Codex review surfaced 1 HIGH, 2 MED, 2 LOW; all addressed:
1. (HIGH, CONTRIBUTING.md violation) Adapter body assertion was raw
text matching `content.includes('<codex_skill_adapter>')` on the
rendered SKILL.md. Replaced with a structural assertion against
the exported builder: `content.includes(getCodexSkillAdapterHeader(name))`.
The expected value is now the full closed adapter block produced
by the production IR — open tag, body, and `</codex_skill_adapter>`
closing tag — so a truncated, empty, or missing-closing-tag adapter
cannot satisfy the assertion. Pre-flight sanity-checks the builder
itself emits the expected open/close shape.
2. (MED, sensitivity) Count-only check on installed skills was
replaceable by a same-count partial install that swapped real
commands for bogus `gsd-*` dirs. Replaced with `deepStrictEqual`
on the sorted full set, computed from `commands/gsd/**/*.md` via
a local `expectedSkillNames()` walker that mirrors the installer's
naming rule (nested dirs collapse to `gsd-<dir>-<file>`).
3. (LOW, flakiness) `runCodexInstallCaptured()` could throw after
creating the temp `codexHome` but before returning, so the
describe-level `afterEach` couldn't see the path and the dir would
leak. Added a try/catch around `install()` that cleans up the
temp dir before rethrowing.
4. (LOW, flakiness) Harness mutates process-global `console.*`,
`process.env`, `process.cwd()` — added `{ concurrency: false }` to
the describe block, matching the existing convention in
`tests/bug-3562-codex-install-skill-surface.test.cjs:45`.
5. (LOW, CONTRIBUTING.md helpers) Local `parseFrontmatter()` was
duplicating the shared helper. Switched to the canonical
`parseFrontmatter` exported by `tests/helpers.cjs`; also adopted
`createTempDir` and `cleanup` from the same module for consistency.
Validation:
- New test still 5/5 pass.
- Full suite via gsd-test-both: 10,682/10,682 on Mac AND Linux Docker,
zero cross-platform diff.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
11 KiB
11 KiB