Guide for assigning tools to agents and configuring `tool_choice` in multi-agent systems. Covers allowlists/blocklists (`tools` vs `disallowedTools`), wildcard expansion, `filterToolsForAgent`, built-in agent scoping, ToolSelector buckets, `tool_choice` modes (auto, any, forced), and agent schemas with `lazySchema` and `BuiltInAgentDefinition`. Use when the user says "agent tool allocation", "tool allocation", "tool_choice", "which tools for which agent", "agent tool scoping", "allowlist blocklist", "filterToolsForAgent", "disallowedTools", "lazySchema", "BuiltInAgentDefinition", "ASYNC_AGENT_ALLOWED_TOOLS", "COORDINATOR_MODE_ALLOWED_TOOLS", or "agent tool boundaries"; also use when designing a new agent's tool access. For tool descriptions use `mcp-tool-design`; for errors use `mcp-error-response`; for MCP optimization use `mcp-tool-enhancement`.
Design input and output JSON Schemas for search or retrieval tools that query the web, databases, code indexes, vector stores, or external services. Use when designing, reviewing, or refactoring a search-result tool, even if the user does not say "schema". Covers required vs optional/nullable fields, empty results, missing metadata, truncation/pagination (`appliedLimit` vs `truncated`), heterogeneous results (`z.union` and discriminated unions), enum `other` vs `unclear`, and HTTP metadata. Trigger on "search tool", "database/web/vector search", "retrieval output schema", "pagination", "truncation", "no results", "nullable fields", "enum other/unclear", or hallucinated missing fields. For general tool description or splitting, use `mcp-tool-design`; for error shapes, use `mcp-error-response`.
Generate or update `.claude/rules/tool-boundary.md` from an approved `docs/tool-routing-report.md`. Use when the developer says "boundary", "load tools", "apply routing", "generate boundaries", or approves a routing report and wants per-tool capability, limitation, chain-pattern, and failure-behavior declarations. Requires a routing report with `status: approved`; if missing or unapproved, stop and tell the user to run `/sentinel:routing` or approve the report first. Do not use for scanning or classifying which skills a project needs; use `/sentinel:routing` for that. Preserve manual edits marked with `<!-- manual -->` when updating an existing boundary file.
所有联网操作必须通过此 skill 处理，包括：搜索、网页抓取、登录后操作、网络交互等。 触发场景：用户要求搜索信息、查看网页内容、访问需要登录的网站、操作网页界面、抓取社交媒体内容（小红书、微博、推特等）、读取动态渲染页面、以及任何需要真实浏览器环境的网络任务。
Guide for locally testing Claude Code CI pipelines with structured JSON output. Assumes CLAUDE_CODE_OAUTH_TOKEN is already configured. Focuses on extracting structured_output from `--json-schema` results using jq, with precheck, output redirection, and CLAUDE.md context loading. Use when the user says "test CI locally", "claude -p", "structured output", "json-schema output", "extract CI results", "jq structured_output", "CI pipeline output", or wants to verify their CI JSON output configuration works before pushing to a real pipeline.
Scaffold a Python FastMCP server for domain-specific authenticated web search (Xiaohongshu, Dianping, Weibo, LinkedIn, NYT, region-locked forums, etc.) where the user's browser session supplies auth. Produces a working server skeleton with four wired-in features: Playwright-MCP browser bridge (with Claude-in-Chrome fallback), EN-to-domain-native query translation, per-search scratchpad directory, and structured writeback with source attribution. ALWAYS use this skill when the user says "build an MCP server for <site>", "scaffold a search MCP", "domain search server", "FastMCP server for <site>", "search scraper MCP", or asks to wrap a site behind an MCP search tool. Do NOT hand-write a FastMCP search server — invoke this skill so attribution, scratchpad layout, and backend selection stay consistent. For generic MCP tool shape guidance, see mcp-tool-design; for search result schema, see search-tool-schema-design.
Guide for designing structured error responses in MCP tools. Covers error classification (transient, validation, business, permission), the isError flag in MCP protocol and Anthropic API, retryable vs non-retryable errors, the critical distinction between access failures and valid empty results, structured error metadata design, and subagent error propagation patterns. Use when the user says "MCP error handling", "error response design", "isError flag", "tool error classification", "retryable errors", "error propagation", "tool returns wrong error", "empty result vs error", "when to retry MCP tool", "subagent error handling", "structured error response", "access failure vs empty result", or wants to design how MCP tools should report and classify errors for LLM agents. For tool description design, see mcp-tool-design instead. For platform-level MCP optimization, see mcp-tool-enhancement instead.
Guide for designing MCP tool content that LLMs can reliably select and use. Covers writing effective tool descriptions, structuring input/output contracts via JSON Schema, splitting tools by single responsibility, avoiding misrouting between similar tools, and coordinating system prompt instructions with tool descriptions. Use when the user says "tool description design", "tool interface design", "MCP tool descriptions", "tool splitting", "input output contract", "tool misrouting", "tool selection", "how to write tool descriptions", or wants to understand how LLMs choose between tools based on description content. For platform-level MCP optimization (searchHint, alwaysLoad, annotations), see mcp-tool-enhancement instead.