Claude Code v2.1.269 release on GitHub showing the plugin eval and output style changes Image: Anthropic / github.com/anthropics/claude-code
by Michael Joiner

Claude Code Adds Plugin Eval Command and Output Style Switching in v2.1.269

Claude Code v2.1.269 ships a plugin evaluation framework that scores plugins against reproducible test suites, plus an /output-style command, Bash diff output, and a round of terminal and permission fixes.

Share

Anthropic released Claude Code v2.1.269 on September 11, 2026, with the headline addition being claude plugin eval — a command that runs a plugin’s test suite against Claude Code and produces scored, reproducible results.

Version 2.1.270 followed on September 12 with a targeted fix for a Bash regression introduced by 2.1.269.

Plugin Eval

The new eval command is designed to give plugin authors a consistent, verifiable way to measure what their plugin actually does.

claude plugin eval runs a plugin against a set of test cases, scores the results using configurable graders, and compares the output against a baseline run without the plugin loaded. The delta shows the plugin’s actual contribution rather than conflating it with general model capability.

claude plugin eval init writes the eval suite for you: Claude reads the plugin, proposes test cases and graders, pilots them against the plugin, and saves the files under evals/. You can then run the suite manually or wire it into CI as a regression gate.

Results come back as JSON and HTML reports. Six grader types are supported, though specific grader names are not documented in the public release notes.

Output Style Switching

The /output-style [name] slash command lets you list and switch between output presentation modes from inside a session. It works across session types, including Remote Control and cloud environments.

Other Changes in v2.1.269

The release also included:

  • Bash diff display: Bash tool results now show a diff when the tool modifies a file, making it easier to track what changed without opening the file separately.
  • Repository telemetry: A new OTEL_METRICS_INCLUDE_REPOSITORY environment variable tags OTEL metrics with the current repository, useful for teams monitoring multiple projects from one dashboard.
  • Prompt cache fix: Interrupted responses and session resumptions now handle prompt cache state more correctly.
  • Permission rule scoping: Rules starting with ! now apply only within their source, fixing a case where negation rules were unintentionally leaking scope.
  • Terminal fixes: Compatibility issues in kitty-protocol, st, rxvt-unicode, and WezTerm terminals were resolved.
  • MCP reconnection: Servers no longer drop reconnection when query parameter order changes.

The v2.1.270 Regression Fix

The update from September 12 fixed one issue introduced in 2.1.269: read-only git commands (things like git status and git log) were unexpectedly prompting for permission after a session had been running for a while. v2.1.270 resolves it.

Earlier in the week, v2.1.268 (September 11) added gateway pricing parity for managed settings, a --remove-session-state flag for self-hosted runners, browser-tab icons for published artifacts, and fixed the WebFetch tool hanging indefinitely by adding a 300-second default timeout.

Sources: v2.1.269 release notes · anthropics/claude-code | v2.1.268 release notes · anthropics/claude-code

Share