Repository navigation
chore(baseline): Re-apply the repo-ci-baseline Template - #253
Conversation
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The Template's review-prompt.md fills in what the host refuses in the reviewer's PR description from {{PR_BODY_RULES}}; this repo's copy wrote those rules by hand, and lib/config.test.mts fails without the placeholder.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…* unset Apply never overwrites a Seed file, so the re-Apply left package.json without the sandcastle:status script that .sandcastle/status.mts documents, and .sandcastle/check.sh without the unset of the GIT_* variables a git hook sets. A test pins both. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #253 +/- ##
==========================================
- Coverage 90.07% 89.82% -0.26%
==========================================
Files 17 17
Lines 393 393
Branches 85 85
==========================================
- Hits 354 353 -1
Misses 15 15
- Partials 24 25 +1 🚀 New features to boost your workflow:
|
|
Copilot was asked to review this head (REST and GraphQL review requests) and neither registered a review request, so it cannot review this PR. Recorded by hand because this PR delivers the PR Auto-Merge that would post it. |
There was a problem hiding this comment.
Claude Review
Overview
This is a repo-ci-baseline Template Apply. It adds:
- a CI job that checks the PR against the base branch's copy of the
gh-repository checker; - a Copilot-instructions sync workflow;
- a Sandcastle workflow with a
record-headjob; watch-pr.sh,sync-copilot-review.shand the hardening-review scripts;- a large rewrite of PR Auto-Merge (a review loop counted in dirty rounds, Claude as Copilot's backup only, and an exhaustion comment);
- the matching Claude Review changes.
I found one real inconsistency, in the new "Skip while Copilot can review" step of claude-review.yml. The rest of what I could read looks sound. The tests, the TypeScript Sandcastle code and the docs were not read (see Not reviewed), so this is not a complete review of the PR.
Changes
| File | What changed |
|---|---|
.github/workflows/ci.yml |
New gh-repo job. It runs the base SHA's check_gh_repo.py over the PR tree and compares the base's Owned files with the head's. Two hook-test steps added. |
.github/workflows/claude-review.yml |
New first step decides whether Copilot can still review, and Claude skips the review while it can. The merge-from-main job can now review a newer head against main's tip. Template-copy dedupe added to the diff splitter. |
.github/workflows/copilot-review-sync.yml |
New workflow running scripts/sync-copilot-review.sh --check. |
.github/workflows/pr-automerge.yml |
labeled trigger added. Review cap replaced by a dirty-round loop with restarts. Copilot is asked once per PR with a recorded ask. Claude is called in only where Copilot can't review. An exhaustion comment is posted. |
.github/workflows/sandcastle.yml |
New Sandcastle host workflow and a record-head job. |
.github/scripts/check_gh_repo.py, scripts/gate.sh, scripts/review.sh, .sandcastle/check.sh |
New checker. The gate runs the checker and the Copilot-instructions sync check. check.sh unsets the GIT_* variables. |
scripts/sync-copilot-review.sh, scripts/watch-pr.sh |
New scripts. |
.claude/skills/hardening-review/scripts/*.py |
New known_open.py and mutate.py. |
.sandcastle/lib/branches.mts, .sandcastle/.env.example |
isHandOffBranch and the SANDCASTLE_MODE note. |
Advisory (not blocking)
sandcastle.yml, header comment: "Its pull request triggers are pull_request_target" is not true ofpull_request_review. That event runs the workflow file from the PR's merge ref, not from main. The job checks outmainitself, and a same-repo pusher could already read the secrets with their own workflow, so this adds no escalation. The comment is just inaccurate.scripts/watch-pr.sh:ours()treats any comment, review or thread reply that contains the session marker anywhere, or starts with<!-- sandcastle:x -->, as the owner's. A third party can therefore silence the watcher for their own comment. Thepersonfilter excludes only Bot accounts, not people without write access.- Conversely, the first 200 characters of any person's comment are printed into the agent session. That is a prompt-injection channel on a public repo. Consider filtering by
authorAssociation.
pr-automerge.yml, exhausted loop: a head nobody reviewed, with unresolved Copilot or Claude threads, now merges once CI is green. This follows from the documented design (dotfiles#378). It does weaken the merge gate for the PR's own author, so confirm it is intended.pr-automerge.yml, ask record: the ask record is read only from the PR's last 100 comments. If it has scrolled off,copilotAskreturns "unread" and a dropped Copilot request (no quota note) never calls Claude in. The PR waits for a manual merge. This only happens on a PR with more than 100 comments.
Not reviewed
- Not read: pieces 009–037. These hold the Sandcastle TypeScript sources and their tests (
.sandcastle/lib/*.mts,main.mts,status.mts,package.json,*.test.mts),pr-automerge.test.mjs, the Python and shell tests, and the docs. - Docs in those pieces:
.claude/rules/sandcastle-hand-off.md, the hardening-review and pre-push-review skill docs,.github/ci/sandcastle-skills.txt,.sandcastle/review-prompt.md,.security/README.mdanddocs/PROCESS.md. - In 014:
.github/hooks/tests/pre-push.test.sh. .gitignorewas excluded and is held for a person.claude-review.ymlandpr-automerge.ymlwere not in the context pieces. I judged them from the diff and could not re-read the checked-out files. The finding's line number comes from counting lines in the diff hunk, not from a read of the file.
Findings outside the diff (no thread to resolve, so they hold the merge until the next push gets a fresh review):
diff:0: Claude Review didn't read all of diff pieces 009.diff, 010.diff, 011.diff, 012.diff, 013.diff, 014.diff, 015.diff, 016.diff, 017.diff, 018.diff, 019.diff, 020.diff, 021.diff, 022.diff, 023.diff, 024.diff, 025.diff, 026.diff, 027.diff, 028.diff, 029.diff, 030.diff, 031.diff, 032.diff, 033.diff, 034.diff, 035.diff, 036.diff, 037.diff: they're past what one review reads, or a Read of them failed or came back in part. The log of the review job's "Check Claude's answer for secrets" step lists their files. Review those files yourself, then merge by hand.diff:0: Claude Review didn't read 1 changed file(s): files Copilot code review excludes (such as dist, bin, vendor or generated files, build configs or lockfiles), files the diff shows only as binary, submodule bumps, symlinks or Git LFS pointers. The log of the review job's "Check Claude's answer for secrets" step lists them. Review them yourself, then merge by hand.
Automated release blog posts for #253, opened by the release workflow. It holds every Release whose post isn't on main yet, rebuilt from main on each run. The [skip-release] title marker keeps its merge from starting another release. Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Summary
Re-applies the repo-ci-baseline Template to TicketManager, bringing mpaulosky/dotfiles#368, mpaulosky/dotfiles#369, mpaulosky/dotfiles#370, mpaulosky/dotfiles#374, mpaulosky/dotfiles#375, mpaulosky/dotfiles#378, mpaulosky/dotfiles#380 and mpaulosky/dotfiles#381.
package.jsongainssandcastle:status,.sandcastle/check.shunsets theGIT_*variables, and.sandcastle/review-prompt.mduses{{PR_BODY_RULES}}..sandcastle/repo-scripts.test.mts.Local pre-push review: 2 rounds, 2 Critical fixed, 2 Advisory left for a follow-up (the Template's
gate.shruns Docker and Playwright test projects; the sandbox gate's comments say it lints butcheck.shdoes not).🤖 Generated with Claude Code