Verification — `fix-claude-config-cache`

Test evidence

Gate 1 — `npm run build`

> colibri@0.0.1 build
> tsc

Exit code: 0. TypeScript compiled cleanly. No errors, no warnings.

Gate 2 — `npm run lint`

> colibri@0.0.1 lint
> eslint src

Exit code: 0. ESLint clean. No new warnings.

Gate 3 — `npm test`

Full Jest suite (matched against the documented baseline):

Test Suites: 30 passed, 30 total
Tests:       1360 passed, 1360 total
Snapshots:   0 total
Time:        21.232 s, estimated 31 s

Exit code: 0. All 1360 tests pass.

Pre-existing flake note

The first run of npm test hit the documented startup — subprocess smoke › tsx src/server.ts boots and logs [Startup] Phase 1 flake (1 fail under heavy parallel suite load). This is a pre-existing pre-Wave-H flake documented in CLAUDE memory: “Pre-existing startup — subprocess smoke flakiness under full-suite load — predates Wave H; all 4 R77 executors hit it once, always green on rerun.”

Re-running the same suite immediately produced 30/30 suites + 1360/1360 tests green, confirming the flake is not introduced by this change.

Targeted test counts

claude.test.ts alone:

Test Suites: 1 passed, 1 total
Tests:       44 passed, 44 total

Pre-change baseline was 41 tests; +3 new tests = 44, confirming all three new test cases are wired and passing.

New test detail

PASS src/__tests__/domains/integrations/claude.test.ts
  default-model resolution from config
    ✓ createCompletion uses config.COLIBRI_ANTHROPIC_MODEL when options.model is omitted
    ✓ createCompletion: injected options.model overrides the config default
    ✓ createCompletionWithTools uses config.COLIBRI_ANTHROPIC_MODEL when options.model is omitted

Each test asserts the request body’s model field against config.COLIBRI_ANTHROPIC_MODEL (the frozen-singleton import) plus a literal 'claude-sonnet-4-5' regression guard for the schema default.

The override test asserts:

body.model === overrideModel — the override wins.
body.model !== config.COLIBRI_ANTHROPIC_MODEL — explicit regression guard against a future regression where overrides are silently swallowed.

Source diff summary

 src/__tests__/domains/integrations/claude.test.ts | 67 +++++++++++++++++++++++
 src/domains/integrations/claude.ts                | 27 +++------
 2 files changed, 74 insertions(+), 20 deletions(-)

src/domains/integrations/claude.ts:

Net: −13 LOC (20 deletions, 7 additions). Two try/catch blocks (~10 LOC each) are gone. Hardcoded fallback literals are gone. The import was swapped (no LOC delta on imports). Replaced with a single options.model ?? config.COLIBRI_ANTHROPIC_MODEL line per entry point plus a clarifying comment block on the first occurrence.

src/__tests__/domains/integrations/claude.test.ts:

+67 LOC (one new describe block with three tests + one new static import line). No existing tests modified.

Acceptance criteria — final check

AC	Status	Evidence
1. Replace try/catch+fallback with direct import from `src/config.ts`	✅	Diff shows `import { config } from '../../config.js'`; both try/catches deleted
2. Preserve `apiKey`, `model`, `timeoutMs` injection seams	✅	`CompletionOptions` interface unchanged; all 41 prior tests still pass with explicit `model: FAKE_MODEL` injection
3. Preserve `AnthropicConfigError` semantics for missing `ANTHROPIC_API_KEY`	✅	Lines 337–342 + 372–378 of `claude.ts` unchanged; test block “API key validation” (lines 309–365) still passes
4. Singleton handles call-time validation; `config.ANTHROPIC_API_KEY` is `string \\| undefined`	✅	`apiKey` resolution path on `process.env['ANTHROPIC_API_KEY']` preserved exactly — same call-time semantics
5. New tests verify model defaults match `config.COLIBRI_ANTHROPIC_MODEL` and override precedence	✅	3 new tests in `default-model resolution from config` describe block, all green
6. All gates pass: build + lint + test	✅	All three exit 0; 1360/1360 tests pass

Risk re-check

Risk from audit	Outcome
Singleton evaluated at module load	Confirmed working — config loads cleanly under jest’s `NODE_ENV=test` and all 1360 tests pass
Test injection seam regression	Zero existing tests modified; all 41 prior `claude.test.ts` tests still pass
`process.env` mutation between calls no longer reflected	Not load-bearing — verified in audit; no test broke
Silent-fallback path was load-bearing	Confirmed not load-bearing — full suite green

Commit chain

Step	SHA	Title
1. Audit	`0fd22e83`	`audit(fix-claude-config-cache): inventory surface`
2. Contract	`0bc19a39`	`contract(fix-claude-config-cache): behavioral contract`
3. Packet	`84011471`	`packet(fix-claude-config-cache): execution plan`
4. Implement	`f3a7830a`	`feat(fix-claude-config-cache): swap loadConfig try/catch for config singleton`
5. Verify	(this commit)	`verify(fix-claude-config-cache): test evidence`

Writeback summary

task_id: fix-claude-config-cache
branch: feature/fix-claude-config-cache
worktree: .worktrees/claude/fix-claude-config-cache
commits: 0fd22e83, 0bc19a39, 84011471, f3a7830a, (verify)
tests: npm run build && npm run lint && npm test
summary: Replaced per-call loadConfig() try/catch+hardcoded fallback in
  src/domains/integrations/claude.ts with direct read from the frozen
  config singleton. Added 3 tests for default-model resolution and
  override precedence. Net -13 source LOC, +67 test LOC. All gates pass:
  build clean, lint clean, 1360/1360 tests green.
blockers: none. Pre-existing 'startup — subprocess smoke' flake observed
  on first run; documented in CLAUDE memory; clean on rerun.

Verification — fix-claude-config-cache