Repository navigation
claim-test: concurrent mkdir CAS race fails intermittently on ubuntu CI #480
Description
Problem
tests/fixtures/quest-log/claim-test.sh has a simultaneous-claims test (rounds 1–10) that launches two background claim-acquire processes and expects exactly one to exit 0 and the other to exit 6. The stub gh uses mkdir as its CAS primitive.
On ubuntu CI runners (azure eastus2, runner 20261021.618), both processes occasionally return exit 0 — the mkdir race fires — causing:
claim-test: round 1: exits 0:0, expected one 0 and one 6
Evidence
Two consecutive failures on PR #479 (only documentation and version-bump changes):
- Run 40908299894, attempt 1: round 1 fails (exits 0:0)
- Run 40908299894, attempt 2 (rerun --failed): round 2 fails (exits 0:0)
Last successful main run on same runner version: Oct 28 (run 40612584652), claim-test passed.
The test passes consistently on macOS CI and locally.
Root cause hypothesis
The stub gh does two separate steps to atomically claim: mkdir -p "$GH_STORE/quest-claim" (parent) then mkdir "$GH_STORE/$name" (leaf, atomic CAS). Under high concurrency or IO pressure, both background processes may complete their mkdir -p (parent, idempotent) before either completes the leaf mkdir, and the kernel mkdir(2) atomicity guarantee occasionally fails on this runner configuration. Alternatively, the two-step sequence creates a window where a third party could interfere.
Proposed fix
Use a single atomic mkdir -p + unique name or a (set -C; >file) pattern for the CAS instead of two separate mkdir calls. Or pre-create the parent directory in new_store() so both background processes skip the mkdir -p step and go directly to the atomic leaf mkdir.
Activity
randomparity commentedon Oct 29, 2026
Scope — issue #480
q480-4ec325bbtests/fixtures/quest-log/claim-test.shpasses deterministically on ubuntu-latest (Ubuntu 26.04) as well as macOS, with its stubghlabel create acting as a true atomic test-and-create.mkdiratomicity (source: issue claim-test: concurrent mkdir CAS race fails intermittently on ubuntu CI #480 problem statement + proposed fix; root cause confirmed by reproduction below)..claude-plugin/plugin.jsonPATCH bump (source: repo rule, ADR 0022).documentationlabel replaced withtype:bug(source: operator decision, this session).ubuntu-26.04, failing) vs run 40612584652 (imageubuntu-24.04, passing); container reproduction:mkdir (uutils coreutils) 0.8.0— two racingmkdirof one leaf both exit 0 in 1991/2000 trials,mkdir (GNU coreutils) 9.40/2000. The issue's kernel-mkdir(2)hypothesis is refuted; the parent-precreate remedy would not fix it. Exclusions approved by the operator interactively this session.tracker.sh/profiles/github.shclaim code — owner: maintainer (no evidence of a defect there; it calls the GitHub API, notmkdir).tests/fixtures/quest-log/claim-test.sh(stubghcreate path and its comments);.claude-plugin/plugin.jsonversion; issue claim-test: concurrent mkdir CAS race fails intermittently on ubuntu CI #480 labels.Tracking: blast radius — one test fixture; change hazards — none (shipped code unchanged; the fix restores the stub's intended atomicity); complexity — S; decompose — one PR; classification — trivial-bugfix; artifact lane — no-spec; review depth — single-pass; design denominator — not-applicable (no-spec).
randomparity commentedon Oct 29, 2026
Trajectory — issue #480
feat/claim-cas-480, PR fix(test): make claim-test stub CAS independent of mkdir #481.just cigreen via pre-push hook at915ae4b; CI green on ubuntu-latest (image ubuntu-26.04) and macos-latest.mkdir0.8.0 is non-atomic under a race; the stub CAS is now a bash noclobber (O_EXCL) create.