Claude Code 2.1.122 team conventions

The situation

Claude Code 2.1.122 is a maintenance release, but a few fixes matter for teams running it in shared repositories. The changelog covers connector visibility, session recovery, structured outputs, telemetry, and workflow edges that show up once Claude is part of day-to-day repo work.

For teams already using CLAUDE.md, scoped rules, MCP connectors, hooks, and review checklists, the question is practical: what got easier to trust, and what still needs guardrails?

The main theme is less ambiguity. This release fixes hidden duplicate connectors, incomplete session recovery, some cloud-provider output issues, and status redraw noise that could interfere with terminal control pipes. None of that is flashy, but it affects whether a team can standardize on one workflow.

If you are mapping this to repo conventions, see Claude Code team conventions.

What changed

1. Shared context stays the same, but the release makes the case for keeping it tight. Claude Code still treats CLAUDE.md as persistent project memory, and the docs still recommend using it for durable rules instead of repeating instructions in chat. Keep it concise and scoped.

2. /mcp now shows claude.ai connectors that are hidden by a manually added server with the same URL, and it adds a hint to remove the duplicate. That is a small UI fix with a clear governance point: avoid shadow connectors and keep one source of truth per integration endpoint.

3. The /mcp message is clearer when a server is still unauthorized after browser sign-in. Treat connector authorization as a review step, not a one-time setup step. Installed does not mean ready.

4. Pasting a PR URL into /resume now finds the session that created that PR across GitHub, GitHub Enterprise, GitLab, and Bitbucket. That improves traceability when reviewers need to jump from a pull request back to the agent session behind it.

5. The release adds ANTHROPIC_BEDROCK_SERVICE_TIER, which selects a Bedrock service tier and sends it as X-Amzn-Bedrock-Service-Tier. It also fixes /model not showing the Effort option for Bedrock application inference profile ARNs, and fixes Vertex AI and Bedrock structured-output failures in some queries. If your team uses cloud inference, verify which tier and provider settings are allowed in repo docs.

6. OpenTelemetry numeric attributes on api_request and api_error are now emitted as numbers, not strings, and a new claude_code.at_mention event was added for @-mention resolution. If your team uses logs for analysis, check dashboards and parsers for type assumptions.

7. Other fixes point to workflow boundaries: /branch on rewound timelines, !exit and !quit in bash mode, image resizing limits, ToolSearch missing late MCP tools in nonblocking mode, and remote-control idle redraws that could flood tmux -CC. These are reminders to keep agent workflows bounded when the CLI is embedded in automation.

A small repo-level fragment can hold most of this:

# CLAUDE.md

## Team conventions
- Use the shared MCP allowlist only; do not add duplicate servers with the same URL.
- Review connector authorization after browser sign-in before relying on a tool.
- Keep cloud provider settings in repo docs and verify Bedrock/Vertex behavior after upgrades.
- Link PRs back to Claude sessions when a review needs provenance.
- Keep durable rules here; put task-specific prompts in the task itself.

A short review checklist keeps the release actionable:

## Claude-authored PR review checklist
- Is the connector duplicated or still unauthorized?
- Does `CLAUDE.md` still hold the right durable rule?
- If Bedrock or Vertex is used, did we verify tier, effort, and structured-output behavior?
- Do logs or dashboards expect numeric OpenTelemetry fields as strings?
- Can a reviewer recover the originating session from the PR URL?

This is a review step, not a redesign. Use the release to verify the conventions you already have, then update the smallest artifact that makes them durable. That is the kind of change we keep aligned with our methodology.

Tradeoffs and limits

The release does not replace local policy. MCP visibility fixes help, but they do not decide which connectors your team should trust. Better /resume matching improves traceability, but it does not replace a review process that records why a session was approved.

The cloud-provider fixes are conditional. If your team is not using Bedrock or Vertex AI, they may not matter. If you are, they deserve a targeted smoke test because structured-output behavior is where small regressions can break automation.

Telemetry changes only help if your observability stack is ready for them. A field that changes from string to number can break filters, alerts, or exports. Treat it as a schema compatibility check.

The changelog is the source of truth for what changed, not a guarantee that every edge case is gone. Keep CLAUDE.md short, scope rules locally, and review connector permissions as part of normal repo hygiene.

Further reading

• https://docs.anthropic.com/en/docs/claude-code/changelog
• https://docs.anthropic.com/en/docs/claude-code/overview
• https://docs.anthropic.com/en/docs/claude-code/memory
• https://docs.anthropic.com/en/docs/claude-code/mcp
• https://docs.anthropic.com/en/docs/claude-code/hooks
• https://docs.anthropic.com/en/docs/claude-code/skills