v2.1.176
2.1.176
(☝️ Click to see full diff.)
+4,360 tokens
- REMOVED: System Prompt: Claude in Chrome skill note — Removes the note telling the agent to invoke the
claude-in-chromeskill (via the Skill tool) before using anymcp__claude-in-chrome__*browser tools. - Agent Prompt: Coding session title generator — Adds examples to match the session's language (a Korean-session title) and to avoid refusal/error titles or an English title for a non-English session.
- Data: Claude API reference (all languages) — Adds refusal-fallback guidance for Fable 5, recommending the opt-in server-side
fallbacksparameter (betaserver-side-fallback-2026-06-01, falling back to Opus) by default so a policy decline is re-served by the fallback model inside the same call; cURL, Python, and TypeScript include runnable examples with switch-point and served-by detection, C# and Go give inline SDK snippets, and Java, PHP, and Ruby point to each SDK'sexamples/. Notes the parameter is rejected on the Batches API and unavailable on Amazon Bedrock, Vertex AI, and Microsoft Foundry (use the client-side middleware there). - Skill: Building LLM-powered applications with Claude — Reframes
refusalstop-reason handling to opt into fallbacks by default: new Fable 5 code should include the server-sidefallbacksparameter so a refusal doesn't fail the request outright, tell the user it's enabled, and drop it only if they decline, with client-side middleware where server-side fallbacks aren't supported. - Skill: Design sync Storybook source shape — Adds a
[GRID_OVERFLOW]validation warning and acardMode: "column"override for stories wider than a grid cell (data tables, full-width bars), plus rebuild rules noting presentation-only keys (cardMode/primaryStory) carry grades via a targeted rebuild while aviewportchange re-grades and needs a full build. - Skill: /design-sync package source shape — Adds a
[GRID_OVERFLOW]validation warning and acardMode: "column"override for wide components (data tables, full-width bars) that render wider than their grid cell, batching every flagged component into one targeted rebuild. - Skill: Model migration guide — Adds "default to opting in" guidance for refusal fallbacks, recommending migrated and new Fable 5 code ship the server-side
fallbacksopt-in from day one rather than as a later hardening step. - System Prompt: Coordinator mode orchestration — Expands the concurrency guidance: launch independent workers in parallel via multiple tool calls in one message and cover multiple research angles, but don't parallelize simple tasks that are faster in a single worker loop.
- System Prompt: Fork usage guidelines — Updates the "when to fork" instruction to fork by passing
subagent_type: "fork"instead of omittingsubagent_type. - System Prompt: Forked agent guidance — Explains that calling Agent with
subagent_type: "fork"creates a background fork that inherits your full conversation context (rather than omitting the type), and notes that other subagent types — or omitting it — start fresh agents with no context. - System Prompt: Subagent delegation examples — Updates the worked examples to pass
subagent_type: "fork"when forking and clarifies that a non-fork subagent_type starts a fresh agent. - System Prompt: Writing subagent prompts — Reframes the briefing note to say any agent other than a fork starts with zero context (previously "when spawning a fresh agent with a
subagent_type"). - Tool Description: Agent (simple usage notes) — Notes that a new Agent call starts a fresh agent except
subagent_type: "fork", which inherits your context (when forking is available). - Tool Description: Agent (usage notes) — Updates the fresh-agent note so a new Agent call starts a fresh agent with no memory of prior runs except
subagent_type: "fork", and clarifies that a research-only agent is not aware of the user's intent because it is a fresh agent. - Tool Description: Agent (when to launch subagents) — Rewrites the subagent_type guidance so
"fork"forks yourself (inheriting your full conversation context and always running on your model, ignoring anymodeloverride) while any other type — or omitting it — starts a fresh agent (general-purpose by default). - Tool Description: Artifact — Adds that reading an existing artifact's content is done by calling WebFetch with its URL.
- Tool Description: claude.ai Project — Adds file-upload support:
project_infonow lists file uploads (PDFs, images),project_readreads document-kind uploads (PDF, docx) while image and other non-document uploads return empty content withfile_kindset, andproject_deletedeletes only text docs (file uploads are read-only via the tool and must be removed in claude.ai). - Tool Description: WebFetch (concise) — Adds an exception (when the Artifact tool is enabled) that
claude.ai/code/artifact/{uuid}URLs ARE fetchable via your claude.ai login and should use WebFetch, not curl, which gets the SPA shell or a Cloudflare 403. - Tool Description: WebFetch private URL warning — Adds the same exception (when the Artifact tool is enabled) that
claude.ai/code/artifact/{uuid}URLs (including preview.claude.ai) are fetchable via the claude.ai login and should use WebFetch, not curl or a headless browser.