Full RPI cycle complete. Infrastructure is solid and non-invasive. Coverage is honest v1 — the harness exists and adding tests is now friction-free, but the original goal of full SPEC.md surface isn't there yet.

Status

63 tests passing across 5 files. npm test runs in ~300ms with no real LLM calls, no Docker.

✓ src/parser.test.ts     (16 tests)
✓ src/loop.test.ts       (17 tests)
✓ src/executor.test.ts   (10 tests)
✓ src/race.test.ts       (15 tests)
✓ src/template.test.ts   (5 tests)

npm run build passes with no TypeScript errors.

What changed

src/executor.ts — ExecutionContext gains one optional field:

poolFactory?: (worktreePath: string, config: CookConfig, runAgents: AgentName[]) => RunnerPool

Seven createRunnerPool(...) calls replaced with (ctx.poolFactory ?? createRunnerPool)(...). parseRalphVerdict exported. In production nothing changes — the field is absent and behavior is identical.

src/testing/test-runner.ts — TestRunner (queue-driven AgentRunner), makeTestPool, testPoolFactory helpers.

vitest.config.ts + "test": "vitest run" in package.json.

Five test files covering: all AST node types and flags (parser), parseGateVerdict table tests and full agentLoop loop (loop), parseRalphVerdict direct + executor integration with mocked Ink (executor), parseJudgeVerdict/buildJudgePrompt/sessionId (race), renderTemplate/loadCookMD (template).

Architecture

graph TD
    A[ExecutionContext.poolFactory?] -->|set in tests| B[TestRunner — queue-driven mock]
    A -->|absent in production| C[createRunnerPool → NativeRunner / Docker / BareRunner]
    B --> D[responses queue controls LLM output in tests]
    E[execute / agentLoop] --> A

Key decisions

1. poolFactory on ExecutionContext over vi.mock or a fake CLI binary — explicit, typed, propagates through all execution paths including composition branches.
2. vi.mock('ink', ...) in executor.test.ts — every path touching a work or review node calls Ink's render(); mocked to no-op to prevent TTY noise or hangs.
3. Composition integration tests deferred — full worktree tests need temp git repo + coordinated mock sequences; pure function coverage of resolver logic is in race.test.ts.

Bug surfaced

parseRalphVerdict / APPROVE inconsistency. parseGateVerdict explicitly handles APPROVE as a DONE keyword. parseRalphVerdict does not — RALPH_DONE_KEYWORDS is ['DONE', 'COMPLETE', 'FINISHED']. Passing 'APPROVE' to parseRalphVerdict returns 'DONE' via the fail-safe default (no keyword matched), not explicit handling. The behaviour is correct today but the two functions are inconsistent. If logic is ever reordered, APPROVE through ralph breaks silently. Tracked in TODO.md.

Coverage gaps (not blocking merge, but honest accounting)

• Parser: complex compositions untested. v3, compare, vs with review branches, nested operators (work x3 review v2 pick). The parser handles all of this — none of it is tested.
• agentLoop with iteratePrompt — when i > 1 && config.iteratePrompt, the loop uses the 'iterate' step instead of 'work'. Code path not exercised.
• agentLoop error handling — when runner.runAgent throws, the loop catches and returns { verdict: 'ITERATE' }. Not tested.
• executeReview with compound inner — the branch where inner.type !== 'work' (e.g. repeat-then-review) is untested. Only the simple inner.type === 'work' path is covered.
• Composition entirely absent — by design (deferred), but was in the original requirement.

Deferred (see TODO.md)

• Composition integration tests (git worktree + poolFactory, infrastructure ready)
• Resolver integration tests (resolvePick/resolveMerge/resolveCompare)
• UI tests (App.tsx, RaceApp.tsx, LogStream.tsx)
• loopEvents singleton refactor to injected EventEmitter

Testing instructions

npm install
npm test       # 63 passed (63), ~300ms
npm run build  # no TypeScript errors

🤖 Generated with Claude Code