Claude Code 2.1.138: team conventions check

A team conventions check is a short audit that asks one question: does your repo, not your chat history, record what changed, how it was verified, and who approved it? If a pull request says "fixed and verified" but nothing is pasted or linked, the check has already found something. Claude Code, Anthropic's coding agent, moves fast and produces a lot of green CI, so the receipts that prove a change is safe have to live somewhere a reviewer can find them later.

This piece walks through four small files that make those receipts real, plus a proof strip you can run against merges you have already shipped. None of it requires new tooling. It requires writing things down in places people will look.

Catch the PR that says "verified" but shows nothing

The expensive failure is verification theater: a green pipeline presented as proof, with a body that describes how the change felt rather than what it touched. The tell is always a missing transcript. Someone claims a regression is fixed, and there is no command output to back it up.

This is easy to miss because the sessions you remember are the ones that worked. The silent failures never make it into the retro, so the habit looks safer than it is. A conventions check exists to surface that gap before merge instead of after the incident.

The fix is not to slow people down. It is to make the proof cheaper to produce than the vibe.

Write down the four files that carry the load

Four common failures each have a one-file answer. A conventions check passes when each of these files already exists in the repo.

Session amnesia. Sessions rotate daily, and shared memory rots when reasoning only lives in transcripts. The fix is a weekly retro note: one appended .md changelog of decisions agents relied on. New teammates inherit the reasoning instead of the rumor.

Skill drift. Skills pile up until a reviewer cannot tell which one governed a diff. The fix is a skill index at skills/README.md that lists each skill's activation cues and what it defers to. On-call engineers resolve mismatches without replaying a session.

Hook thunder. Hooks that fire everywhere drown out the alerts that matter. The fix is a hook budget: cap active hooks per repo, and document each trigger plus its rollback. Alerts get their urgency back.

CLAUDE.md bloat. An encyclopedic CLAUDE.md gets skimmed, and an unread file behaves like an absent one, so the permissions it describes stop applying in practice. The fix is a short table of contents: keep the top 15 lines as the rules that win, then link out to deeper fragments.

Here is what those top lines can look like:

# CLAUDE.md supremacy fragment

- Hooks win over informal chat agreements; document each hook's rollback path.
- Skills defer to this file on security-sensitive folders.
- Bash approvals never bypass the red-folder list maintained here.

A /receipt slash command in .claude/commands that pastes the proof strip into the PR body keeps the ritual cheap. If your team runs these conventions formally, our team conventions cluster carries the standard, and the Claude Code 2.1.136 fixes note is the companion piece.

Run the proof strip against your last five merges

Audit merges you already shipped, not the next one. Old merges tell you what your team actually produces under normal pressure. Walk each one through these four gates.

Then check the PR body itself against this strip:

• MCP connectors are named, and each lists an owner.
• Verification command output is pasted or linked.
• Forked agent work lists parent and child responsibilities.
• Red-folder paths got explicit human acknowledgement.

If your repo cannot state its boundaries plainly, agents will guess at them. Guessing scales badly.

Common questions

• What is a Claude Code 2.1.138 team conventions check?

  It is a short audit that asks whether your repo, not your chat, records scope, verification, and approval. It runs four fixes against your current setup: a weekly retro note, a skill index, a hook budget, and a CLAUDE table of contents. Then it applies a proof strip to your recent merges to see which receipts your team actually produces.

• What counts as verification theater?

  Verification theater is green CI presented as proof while the PR body describes vibes instead of boundaries. The tell is a missing transcript: someone claims "fixed and verified" but nothing is pasted or linked. The receipt-match gate exists to make that gap visible before merge, rather than after an incident forces the question.

• Which receipts belong in a PR body?

  Scopes plus a verification transcript, with command output pasted or linked. The strip adds three more: MCP connectors list their owners, forked agent work names parent and child responsibilities, and red-folder paths carry explicit human acknowledgement. Together these let a reviewer proceed without replaying the chat.

• How does the weekly retro note support the check?

  The retro note is the memory the check audits against: one appended .md changelog of decisions agents relied on. Without it, sessions rotate daily and the reasoning behind your conventions evaporates. The next check then re-litigates the same questions from rumor instead of from a record.

• Where should this fit in our workflow?

  Treat the conventions check as a handoff gate, before a change reaches review. The point is that the work survives without the original operator in the room. If you want to build the habit with your team, our training covers the proof strip with real merges.

Next move

If your last five merges fail the proof strip, that is your starting metric: pick one of the four files and add it this week. Contact us and we will run the check with your leads in one session.

Further reading

• Claude Code getting started
• Model Context Protocol specification
• Anthropic Claude Code product page
• anthropics/skills on GitHub