Claude Code 2.1.233

• [new] Settings support vimMode for vim-style editing in the CLI.
• [changed] WebFetch response caching is no longer hard-coded to a fixed 15-minute window; cache lifetime is controlled with CLAUDE_CODE_WEBFETCH_CACHE_TTL_MS.
• [changed] Plugin evaluation workflow (claude plugin eval / claude plugin eval init) is clearer for early-access users: pilot runs should use --no-publish, check evals/results/*/aggregate-result.json, and treat references/plugin-eval.md as the source of truth for flags and availability—never invent flags or report shape from memory.
• [changed] PR/comment thread handling: after acting on a thread, resolve it with the Artifact tool (resolve); leave it open only when the conversation is still active. Duplicate “sent to Claude” drafts are discarded unless a further reply adds something new (acknowledge_duplicate: true).
• [changed] Subagents default to background execution with completion notifications; use run_in_background: false only when the next step truly depends on the result. Never invent pending agent output.
• [changed] Watches that the user explicitly requested reconnect on their own without turning auto-replies back on; use the kill-all-agents gesture to disarm auto-replies for the session.
• [changed] Stronger correction → memory behaviour: when the user corrects preference or points out a real mistake, write durable memory in the same turn before finishing the reply.
• [changed] Focus mode guidance: only the final text of each response is shown—skip mid-turn progress chatter and put everything needed in the last message.
• [new] Todo-list tools can be gated on via CLAUDE_CODE_ENABLE_TODO_TOOLS.
• [new] Optional controls around permission UX: disable permission-prompt notify hooks (CLAUDE_CODE_DISABLE_PERMISSION_PROMPT_NOTIFY_HOOKS) and retire unanswered parked permissions (CLAUDE_CODE_RETIRE_UNANSWERED_PARKED_PERMISSION).

Models & prompts

• Built-in guidance now covers claude plugin eval and /skill-doctor as early-access surfaces (no public docs page yet; do not guess URLs). The claude-code-guide agent scope includes both.
• Plan mode, permission-denial, classifier “err on the side of blocking,” git-stash safety across worktrees/sessions, Chrome-extension decline, and Slack-on-PR flows are restated more precisely for the model.
• Computer-use MCP (mcp__computer-use__*) and tab-context recovery after closed tabs / navigation errors remain documented for desktop control sessions.
• Context-management copy clarifies that long conversations may be summarized mid-work and that the agent should keep going rather than wrapping up early.

Under the hood

• Tool process limits: CLAUDE_CODE_TOOL_MEMORY_LIMIT plus related cgroup feature gates for bounding tool memory use.
• MCP listen path can park/reopen connections (tengu_mcp_listen_reopen_park and tuning gate) for more resilient server lifecycles.
• Skill-proposal plumbing (CLAUDE_CODE_SKILL_PROPOSALS) and plugin ratings store configuration gates land for upcoming plugin UX.
• Eval/session hooks: CLAUDE_CODE_EVAL, CLAUDE_CODE_EVAL_INTERVIEW_SESSION, and related gates for unrecognized-model and checkpoint/auto-continue edge cases.
• Permission pre-park timing via CLAUDE_AX_PREPARK_MS.
• Internal cleanup of stale or unused env-var and route strings; notification-preferences and session/environment API routes adjusted.

Upgrade notes

• If you relied on WebFetch’s old fixed ~15-minute cache, set CLAUDE_CODE_WEBFETCH_CACHE_TTL_MS explicitly to the TTL you want.
• Enable vimMode in settings if you want vim keybindings.
• Organizations without plugin-eval early access may still see “currently in early access” from claude plugin eval; that is expected—do not treat the command as missing.
• /skill-doctor remains early-access-only in sessions where it is not enabled; describe it if asked, but do not instruct users to run it when unavailable.