Repository navigation
fix(ci): pin Kitchen-sink E2E job to ubuntu-24.04 (FIX-2438) - #4431
Conversation
ubuntu-latest is rolling to ubuntu-26.04, where the installed Playwright refuses to install chromium-headless-shell, so the e2e job fails in setup on unrelated PRs. Fixes FIX-2438 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiFgEDbwFrhqAkyz6pH
|
The latest updates on your projects. Learn more about Vercel for GitHub.
3 Skipped Deployments
|
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
|
Second-look / architecture review: no blocking changes. Approach is right-sized. Is there a better approach? Not really. Alternatives considered:
Minor, non-blocking notes for the implementer:
Architecture/refactor opportunities: none worth taking in this PR. It's a one-line CI change, and widening it would break the surgical-change rule. Changeset skip is correct (CI-only, BP-022). Generated by Claude Code |
There was a problem hiding this comment.
/simplify review
Scope: One commit — pin the Kitchen-sink E2E job to ubuntu-24.04 because ubuntu-latest now resolves to 26.04 and Playwright’s headless-shell install path fails there (FIX-2438).
Did this PR over-engineer?
No. The diff is a single job label plus a rationale comment. Parallel read-only passes (code quality, performance, reuse) all landed on the same conclusion: this is the smallest fix that restores the existing custom Playwright pipeline (version-keyed cache, --only-shell chromium, conditional install-deps) without introducing a second setup path.
Alternatives considered (and why I’m not asking for them here)
| Approach | Verdict |
|---|---|
Bump @playwright/test |
Right exit for FIX-2438, but a separate lockfile/verify change; doesn’t belong in this one-line hotfix unless someone has already confirmed a newer release fixes 26.04. |
microsoft/playwright-github-action |
Official, but would replace or duplicate the job’s deliberate caching/timeouts/--only-shell design — a workflow rewrite, not simplification of this incident. |
Playwright Docker image / job container: |
No container jobs elsewhere in CI; still need Node 22 + pnpm + monorepo build — much heavier than runs-on. |
| Pin all CI jobs to 24.04 | Over-broad; failure is scoped to browser install on the e2e job only. |
Reuse / consistency
This matches the Redis pin pattern in the same workflow: pin the rolling upstream piece that broke one narrow job, explain in YAML, track unpin. Optional polish (not blocking): a short “E2E runner in CI” note in docs/contributing/development-setup.md mirroring the Redis section would help the next person who wonders why only e2e is pinned.
Performance
Pinning 24.04 is a net win — failed installs on 26.04 wasted queue time without populating the browser cache. Cache key playwright-${{ runner.os }}-… is unchanged (Linux), so no accidental cache bust.
Nits (optional, not blocking)
- FIX-2438 appears only in this comment — fine if Linear is filed; no retained spec in-repo yet (unlike some other FIX-* CI follow-ups).
- Comment phrase “the installed Playwright” could optionally name the lockfile-resolved version if you want faster future debugging — purely cosmetic.
Approval: Proportionate, low-risk, and simpler than any alternative that would actually ship in the same PR.
Sent by Cursor Automation: Review For Best Approach
There was a problem hiding this comment.
Code Snob — already tight. No POC.
Head ca856e8e326f1c978c350d8cdc58ccc96999312c. I read .github/workflows/ci.yml: the e2e job and every other runs-on in that file.
A rewrite would not shrink the contract. The change is one pin: Kitchen-sink E2E runs on ubuntu-24.04 because the installed Playwright cannot install chromium-headless-shell on ubuntu-26.04, which ubuntu-latest now is. The three-line comment is the unpin condition (FIX-2438). Folding it into one line still leaves the pin. A shared runner label would touch the other jobs this PR leaves on ubuntu-latest on purpose. I am not asking for a change.
Sent by Cursor Automation: Code Snob
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGRHaqqMGz8cJxCeoQ
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiDfgRv9riWiVVdj8JG
ubuntu-latest now resolves to 26.04, where the installed Playwright can't install chromium-headless-shell, so the E2E job fails before any test runs. Same change as #4431, so the two merge cleanly in either order. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGyofiiBhyj3poEM5L
Copied verbatim from #4431 so this PR's E2E check runs; merges cleanly when #4431 lands. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiH8rtgw7Jgm92Gbhze
Picks up the E2E runner pin (#4431). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiDbo1kHJMahMHLh61N
… (FIX-2388) (#4421) * spec(FIX-2388): coding work that knows which issue it is for Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGyofiiBhyj3poEM5L * feat(workspace): start a run on an existing or fresh remote branch; list and remove worktrees (FIX-2419) A base naming a pinned remote's branch is fetched before the cut, and refused when the fetch fails. PlaceRequest.existing checks a branch out rather than cutting one, making it from the remote's when absent. WorkspaceHost.list and .remove are the one listing and removal path: removal refuses uncommitted work, a place mid-provision, and (unless forced) deleting a branch that holds the only copy of a commit. pullRequestHead resolves a same-repository PR to its head branch. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiDQ6JwfEn1UWLepwe2 * feat(harness-manager): a session's start, picked at create, and removeWorktree (FIX-2419) Session state gains a read-only start ({ base } | { branch } | { pullRequest }) handed to the workspace's provision, and a workspace record holding the branch the workspace gave the session, which the turn-end pull request check reads. removeWorktree removes the session's worktree through the workspace's remove, refused while a turn runs. resolveHarnessSessionStart maps a start to a workspace host's request. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiDQ6JwfEn1UWLepwe2 * feat(fsdev): fsdev worktrees lists and removes a workspace host's worktrees (FIX-2419) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiDQ6JwfEn1UWLepwe2 * test(FIX-2419): goal check: a session starts where asked; removal never loses work Model-free, over the engine's HTTP router and a real clone of a bare remote. Controls no-fetch, cut and lose-work each redden only their legs. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiDQ6JwfEn1UWLepwe2 * chore(FIX-2419): changeset Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiDQ6JwfEn1UWLepwe2 * feat(harness-manager): coding runs that know which issue they are for (FIX-2388) A harness session can be born with a tracker issue (state.issue, fixed like the model), handed to the workspace on its place. A board task names its issue under harnessFeature with the phase now optional; on a board read from the claim its branch leads with the issue, and a run already under way keeps its recorded branch. An opt-in issueSpec option on both entries opens the brief with specs/issues/<ISSUE>/SPEC.md from the run's checkout. Phases see the issue as trackerIssue. Goal check goals/harness-manager/issue-aware-run (model-free), with the no-brief and task-branch controls failing their named assertions. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGyofiiBhyj3poEM5L * docs(harness-manager): telling a coding run its issue (FIX-2388) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGyofiiBhyj3poEM5L * docs(harness-manager): say user, not person Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGyofiiBhyj3poEM5L * FIX-2419: refuse removing a worktree a run is working in A turn now holds its worktree in use (`<dir>.lock`, the same lease the board's runs take) from provision until the turn ends, so the host's one `remove` refuses `in-use` whoever calls it: the session action, the `fsdev worktrees --remove` command, or later automated cleanup. Docs for start, list and remove land with it, edited against the code. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiDQ6JwfEn1UWLepwe2 * FIX-2388: refuse a spec linked out of the checkout; record a brief-step provision failure Codex review on #4421: containment now checks where the spec resolves, so a SPEC.md (or a spec directory) linked outside the checkout is refused, not read into the prompt. The issue-spec brief step provisions the checkout, so it now sits under the same rescue as the harness and a failure there records the turn's error. The goal imports runBoardSegment from the package. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGyofiiBhyj3poEM5L * chore(FIX-2388): name the issue in the changeset Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGyofiiBhyj3poEM5L * test(FIX-2388): a phase with no issue is the half feature payload now An issue alone is a whole feature key since harnessFeature's phase became optional; the coder lab's refusal test now files a phase with no issue. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGyofiiBhyj3poEM5L * docs(FIX-2388): README example path off the retired docs/specs tree Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGyofiiBhyj3poEM5L * ci: pin Kitchen-sink E2E to ubuntu-24.04 (port of FIX-2438, #4431) ubuntu-latest now resolves to 26.04, where the installed Playwright can't install chromium-headless-shell, so the E2E job fails before any test runs. Same change as #4431, so the two merge cleanly in either order. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGyofiiBhyj3poEM5L * FIX-2419: attach a session to another session's worktree start: { worktreeOf: sessionId } works in the owner session's worktree on its branch and makes neither. Each turn of owner and attached sessions holds the same in-use lock, so two harnesses never write there at once. The record carries owner; only the owner may removeWorktree. Attaching is refused at create (404) when the worktree isn't there in the caller's tenant, and an attached turn that leaves the worktree on another branch is recorded failed. workspace gains checkedOutBranch. The goal gains leg d and a no-attach control that reddens exactly its legs. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiDQ6JwfEn1UWLepwe2 * feat(workspace): one shared list of worktrees, written by the host (FIX-2439) worktreesCollection: a user-scoped, llmReadable collection with one row per worktree a workspace host made (name, path, branch, repo, baseCommit, owner, attached, state, commits, lastTurnAt, pullRequests). Only ./worktrees writes rows; the host calls it on provision, attach, recordCheckout and remove, and recordBranchPullRequests adds pull requests. The harness session flow declares the collection, hands it to its workspace and stores the row key as state.workspace.worktree; the turn-end handler calls recordCheckout while the turn still holds the worktree. The board manager writes a row for each run's checkout. Goal check goals/workspace/worktrees-resource, with controls no-row, no-checkout and keep-ready each reddening exactly their legs. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGY67PxEEvoXH9cZeZ * FIX-2386: a session's workspace record is a pointer to its worktree row state.workspace is { worktree } only. Path, branch, owner and removal live on the row (FIX-2439); the turn-end check reads the row's branch, and removal is refused for an attached session by its start. Stored records with the old copies parse to { worktree: null } and are pointed at the row on the next turn. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGU4FtTXJaQZSt9vLg * FIX-2439: attached entries with last-used time, recordCheckout returns its read, rows read-only to models From the architect's review of the spec (#4442): attached is { kind, id, lastTurnAt }[] so follow-up runs attach too and each entry says when it was last used; recordCheckout answers its snapshot so a run's output uses the same read; the session flow marks worktrees read-only in its resourceVisibility. The goal gains e:other-org and the one-org control. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGY67PxEEvoXH9cZeZ * FIX-2386: show where a run's work went — output fields, workPlace, the chip Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGU4FtTXJaQZSt9vLg * FIX-2386: changeset bumps only harness-manager (ui is private) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGU4FtTXJaQZSt9vLg * FIX-2386: install the work-place chip in Kitchen Sink so the registry drift check covers it Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGU4FtTXJaQZSt9vLg * feat(workspace): one shared list of worktrees, written by the host (FIX-2439) (#4453) * feat(workspace): one shared list of worktrees, written by the host (FIX-2439) worktreesCollection: a user-scoped, llmReadable collection with one row per worktree a workspace host made (name, path, branch, repo, baseCommit, owner, attached, state, commits, lastTurnAt, pullRequests). Only ./worktrees writes rows; the host calls it on provision, attach, recordCheckout and remove, and recordBranchPullRequests adds pull requests. The harness session flow declares the collection, hands it to its workspace and stores the row key as state.workspace.worktree; the turn-end handler calls recordCheckout while the turn still holds the worktree. The board manager writes a row for each run's checkout. Goal check goals/workspace/worktrees-resource, with controls no-row, no-checkout and keep-ready each reddening exactly their legs. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGY67PxEEvoXH9cZeZ * FIX-2439: attached entries with last-used time, recordCheckout returns its read, rows read-only to models From the architect's review of the spec (#4442): attached is { kind, id, lastTurnAt }[] so follow-up runs attach too and each entry says when it was last used; recordCheckout answers its snapshot so a run's output uses the same read; the session flow marks worktrees read-only in its resourceVisibility. The goal gains e:other-org and the one-org control. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGY67PxEEvoXH9cZeZ * FIX-2439: board settle reads the row key from recordCheckout's return, not a second store read Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGY67PxEEvoXH9cZeZ * FIX-2439: row writes against the stored row; a late ready never overwrites a removal; known pull requests refresh - markWorktreeReady and addWorktreePullRequests use updateState, so the write is computed from the stored row and re-runs on a conflict. Two turns ending together both keep their pull requests. - A provision's ready write skips a row a removal has marked removed since the provisioning write. - recordBranchPullRequests refreshes the row's known pull requests from the whole listing, so a merge or close after the turn cutoff still lands. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGY67PxEEvoXH9cZeZ * FIX-2439: a provision's row writes run under the place lock removal takes Replaces the removed-row guard, which misread a removal that finished before a re-provision. Under the lock a removal lands wholly before or after the provision, so the row always ends matching the checkout. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGY67PxEEvoXH9cZeZ * FIX-2439: every row write runs against the stored row One helper, change(), runs each writer's guard inside updateState, so a write computed from a stale read re-runs. A turn-end read that finds the directory gone no longer overwrites a removal that committed first. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGY67PxEEvoXH9cZeZ * FIX-2439: a checkout made again of a kept branch keeps the row's base created is true for any new checkout, including one of a branch a removal kept. The place now says whether this call cut the branch (branchCut), and only a cut branch replaces the row's base. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGY67PxEEvoXH9cZeZ --------- Co-authored-by: Claude <noreply@anthropic.com> * FIX-2419: keep fsdev worktrees' helpers module-private Fallow flagged pinnedRepositories and formatWorktrees as exports with no consumer; only the command uses them. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiDQ6JwfEn1UWLepwe2 * FIX-2419: review fixes on #4472 - Attaching checks whose worktree it is: the owner's place names no user, so the first turn attaches through the user's own list of worktrees and is refused (HARNESS_SESSION_ATTACH_NOT_YOURS) when it holds no row, before the harness gets the path. Goal leg d:other-user-refused, control any-user. - remove() refuses a detached HEAD holding commits no branch, tag or remote holds (unique-commits), unless forced. - list() skips a bare clone's own porcelain block, so list and remove work once the host has cloned a remote. - Both in-use lock waits take the turn's abort signal. - recordCheckout reads the branch through checkedOutBranch. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiDQ6JwfEn1UWLepwe2 * fix(workspace): remove re-reads HEAD under its lease; list reads layout, not the name FIX-2419 review round 2 on #4472: - remove() judged the HEAD list() saw. A run ending between the listing and the lease could leave a detached HEAD with commits, removed without force. Branch and HEAD are now read again once the lease is held. - list() took any worktree named checkout/ for a nested one, so a listed repository's place named ["checkout"] listed as the root. A clone's worktree is always nested; a listed repository's only beside project/. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiDQ6JwfEn1UWLepwe2 * fix(harness-manager): an owner turn takes the worktree lease before provisioning FIX-2419 review round 3 on #4472. Provisioning can move an off-branch checkout back to the owner's branch, so it ran under an attached session's running turn. The owner's turn now locates its worktree and holds the in-use lease first, as the board run does. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiDQ6JwfEn1UWLepwe2 * feat(harness-manager): show where a coding run's work went (FIX-2386) (#4491) * test(kitchen-sink): registry drift counts 39 installed copies main's review-hold (FIX-2380) and FIX-2386's work-place each took the count from 37 to 38; together, after the main merge, 39 are installed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NczDiGyofiiBhyj3poEM5L --------- Co-authored-by: Claude <noreply@anthropic.com>


Requested by Jake · project thread
What
ubuntu-latestis rolling to ubuntu-26.04. On those runners theKitchen-sink E2E (Playwright)job fails in setup, before any test runs:This pins only the e2e job to
ubuntu-24.04. Playwright is not bumped here; unpinning once the installed Playwright supports 26.04 is the follow-up, noted in the workflow comment. No other job in the workflows installs Playwright or a browser, so nothing else needs the pin.Fixes FIX-2438.
Evidence
playwright installwith the error above.ubuntu-24.04. The browser installed, and the suite reported41 passed (1.8m).No changeset: CI-only change, no published package affected (BP-022).
🤖 Generated with Claude Code
https://claude.ai/code/session_01NczDiFgEDbwFrhqAkyz6pH