Claude Code 2.1.221

• [new] Added --autocompact to control automatic context compaction from the CLI.
• [changed] Memory saving is stricter and more deliberate: only durable, applicable, legible lessons are kept; you can pin up to four memories that apply to every future session; saves must happen in the same turn as the correction, before the reply finishes.
• [changed] Focus mode behaviour is clearer — only the final text of each response is shown, so intermediate tool chatter and progress lines are skipped and everything the user needs goes in that final message.
• [changed] Subagent guidance tightened: do not spawn agents unless the user asks or names an agent type; multi-part or “thorough” tasks are handled inline by default. Background runs remain the default where available, with optional synchronous hand-back.
• [new] Artifact comment threads: reply-only and edit-capable composers so Claude can participate in artifact discussion (and, when enabled, update the artifact) without inventing actions it did not take.
• [changed] Browser automation cleanup: tabs Claude opens are closed when finished, and tab context is refreshed after closes or navigation errors.
• [changed] Plan mode blocks page-data reads and publish attempts that need a consent surface when no one in the session can answer — planning stays in the plan file and open choices go to chat.
• [changed] Auto-mode / permission evaluation covers subagent hand-backs as real actions, expands adversarial-pattern clearing rules, and treats agent-inferred or relayed instructions as non-user intent for must-name bars.
• [changed] PR flow still offers optional Slack posting when CLAUDE.md mentions it; plugin and skill listing tools can filter by keyword and distinguish “what you have” from suggest/search for new ones.
• [changed] Prompt-audit path (/claude-api prompt-audit and related guidance) is much more complete: scoped inventory, confidence rubric, concrete remove/rewrite/move/add proposals, and migration notes for 4.5 → 4.6 prompt habits.

Models & prompts

• Built-in prompts push harder on not gold-plating, not fabricating pending subagent results, and treating memories as snapshots to verify—not source of truth.
• Computer-use MCP and coordinator/worker copy refreshed; git stash remains called out as shared across worktrees/sessions (prefer WIP commits or tagged stash apply/drop).
• New API beta header pre-2026-07-28 is available for eligible requests.

Under the hood

• MCP: protocol negotiation paths (stdio/HTTP/claude.ai), discovery cache/source, listen reopen, and skills funnel wiring; related gates for refresh intervals and malformed-imark handling.
• Artifacts stack: token, asset/live base URLs, comments, and auto-react controls via env (CLAUDE_CODE_ARTIFACTS_* / CLAUDE_CODE_ARTIFACT_*).
• Session/control knobs: child sessions, fork subagent, force bridge / mid-conversation system, disable mouse, SDK file checkpointing, token-usage attachment, OAuth scopes, feedback send, thrifty-sonic, workflow size warning tokens, and assorted host/AWS timeout overrides.
• Org-memory write opt-in / outcome / discard tracking; plugin install auto-activate; large cleanup of retired dead-probe and auto-mode classifier queue flags.
• Backend routes for frame comments and pull helpers; SDK compat list extended through newer 0.3.x clients.
• Binary size up modestly across platforms with the above surface area.

Upgrade notes

• If you rely on custom automation around context size, try --autocompact where you previously depended only on default compaction timing.
• Memory pins are capped at four session-wide entries — review existing saved memories if you depend on aggressive auto-saving of preferences.
• Optional env vars above are off/default unless set; no action required for typical upgrades.
• SDK users: newer @anthropic-ai/claude-code patch clients in the 0.3.210+ range are marked compatible in this build.