Add dashboard scoped-model settings editor

README.md

@@ -129,7 +129,9 @@ The dashboard is the visual face of dreb: every agent session on the host, live
 
 **Fleet overview.** Home base is every session across every project: live sessions with status chips (running / needs-attention / idle / error), activity lines, running subagents, task progress, context usage, and model — plus past sessions grouped by project, resumable with one tap. Terminal provider/API failures show their reason on the fleet card; transient failures clear that terminal card state when automatic retry begins. Live cards update through compact SSE snapshots instead of re-fetching the full cross-project inventory on every turn, keeping weak mobile links responsive. When a session needs input, the browser tab badges and (opt-in) sends a service-worker notification (installable PWA, works on Android and iOS). For opt-in local mobile transport profiling, see the [dashboard docs](packages/coding-agent/docs/dashboard.md#mobile-transport-profiling).
 
-**Full-parity session view.** Not a reduced chat client: streaming markdown, thinking blocks, bespoke tool cards (read/write/edit/bash and markdown-rendering tools), sanitized inline PNG/JPEG/GIF/WebP images returned by any tool, task panels, queued-message chips, image attach/paste with sent-image transcript previews, built-in slash-command autocomplete and execution (including model/settings, session tree, fork, compact, import/export, dream, resume/reload, and new/quit), model/thinking switchers, fork-from-message, HTML export. Built-ins are intercepted generically and rejected fail-closed at the RPC prompt boundary, so unsupported or future commands show guidance instead of leaking into model input. Provider/API failures render inline on the failed assistant attempt with any partial output preserved, including after refresh or recovery; transient failures then switch the session status to retrying without erasing that history. Tool-result images remain visible to the human even when the active model is text-only. While the agent works you can **steer** (inject into the running turn), **queue follow-ups**, or **stop** — the same queue semantics as the TUI.
+**Full-parity session view.** Not a reduced chat client: streaming markdown, thinking blocks, bespoke tool cards (read/write/edit/bash and markdown-rendering tools), sanitized inline PNG/JPEG/GIF/WebP images returned by any tool, task panels, queued-message chips, image attach/paste with sent-image transcript previews, built-in slash-command autocomplete and execution (including model/settings, scoped-models, session tree, fork, compact, import/export, dream, resume/reload, and new/quit), model/thinking switchers, fork-from-message, HTML export. `/scoped-models` opens the Settings editor for the session's current project context. Built-ins are intercepted generically and rejected fail-closed at the RPC prompt boundary, so unsupported or future commands show guidance instead of leaking into model input. Provider/API failures render inline on the failed assistant attempt with any partial output preserved, including after refresh or recovery; transient failures then switch the session status to retrying without erasing that history. Tool-result images remain visible to the human even when the active model is text-only. While the agent works you can **steer** (inject into the running turn), **queue follow-ups**, or **stop** — the same queue semantics as the TUI.
+
+**Scoped models.** Dashboard Settings includes an editor for the persistent model-cycling scope. It searches provider-grouped available models; offers model, provider, and all-model toggles; shows accessible up/down controls for the ordered partial scope; and has save/reset actions that work on mobile. An absent `enabledModels` value means implicit all models in registry order, including future registry additions, and that all-model view cannot be reordered. A saved partial scope is a non-empty ordered list of exact canonical `provider/model` references; editing a legacy glob, fuzzy, or thinking-suffix scope normalizes it to that form. The selected project context reads effective global plus project settings, but dashboard writes remain global and warn when a project value shadows the result. Changes seed new sessions only and never mutate running sessions. See [Settings](packages/coding-agent/docs/settings.md#model-cycling) and the [RPC settings contract](packages/coding-agent/docs/rpc.md#settings).
 
 **Low-data transcript images.** Tool results and images uploaded with user turns remain visible through content-addressed browser references rather than full base64 in events and transcript JSON. The default browser-local mode lazily requests a preview bounded to 1024 × 1024 and 256 KiB; clicking enlarges that same preview without downloading the original. Settings also offers request-free placeholders and informed-opt-in automatic originals. Explicit originals disclose their size and confirm above 1 MiB. Authenticated same-origin routes accept only signature-matching PNG/JPEG/GIF/WebP, send `nosniff`, and recover evicted entries from authoritative session data; static GIF previews preserve animated originals behind the original route. Image bytes therefore cannot cause an SSE oversized-event resync. Full-resolution HTML exports remain self-contained and unchanged.
 

package.json

@@ -39,7 +39,7 @@
 		"node": "22.x"
 	},
 	"packageManager": "npm@11.5.1",
-	"version": "2.53.1",
+	"version": "2.54.0",
 	"dependencies": {
 		"@dreb/coding-agent": "*",
 		"@mariozechner/jiti": "^2.6.5",

packages/agent/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@dreb/agent-core",
-	"version": "2.53.1",
+	"version": "2.54.0",
 	"description": "General-purpose agent with transport abstraction, state management, and attachment support",
 	"type": "module",
 	"main": "./dist/index.js",

packages/ai/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@dreb/ai",
-	"version": "2.53.1",
+	"version": "2.54.0",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"type": "module",
 	"main": "./dist/index.js",

packages/ai/test/abort.test.ts

@@ -172,7 +172,7 @@ describe("AI Providers Abort Tests", () => {
 	describe.skipIf(
 		process.env.DREB_SKIP_LIVE_API === "1" || (!process.env.ANTHROPIC_API_KEY && !process.env.ANTHROPIC_OAUTH_TOKEN),
 	)("Anthropic Provider Abort", () => {
-		const llm = getModel("anthropic", "claude-opus-4-1-20250805");
+		const llm = getModel("anthropic", "claude-opus-4-6");
 
 		it("should abort mid-stream", { retry: 3 }, async () => {
 			await testAbortSignal(llm, { thinkingEnabled: true, thinkingBudgetTokens: 2048 });

packages/ai/test/supports-xhigh.test.ts

@@ -37,9 +37,8 @@ describe("supportsXhigh", () => {
 	});
 
 	it("returns false for Opus 4.1 (below threshold)", () => {
-		const model = getModel("anthropic", "claude-opus-4-1");
-		expect(model).toBeDefined();
-		expect(supportsXhigh(model!)).toBe(false);
+		const base = getModel("anthropic", "claude-opus-4-6");
+		expect(supportsXhigh({ ...base, id: "claude-opus-4-1" })).toBe(false);
 	});
 
 	it("returns false for Opus 4.5 (below threshold)", () => {

packages/coding-agent/README.md

@@ -168,7 +168,7 @@ Type `/` in the editor to trigger commands. [Extensions](#extensions) can regist
 |---------|-------------|
 | `/login`, `/logout` | OAuth authentication |
 | `/model` | Switch models |
-| `/scoped-models` | Enable/disable models for cycling |
+| `/scoped-models` | Set the ordered model-cycling scope; in dashboard sessions, opens the scoped-models Settings editor for the current project context |
 | `/settings` | Thinking level, thinking summaries (adaptive Claude models), theme, message delivery, transport |
 | `/resume` | Pick from previous sessions |
 | `/new` | Start a new session |
@@ -278,7 +278,7 @@ Use `/settings` to modify common options, or edit JSON files directly:
 | `~/.dreb/agent/settings.json` | Global (all projects) |
 | `.dreb/settings.json` | Project (overrides global) |
 
-See [docs/settings.md](docs/settings.md) for all options. Dashboard Settings re-entry reloads durable global and project settings after flushing pending writes, so it sees external file edits; unreadable, invalid, or failed writes are surfaced as errors rather than showing stale values. See the [RPC settings contract](docs/rpc.md#get_settings).
+See [docs/settings.md](docs/settings.md) for all options. Dashboard Settings re-entry reloads durable global and project settings after flushing pending writes, so it sees external file edits; unreadable, invalid, or failed writes are surfaced as errors rather than showing stale values. Its scoped-models editor provides provider-grouped search, model/provider/all toggles, accessible ordered partial-scope controls, and explicit save/reset on desktop and mobile. Absent `enabledModels` means future-inclusive all models in registry order; saved partial scopes are non-empty ordered canonical references, and editing a legacy pattern scope normalizes it. A selected project context shows effective merged settings, while writes remain global and warn when project settings shadow them. Changes seed new sessions only; `/scoped-models` opens the editor for the current dashboard session cwd. See the [RPC settings contract](docs/rpc.md#get_settings).
 
 ---