Patch release: in-session feedback drafting (/feedback), skill proposals from observed workflows, and more proactive memory of corrections and preferences.
Changelog
[new] Claude can draft product feedback (bugs, ideas, or missing capabilities) into a local queue for you to review and send with /feedback. Drafts are never submitted without your approval, and drafting does not interrupt the conversation.
[new] Skill proposals: Claude can surface new or improved skill drafts (with a full SKILL.md body) based on procedures observed in memory, for you to accept or discard without blocking the current task.
[changed] Memory updates are more proactive. User corrections, durable preferences, and environment learnings are written before the turn finishes so later sessions do not re-learn the same lessons; memories are treated as point-in-time context to validate, not absolute truth.
[changed] The teammate/subagent mode field is deprecated and ignored. Subagents inherit the parent session’s permission mode; agent-definition frontmatter may still override it.
[changed] Background agent metadata now records the model at the backgrounding transition and an ordered modelsUsed list when a mid-run model switch occurred.
[new] Review findings accept an optional short_summary (≤60 chars) for compact UI labels, separate from the full claim text.
[new] Optional session limits via CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION and CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION.
Models & prompts
Stronger guidance on when to write memory: corrections (including skeptical pushback), stated preferences, and non-transient environment discoveries; explicit “remember” / “forget” requests are handled immediately.
Feedback drafting is rate-limited per session; once the cap is hit, further draft calls are skipped while already-queued drafts remain available in /feedback.
Focus mode wording is tightened so interim progress chatter is skipped and everything the user needs lands in the final message.
Long-running work continues to background with a completion notification; TaskStop remains the way to cancel.
Artifact force overwrite behavior is clarified: baseVersion is always sent, and force: true treats it as informational so a reconciled overwrite does not 409.
Under the hood
New beta header per-turn-control-2026-07-01 for finer per-turn control on the API path.
Background-task plumbing for respawn/rescue and MCP tools that can be auto-backgrounded mid-session (CLAUDE_BG and related gates).
Memory sync options around push/delete handling (CLAUDE_CODE_MEMORY_PUSH_DELETE_MODE).
Live mid-session model switch tracking, session-fork support, and an authenticated eval endpoint (/api/eval-authed).
Cleanup of older fast-mode / plan-interview / warm-resume internal flags and the team-usage discovery route.
Upgrade notes
If you previously set mode when spawning teammates/subagents, remove it; the value is ignored and the parent session’s permission mode applies instead.
Feedback drafts stay local until you explicitly send them from /feedback; Claude will not announce or ask about drafts mid-task.
content/
Browse captured content/
Same tree as claude-code/latest/content for this version. Open the full browser for sidebar navigation between docs and other files.
