🤖 fix: make no provider model calls under XUM_MOCK_AI

[ThomasK33]

Summary

With XUM_MOCK_AI=1, model creation now refuses before it builds a provider model, so background features (workspace title, sidebar status, branch summary, memory consolidation and harvest, continuous compaction summary, refine, workflow evaluate, auto model routing) make no provider request. Chat turns still play through the mock stream player, which creates no models.

Fixes #5604

Background

Mock mode replaced only the chat turn player. Every other model caller still created a real model from providers.jsonc, so a mock-mode server sent claude-haiku requests from the status timer and title generation. The remote UAT for #5602 saw these as repeated AI_RetryError logs.

Implementation

Every inference caller in src/node gets its model from one of three functions. A read-only inventory of all generateText/streamText/evaluate callers found no other path. The perf replay mode (XUM_REPLAY_TAPES) already refuses at the same three points, and this change adds the mock check next to it:

1. ProviderModelFactory.createModelEffect: chat, title, status, summary, memory, compaction summary, refine, advisor and intuition models. It returns { type: "unknown", raw: "Model calls are disabled in mock AI mode (XUM_MOCK_AI=1)" }.
2. ProviderModelFactory.createEvaluationModelEffect: workflow evaluate() and intuition evaluation.
3. evaluationModelFactory.createEvaluationModel: auto model routing, which does not use ProviderModelFactory.

isMockAiMode() (new src/node/services/mock/mockAiMode.ts) reads XUM_MOCK_AI (or legacy MUX_MOCK_AI). AIService now uses it too, so the flag has one definition. mockAiStreamPlayer.ts is unchanged.

Decision (conservative option): refuse instead of routing background generators through a mock model. A mock model would make memory consolidation write mock output into memory files. Callers already handle a refused creation, as they do for a missing API key: titles fall back, and status generation retries on its normal schedule with a debug log.

Scope note: the server makes some network calls that are not inference, and this change leaves them alone. Model discovery runs on user request. Coder and Copilot catalog refreshes run after OAuth login. Voice transcription runs when the user records.

Validation

src/node/services/mockAiModelGate.test.ts points the anthropic and openai providers at a loopback HTTP server that records every request.

1. The control case (no mock env) shows the title generator reaching the loopback server, and every model entry point succeeding.
2. The mock case sets XUM_MOCK_AI=1. It runs the title and status generators end to end and calls every model entry point (including the auto-routing factory). It then asserts the loopback server received no request.

Pre-fix result of this test on origin/main (the status generator reached the provider):

129 |     // Background generators, end to end.
130 |     const title = await generateWorkspaceIdentity("Fix the login bug", [{ model: MODEL }], service);
131 |     expect(title.success).toBe(false);
132 |     const status = await generateWorkspaceStatus("user: fix the login bug", [MODEL], service);
133 |     expect(status.success).toBe(false);
134 |     if (!status.success) expect(status.error.reachedProvider).toBe(false);
                                                                    ^
error: expect(received).toBe(expected)
Expected: false
Received: true
(fail) XUM_MOCK_AI model gate > with XUM_MOCK_AI=1 no background model caller reaches the provider [18.00ms]
 1 pass
 1 fail

Risks

Low, and limited to mock mode. E2E runs set XUM_MOCK_AI=1 but configure no real provider, so their title and status generation already failed at model creation. In-process tests that call enableMockMode() without the env var are unaffected.

---

Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: high

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-04T19:01:30.097538Z	afae467	PR opened
🔒 Security Review	✅ Completed 2026-10-04T19:03:15.474285Z	afae467	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[ThomasK33]

Readiness record:

• Final commit: afae4677426a2f5fdd48f42ec583d576580fb8ad
• CI: Required passed on this commit. Smoke / Docker was skipped by its path filter.
• Reviews: 3 of 6 assessments used (automatic normal review, automatic security review, final independent check). Both automatic reviews had no findings.
• Final independent check: READY.
• Deferred: voice transcription gate and the enableMockMode() / env-gate mismatch, tracked in 🤖 fix: close the remaining XUM_MOCK_AI provider-call gaps (voice, enableMockMode) #5632.
• Decision: ready. Merging with --match-head-commit.

---

Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: high