fix(acp): disable LSP for MCP discovery config - #12256

E2E verification report — fix(acp): disable LSP for MCP discovery config

Verified at head 44bc36e8cd05cc1bb66b1afae42b55af58c28d95 (read 2026-09-19T14:3xZ). No Critical found. Recommendation: mergeable.

Why not a tmux arm

A tmux/TUI run cannot reach this code. createWorkspaceMcpDiscoveryConfig() only executes inside a qwen serve ACP child, so the executed arm here is a build + focused-regression + mutation A/B in a worktree at the PR head, with the multi-process claim reported as NOT MEASURED and the reason given below.

Arm 1 — compile (r179's mandatory gate: a red build voids every test argument)

Worktree at head, node_modules hardlinked (17 package-level dirs, zero symlinks, @qwen-code/qwen-code-core resolving inside the worktree):

• npm run generate → GEN_rc=0
• npm run build → BUILD_rc=0, and grep -c 'error TS' over the build log = 0

Blob fidelity was certified structurally, not by grep — sha1('blob ' + len + '\0' + bytes) against the API's own files[].sha:

⚠️ Worth noting for anyone re-verifying: acpAgent.test.ts is 1,189,457 bytes, i.e. over the ~1 MB inline cap on GET contents/…, which returns HTTP 200 with encoding: "none" and an empty content field. Reading it that way decodes to one empty line and fabricates a build break. The git fetch upstream pull/12256/head route used here is byte-faithful.

Arm 2 — focused regression at head, and its non-vacuity oracle

npx vitest run src/acp-integration/acpAgent.test.ts \
  -t "manages the workspace MCP discovery config used by status" --reporter=dot

Test Files  1 passed (1)
     Tests  1 passed | 766 skipped (767)

Exactly one test was selected, so the green cannot be an empty selection.

Arm 3 — mutation witness (delete the single production line)

Removing experimentalLsp: false, from acpAgent.ts (assert count == 1 before removing; Python len 583,019 → 582,983 characters, −36, the line being pure ASCII):

Test Files  1 failed (1)
     Tests  1 failed | 766 skipped (767)

AssertionError: expected { experimentalLsp: undefined, …(4) } to deeply equal ObjectContaining{…}
-   "experimentalLsp": false,
+   "experimentalLsp": undefined,
   30771|       expect.objectContaining({ experimentalLsp: false }),

Restoration witnesses: marker census back to 1, and git diff --exit-code <head-ref> -- packages/cli/src/acp-integration/ → rc=0.

Mechanism, precedence and completeness — read in the head blob, not in the patch

• config/config.ts:1740 — lspEnabled = !provisionalWorkspace && !bareMode && argv.experimentalLsp === true, a strict === true, so false disables.
• config/config.ts:2687–2716 — inside if (lspEnabled), loadCliConfig constructs NativeLspService and awaits discoverAndPrepare() and then start() before returning. So the cached discovery config (if (this.workspaceMcpDiscoveryConfig) return;) really did own a second started stack for the ACP process lifetime — the mechanism in bug(lsp): MCP workspace-discovery config starts a second, unused LSP server set per ACP process (inherits --experimental-lsp) #12216 is confirmed, not assumed.
• Precedence: the override sits after ...this.argv (acpAgent.ts:4159), so object-literal semantics guarantee it wins over a truthy --experimental-lsp on the child argv. (Had it been placed before the spread it would have been a silent no-op — that is the one defect shape available here, and it is not present.)
• Completeness: workspace-mcp-discovery occurs at exactly two places repo-wide in this file (:4151 the sessionIdContext.run label, :4158 the argv override), i.e. there is only one discovery-config construction site, so no second site is left inheriting the flag.

Deref-safety enumeration (does anything need the client the discovery config no longer has?)

getLspClient() has exactly 4 non-test call sites, and every one is guarded:

⇒ A config with no LSP client cannot throw on this path.

Base arm (r194 discipline — is the hole the diff's?): acpAgent.ts:14900 already ships ...(provisionalWorkspace ? { experimentalLsp: false } : {}) for ACP session configs. "An ACP-loaded config with LSP disabled" is therefore a pre-existing, already-exercised state, and this PR extends an existing pattern rather than introducing a new one.

NOT MEASURED, with the precondition that makes it unmeasurable here

• The multi-process claim (doubled language-server processes and memory). This host has no LSP server binary at all — typescript-language-server, vscode-langservers-extracted, pylsp, gopls, rust-analyzer, clangd all ABSENT — so discoverAndPrepare() would find zero servers on both arms and a process-count A/B would be vacuous by construction, painting both arms green. Not attempted rather than reported as a null result.
• A live qwen serve ACP child process tree. Not driven.

Corroboration of existing feedback — no new finding filed

The bot's non-blocking note that the regression test pins undefined → false rather than the true → false override the description advertises (the mockArgv.experimentalLsp = true line sits in a different it block, and the beforeEach this PR touched resets it first) is independently confirmed by Arm 3's own output: the received value on the mutated arm is literally experimentalLsp: undefined. That agrees with the bot's classification — the test is load-bearing, not decorative, and the gap is a Suggestion. Nothing is added to it here, and it is not treated as blocking.

Gate state at post time (re-read inside the pre-write guard)

state=open merged=False draft=False; qwen-code-ci-bot APPROVED at head 14:10:23Z is the newest and only verdict row; CHANGES_REQUESTED at any commit = 0; [Critical] at head by original_commit_id = 0; check-runs total_count=54 fetched=54 MATCH=True, raw {success 24, skipped 29, in_progress 1} over 45 distinct names, product lanes 40, non-green 0, in-flight 0, and the single unsettled run identified by name as the bot's own review-pr lane (in_progress since 13:40:01Z) rather than assumed to be a bot lane; check-suites total_count=11 fetched=11 MATCH=True at suite level {success 7, in_progress 1, skipped 3} (latest_check_run is null in the list route, so the suite-level pair is the sound field); ms=blocked, autoMergeRequest=None.

As of the state read immediately before posting, this comment carries no approval.