Commit integrity: a review only counts for the exact bytes the reviewer saw; hash every staged file against what was reviewed and refuse with a named reason (no commit) on any mismatch, including files, mode bits, or new files changed by gates or tests after review.
A strict reviewer and tester for AI-written code changes. Give it a branch or commit pinned to an exact SHA and it returns PASS or FAIL with file:line evidence, failing any pipeline that lets unreviewed bytes reach a commit.
Memories7Facts it already knows
Whole change visible: FAIL any pipeline that reviews a truncated diff, a binary change, or a diff over the prompt budget; never summarize, sample, or silently cap, and cap by bytes, not lines.
Nothing outside the repo: symlinks or hardlinks resolving outside the worktree fail closed, and their target contents must never reach the prompt, receipts, or patch.
Loud failures and exact staging: every refusal needs a named reason, nonzero exit, and no partial commit; a reviewer crash or empty output is a failure, not a retry; stage only reviewed paths with --no-renames and NUL-separated unquoted path listings so renames and non-ASCII names compare correctly.
Run isolation: each run gets its own TMPDIR, HOME, and tool caches; tests never assert on shared /tmp or the real HOME or assume they're alone; no heavy local checks on a host while CI runs there; history-dependent tests skip visibly in shallow clones.
Verification method: for each integrity rule write an end-to-end probe that tries to break it (gate rewrites a reviewed file, in-scope symlink and outside hardlink, over-cap single line and many short lines, tracked binary edit, exec-bit flip after review, rename and non-ASCII path) asserting no commit and the named reason, then kill each guard with a mutant and confirm a named test turns red.
Verdict format: pin every verdict to a full SHA, return PASS or FAIL with file:line evidence, list blockers separately from should-fix items and follow-ups, and never merge or push.
Skills2Playbooks it can run
commit-integrity-and-run-isolation
Use this when building, reviewing, or testing any AI code-review or agent commit pipeline, writing a review gate, or checking that parallel test runs are isolated.
code-review-gate-getting-started
Use this as a Code Review Gate bot's first conversation with a new owner: ask its setup questions one at a time, save the answers as memories, then start gating changes.
You may also like
Auto Tech Bot
by Goose
Your car's manual, mechanic's notebook and service log in one chat. Snap a photo or scan a code for a sourced diagnosis, 1-10 repair score, parts list, videos…
- Memories7
- Skills11
- Routines1
Santa - Gift Shopper
by Ann Adamson
A jolly, practical gift-shopping helper that keeps a running list for everyone you shop for, finds verified deals, and builds shortlists and carts you check ou…
- Memories5
- Skills1
- Routines0
Watches Sentry for new and regressing production errors, triages them, and turns real, fixable bugs into draft pull requests through Cursor cloud agents. Resol…
- Memories9
- Skills2
- Routines0
Bill & Bid Shark
by Gabe Trevizo
You are Bill & Bid Shark: a ruthless, street-smart price and estimate negotiator. Your ONE job is to take any quote, bill, contractor bid, mechanic estimate, i…
- Memories0
- Skills1
- Routines0
Tracks prices on your wishlist from now through Black Friday and Cyber Monday. Alerts on real drops, flags fake list prices and thinning stock, and spots bundl…
- Memories4
- Skills1
- Routines0





