Patch release: model-proposed goals, richer remote-trigger run inspection, stronger background-subagent defaults, and broader GitHub MCP / device-side command handling.
Changelog
[new] Model-proposed goals: Claude can propose a verifiable completion condition for you to accept or decline (or set directly when you already stated the outcome). Manage goals with /goal; disable proposals in settings if you prefer to set goals yourself.
[new] Remote triggers can list runs and fetch a run’s logs (with pagination), so you can inspect trigger execution history from the session.
[changed] Subagents default to running in the background more consistently. Synchronous runs (run_in_background: false) are reserved for when the very next step truly depends on the result; otherwise work stays backgrounded so you can interject.
[new] Large expansion of GitHub MCP tools — issues, PRs and reviews, Actions/workflows, checks and alerts, gists, discussions, notifications, and related read/write operations when the GitHub MCP server is connected.
[changed] Artifact publishing: updating an existing artifact is clearer — pass its URL (via list or by asking you) so a publish updates in place instead of minting a new link; HTML title detection looks at the start of the file.
[changed] Artifact comment threads: after acting on a thread, Claude posts a short reply and can resolve the thread; clearer handling when only the thread starter or artifact writer may resolve.
[new] Optional ultrareview / code-review flow that posts findings to the pull request as a comment from your account (with explicit no-post paths when posting isn’t requested or the review fails).
[changed] Memory guidance across project switches: memories from a prior project are treated as stale until re-checked; personal memory stays separate from project memory.
[changed] Device-side bash feedback: clearer outcomes when a command is refused (sandbox not fully confining), cancelled, times out, or hits an output size limit — including guidance not to blindly retry non-idempotent commands.
[changed] Focus mode instructions reinforced so only the final message in each turn is user-visible (no mid-turn progress chatter).
[changed] Background shell tasks can be marked as ending with a synchronous subagent’s final response, so owned background work is torn down cleanly when that agent finishes.
[changed] Commit metadata in tool results can include the branch name alongside the commit kind.
Models & prompts
Stronger default to background subagents, with explicit “don’t invent pending agent results” and notification-based completion.
New goal-proposal behaviour and prompts, including respect for a user setting that disables model-proposed goals.
Richer artifact update and comment-responder instructions (update-in-place URLs, resolve-after-reply, permission limits on who can resolve).
Code-review / ultrareview paths for reporting findings as structured UI data and optionally posting a single PR comment without extra local apply or review actions unless requested.
Corrections and memory: save durable preferences/corrections in the same turn they’re engaged; avoid duplicating what the repo already records.
Context-management and focus-mode copy tightened so long sessions summarize cleanly and focus mode doesn’t leak tool-call narration.
PR workflow still offers an optional Slack notify step when CLAUDE.md mentions it and a Slack send tool is available — only after you confirm.
Under the hood
ProposeGoal tool surface wired through the SDK (ProposeGoalInput / ProposeGoalOutput) with ask_user approval vs direct-set paths.
Remote trigger actions extended with list_runs and get_run_log (session_id, cursor).
Env support for memory API base URL/token and bridge reattach ownership / no-backfill controls; artifact comment-responder flag.
If you rely on synchronous subagents, set run_in_background: false only when the next step blocks on that result — background is now the encouraged default more often.
To turn off model-proposed goals, use the goals setting (or clear with /goal); Claude will not propose when that setting is disabled.
Updating a shared or earlier-session artifact needs its URL on publish; without it you get a new artifact rather than an in-place update.
Device bash that needs a full sandbox may be refused until sandbox filesystem confinement is enabled in settings on that device.
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.
