BH-M87

教 mainagent 何时 / 如何把任务分发给 subagent 来控制主线程上下文预算，并判断哪些独立任务适合并行、如何回收和验证结果。当任务命中「读超过 5 个文件 / 包含搜索 / 输出超过 500 行 / 主线程已 30+ tool call / 用户说调研/分析/扫一遍/梳理」任一条时启动。触发词：分发任务、上下文预算、主线程上下文、要不要 dispatch、subagent 分发、长任务、调研、扫一遍、避免上下文爆炸、context budget、控制上下文、mainagent context、上下文要爆了。

Use before implementation when a request is ambiguous, has materially different solution paths, changes architecture or public behavior, or involves costly or irreversible decisions. Skip for small well-scoped fixes, mechanical edits, and execution of an already approved plan.

Use when code review feedback has been received and needs to be assessed, answered, or implemented. Verify each suggestion against the actual code and requirements before changing anything.

Use when diagnosing a bug, test or build failure, performance issue, or unexpected technical behavior. Gather evidence and confirm a root-cause hypothesis before implementing a fix.

Use immediately before claiming work is complete, fixed, correct, or passing, and before publishing changes. Requires fresh, claim-specific evidence and accurate reporting of what was and was not verified.

Use when computing realized P&L, dividends, interest, or tax due from a Longbridge (長橋證券) account — for individual income tax on foreign income (个税 境外所得), CRS reporting, accounting, or broker P&L reconciliation. Pulls the official monthly statements via the Longbridge CLI and reconstructs the figures the statement does not provide. Companion to the Futu statement skill.

Use when extracting trades, dividends, interest, fees, or realized P&L from Futu (富途) HK statements — the annual statement (年度账单 .xlsx) or the monthly e-statements (證券月結單 .pdf) — for individual income tax on foreign income (个税 境外所得), CRS reporting, accounting, or broker P&L reconciliation. Covers where dividends and realized P&L actually live (they are not where you expect) and how to reconstruct figures the statement does not compute. Reads official statement files only; does not use the Futu OpenD/API.

Install, initialize, configure, repair, and verify this repository's Zed Project Manager extension and companion zpm CLI. Use when a user asks an agent to set up Project Manager from source, choose Git scan roots, enable cmd-alt-o shortcuts, register Zed Tasks or MCP settings, diagnose a missing project picker, or validate an existing installation.

Use when Codex repeatedly shows "Reconnecting... 2/5" through "5/5", WebSocket connection attempts fail before HTTP fallback, or the user wants to force the Responses API to use HTTP only.

多视角智能路由顾问。内置 14 位思维顾问，根据用户问题自动选出最合适的 1-3 位， 直接以该顾问的身份回答（或多人对话）。 触发词：「多视角」「找个顾问」「谁来回答」「multi-perspective」「顾问团」 「帮我找最合适的角度」「让专家来」「谁最适合」「多角度分析」「圆桌讨论」「让他们辩一辩」。 用户说「开个圆桌」「多位顾问」「几个人来看」时，启动多顾问模式。 用户说「让芒格来」「用费曼方式」「切换到乔布斯」等时，直接激活指定顾问。

Use this skill whenever the user wants to read news, get articles, browse RSS feeds, or discover and subscribe to content through Folo (app.folo.is) — an AI-powered RSS reader. Invoke when the user says things like "看一下 folo 上的新闻", "帮我看看 folo 里有什么", "fetch my folo timeline", "search for feeds about X in folo", "subscribe to a feed", "add RSS feed", "folo 资讯", "show me articles from folo", "folo 上搜索", "folo 订阅", "rss 阅读", or any mention of checking/browsing RSS reader content. This skill uses the Playwright browser tool to navigate the Folo web app and extract content.

确保技术文档符合 Why-What-How 三段式结构，但用自然的标题呈现（问题-目标-方案）。检查文档逻辑是否完整，不符合则修改。当生成或审查 README、设计文档、技术方案、功能说明时使用。

生成更自然、更像人类撰写的技术文档。避免 AI 生成文本的典型特征：过度结构化、冗余词汇、模板化表达。当生成 README、API 文档、设计文档、技术博客时使用。