Flaky test: codex/claude adapter skill-injection tests collide on shared tmpdir paths

[tadasant]

Symptom

CI intermittently fails with:

× CodexAdapter > prepareSession > injects skills into .agents/skills/
  → expected '# Catalog version' to contain '# Deploy'

Observed on PR #155 — failed on test (18) and test (20), passed on test (22), passed on rerun. Not reproducible reliably; passes locally.

Root cause

Vitest runs test files in parallel workers. Two adapter test files write skill-source SKILL.md files to shared, non-unique paths under the OS temp dir:

• packages/extensions/adapter-codex/tests/codex-adapter.test.ts — createTempDir() makes a unique air-codex-test-<ts>-<rand> dir, but the test then writes its skill source to join(dir, "..", "skills", "deploy") — i.e. a sibling path (<tmpdir>/skills/deploy) that is not unique to the test. It writes # Deploy, then asserts the injected file contains # Deploy.
• packages/extensions/adapter-claude/tests/claude-adapter.test.ts:542 writes # Catalog version to its own catalog skill source.

When the two files run concurrently and their skill-source paths overlap (shared <tmpdir> ancestry with a fixed skills/deploy leaf), one worker's # Catalog version clobbers the other's # Deploy between write and read, so the codex assertion sees # Catalog version.

Fix direction

Keep every test's skill-source files inside that test's unique temp dir, never in a shared sibling path. E.g. write skill sources under join(dir, "catalog", "skills", "deploy") (a child of the unique temp dir) instead of join(dir, "..", "skills", "deploy"). Audit both adapter test files (and any others using the join(dir, "..", …) pattern) for the same shared-path issue.

Notes

Pre-existing; surfaced on PR #155 but unrelated to that change (which is scoped to plugin manifest resolution). Low severity (rerun clears it) but recurring CI noise.

[tadasant]

🚦 Issue gate

Kind: test-flake-fix (🟢 fast lane) · Scope: convergent · Facets: none

Verdict: AUTO-PROCEED — convergent, no hold test fired. A named flake with a named root cause: test files write skill sources to join(dir, "..", "skills", "deploy"), a path outside their own unique temp dir, so parallel vitest workers clobber each other. Fixing it closes a gap between what createTempDir() is for and what the tests actually do, and adds no surface.

Live pages: none — not alert-shaped, and nothing in pulsemcp/air is deployed (the surface doc: "nothing here is deployed, so nothing pages"), so anchor 2a has no rule to resolve; anchor 2b — quick_search_sessions for pulsemcp/air issue sessions with show_archived: true — returns no alert-genesis session. Normal dispatch.

Rating detail

Axis	Rating	Why
Requirement complexity	small	The issue names the files, the lines, the colliding paths and the fix direction.
Requirement impact	small	Test-only. The shipped packages are unaffected; the cost is recurring CI noise and a rerun.
Solution complexity	small (est.)	Guessing: move the skill/hook sources under a child of the unique temp dir in two adapter test files.
Implementation risk	small (est.)	Guessing: test-only, failure is visible instantly in CI, trivially revertible.
Estimated cost	small (est.)	Guessing: mechanical, but there are ~14 join(dir, "..", …) sites in the codex test file alone, so it is a sweep rather than one line.

Note: the three solution-side axes are estimates — there is no diff yet, and under this surface's posture they are recorded rather than decisive.

Staleness check: live. packages/extensions/adapter-codex/tests/codex-adapter.test.ts on main still contains join(dir, "..", "skills", "deploy") at lines 760, 790, 836 and join(dir, "..", "hooks", …) at 875, 920, 944, 971, 1005, 1168, 1246 — the shared-sibling pattern is unchanged. No merged PR since 2026-06-12 touches these tests (#158, #159 are an AGENTS.md symlink and per-catalog failure isolation). No open issue duplicates it.

[tadasant]

Queued on the work backlog at precedence 5960 (small band, position 5 of 7 at the time of writing).
The backlog groomer pulls from the top of that queue — on its 04:00 run, or sooner if the fleet runs out
of work first. A human can pin or re-place this in Zimmer's Issues view.

[tadasant]

Pulled from the work backlog and started: https://zimmer.tadasant.com/sessions/15293 — a spot implementing session. Started by the fleet-maintenance backlog pull (session https://zimmer.tadasant.com/sessions/15284), 2026-09-06.