[CI] The review prompt points at a CLAUDE.md that does not exist on any branch

[erikdarlingdata]

What

.github/workflows/claude-review.yml opens its prompt with:

Review this pull request. Follow the conventions in CLAUDE.md and the T-SQL style guide it points to.

Neither file exists.

$ git ls-files | grep -i claude
.github/workflows/claude-review.yml     # the only match, on any branch

$ git ls-tree -r origin/main --name-only | grep -i 'claude\.md'
(nothing)

There is no CLAUDE.md on dev, none on main, and no T-SQL style guide for it to point to. The repo's agent-facing documentation is CONTRIBUTING.md and llms.txt.

Why it matters

Every review this repo runs is told, as its first instruction, to follow conventions in a file it cannot open. Two consequences:

1. The conventions are not actually applied. The house rules that do exist (T-SQL style, Lite/Darling parity, the no-missing-index-DMV rule that the prompt has to restate inline precisely because there is no file to carry it) reach the reviewer only insofar as the prompt repeats them.
2. It burns a turn on every run. The reviewer's first move is to look for the file. With allowedTools granting no file-reading tool at all (see [CI] Review workflow can execute fully (real token spend) and post nothing - green check with vanished output #2229), that attempt is also a permission denial.

Fix shape

Either direction is cheap; they are not equivalent:

• Point the prompt at what exists. Swap CLAUDE.md for CONTRIBUTING.md (and drop the dangling "and the T-SQL style guide it points to", or name the guide directly if one should exist). Smallest change, no new file to maintain.
• Add a real CLAUDE.md. Worth it if the intent was always a single agent-facing conventions file — it would also give the review a place to carry rules that currently have to be inlined in the workflow prompt, which matters because editing that prompt costs a repo-wide review outage until the next release (mechanism in [CI] Review workflow can execute fully (real token spend) and post nothing - green check with vanished output #2229).

Note the second option has a side benefit: rules that live in CLAUDE.md can be changed without touching claude-review.yml, so they cost no drift window.

Relations

• [CI] Review workflow can execute fully (real token spend) and post nothing - green check with vanished output #2229 (the review can spend real money and post nothing; the same allowedTools list that blocks reading this file is that issue's root cause)

[erikdarlingdata]

Correcting the premise, because it decides which fix is right. CLAUDE.md is not merely missing — it is gitignored, and has been since the v1.0.0 initial release:

# Internal development files (not for public release)
.internal/
.claude/
CLAUDE.md
AGENTS.md
Dashboard/todo.md

So it is deliberately unpublished, alongside .claude/ and AGENTS.md. That removes the issue's second option — "add a real CLAUDE.md" — because committing it would publish a file intentionally kept out of a public repo. I found this by trying it: git add refused. The references are the bug, not the absence.

Also: there are two references, not one, and the missed one is worse.

CONTRIBUTING.md:233 says "The full T-SQL style guide is in CLAUDE.md." That is a public, outward-facing file telling contributors the guide is one click away, in a file they can never see. #2268 fixes it by putting the rules in the public document instead of pointing elsewhere — including the three prohibitions that previously had no public home at all (no missing-index DMV recommendations, no full-text search, no ASCII-art headers), which is the same homelessness you identified as forcing the prompt to restate them inline.

Leaving this open for the workflow half, with a recommendation rather than a change:

Repoint the prompt's first sentence at CONTRIBUTING.md, which after #2268 actually holds the guide.

I am not making that edit tonight because it costs a repo-wide review outage until the file matches the default branch — claude-code-action validates it server-side during the OIDC exchange and skips with conclusion=success when it differs, silently. So it should ride with the #2229 allowedTools cure, which needs a release anyway. Bundling them means one outage window instead of two.

One thing worth noting for whoever does that edit: if it lands with ALWAYS POST in the prompt text, the #2262 review guard automatically becomes strict — it currently treats a silent review as a warning precisely because nothing obliges the reviewer to post. That is a free upgrade if the sentinel goes in at the same time.

[erikdarlingdata]

Scoping note before anyone burns an afternoon on this as a quick fix: the prompt lives inside claude-review.yml, and claude-code-action refuses to run whenever a PR branch's copy of that file differs from the DEFAULT branch's (main). Since every feature PR is based on dev, landing this fix on dev means every dev PR's review silently refuses to run until the next release syncs main — the exact repo-wide hazard claude-review-guard.yml's header documents (and the guard now surfaces as a warning per #2309's fixes, so at least it won't be silent).

So the fix is release-coupled, and there are two clean shapes:

1. Batch it into the next dev→main release: land the prompt change (point at CONTRIBUTING.md, which is the real authority, and drop the reference to the gitignored CLAUDE.md) as the LAST merge before cutting the release, accepting a short window of review-refusals on any PR that races it.
2. Fix-and-release immediately: land it and cut a dev→main release right after, shrinking the window to minutes.

Either way the edit itself is two lines. Erik's call on timing — flagging rather than fixing so a well-intentioned quick win doesn't turn off review for the whole repo.

[erikdarlingdata]

Fixed in 6ce8413 (#2334), riding the 3.5.0 release window as planned: the review prompt now points at CONTRIBUTING.md — the agent-facing doc that actually exists on every branch — and names the T-SQL style rules inline instead of dangling a reference through a gitignored file. Takes effect on the next PR's review run.