Claude Code adds plugin eval — scored, reproducible test suites
Same release raises the Workflow agent cap to 256, adds live /output-style switching, and shows Bash edits as diffs; Codex adds a max-effort mode.
Copy markdown
plugin eval scores your plugins for real
`claude plugin eval` runs a plugin's eval suite against Claude Code and emits scored, reproducible results as JSON plus an HTML report. If you ship or maintain a plugin, that's a regression check you can wire into CI instead of eyeballing every change.
Workflows can fan out to 256 agents
A new env var, `CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS` (1–256), raises the Workflow tool's per-run concurrent-agent cap. Fan-out jobs like reviewing every changed file or sweeping a monorepo finish in fewer waves.
Switch output styles mid-session
`/output-style [name]` lists and switches output styles live — including over Remote Control and in cloud sessions. Flip from terse to explanatory without editing settings and relaunching.
Bash edits now show as diffs, plus a git-permission hotfix
With `bashEditDiffEnabled`, a Bash command's result includes a diff of the files it touched, so agent edits are visible inline. The 2.1.270 patch a day later also stops read-only git commands from re-asking for permission mid-session.
Elsewhere: Codex adds max-effort, drops an entry point
OpenAI's Codex 0.154.0 adds `max` and `ultra` reasoning-effort levels for harder tasks and deprecates the `mcp-server` entry point — migrate any scripts that call `codex mcp-server` before it's gone.