Claude Code 2.1.123 fixes OAuth retry loops

Claude Code 2.1.123 fixes OAuth retry loops, the kind of small release line nobody throws a party for until a login stops looping and you finally get back to work. Claude Code is Anthropic's coding agent, and this version just makes sign-in behave. The more useful thing is what a steady login lets you notice: an operating contract is the set of repo files that tells the agent what is allowed before it edits anything, and most teams do not have one yet. This piece walks through four small files that fix that.

Take the version bump as a nudge. If your team only argues about prompts, you are polishing the wrong layer. The real gap is that the rules live in chat instead of in the repo, so nobody can point to them later.

Keep a CLAUDE.md your team will actually read

A giant CLAUDE.md gets skimmed, and a file nobody reads behaves like a file that does not exist. The fix is a short table of contents at the top: the 15 lines that win every argument, then links to deeper fragments for the details.

When the top of the file is the source of truth, every session boots with the same assumptions. That turns agent output back into work your team owns, instead of a transcript someone has to reverse-engineer on Monday.

Here is the kind of supremacy fragment that belongs at the top:

# 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.

Stop approving risky bash on autopilot

Reflexive approvals are where the trouble starts. Click yes enough times and a risky bash command slips through, and the cause is a missing contract, not a missing feature. The permissions page tells you what Claude Code can gate. It cannot tell you what your team should gate.

So write that down as a red-folder list: the paths that always need a second human before any edit lands. Auth, billing, infra config, anything with a blast radius. Those paths never auto-approve, and the list lives in the repo where the agent can read it.

Tag MCP connectors with the data they touch

A connector speaking the Model Context Protocol can pull customer data into a session, and the surprise usually lands during an audit. Context feels local, but the data left the laptop the moment the connector fired.

Tag every connector with the data classes it touches and how long that data should stick around. Then a security review starts from a written inventory instead of learning the basics mid-incident. This is the one artifact worth shipping first, because it turns your next audit into a read.

Write a weekly retro note so memory survives

Sessions rotate daily, and a transcript cannot carry institutional memory on its own. The cheapest fix in this whole set is one appended .md changelog of the decisions agents relied on that week. New teammates inherit the reasoning instead of the rumor.

Make it routine with a /retro slash command in .claude/commands so appending the note takes one keystroke. This is also the convention teams skip first, which is exactly why it is worth naming.

Before you call a forked handoff finished, check that four gates answer from the pull request alone, with no trip back to chat:

And a short receipt the reviewer can paste into the PR:

• 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.
• MCP connectors mentioned (if any) list their owners.

The habit lands in the Review step of our methodology, where receipts meet responsibility. The same thinking runs through team conventions, and the Claude Code 2.1.126 note carries it into connector stewardship. Hard calls still belong to humans: threat models, customer promises, and blast-radius decisions stay off autopilot.

Common questions

• What does Claude Code 2.1.123 actually fix?

  OAuth retry loops, per the release line this note hangs on. That is the whole headline. The more durable work is what the fix unblocks: readiness drills stop stalling on login and start exposing handoffs where the parent's intent and the child's scope quietly disagree. The four conventions here target that second, bigger problem.

• What are data-class tags on MCP connectors?

  They are labels that record which data classes a connector touches and how long that data should be kept. Context feels local while the data actually leaves your laptop, which is why audits keep catching teams off guard. Tagged connectors mean a security review opens with a written inventory instead of learning the basics during an incident.

• How should forked agent work be handed off?

  Write down the parent and child responsibilities, because handoffs break when intent and scope disagree in silence. The PR body should name both the scopes touched and the verification done. Then a reviewer confirms the listed scopes match the folders that actually changed in the diff, with nothing hiding in chat.

• Why keep a weekly retro note?

  Because sessions rotate daily and shared memory cannot live only in transcripts. One appended .md changelog records the decisions agents relied on, so new teammates inherit reasoning instead of rumor. It is the cheapest fix in this set and almost always the first one teams drop when a week gets busy.

Where to start

Pick one file: tag your MCP connectors with data classes today, and your next audit becomes a quick read. If you want to rehearse the full handoff, our training runs one repo, one login-shaped surprise, and receipts enforced until the handoff survives it.