fix(cli-registry): merge-time fixes for the run-menu consolidation (#458)

- test/opencode-resize.test.ts: retarget the launcher guard at the real code (this.selectSession(firstSessionId), any this.activeSessionId assignment) with an anti-vacuity check; the old strings existed nowhere, so it could never fail
- session-ui.js: restore as comments the two invariants the merged bodies lost (deepseek leaves statusReporting unset, i.e. ON; no effort field for external CLIs, it is Claude-specific)
- docs/cli-registry.md: move the frontend-guard paragraph below the two backend-guard paragraphs so they keep their antecedent, and note the widened comparison shape
- test/frontend-cli-no-id-branching.test.ts: the comparison shape accepts any left-hand identifier (const m = this._runMode; m === 'codex' was invisible), normalized to `mode`; the two `m !== 'shell'` display filters are allowlisted and the remaining blind spots documented
- test/run-mode-dispatch.test.ts: table-driven pin of run() dispatch (claude to runClaude, each RUN_MODE_LAUNCH id to _runCliMode(id), shell to runShell, unknown to runClaude, lock held and released)
- CLAUDE.md: name the second CI-gated guard next to the backend one
- server.ts: every </head> injection passes a replacer function; a clis.json label containing $' re-injected the rest of the document past escapeScriptJson (two render tests pin it, proven failing on the string form)
- _isAltCliMode(): no reference anywhere in the tree, nothing to fix

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
(cherry picked from commit 1ea363ff808a62861559bc141e724b163cc1c56e)

test/frontend-cli-no-id-branching.test.ts

@@ -82,6 +82,17 @@ const ALLOWED_BRANCHES: Record<string, { count: number; reason: string }> = {
   "session-ui.js::mode === 'deepseek'": { count: 2, reason: 'button-label ternary + runMode setter validity check' },
   "session-ui.js::mode === 'omp'": { count: 2, reason: 'button-label ternary + runMode setter validity check' },
 
+  // The docker adopt-preflight status line and the docker link/adopt toast
+  // both list the agent CLIs probed INSIDE the container and leave `shell`
+  // out of that human-readable "found ..." summary (it is always present and
+  // is not an agent CLI). Written as `(m) => m !== 'shell'`, the naming the
+  // original named-variable pattern could not see; the widened pattern
+  // normalizes the `m` to `mode` (see BRANCH_PATTERN below).
+  "session-ui.js::mode !== 'shell'": {
+    count: 2,
+    reason: 'display filter: the "CLIs found inside the container" summaries omit shell, which is not an agent CLI',
+  },
+
   // mobile-overview.js: shell is exempt from the isCliAvailable() gate the
   // same way the toolbar's #runModeMenu exempts it (shell needs no CLI).
   "mobile-overview.js::mode !== 'shell'": {
@@ -94,11 +105,35 @@ const ALLOWED_BRANCHES: Record<string, { count: number; reason: string }> = {
 const IDS = STOCK_CLIS.map((e) => e.id as string);
 const ID_ALT = IDS.join('|');
 
-/** Same four shapes as the backend guard — see its own comment for why all four matter. */
+/**
+ * The backend guard's four shapes (see its own comment for why all four
+ * matter), with ONE deliberate widening on the first.
+ *
+ * The backend pattern accepts a comparison only when its left-hand side is
+ * literally named `mode`, `id` or `agentType`, so both
+ * `const m = this._runMode; if (m === 'codex')` and
+ * `if (this._runMode !== 'gemini')` slip past it, and `session-ui.js` already
+ * uses exactly that naming (`(m) => m !== 'shell'`, twice). The review of
+ * PR #458 surfaced that blind spot, so here the left-hand side is ANY
+ * identifier (`[\w$]+`, the leaf of a member chain), normalized to `mode` in
+ * the allowlist key by `scan()` so a local rename never churns the entries.
+ * Measured over both scanned files before widening: every extra hit was a
+ * genuine mode comparison (the two `m !== 'shell'` filters, allowlisted
+ * above), so the widening added no false positive; a future one gets an
+ * allowlist entry with its reason like any other. The backend guard keeps
+ * its narrower form and is deliberately not changed here.
+ *
+ * Still unseen, and worth knowing: a Yoda comparison (`'codex' === mode`),
+ * and an id list held in a variable (`EXTERNAL.includes(mode)`), since the
+ * third shape needs the literal list inline.
+ */
 const BRANCH_PATTERN = new RegExp(
   [
-    `\\b(?:mode|id|agentType)\\s*[!=]==\\s*'(?:${ID_ALT})'`,
+    // <identifier> === 'codex'  /  <identifier> !== 'codex' (any left-hand identifier, see above)
+    `\\b[\\w$]+\\s*[!=]==\\s*'(?:${ID_ALT})'`,
+    // case 'codex':
     `\\bcase\\s+'(?:${ID_ALT})'\\s*:`,
+    // ['codex', 'gemini'].includes(mode) — the id list IS the branch, wherever `mode` sits
     `'(?:${ID_ALT})'\\s*(?:,\\s*'(?:${ID_ALT})'\\s*)*\\]\\s*\\.includes\\(`,
   ].join('|'),
   'g'
@@ -126,7 +161,10 @@ function scan(): Finding[] {
     lines.forEach((line, i) => {
       BRANCH_PATTERN.lastIndex = 0; // shared /g regex — see utils/regex-patterns.ts
       for (const match of line.matchAll(BRANCH_PATTERN)) {
-        const expression = match[0].replace(/\s+/g, ' ').replace(/^(?:id|agentType)/, 'mode');
+        // Normalize the comparison shape's left-hand identifier (whatever the
+        // local is called: `id`, `agentType`, `m`, `_runMode`) to `mode`; the
+        // lookahead leaves the `case`/`.includes(` shapes untouched.
+        const expression = match[0].replace(/\s+/g, ' ').replace(/^[\w$]+(?=\s*[!=]==)/, 'mode');
         findings.push({ file, expression, line: i + 1, key: `${file}::${expression}` });
       }
     });
@@ -162,6 +200,9 @@ describe('no NEW CLI-id branching in session-ui.js / mobile-overview.js (PR B2)'
       "if (mode !== 'shell' && mode !== 'deepseek') { doSomething(); }",
       "switch (mode) { case 'gemini': return 1; }",
       "if (['codex', 'gemini'].includes(mode)) { doSomething(); }",
+      // The two forms the named-variable pattern was blind to (see BRANCH_PATTERN).
+      "const m = this._runMode; if (m === 'codex') { doSomething(); }",
+      "if (this._runMode !== 'gemini') { doSomething(); }",
     ];
     for (const sample of samples) {
       BRANCH_PATTERN.lastIndex = 0;

test/opencode-resize.test.ts

@@ -68,23 +68,33 @@ describe('OpenCode session initial resize', () => {
     ({ context, page } = await freshPage());
     await navigateAndWait(page);
 
-    const hasPreAssignment = await page.evaluate(() => {
+    const { selectIdx, assignIdx } = await page.evaluate(() => {
       const app = (window as unknown as { app: { _runCliMode: { toString: () => string } } }).app;
       const source = app._runCliMode.toString();
 
-      // Check: the source should NOT have activeSessionId = ... before selectSession
-      // Find positions of both patterns
-      const assignIdx = source.indexOf('this.activeSessionId = data.sessionId');
-      const selectIdx = source.indexOf('this.selectSession(data.sessionId)');
-
-      // If assign doesn't exist at all, that's the correct fix
-      if (assignIdx === -1) return false;
-
-      // If assign comes before select, that's the bug
-      return assignIdx < selectIdx;
+      // The launcher hands the FIRST created session to selectSession
+      // (`_launchQuickStartInstances()` returns `firstSessionId`). An earlier
+      // version of this check looked for `this.selectSession(data.sessionId)`,
+      // a string that exists nowhere in session-ui.js, so both lookups came
+      // back -1 and the assertion could never fail. Hence the anti-vacuity
+      // check below: the select call itself must be found.
+      const selectIdx = source.indexOf('this.selectSession(firstSessionId)');
+      // ANY assignment to activeSessionId (whatever the right-hand side is
+      // called), not `==`/`===` comparisons and not the comment that mentions
+      // pre-setting it without a `this.` prefix.
+      const assign = /this\.activeSessionId\s*=(?!=)/.exec(source);
+      return { selectIdx, assignIdx: assign ? assign.index : -1 };
     });
 
-    expect(hasPreAssignment).toBe(false);
+    // Anti-vacuity: if the select call is renamed again, fail here rather
+    // than pass on two -1s.
+    expect(selectIdx).toBeGreaterThan(-1);
+    // Correct: no assignment at all. Bug: an assignment that lands BEFORE
+    // selectSession runs, which makes selectSession early-return.
+    expect(
+      assignIdx === -1 || assignIdx > selectIdx,
+      `activeSessionId is assigned at ${assignIdx}, before selectSession at ${selectIdx}`
+    ).toBe(true);
   });
 
   it('sends resize to server after creating a session via quick-start', async () => {

test/render-index-html.test.ts

@@ -23,6 +23,8 @@ import { isDeepSeekAvailable, isDeepSeekRunnable } from '../src/utils/deepseek-c
 import { isOmpAvailable } from '../src/utils/omp-cli-resolver.js';
 import { isCloudflaredAvailable } from '../src/utils/cloudflared-resolver.js';
 import { isGitAvailable } from '../src/git-clone.js';
+import { enabledClis } from '../src/config/cli-registry/registry.js';
+import { STOCK_CLIS } from '../src/config/cli-registry/stock.js';
 
 // renderIndexHtml probes the real PATH for every CLI, which would make the
 // assertions below depend on whatever happens to be installed on the machine
@@ -79,6 +81,14 @@ vi.mock('../src/utils/cloudflared-resolver.js', () => ({
 vi.mock('../src/git-clone.js', () => ({
   isGitAvailable: vi.fn(() => false),
 }));
+// The custom-model list carries `label`, a string a user's own clis.json can set.
+// Wrap enabledClis so ONE test below can hand renderIndexHtml a label with `$'`
+// in it while every other test still reads the real stock registry through the
+// real implementation.
+vi.mock('../src/config/cli-registry/registry.js', async (importOriginal) => {
+  const actual = await importOriginal<typeof import('../src/config/cli-registry/registry.js')>();
+  return { ...actual, enabledClis: vi.fn(actual.enabledClis) };
+});
 
 const TEMPLATE = [
   '<head>',
@@ -222,6 +232,35 @@ describe('WebServer.renderIndexHtml', () => {
     expect(eval(escaped)[0].label).toBe('</script><script>alert(1)</script>');
   });
 
+  it("inserts a label containing $' verbatim instead of splicing the document into the script", async () => {
+    // `String.replace` with a STRING replacement interprets `$'` as "the text
+    // after the match", so a clis.json label carrying it used to re-inject the
+    // rest of the document (the whole <body>) into the inline script, past
+    // escapeScriptJson, which only neutralizes `<`. Every `</head>` injection
+    // passes a replacer FUNCTION instead, whose return value is inserted
+    // verbatim. The other `$` forms ride along so a partial escape cannot pass.
+    const claude = STOCK_CLIS.find((e) => e.id === 'claude')!;
+    const label = "Claude $' $& $` $1 $$";
+    const real = vi.mocked(enabledClis).getMockImplementation()!;
+    vi.mocked(enabledClis).mockImplementation(() => [{ ...claude, label }]);
+    try {
+      const { server } = makeServer({});
+      const html = await render(server);
+      expect(html.match(/<body>/g)).toHaveLength(1);
+      const clis = JSON.parse(html.match(/window\.__codemanCustomModelClis=(\[.*?\]);/)![1]);
+      expect(clis).toEqual([{ id: 'claude', label }]);
+    } finally {
+      vi.mocked(enabledClis).mockImplementation(real);
+    }
+  });
+
+  it("inserts a solo id containing $' verbatim, under the same replacer rule", async () => {
+    const { server } = makeServer({});
+    const html = await render(server, "sess$'x");
+    expect(html.match(/<body>/g)).toHaveLength(1);
+    expect(html).toContain(`window.__CODEMAN_SOLO__="sess$'x"`);
+  });
+
   it('still emits the object when nothing at all is installed', async () => {
     // The all-false case is the one that matters most and the easiest to get
     // wrong by only injecting when something resolves.

test/run-mode-dispatch.test.ts

@@ -0,0 +1,147 @@
+/**
+ * @fileoverview Table-driven pin for `run()`'s dispatch in session-ui.js
+ * (PR #458). Before the run-menu consolidation, `run()` was an eight-arm
+ * `if (mode === 'codex') return this.runCodex(); ...` chain and each arm was
+ * pinned only by the name it called; after it, every non-claude, non-shell
+ * mode reaches ONE shared launcher, `_runCliMode(mode)`, gated on
+ * `EXTERNAL_CLI_MODES` (the key set of `RUN_MODE_LAUNCH`). Nothing pinned that
+ * gate: a mode dropped from the table would fall through to `runClaude()` and
+ * start a Claude session under a Codex label with no error, while an unknown
+ * mode reaching `_runCliMode()` would throw on `entry.label` of an undefined
+ * entry.
+ *
+ * The external ids are read off `RUN_MODE_LAUNCH` itself (same JSDOM
+ * extraction as test/run-mode-launch-table-drift.test.ts, which separately
+ * pins that key set against stock.ts), so a ninth CLI is covered the day it
+ * lands in the table.
+ *
+ * Port: none.
+ */
+import { readFileSync } from 'node:fs';
+import { JSDOM } from 'jsdom';
+import { describe, expect, it, vi } from 'vitest';
+
+const SESSION_UI_JS = readFileSync(new URL('../src/web/public/session-ui.js', import.meta.url), 'utf-8');
+
+interface HarnessApp {
+  _runMode?: string;
+  _runInFlight?: boolean;
+  _runMinLockMs?: number;
+  run: () => Promise<unknown>;
+  runClaude: ReturnType<typeof vi.fn>;
+  runShell: ReturnType<typeof vi.fn>;
+  _runCliMode: ReturnType<typeof vi.fn>;
+}
+
+interface Harness {
+  app: HarnessApp;
+  runBtn: HTMLButtonElement;
+  externalIds: string[];
+}
+
+function loadHarness(): Harness {
+  const dom = new JSDOM('<!doctype html><body><button id="runBtn"></button></body>', {
+    url: 'http://localhost/',
+    runScripts: 'dangerously',
+  });
+  const win = dom.window as unknown as {
+    eval: (s: string) => void;
+    document: Document;
+    CodemanApp: new () => HarnessApp;
+    __TEST_RUN_MODE_LAUNCH: Record<string, unknown>;
+  };
+  win.eval('window.CodemanApp = function CodemanApp() {};');
+  // The assignment rides in the SAME evaluated string as the module:
+  // RUN_MODE_LAUNCH is a bare top-level `const`, visible only to this eval call
+  // (see test/run-mode-launch-table-drift.test.ts for the measurement).
+  win.eval(`${SESSION_UI_JS}\nwindow.__TEST_RUN_MODE_LAUNCH = RUN_MODE_LAUNCH;`);
+  const app = new win.CodemanApp();
+  app._runMinLockMs = 0; // run() otherwise holds its lock for >= 500ms per call
+  app.runClaude = vi.fn(async () => 'claude');
+  app.runShell = vi.fn(async () => 'shell');
+  app._runCliMode = vi.fn(async (mode: string) => `cli:${mode}`);
+  return {
+    app,
+    runBtn: win.document.getElementById('runBtn') as HTMLButtonElement,
+    externalIds: Object.keys(win.__TEST_RUN_MODE_LAUNCH),
+  };
+}
+
+describe('run() dispatch (session-ui.js)', () => {
+  const { externalIds } = loadHarness();
+
+  it('reads at least the eight external CLIs off RUN_MODE_LAUNCH (anti-vacuity)', () => {
+    expect(externalIds.length).toBeGreaterThanOrEqual(8);
+    expect(externalIds).not.toContain('claude');
+    expect(externalIds).not.toContain('shell');
+  });
+
+  it("'claude' reaches runClaude() and nothing else", async () => {
+    const { app } = loadHarness();
+    app._runMode = 'claude';
+    await expect(app.run()).resolves.toBe('claude');
+    expect(app.runClaude).toHaveBeenCalledTimes(1);
+    expect(app._runCliMode).not.toHaveBeenCalled();
+    expect(app.runShell).not.toHaveBeenCalled();
+  });
+
+  it("'shell' reaches runShell(), which needs no CLI probe at all", async () => {
+    const { app } = loadHarness();
+    app._runMode = 'shell';
+    await expect(app.run()).resolves.toBe('shell');
+    expect(app.runShell).toHaveBeenCalledTimes(1);
+    expect(app.runClaude).not.toHaveBeenCalled();
+    expect(app._runCliMode).not.toHaveBeenCalled();
+  });
+
+  for (const id of externalIds) {
+    it(`'${id}' reaches _runCliMode('${id}') and never runClaude()`, async () => {
+      const { app } = loadHarness();
+      app._runMode = id;
+      await expect(app.run()).resolves.toBe(`cli:${id}`);
+      expect(app._runCliMode).toHaveBeenCalledTimes(1);
+      expect(app._runCliMode).toHaveBeenCalledWith(id);
+      expect(app.runClaude).not.toHaveBeenCalled();
+      expect(app.runShell).not.toHaveBeenCalled();
+    });
+  }
+
+  it('an unknown mode lands on runClaude(), never on the shared launcher', async () => {
+    // `_runCliMode(mode)` reads `RUN_MODE_LAUNCH[mode].label` unguarded, so an
+    // unknown id reaching it would throw rather than launch anything.
+    for (const mode of ['nope', 'CLAUDE', 'code x']) {
+      const { app } = loadHarness();
+      app._runMode = mode;
+      await expect(app.run()).resolves.toBe('claude');
+      expect(app.runClaude).toHaveBeenCalledTimes(1);
+      expect(app._runCliMode).not.toHaveBeenCalled();
+      expect(app.runShell).not.toHaveBeenCalled();
+    }
+  });
+
+  it('an unset or empty _runMode defaults to claude', async () => {
+    for (const mode of [undefined, '']) {
+      const { app } = loadHarness();
+      app._runMode = mode;
+      await expect(app.run()).resolves.toBe('claude');
+      expect(app.runClaude).toHaveBeenCalledTimes(1);
+      expect(app._runCliMode).not.toHaveBeenCalled();
+    }
+  });
+
+  it('holds the launch lock for the whole launch and releases it afterwards', async () => {
+    const { app, runBtn } = loadHarness();
+    app._runMode = 'codex';
+    const pending = app.run();
+    expect(app._runInFlight).toBe(true);
+    expect(runBtn.disabled).toBe(true);
+    expect(runBtn.getAttribute('aria-busy')).toBe('true');
+    // A second click while the first launch is in flight is a no-op.
+    await expect(app.run()).resolves.toBeUndefined();
+    await pending;
+    expect(app._runCliMode).toHaveBeenCalledTimes(1);
+    expect(app._runInFlight).toBe(false);
+    expect(runBtn.disabled).toBe(false);
+    expect(runBtn.hasAttribute('aria-busy')).toBe(false);
+  });
+});