integration-smellz
Skill1 starsChanged 36 days agoMITView on GitHub
Deep-reviews every PR in a PM issue or board's multi-repo integration for defects, cross-PR contract breaks, and work that can be cut or simplified. Read-only. Only use when explicitly requested.
In this repository17 other files it gives its agents.
- AGENTS.mdAGENTS.md
- attentionSkill
- audit-debian-serverSkill
- frontend-qaSkill
- goalSkill
- implementSkill
- invariant-minerSkill
- llm-providersSkill
- paretoSkill
- pr-describeSkill
- pr-smellzSkill
- red-teamSkill
- rust-evmSkill
- shrinkraySkill
- skill-labSkill
- student-counselSkill
- vfs-docsSkill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 1 is distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 36
Check every cross-repo pin each PR records (manifests, lockfiles, compatibility files, workflow defaults, test literals): full, on the …
- Very specific: line 48
Before review, write seams.md: one row per interface a PR changes (wire types, protocol versions, persisted state, paths, file formats, …
- Very specific: line 64
Wave 2, one seam reviewer: when more than one repo is involved: takes seams.md, wave 1 results, and the cross-PR checklist.
- Very specific: line 97
When a host supplies scope.json and a stage JSON contract, load references/watcher-contract.md.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Integration Smellz
Find what is wrong with a set of related PRs before they land: defects in each
PR, breaks between PRs and repos, and work to cut. The user knows which PRs
are open. Never report open PRs, landing order, or pending release steps as
results. Read-only: no merges, retargets, pushes, comments, or board edits.
Load ../common/patterns/execution-contract.md and
references/review-contract.md.
1. Resolve the set
From a PM issue, board view, or PR URLs, refresh every PR and record repo,
number, base, head SHA, and stack parent. Add merged PRs the open ones build
on and repos that consume what they change (e.g. a service reading a path
another repo writes). Do not expand to the whole org. Host scratch policy decides WORKSPACE's
location. Work in full clones under WORKSPACE
(git clone --shared, unshallow, fetch pull/<n>/head); never fetch into,
check out, or build in a shared checkout. Access failures are coverage gaps.
2. Compose against today's bases
Review each PR as it will land, not as it was written.
- Per repo, merge open PRs in stack order onto the current base tip with
git merge-tree --write-tree. A conflict is a finding. Base commits since the merge-base that touch the same files are semantic-merge risks. - Check every cross-repo pin each PR records (manifests, lockfiles, compatibility files, workflow defaults, test literals): full, on the upstream default branch or a PR head in the set, and consistent. A pin to a PR head in the set is expected, not a finding.
- Build the composition in a WORKSPACE clone; run the locked check and the PR's tests. A failure is a finding; an unrun build is a gap. Never resolve a code conflict: test the head on its own base and mark the composition untested. Only the orchestrator builds; finish this step first so reviewers start with its results.
3. Map the seams
Before review, write seams.md: one row per interface a PR changes (wire
types, protocol versions, persisted state, paths, file formats, manifest or
config fields, CLI, policy and idempotency keys, error codes, enums, agent
docs). List its producer and every consumer at the composed revisions,
including released peers that still read it. Each row ends agrees,
breaks: <id>, or unchecked: <why>. The orchestrator owns this file.
4. Review
Fresh-context reviewers, at most 6 at once, each writing only its own file
under reviews/:
- Wave 1, one per open PR: its diff against its stack parent, the PM goal, its composition result, and its seam rows. Reads its repo at the head and peers at composed revisions. Applies the lenses and recurring patterns; returns findings, seam results, and whether it read every changed hunk.
- Wave 2, one seam reviewer when more than one repo is involved: takes
seams.md, wave 1 results, and the cross-PR checklist; closes open rows and traces each goal producer -> consumer -> user effect, including failure, retry, upgrade, and rollback.
Without agents, run the roles in sequence and say so.
5. Verify, then report
Re-check every finding against source; keep it only if code shows it or a focused check reproduces it. Merge shared root causes. For each medium or higher, run a skeptic pass that tries to refute it, including whether the behavior is a documented, user-disclosed design. Drop refuted findings and findings on heads that moved.
Write INTEGRATION_REVIEW.md:
- Findings by severity: PR(s),
repo:path:lineat the SHA, what breaks for whom, evidence, fix. None? First line says "No defects found". - Cuts: what to delete or simplify, saved cost, lost benefit. Moving a change to another PR is not a cut.
- Coverage: per PR, SHA, composed or not, finding count, what was not read (fully reviewed = every changed hunk read); the seam tally.
- Questions: open points that could hide a defect.
Also write findings.json. In chat give the counts and one line per critical,
high, or medium finding and per cut; point to the report for the rest.
Final line:
DONE: <WORKSPACE>/INTEGRATION_REVIEW.md — <N> findings (<critical>/<high>), <M> cuts, <K>/<T> PRs fully reviewed, <S> seams unchecked
Watcher mode
When a host supplies scope.json and a stage JSON contract, load
references/watcher-contract.md. Its rules replace conflicting ones above.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

