What

Brings the Ollama drift surface into live CI coverage. Today src/__tests__/drift/ollama.drift.ts has a live leg gated on OLLAMA_HOST, but no Ollama daemon exists in CI and there is no static fallback, so the leg is silently skipped — zero CI signal. Ollama needs no API key; the only "credential" is a running local daemon, so we can provision one in CI for free.

Changes

• test-drift.yml (scheduled/dispatch drift job) and fix-drift.yml (fix job): added an additive Provision Ollama daemon step that installs Ollama, starts ollama serve in the background, waits for readiness, and pulls the tiny qwen2:0.5b (~350 MB, smallest model supporting both /api/chat and /api/generate). Exported OLLAMA_HOST=127.0.0.1:11434 + OLLAMA_MODEL=qwen2:0.5b into the collector/verify steps so the leg runs against the real daemon. fix-drift.yml's post-fix verify + authoritative re-collect steps are also un-gated so the auto-fix path can actually verify an Ollama fix (daemon is already running in that job).
• ollama.drift.ts: model is now OLLAMA_MODEL ?? "llama3.2" (lets CI use the tiny model while a local dev run defaults unchanged); OLLAMA_HOST is normalized to accept both a full base URL and Ollama's native bare host:port form. No assertion logic changed.

⚠️ CI-time cost (for merge consideration)

The daemon install + model pull adds ~2-4 min to every drift run (daily scheduled test-drift + every fix-drift invocation). This is the tradeoff for real Ollama coverage. Weigh at merge — if too costly, options are caching the model or gating the leg to a less frequent cadence.

Red-green proof (LIVE, real Ollama daemon)

Ran locally against a real ollama serve + qwen2:0.5b (no key needed).

GREEN (baseline, unmodified mock):

OLLAMA_HOST=127.0.0.1:11434 OLLAMA_MODEL=qwen2:0.5b npx vitest run src/__tests__/drift/ollama.drift.ts --config vitest.config.drift.ts
 ✓ src/__tests__/drift/ollama.drift.ts (3 tests) 1256ms
 Test Files  1 passed (1)
      Tests  3 passed (3)

RED (temporarily dropped done_reason from the mock's non-streaming /api/chat builder — a genuine mock/real divergence):

 × Ollama drift > /api/chat response shape matches
API DRIFT DETECTED: Ollama /api/chat
  Surface: ollama

  1. [critical] LLMOCK DRIFT — field in SDK + real API but missing from mock
     Path:    done_reason
     SDK:     string
     Real:    string
     Mock:    <absent>

The Surface: ollama marker is emitted → the collector maps it to exit 2 (auto-fixable), confirming the leg has teeth against the real API.

GREEN again (reverted): 3/3 pass; working tree clean (only the test + workflows changed).

Skip-safe: with no OLLAMA_HOST (local dev / non-CI), the leg still cleanly skips (3 skipped) — no regression.

Other checks

• Tests-inclusive ES2022 typecheck on the touched test file: 0 new errors (pre-existing errors in other test files unchanged).
• prettier ✓, eslint ✓ (exit 0), pnpm build ✓, full unit suite ✓ (4684 passed).

🤖 Generated with Claude Code

https://claude.ai/code/session_01KamK73Wu5heLxJEogM6hHC