpenAICodexAuthorizationInput,
@@ -18,6 +19,7 @@ export {
 } from './openai-codex-oauth';
 export type {
   ApplyOpenAICodexOAuthResult,
+  OpenAICodexCallbackResult,
   FetchOpenAICodexModelsOptions,
   OpenAICodexCallbackServer,
   OpenAICodexConfigShape,
diff --git a/packages/oauth/src/openai-codex-oauth.ts b/packages/oauth/src/openai-codex-oauth.ts
index 3b1ebe842..e2a4cc567 100644
--- a/packages/oauth/src/openai-codex-oauth.ts
+++ b/packages/oauth/src/openai-codex-oauth.ts
@@ -63,6 +63,27 @@ export interface OpenAICodexTokenBundle {
   readonly accountId: string;
 }
 
+/**
+ * `{ denied: true }` is the consent page answering `error=access_denied` — the
+ * user pressed Deny, so there is nothing to retry and no URL worth pasting.
+ * `null` is every other dead end, where manual input can still rescue the flow.
+ */
+export type OpenAICodexCallbackResult =
+  | { readonly code: string }
+  | { readonly denied: true }
+  | null;
+
+/**
+ * The user denied the authorization request on the consent page. Callers
+ * report it as a cancellation, not a failure.
+ */
+export class OAuthAccessDeniedError extends Error {
+  constructor(message = 'Authorization denied.') {
+    super(message);
+    this.name = 'OAuthAccessDeniedError';
+  }
+}
+
 export interface OpenAICodexCallbackServer {
   readonly redirectUri: string;
   /**
@@ -73,7 +94,7 @@ export interface OpenAICodexCallbackServer {
   waitForCode(opts?: {
     readonly signal?: AbortSignal | undefined;
     readonly timeoutMs?: number | undefined;
-  }): Promise<{ readonly code: string } | null>;
+  }): Promise<OpenAICodexCallbackResult>;
   cancelWait(): void;
   close(): void;
 }
@@ -246,10 +267,10 @@ export async function exchangeOpenAICodexAuthorizationCode(
 export async function startOpenAICodexCallbackServer(
   expectedState: string,
 ): Promise<OpenAICodexCallbackServer> {
-  let settleWait: ((value: { code: string } | null) => void) | undefined;
+  let settleWait: ((value: OpenAICodexCallbackResult) => void) | undefined;
   let server: Server | undefined;
 
-  const waitPromise = new Promise<{ code: string } | null>((resolve) => {
+  const waitPromise = new Promise<OpenAICodexCallbackResult>((resolve) => {
     settleWait = resolve;
   });
 
@@ -283,7 +304,7 @@ export async function startOpenAICodexCallbackServer(
           res.writeHead(400, { 'content-type': 'text/html; charset=utf-8' }).end(
             errorPage,
           );
-          settleWait?.(null);
+          settleWait?.(errorParam === 'access_denied' ? { denied: true } : null);
           settleWait = undefined;
           return;
         }
@@ -325,7 +346,7 @@ export async function startOpenAICodexCallbackServer(
           loopback: true,
           waitForCode: async (opts = {}) => {
             const timeoutMs = opts.timeoutMs ?? DEFAULT_CALLBACK_TIMEOUT_MS;
-            return new Promise<{ code: string } | null>((resolveWait, rejectWait) => {
+            return new Promise<OpenAICodexCallbackResult>((resolveWait, rejectWait) => {
               let timer: NodeJS.Timeout | undefined;
               const onAbort = (): void => {
                 // Tear down here rather than leaning on the `waitPromise.then`
@@ -407,7 +428,10 @@ export async function runOpenAICodexOAuthFlow(
       signal: options.signal,
       timeoutMs: options.timeoutMs,
     });
-    if (callbackResult?.code !== undefined) {
+    if (callbackResult !== null && 'denied' in callbackResult) {
+      throw new OAuthAccessDeniedError();
+    }
+    if (callbackResult !== null) {
       code = callbackResult.code;
     } else if (options.onManualInput !== undefined) {
       const manualInput = await options.onManualInput();
diff --git a/packages/oauth/test/openai-codex-oauth.test.ts b/packages/oauth/test/openai-codex-oauth.test.ts
index deeeabffd..082a9217a 100644
--- a/packages/oauth/test/openai-codex-oauth.test.ts
+++ b/packages/oauth/test/openai-codex-oauth.test.ts
@@ -10,6 +10,9 @@ import {
   fetchOpenAICodexModels,
   OPENAI_CODEX_AUTH_INPUT_MAX_LENGTH,
   parseOpenAICodexAuthorizationInput,
+  OAuthAccessDeniedError,
+  OPENAI_CODEX_REDIRECT_URI,
+  runOpenAICodexOAuthFlow,
   startOpenAICodexCallbackServer,
   type OpenAICodexConfigShape,
 } from '../src/openai-codex-oauth';
@@ -311,6 +314,85 @@ describe('startOpenAICodexCallbackServer', () => {
     }
   }
 
+  it('reports a denied consent page as a denial, not as a dead end', async () => {
+    const server = await startOpenAICodexCallbackServer('state-abc');
+    try {
+      if (!server.loopback) return;
+      const pending = server.waitForCode({ timeoutMs: 10_000 });
+      const url = new URL(OPENAI_CODEX_REDIRECT_URI);
+      url.searchParams.set('error', 'access_denied');
+      url.searchParams.set('state', 'state-abc');
+      await fetch(url);
+      await expect(pending).resolves.toEqual({ denied: true });
+    } finally {
+      server.close();
+    }
+  });
+
+  it('fails the whole flow with OAuthAccessDeniedError when the user denies consent', async () => {
+    // The flow owns its callback server, so drive that one rather than probing
+    // with a second server first: `close()` returns before the socket is
+    // actually released, and the probe would race the flow for the port.
+    const abort = new AbortController();
+    let authorizeUrl: string | undefined;
+    const pending = runOpenAICodexOAuthFlow({
+      timeoutMs: 10_000,
+      signal: abort.signal,
+      openBrowser: (url) => {
+        authorizeUrl = url;
+      },
+    });
+    const settled = expect(pending).rejects.toBeInstanceOf(OAuthAccessDeniedError);
+
+    await vi.waitFor(() => expect(authorizeUrl).toBeDefined());
+    const state = new URL(authorizeUrl!).searchParams.get('state');
+    expect(state).toBeTruthy();
+
+    const url = new URL(OPENAI_CODEX_REDIRECT_URI);
+    url.searchParams.set('error', 'access_denied');
+    url.searchParams.set('state', state!);
+
+    // Reaching the callback is the whole point of this case, so retry rather
+    // than skip on the first refusal; only an environment that never accepts a
+    // loopback listener gets a pass, and it says so.
+    const deadline = Date.now() + 5_000;
+    let delivered = false;
+    while (!delivered) {
+      const remaining = deadline - Date.now();
+      if (remaining <= 0) break;
+      try {
+        // Cap each attempt too: a listener that accepts but never answers
+        // would otherwise hold this past the deadline.
+        await fetch(url, { signal: AbortSignal.timeout(remaining) });
+        delivered = true;
+      } catch {
+        await new Promise((resolve) => setTimeout(resolve, 50));
+      }
+    }
+    if (!delivered) {
+      abort.abort();
+      await pending.catch(() => undefined);
+      console.warn('skipped: no loopback listener on the OAuth callback port');
+      return;
+    }
+
+    await settled;
+  });
+
+  it('reports any other callback error as a dead end the user can still paste past', async () => {
+    const server = await startOpenAICodexCallbackServer('state-abc');
+    try {
+      if (!server.loopback) return;
+      const pending = server.waitForCode({ timeoutMs: 10_000 });
+      const url = new URL(OPENAI_CODEX_REDIRECT_URI);
+      url.searchParams.set('error', 'server_error');
+      await fetch(url);
+      await expect(pending).resolves.toBeNull();
+    } finally {
+      server.close();
+    }
+  });
+
   it('tears the wait down when it is aborted before it starts', async () => {
     const server = await startOpenAICodexCallbackServer('state-abc');
     // Fake timers only count what is scheduled while they are installed, so
diff --git a/packages/protocol/src/__tests__/rest-task.test.ts b/packages/protocol/src/__tests__/rest-task.test.ts
index f0dfeed72..e05f64965 100644
--- a/packages/protocol/src/__tests__/rest-task.test.ts
+++ b/packages/protocol/src/__tests__/rest-task.test.ts
@@ -2,6 +2,7 @@ import { describe, expect, it } from 'vitest';
 
 import {
   cancelTaskResultSchema,
+  detachTaskResultSchema,
   getTaskQuerySchema,
   getTaskResponseSchema,
   listTasksQuerySchema,
@@ -62,6 +63,21 @@ describe('cancelTaskResultSchema', () => {
   });
 });
 
+describe('detachTaskResultSchema', () => {
+  it('accepts a detached result with a wire status', () => {
+    expect(detachTaskResultSchema.parse({ detached: true, status: 'running' })).toEqual({
+      detached: true,
+      status: 'running',
+    });
+  });
+  it('rejects engine-only statuses and missing flags', () => {
+    expect(detachTaskResultSchema.safeParse({ detached: false, status: 'killed' }).success).toBe(
+      false,
+    );
+    expect(detachTaskResultSchema.safeParse({ status: 'running' }).success).toBe(false);
+  });
+});
+
 describe('taskAlreadyFinishedDataSchema (40904 envelope data)', () => {
   it('requires cancelled: false literal', () => {
     expect(taskAlreadyFinishedDataSchema.parse({ cancelled: false })).toEqual({
diff --git a/packages/protocol/src/rest/task.ts b/packages/protocol/src/rest/task.ts
index 9893bd2dd..0f649c65f 100644
--- a/packages/protocol/src/rest/task.ts
+++ b/packages/protocol/src/rest/task.ts
@@ -10,6 +10,14 @@
  *     Body: empty
  *     Response data: `{ cancelled: true }`
  *     Errors: 40406 (task.not_found), 40904 (task.already_finished)
+ *
+ *   POST /v1/sessions/{session_id}/tasks/{task_id}:detach
+ *     Body: empty
+ *     Response data: `{ detached: boolean, status: TaskStatus }`
+ *     (`detached: true` only when the call converted a running foreground
+ *     task; already-background or terminal tasks answer `detached: false`
+ *     with their current status)
+ *     Errors: 40406 (task.not_found)
  */
 
 import { z } from 'zod';
@@ -40,6 +48,12 @@ export const cancelTaskResultSchema = z.object({
 });
 export type CancelTaskResult = z.infer<typeof cancelTaskResultSchema>;
 
+export const detachTaskResultSchema = z.object({
+  detached: z.boolean(),
+  status: taskStatusSchema,
+});
+export type DetachTaskResult = z.infer<typeof detachTaskResultSchema>;
+
 export const taskAlreadyFinishedDataSchema = z.object({
   cancelled: z.literal(false),
 });
diff --git a/packages/protocol/src/task.ts b/packages/protocol/src/task.ts
index 159d7a42e..593ab1382 100644
--- a/packages/protocol/src/task.ts
+++ b/packages/protocol/src/task.ts
@@ -31,6 +31,7 @@ export const taskSchema = z.object({
   /** Subagent tasks only: the child's effective thinking effort at spawn. */
   thinking_effort: z.string().optional(),
   agent_id: z.string().optional(),
+  parent_tool_call_id: z.string().optional(),
   /** Whether the task runs detached from the caller's turn (background).
    *  Optional: producers that predate the field (e.g. the agent-core v1
    *  task service) omit it — consumers apply the foreground fallback. */
diff --git a/packages/transcript/src/history/groupTurns.ts b/packages/transcript/src/history/groupTurns.ts
index 99099643b..f692c6d37 100644
--- a/packages/transcript/src/history/groupTurns.ts
+++ b/packages/transcript/src/history/groupTurns.ts
@@ -198,7 +198,11 @@ export function groupMessagesIntoSnapshot(
     if (message.role === 'user') {
       if (originKind !== undefined && HIDDEN_USER_ORIGINS.has(originKind)) {
         if (opensOwnTurn(message)) {
-          startTurn(mapOrigin(message));
+          const opening =
+            (message.origin as { name?: unknown }).name === 'subagent'
+              ? foldTurnOpeningInput(message)
+              : undefined;
+          startTurn(mapOrigin(message), opening?.text || undefined, opening?.attachmentIds);
         }
         continue;
       }
diff --git a/packages/transcript/test/layers.test.ts b/packages/transcript/test/layers.test.ts
index 890c941eb..ae02037e3 100644
--- a/packages/transcript/test/layers.test.ts
+++ b/packages/transcript/test/layers.test.ts
@@ -989,7 +989,7 @@ describe('groupMessagesIntoSnapshot (cold path)', () => {
     if (subTurn?.kind !== 'turn') throw new Error('expected turn');
     expect(subTurn.ordinal).toBe(1);
     expect(subTurn.origin.kind).toBe('other');
-    expect(subTurn.prompt).toBeUndefined();
+    expect(subTurn.prompt).toBe('scan the repo');
     expect(subTurn.steps).toHaveLength(1);
   });