Claude Code maintenance

Claude Code 2.1.321: OAuth and permission security fixes

For an engineering team, dependable permission boundaries are part of getting work done. This guide translates the official release notes into Werkverstand's maintenance proposal. It does not repeat existing model-launch coverage or claim complete security. Sources and version information were checked on November 11, 2026; the version actually present in the organization's systems remains decisive.

LOL.MARKETING / CONNECTING INTELLIGENCE

The essential answer

Claude Code 2.1.321 includes security fixes that justify a controlled update and acceptance step. Review the actual installed version, approved workflows and existing permission boundaries. Successful installation alone does not establish that the organization's working process remains functional.

Treat the security update as a distinct event

The official 2.1.321 notes list fixes involving MCP OAuth redirects and sandbox and file-permission checks. Version 2.1.322 adds Vertex fast mode. The security fixes are a maintenance event in their own right; existing model-launch coverage does not need to be repeated. Check which supported version the team actually uses and how updates are managed. A local version display does not show that every workstation or remote runtime has the same release. Begin with a short inventory covering ownership, installation method and the main approved workflows, without collecting credentials or confidential session content. Keep actual installed versions separate from the version merely available for installation.

Inventory workstations, runtimes and policies together

A development environment includes more than the visible terminal. Acceptance also concerns editor integration, connectors, project settings and organization policy. Collect only information needed for the update plan, paying particular attention to workflows with file writes and external tools. Avoid introducing several unrelated changes at once, because failures then become harder to attribute. A representative workstation is a better first test than replacing everything across the organization immediately. Record who decides when an unexpected interruption occurs and which work can continue manually or through an existing approved route. This keeps maintenance manageable without treating every difference between machines as a reason to broaden access.

Test normal work and refused actions

A good acceptance set includes approved routine tasks and deliberately prohibited actions in an isolated environment. For example, a sample file may be editable in an approved project area while an excluded area must remain untouched. Use harmless files. The aim is to verify existing rules, not circumvent safeguards. Inspect the actual file state and the decision displayed for each case. A correctly refused action is a passing result. If a previously necessary approval no longer appears, assess that change professionally rather than automatically treating it as a convenience improvement. Keep expected permissions explicit so the reviewer can distinguish a product change from a misunderstanding of the test.

Review connector sign-ins without collecting secrets

The release notes are a reason to check affected connector workflows through existing safe procedures. Do not put tokens, passwords or complete sign-in responses into a test log. Record whether the connection works through the intended service and whether an unclear redirect is handled appropriately. Connecting additional services or creating permanent access is not an incidental part of updating software. If a check requires further consent, keep that item visibly unresolved. The responsible administrator decides on subsequent measures. Avoid precautionary account switching or wholesale credential resets without a concrete finding and an authorized reason. Record operational success without turning the verification exercise into a new collection of sensitive information.

Coordinate the rollout with active work

Do not let an update silently interrupt active assignments. Choose an appropriate time for the representative workstation and record the tested state. After acceptance, the owner can plan the wider rollout. Keep a supported recovery or transition route available and understand its limits beforehand. A fallback should not simply mean permanently omitting a known security fix. If an important function fails after updating, the team needs a bounded temporary process and a clear issue description. Include the version, reproducible steps and observed result, while excluding confidential project files and secret access information. State which work is paused and which approved activity can safely continue.

Record the final state and remaining limits

Completing the update means more than a successful installation. Record which version was checked on each relevant runtime, which cases passed and which questions remain open. Rerun affected tests against the final state if settings change after acceptance. A vendor fix does not justify a promise of complete security. Operationally, the question is whether approved workflows function within their intended limits and owners can identify failures. Preserve the small acceptance set for future updates. This turns a long release-note list into an accountable maintenance step whose outcome the team can assess without extra assumptions. Distinguish a completed check from one that was planned but could not run.

Keep it verifiable

Primary sources

The next sensible step

Which AI system fits your business?

Eight steps from a general interest in AI to a clearer decision for your business.

Start AI System Check
FreeProvider-neutralNo credentials