Patch release: clearer Artifacts workflows (listing, comments, watch/resolve), tighter subagent and focus-mode behaviour, and Managed Agents session-tool updates.
Changelog
[changed] Artifacts guidance is more complete end-to-end: list earlier published artifacts when the URL is unknown, redeploy in-session by path, and point users at /artifacts, the default ctrl+] reopen shortcut, and the web gallery for owned/shared work.
[changed] Artifact comment threads can be marked resolved after you act on them (action: "resolve" with url and thread_id); keep existing block ids on edit so pinned threads stay attached, and do not invent ids unless you need an in-page link target.
[changed] Sessions that cannot read or reply to artifact comments say so plainly instead of offering to watch for them; failed Artifact calls should be fixed or stopped rather than retried unchanged.
[changed] Artifact watch / idle / interrupt behaviour is spelled out more carefully (re-arm after Stop/Ctrl+C, hold of inbound peer traffic, when idle notices will or will not arrive) so the agent does not poll or claim a watch it does not have.
[changed] Subagents are for explicit user requests only — multi-part or “thorough” tasks are handled inline; spawning is called out as the expensive path when it re-derives context you already have.
[changed] Focus mode: only the final text of each turn is user-visible; skip mid-turn progress chatter and put everything needed in that final message.
[changed] On corrections or stated preferences, memory writes happen in the same turn you engage the feedback — before treating the turn as finished.
[changed] Managed Agents sessions.update() can replace agent.tools and agent.mcp_servers on an idle session, including per-tool web settings such as allowed_domains / blocked_domains (session-local; full array replace; vault_ids remains create-only).
[changed] Clarified that networking / egress limits do not apply to web_search and web_fetch (those run on Anthropic’s servers); restrict them via the tool’s own domain config instead.
[changed] Permission-denied and plan-mode / consent edges for Artifacts (including plan-mode and Cowork sessions with no one to answer a consent surface) fail closed with a clear user-facing explanation instead of silent retries.
[changed] Asset tool typings document out_dir (read destination) and asset_id (read/delete) more clearly for read_asset / delete_asset.
Models & prompts
Stronger default against unsolicited subagent spawns; background vs synchronous subagent notes unchanged in spirit, with remote isolation: "remote" still always background where available.
Focus-mode, correction/memory timing, and “follow-up ≠ you were wrong” guidance tightened so the model answers what was asked and only rewrites when the user points at a real error.
Artifact publish/list/watch/comment/resolve instructions expanded; Chrome-extension decline still suppresses re-upsell for the session (/chrome to revisit).
Managed Agents / self-hosted notes restated (autonomous session cap, no memory stores on self-hosted, IAM/SigV4 or console API key + managed policy) and session tool/MCP update semantics aligned with domain allow/block lists.
Auto-mode classifier stage still errs on the side of blocking with a strict <block>-first response shape; read-only work remains outside that path.
Under the hood
ANTHROPIC_DEFAULT_MODEL and several internal CLAUDE_CODE_* switches (including silent-turn reminder knobs, artifact type hints, TUI trial, and pacing-related flags) landed for configuration and experiments.
Request-path hardening around gzip body fallback/latch behaviour, tool-use interrupt accounting, MCP input/protocol negotiation, plugin sync refusal handling, and UDS startup bind.
Frame-oriented API surface expanded (favorite, retitle, types, watch-token) in place of older agent-proxy artifact/frame routes.
PyPI is reachable as a network endpoint (e.g. package/index checks).
Hook callback payloads can carry an optional sender issued_at (epoch ms) so device clients can drop stale or replayed deliveries.
SDK compatibility matrix no longer lists wrapper 0.3.197 among tested versions (0.3.198+ retained).
Upgrade notes
If you automate Artifacts or comment workflows, prefer list → update by URL, preserve block ids on edit, and use resolve only for threads you actually handled.
Managed Agents callers that patch tools mid-session should GET, modify, and POST full agent.tools / agent.mcp_servers arrays on an idle session; domain list changes apply for the rest of that session only.
SDK users reading assets should set out_dir / asset_id per the updated field docs; optional artifact_id on publish-related payloads is additive.
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.
