Patch release focused on artifact assets (upload, list, read, delete media and related files on published artifacts), clearer artifact edit/save and watch behaviour, and smoother handling when usage limits clear.
Changelog
[new]Artifact assets — Artifacts support an asset store for local media, PDFs, and fonts. New actions: upload_asset (attach a file to an existing artifact), list_assets (enumerate with usage caps and pagination via after), read_asset (save an asset locally by id, optional out_dir), and delete_asset (permanent remove). Reads of your own artifacts need no extra approval; others’ ask once per artifact.
[changed]Document, slides, and spreadsheet artifacts — Edit existing artifacts against their current HTML; leave toolbar/rail/KIT: regions intact and don’t re-apply creation templates. Save is explicit (toolbar unsaved state, Save or Ctrl/⌘+S) and publishes a full new version including comments. Don’t embed how-to-edit instructions in the content — the surface is self-evident.
[changed]Artifact watching — Publish still arms a live watch when applicable; watch / status / unwatch cover explicit control. Watches are session-local (not across restart/--resume) and show under /tasks. Remote sessions use durable wake subscriptions for republishes (and eligible comments); continuous comment watching isn’t supported remotely — use action: "comments" when asked, or suggest claude --watch-artifact <url> on a local machine.
[new]Usage-limit auto-resume — When claude.ai usage becomes available again (early or at reset), the session can continue the interrupted task without redoing finished work. Rate-limit UI includes options to select or cancel auto-resume.
[changed]Focus mode — With focus mode on, only the final text of each response is shown; skip mid-turn progress chatter and put everything needed in the last message.
[changed]Session titles — Titling prefers a short noun phrase (about two to five words, sentence case) led by the most specific named thing (file, symbol, issue/PR, error), without request verbs or trailing “-ation” task labels; language follows the description, branch names stay English.
[changed]Subagents — Background-by-default guidance refined: keep work in the background unless the next step truly depends on the result; never invent pending-agent outcomes; remote isolation: "remote" stays always-background.
[changed]Corrections and memory — Stronger push to write durable memories in the same turn you engage a correction or preference, not after the conversation moves on.
[changed]PR workflow — After create/update PR, if CLAUDE.md mentions Slack, search for Slack send-message tools and only post the PR URL after explicit confirmation; failures stay silent. Push/notification handling recognizes reopened PRs.
[changed]Plugin eval calibration — Pilot runs keep --no-publish; aggregate checks now flag manifest_invalid, disabled_by_default, and will_not_load (and related load problems) so a with-arm that never loaded the plugin isn’t treated as a valid pilot.
[changed]Permission denials — Denied tool uses surface clearer rejection text, including optional user rationale when provided.
[fix]Plan / Cowork consent dead-ends — Asset upload/delete and auto-reply resume that need a consent surface no one can answer in-session fail closed with a clear message instead of retry loops.
[new]Async script dispatch marker — Tool output can report asyncDispatched when work was queued (or refused at the queue cap) so resume/replay doesn’t treat the receipt as execution output.
Models & prompts
Artifact guidance distinguishes creation-only templates from in-place edits, documents asset list/upload/read/delete rules, and sets expectations for watches vs. remote limitations.
Usage-limit recovery copy tells the agent to resume the prior task and avoid repeating completed work.
Focus mode, session naming, stash safety across worktrees/sessions, coordinator/subagent background rules, and correction→memory timing all got clearer model-facing instructions.
Auto-mode classifier review and high-effort review prompts remain; blocking classifiers still require a <block>-first response with no preamble.
Under the hood
New environment knobs for artifact assets, optional narration, goal check-in interval, eval artifact publish/stub/flag overrides, policy-helper PowerShell path settings, project dir naming, and Slack tag tokens — mostly for gated or internal workflows.
Feature gates for quota auto-resume arming/cancel/fire/stale paths, MCP listen/negotiation server denylists, account skills sync, device↔account bind, bridge owner/command lifecycle, and several experimental subsystems.
Agent-proxy routes for artifacts and frames; notification preferences path refreshed.
SDK tool typings extended for artifact asset actions/results, PR reopened, and asyncDispatched.
SDK compatibility matrix drops wrapper 0.3.195 from the tested list (still covers newer 0.3.196+ builds).
Upgrade notes
If you automate against the Artifact tool schema, allow the new upload_asset / list_assets / read_asset / delete_asset actions and their result shapes.
Artifact watches still do not survive process restart or --resume; re-arm with watch or publish if you need them again.
Remote sessions won’t stream artifact comments live — poll with comments or run --watch-artifact locally when that matters.
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.
