Claude Code 2.1.137 fixes Windows activation

Claude Code 2.1.137 fixes Windows activation, so the install just works for the engineer who picks up an unfamiliar repo on a Monday. That is the easy part. The hard part starts after the tool launches, when the repo's conventions turn out to be vaguer than the code. Claude Code is only as predictable as the files that tell it how your team works. A team convention is a written rule, kept in a checked-in file, that says how the agent should behave here. This piece is about four small files that carry that weight.

The good news: none of these are meetings or process. They are four .md files a reviewer can open instead of a teammate they have to track down.

Why a working install exposes the real problem

A fast agent without a paper trail is not faster delivery. It is faster guessing.

When the setup friction disappears, the next thing the reviewer hits is skill drift. Skills accumulate small, undocumented exceptions, and each one feels safe at the time. Add them up over a few weeks and nobody can say which rule governed a given change. The cost lands on review bandwidth first, because the reviewer is the one who has to reconstruct intent.

So the question underneath 2.1.137 is not "what is new." It is "can someone unfamiliar trust this repo by lunch." The four files below are built to answer yes.

Write four files your team will actually read

Each fix targets one common failure. Open the file, not a session replay.

Session amnesia. Daily session rotation wipes shared memory, so reasoning from last week vanishes. The fix is a weekly retro note: one .md changelog of the decisions the agent relied on. New teammates inherit the why instead of guessing. Treat it as a rule, not a nice-to-have.

Skill drift. Skills pile up exceptions until a reviewer cannot tell which one applied. The fix is a skill index, a skills/README.md that lists every skill's activation cues and what it defers to. On-call engineers resolve mismatches by reading one file.

Hook thunder. Hooks that fire everywhere train people to ignore them. The fix is a hook budget: cap active hooks per repo, and document each trigger plus its rollback. The alerts that survive the cap stay loud.

CLAUDE.md bloat. An encyclopedic CLAUDE.md gets skimmed, and a skimmed file behaves like an absent one. The fix is a CLAUDE TOC: keep the top 15 lines as the rules that win, then link to deeper fragments.

Here is what the top of that file 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 slash command in .claude/commands that opens the index plus the latest retro note gives the inheriting reviewer a one-keystroke briefing.

Prove the convention held in the pull request

Drift shows up at the review gate before it shows up in an incident. Run a PR through these four questions:

And the short checklist a reviewer can paste into a PR template:

• Forked agent work lists parent plus child responsibilities.
• Red-folder paths got explicit human acknowledgement.
• Scopes in the PR body match the folders in the diff.
• Primary-doc links were smoke-checked after publishing edits.

None of this replaces architecture judgement. The agent speeds up execution; you still own the decision. Our methodology draws the line at review: a test proves behavior, a review proves the team can explain it.

Where to go next

Pick the one file that hurts most right now, usually the skill index, and write it before your next handoff. Our training takes one unfamiliar repo and one inheriting reviewer and rebuilds these briefing files live, so the Monday handoff stops feeling like an initiation rite.