Patch release: clearer handling when device tools can’t run because another cloud session owns the machine, tighter artifact and Cowork guidance, plus agent/UX polish (focus mode, subagents, structured review findings).
Changelog
[new] When device tools are refused because this machine is already bound to a different Claude Code cloud session, the agent explains the conflict, keeps working with in-environment tools, and avoids retry loops until the other session ends.
[changed] Document and spreadsheet artifact creation vs editing is stricter: templates apply on create only; existing artifacts are edited in place; toolbar/KIT: regions and server-owned data-id attributes are left alone; the UI is treated as self-evident (no how-to-edit instructions baked into the content).
[changed] Artifact titles should be short, distinctive noun phrases—not generic labels, summaries, or names with explainers tacked on.
[changed] Cowork sessions that need an approval card (database writes, comment replies, publishing, resolving threads) are instructed not to retry those actions in-session when no one can answer the card.
[changed] Focus mode behavior is spelled out more clearly: only the final text of each response is shown—no mid-turn progress chatter assuming the user saw tool calls or interim updates.
[changed] Subagent defaults: background by default when the user can usefully interject; run_in_background: false only when the next step truly depends on the result; never invent pending-agent outcomes.
[changed] Safer multi-session git stash guidance: the stash stack is shared across worktrees/sessions—prefer WIP commits; if stashing, use a uniquely tagged push, apply by SHA, and drop by retargeted ref (no bare stash/stash pop).
[changed] Code-review results can be reported as a typed findings list for the host UI (severity-ordered, optional fix outcome), instead of duplicating them as plain text when that format is required.
[changed] Correction handling pushes durable memory writes in the same turn the correction is engaged—not deferred until later.
[changed] After opening or updating a PR, if CLAUDE.md mentions Slack, the agent may offer to post the PR URL via a discovered Slack tool—only after user confirmation; missing Slack tools fail silently.
[changed] Declining the Claude in Chrome extension suppresses further install prompts for the rest of the session; read-only web tools or manual browser steps remain the path forward (/chrome to revisit).
[new]alt+down keybinding support.
[changed] SDK usage payloads can include output_tokens_details.thinking_tokens for more detailed token accounting.
Models & prompts
Stronger “plausible by default” and corrections guidance: don’t treat ordinary follow-ups as proof of earlier failure; on real mistakes, correct plainly; if the user reaffirms after a concern, proceed without moralizing—refusals stay limited to genuinely harmful or prohibited requests.
Permission-denied tool results distinguish generic rejection from denial with the user’s reason, and push an alternate approach rather than stalling.
Context-management copy clarifies that long conversations may be summarized mid-work and that the agent should continue rather than wind down early; separate summarizers cover full-session vs recent-portion compaction.
Computer-use MCP (mcp__computer-use__*) instructions cover desktop screenshots and input control, including recovering tab context after closes or navigation errors.
Task-list updates remain available via the dedicated update-task tool; coordinator-style multi-worker orchestration prompts are refreshed in place.
Under the hood
New API beta surface area for agent memory (agent-memory-2026-07-22), dreaming (dreaming-2026-04-21 / /v1/dreams), and MCP tunnels (mcp-tunnels-2026-06-22 / /v1/tunnels).
Early directory sync (pull/push), plan/frame artifact plumbing, and optional plugin sync/buffered download controls.
New configuration knobs for deployments and power users, including Bedrock/Foundry base URLs, small-fast and subagent model overrides, effort level, gateway model discovery, max web searches per session, tmux session/truecolor, scroll speed, Perforce mode, cron disable, bash working-directory retention, tasks enablement, and related session/metadata flags.
If device tools suddenly refuse on a machine that also runs Claude Code cloud sessions, check for another active session holding the device binding before retrying.
Cowork flows that depend on approval cards will not keep retrying blocked publish/write/comment actions in the same session—complete approvals from a context that can show the card, or continue with other work.
Custom keybinding users: alt+down is now part of the default set; move or unbind explicitly if it conflicts with your map.
SDK consumers that inspect usage may see output_tokens_details.thinking_tokens populated when available.
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.
