v2.1.251

2.1.251

(☝️ Click to see full diff.)

-743,949 tokens

• NEW: Agent Prompt: Artifact editor thread follow-up — Routes a tagged comment that arrives while an editor owns an Artifact to that worker, requiring a page-scoped edit and republish for relevant requests and no change for unrelated ones.
• NEW: Agent Prompt: Schedule local MCP server limitation — Explains that MCP servers configured directly in Claude Code cannot be attached to cloud routines, limits routines to Claude.ai connectors, and adapts connector advice to why the connector list was unavailable.
• NEW: Data: SDK Remote Control availability field — Documents the optional internal initialize-response field for stable deployment eligibility, lets IDE hosts hide an impossible Remote Control affordance, and tells older-CLI consumers to treat an absent field as available.
• NEW: Skill: Multiplayer whiteboard description — Adds trigger guidance for live shared boards with synchronized strokes and cursors, session presence, pasted images, and Claude drawing in real time.
• System Prompt: Interactive agent intro — Splits the configured Output Style opening into variants that say the style is “below” only when it is actually included below the intro.
• NEW: System Prompt: Reporting outcomes — Requires completion claims to rest on observed results, puts failures, skipped steps, and incomplete work first, and forbids presenting unchecked or partial work as done.
• NEW: System Reminder: Cross-session peer message authority warning note — Adds a note-only peer-message variant that identifies the sender as a session-owned subagent or teammate while preserving the rules against escalation, approval transfer, and permission laundering.
• NEW: System Reminder: Memory updates — Marks earlier memory reads stale when memory files change, establishes the refreshed turn-start state, and forbids exposing the machine-generated update block to the user.
• NEW: System Reminder: Session context — Supplies selected session values, explicitly replaces earlier values when they change, and tells Claude not to respond to them unless highly relevant.
• Agent Prompt: Status line setup — Adds the optional Claude gateway rate_limits.spend_limit object, including percentages above 100 and reset timing, plus a shell example for displaying it.
• Agent Prompts: Web fetch usage guidance and Web reading specialist — Make foreground result delivery explicit, cap the specialist at 15 turns, and require failed or denied fetch reports to name the URL and HTTP status or error so the caller can retry directly.
• Data: Claude Code agent proxy troubleshooting guide — Adds diagnosis for mid-transfer resets and RPC disconnects, directing Claude to inspect recentRelayFailures before blaming the remote service.
• Data: Files API references (C#, Go, Java, PHP, Python, TypeScript) and Platform availability — Mark the Files API as out of beta on the first-party Claude API and Claude Platform on AWS, warn that current stable SDK shapes break from the older beta namespaces, and mark Agent Skills as generally available on those same two platforms.
• Skill: Plugin eval authoring interview — Requires local mocks for every MCP tool an evaluated flow can call, supports canned, argument-sensitive, error, and stateful multi-call mocks, and adds the mock layout to the generated eval suite.
• Skill: Artifact document, Skill: Artifact report, Skill: Data Visualization description, Tool Description: Artifact, and Tool Description: Artifact type discovery guidance — When a host-designated first-party document connector is attached, route shared document work there instead of to Artifact types or document skills, keep third-party tools from triggering that route, retain Artifacts for explicit Artifact/HTML/Markdown requests and app-like pages, and pass chart rows rather than rendered images to connectors with live charts.
• Tool Description: Artifact comments guidance, Tool Parameter: Artifact comment actions guidance, System Prompts: Artifact comment list/thread framing and result guidance, and System Reminder: Artifact comment reply activation failure — Explain how writers send threads to Claude, distinguish comments sent to another Claude session, add stale anchor-label and source-snippet context, and tell Claude how to report threads it cannot reply to or resolve.
• Tool Description: Artifact publishing and update guidance — Adds app-specific post-publish handoff through the conversation card without repeating the URL or terminal shortcuts.
• Tool Description: Artifact — Spells out the publisher-supplied HTML reset and hidden behavior and requires page-owned title and styles.
• NEW: Tool Description: Artifact nested runtime cleanup error — Diagnoses repeatedly nested runtime markers, <base> tags, and page skeletons and directs Claude to retain and republish the innermost clean document.
• Tool Description: Artifact live room guidance and Tool Parameter: Artifact watch actions guidance — Clarify approval and lifetime boundaries across auto mode, /clear, conversation switches, and resumes, and state that only interactive or SDK main-loop sessions—not subagents, teammates, background work, or print sessions—can hold Artifact watches.
• Tool Description: Artifact database guidance — Clarifies that private data/users/me paths are expressed through the collection path and resolve to the current user only when the published Artifact also declares the user capability.
• Skill: Claude Code configuration guide — Sends users to GitHub issues instead of /feedback when feedback is disabled by policy or a kill switch, as well as on Bedrock, Vertex, and Foundry.
• System Prompt: REPL tool usage and scripting conventions — Resolves top-level Edit and Write names through their available aliases and documents NotebookEdit, rather than Write, in the direct REPL-call examples.
• REMOVED: Agent Prompt: Security monitor for autonomous agent actions (second part) — Removes the standalone second policy segment covering environment context, block and allow rules, consent evaluation, and classification output; the first part, host-context guidance, and remote-machine rules remain.
• REMOVED: System Reminder: Permission prompt auto-denied after timeout — Removes the standalone reminder that an unattended timeout neither performed nor judged the action and that approval-gated variants should not be retried until the user returns.
• REMOVED: Managed Agents onboarding and reference suite — Removes the onboarding flow, ant CLI guide, overview, client and core-concept references, endpoints, environments and resources, events, memory stores, multi-agent sessions, scheduled deployments, self-hosted sandboxes, tools and skills, webhooks, and the cURL, Go, Java, PHP, Python, Ruby, and TypeScript references; the Managed Agents outcomes reference remains.
• REMOVED: Claude API and application-building reference suite — Removes the Admin API guide, Claude API references (C#, cURL, Java, Python, TypeScript), model catalog, HTTP error and live-source guides, prompt-caching guide, tool-use concepts and references (Go, Java, Python, TypeScript), Agent Design Patterns, Python SDK migration, application-building, cost-optimization, model-migration, and prompt-audit skills.
• REMOVED: Artifact and visual-authoring bundle — Removes the decision-component assets, workshop and plan templates, Artifact design and workshop skills, both Artifact PR-review flows, the Design, Prototype, and Whiteboard skills, and the data-visualization reference palette; the original Whiteboard trigger description remains alongside the new multiplayer-specific description.
• REMOVED: Plugin, design-sync, run, and verification bundle — Removes the Cowork plugin schemas and authoring skill, plugin eval and skill-doctor references, Design sync and its package/Storybook variants, the Electron run example, Run skill generator, and Verify skill.