Claude Code v2.1.269–v2.1.270: plugin evals, repository telemetry, and a git-permission hotfix

Claude Code v2.1.269 adds a real test loop for plugin authors and closes several permission and local-file handling gaps. The release also improves long-session responsiveness, prompt-cache continuity, terminal compatibility, and visibility into background work. v2.1.270, released the following day, fixes a v2.1.269 regression that could make read-only git commands unexpectedly request permission after a session had been running for a while.

What changes for users

• Plugins can ship behavioral evals. claude plugin eval runs realistic prompts against a plugin, scores the results with deterministic or model-judged graders, compares them with a no-plugin baseline, and writes JSON plus HTML reports. claude plugin eval init can scaffold the suite. Eval runs use your normal model credentials and count against plan limits or API billing.
• Telemetry can identify repositories. OTEL_METRICS_INCLUDE_REPOSITORY=true adds vcs.* repository attributes to metrics and events. This is opt-in because repository labels increase telemetry cardinality and may reveal organization or project names to the configured collector.
• Bash edits are easier to review. The v2.1.269 release notes say Bash-handled file edits can attach a diff of affected files to the tool result through bashEditDiffEnabled.
• Workflow fan-out gets a higher ceiling. Operators can now raise a Workflow run's concurrent-agent limit from 1 up to 256 for inference-bound fan-outs. The release notes document the configuration variable for teams that need it.

v2.1.270 hotfix

If you adopted v2.1.269, move to v2.1.270 before relying on long-running unattended sessions. Anthropic says v2.1.270 fixes a regression where read-only git commands executed through Bash could unexpectedly start asking for permission after the session had been running for a while. No configuration change is required; updating is the fix.

Documentation caveat: /output-style

The v2.1.269 changelog announces /output-style [name] for switching styles, including Remote Control and headless sessions. However, the current Output styles reference still says the standalone /output-style command was removed in v2.1.91 and directs users to /config or the outputStyle setting instead.

Because those two official sources currently conflict, /output-style is not included in the cheatsheet. Treat the changelog item as documentation drift until the reference documentation is updated or Anthropic clarifies the supported interface.

Security and reliability details

Two file-boundary fixes deserve attention for teams that depend on deny rules. A deny or ask rule beginning with ! is now limited to the settings source that defined it, and a bare negation is ignored. Write-path checks also cover files written through Bash tee; an allow rule for the tee command no longer authorizes destinations outside the working directories.

Plugin archives are extracted with tighter local permissions, without preserving world-writable bits, and stale extracted files no longer survive a re-extraction. The release also keeps organization-managed plugins available in headless and Desktop sessions and repairs synced MCP reconnection after remote-session resume.

Prompt-cache continuity and long-session behavior received another repair pass, including interrupted-turn resume, output-token auto-resume, cloud-session startup configuration, background-agent status, and transcript responsiveness.

What to do

Update first, then add evals only to plugins you trust:

claude update
claude --version
claude plugin eval --help

Treat an eval suite as a quality signal, not a security scan. Running an untrusted plugin still loads its code and components with your local identity. Administrators should also regression-test path-scoped deny rules that protect sensitive directories, especially if existing automation writes through shell utilities.

Official Claude Code v2.1.269 release notes · Official Claude Code v2.1.270 release notes · plugin eval documentation · monitoring documentation · output styles reference