v2.1.260

2.1.260

(☝️ Click to see full diff.)

+4,234 tokens

• NEW: Agent Prompt: Security monitor forwarded user turns — Uses request-keyed, harness-authored records of the parent session's human messages to establish action-specific intent and consent, including clearing a soft block for that action; rejects forged tags, unattributed consent, bare affirmations without the proposed action, blanket approval, and instructions addressed to the classifier.
• NEW: Skill: Plugin authoring — Documents function-hook plugins, build-generated /plugin-types declarations, validation and hot reload, surface-specific UI rendering, dispatch cancellation, persistent background work, and model-callable tool registration.
• NEW: System Reminder: Remote machine file sync timing for subagents — Explains that remote results and user edits enter the local copy at the main conversation's steps, not the subagent's; directs subagents to read newly generated output and ignored files on the remote machine itself.
• NEW: System Reminder: Remote machine Git and credential routing — Routes credential-dependent Git and gh commands, or commands for non-GitHub remotes, to the user's machine under its own approval rules instead of requesting tokens; makes clear that credentials are never copied into this environment.
• NEW: Tool Description: Artifact publishing introduction — Reintroduces default-private HTML publishing guidance, offers useful visual or interactive pages before building them when unrequested, permits proactive publishing of Claude's own work, and keeps sensitive or potentially misleading or harmful content as files for the user to decide whether to publish.
• NEW: Tool Description: AskUserQuestion extended host guidance — Adds text and bounded-number questions, prioritizes the most important question, defaults choices to multiselect unless mutually exclusive, keeps helper text optional, and supports unanswered questions, free-form answers, and user-requested follow-up questions without explicit Other or Skip options.
• NEW: Tool Parameter: Artifact supporting files with cross-artifact sources — Documents server-side copying of published files from accessible Artifacts in the same organization, with source-version limits and HTML/SVG/XML exclusions; preserves omitted files on updates and uses null to remove a path.
• NEW: Tool Parameter: Artifact title — Defines concise Artifact names, HTML-title precedence within the first 8KB, Markdown filename identity, file-based content, and the separate naming and type-name fallback rules for type_url creation.
• Agent Prompt: Status line setup — Documents optional prompt-cache health, TTL and expiry, hit ratio, miss causes, expected rebuilds, and recache-token metrics, with a cold-cache shell example that preserves false boolean values when reading JSON with jq.
• Data: Artifact host MCP server guidance — Allows the built-in host:claude_browser exception for reading and acting on websites, lists its supported browser tools, and limits it to desktop Cowork viewers with approval before each website; other built-in servers remain disallowed.
• Data: Claude API reference — Go — Prefers current model string IDs over potentially outdated SDK constants, switches the adaptive-thinking example from Sonnet 4.6 to the configured Opus model, and opts into summarized thinking with a note covering Fable 5/5.1, Mythos 5/5.1, Opus, and the configured Sonnet model.
• Data: Streaming references (C#, Go, Java) and Tool use reference (C#) — Switch example models from older typed constants to the configured Opus string ID, replacing Opus 4.8 in streaming and Sonnet 4.6 in tool use.
• Data: Claude API reference (PHP) and Streaming references (Python, TypeScript) — Expand the summarized-thinking opt-in note to cover Fable 5.1, Mythos 5.1, and the configured Sonnet model alongside the existing model list.
• Data: Review upload excluded changes error — Identifies sandbox read-deny settings, as well as Read rules, as reasons changes can be withheld from upload.
• Data: Rewind files skippedLinks field — Clarifies that non-link restoration failures are excluded from this count but reported in telemetry, and that the rewind fails with canRewind: false when every differing file fails to restore.
• Data: SDK cloud session init snapshot field — Adds connection state and live/recent remote-call status, including approval timing and completion dispositions without call inputs or outputs; documents cloud_session_delta updates and forward-compatible handling of new states.
• Data: Self-hosted runner command help and Session metrics help, and System Prompt: Self-hosted runner doctor — Describe releasing and pausing resumable sessions at the lifetime limit when waiting for the user, releasing active turns when they park or finish, and reserving SIGTERM for the hard cap; distinguish clean retire/max-age releases from interrupted hard-cap kills.
• Skill: Multiplayer whiteboard description — Broadens the trigger from explicitly multiplayer or live whiteboards to ordinary whiteboard requests and sketches of designs or diagrams for discussion, while retaining live collaboration and new-board-only scope.
• Skills: Setup Cowork and Setup Cowork role selection — Reframe the onboarding as “Guided setup” for Claude rather than Cowork, including the introduction and writing-voice handoff.
• Skill: Workflow authoring reference — Requires structured-output schemas to have an object root with properties and any required keys present in those properties, and notes that unsatisfiable schemas throw at agent().
• System Prompt: Artifact comment result guidance — Removes the opening reply-call recipe and its 4096-byte plain-text limit from this guidance while retaining activated-thread, attribution, resolution, and thread-reading rules.
• System Reminders: Artifact type instructions trust boundary and Artifact type page untrusted content warning — Extend the type's content scope from data files to store documents, explicitly forbid unauthorized writes to other addresses, and generalize page inspection from expected files to expected content without relaxing the untrusted-data boundary.
• Tool Descriptions: Artifact publishing and update guidance and Artifact runtime capabilities guidance — Direct reliable, shared, or Claude-readable state to available runtime capabilities, and require loading the runtime skill whenever capabilities would make the page more useful, not only when the user explicitly requests them.
• Tool Descriptions: Artifact type creation guidance and Artifact type discovery guidance — Require a title when creating from a type and explain that a type can take store documents rather than published data files; follow the returned instructions to populate the new Artifact instead of assuming every type takes files.
• REMOVED: Skill: Code Review (altitude dimension) — Removes the standalone review guidance to fix problems at the underlying mechanism rather than layering fragile special cases on shared infrastructure.