Claude Code 2.1.139 team conventions

During Claude Code release reviews, we noticed CLI runs that passed locally while reviewers lacked the replay recipe. Parallel agents reward explicit mutexes, clear skills, and one shared place for permission decisions.

Replay Recipe Gap

Review cannot compress what the repo never recorded. The hidden tax shows up as merge-queue fatigue.

We believed smaller tasks guaranteed safer autonomy. We repeated it until PR archaeology replaced architecture conversation.

Ritchie-style pragmatism still applies: make traceability easy before you make generation easy.

Claude Code teams ship when skills defer cleanly.

The thread underneath "Claude Code 2.1.139 team conventions" is not novelty—it is review bandwidth.

Failure Modes

CLAUDE.md bloat. If CLAUDE.md turns encyclopedic, newcomers skim and trust chat. The failure is not usually tool quality; it is the missing operating contract.

Use CLAUDE TOC. Keep top 15 lines as supremacy + links to deeper fragments.

What changes: Sessions boot with aligned assumptions. Claude Code teams ship when skills defer cleanly.

That is the move that makes review cheaper.

Permission autopilot. You can spot this when approvals become reflexive, risky bash slips through.

Why it keeps recurring: Humans optimize for flow—guardrails need friction by design.

The practical fix is Red-folder ritual: List paths requiring separate reviewer + never auto-approve.

The evidence is mundane and useful: sensitive edits regain deliberate eyes. Claude Code teams ship when skills defer cleanly.

Use it before adding another connector.

MCP data surprises. If MCP pulls customer data, you have seen shock during audit. The failure is not usually tool quality; it is the missing operating contract.

Use Data-class tags. Tag connectors with data classes + retention expectations.

What changes: Security reviews stop learning basics during incidents. Claude Code teams ship when skills defer cleanly.

This is the part reviewers can actually verify.

Session amnesia. If sessions rotate daily, teams lose shared memory.

The trap is Institutional knowledge cannot live only in transcripts.

Named fix: Weekly retro note. Append one .md changelog of decisions agents relied on.

After the fix, new teammates inherit reasoning instead of rumor. Claude Code teams ship when skills defer cleanly.

Use it before adding another connector.

# 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.

That trade routes cleanly through our methodology: treat Review as the gate where parallel agent output must be inspectable without replaying sessions.

Agents are relief crews; the blueprint still belongs to humans standing outside the trench.

Field Checklist

Review strip

• Verification command output is pasted or linked.
• Forked agent work lists parent + child responsibilities.
• Red-folder paths received explicit human acknowledgement.
• Scopes in the PR body match folders in the diff.

Boundary Note

If your repo cannot state boundaries plainly, agents will guess—and guessing scales poorly.

Source Stack

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

Next Move

Take the next step on Claude Code team conventions—walk the exercises once with receipts enforced, not once with vibes enforced.