ha-multi-agent-coding
Hope-native multi-agent coding orchestration: fan out only independent valuable work, enforce bounded scope and isolation, consume structured results progressively, steer or cancel, and keep synthesis with the main Agent.
git clone --depth 1 https://github.com/shiwenwen/hope-agent /tmp/ha-multi-agent-coding && cp -r /tmp/ha-multi-agent-coding/skills/ha-multi-agent-coding ~/.claude/skills/ha-multi-agent-codingSKILL.md
# Hope Multi-Agent Coding Use multiple Agents when parallel evidence or implementation meaningfully outweighs coordination cost. Do not make delegation the default for complex-looking work. ## Fan-Out Decision Good candidates: - Similar independent investigations across modules. - Distinct review angles with a shared structured output. - Independent implementations with non-overlapping ownership. - A bounded set of alternatives that the main Agent will compare. Keep work with one Agent when: - The task is small or one search path is likely sufficient. - Steps depend on prior results. - Agents would edit the same files or shared generated state. - A single broad investigation needs coherent context. - Coordination, token, or merge cost exceeds expected parallel gain. ## Define Each Child Contract Provide: - One concrete objective and bounded scope. - Relevant context already known by the parent. - Allowed and forbidden actions. - File ownership or read-only isolation. - Required output schema, evidence, and stop condition. - Verification expected from the child. Do not re-delegate the entire parent assignment to one child. Children do not own final user communication or Goal closure. ## Isolation - Prefer `shared_read_only` for research, discovery, and verification. - Use separate worktrees for independent writes. - If writes cannot be isolated, serialize them or assign mutually exclusive file ownership. - Permission mode, protected paths, approval surfaces, and tool restrictions remain runtime-enforced. A child prompt cannot grant access. ## Bounded Execution Set explicit limits for fan-out count, depth, turns, tokens, and time. Respect runtime queues and backpressure. Never recursively create Workflow runs or an unbounded Agent tree. ## Progressive Control The main Agent may choose based on task needs: - Consume the first useful results and adapt (`waitAny` / checkpoint). - Query status without consuming output. - Read one structured result, then steer or cancel remaining work. - Add a follow-up child when new evidence changes the decomposition. - Wait for all children only when a true barrier is required. Background work must not block the user's conversation. Use runtime completion or checkpoint injection rather than polling loops. ## Synthesis The main Agent must: 1. Check which children completed, failed, timed out, or returned no evidence. 2. Resolve conflicts using source evidence, not majority vote. 3. Preserve partial failures and uncertainty. 4. Integrate or review writes in the parent worktree deliberately. 5. Run parent-level verification for the combined outcome. "All Agents completed" is orchestration state, not a user result. Do not finish until the parent has synthesized and answered the actual task. ## Workflow Boundary This skill decides delegation strategy. Use `ha-workflow-script` when execution must be durable, replayable, observable, or script-controlled. Simple bounded subagent work does not require a Workflow; a Workflow may use this strategy without surrendering its runtime contracts. ## Smoke Prompts - "Investigate these six independent modules in parallel and synthesize." - "Use staged child results; do not wait for every slow reviewer." - "Decide whether this implementation should stay single-Agent or use worktrees."
>
Use when the user asks to draft, polish, translate, or reply to an email. Produces a clean draft with subject line, greeting, body, and sign-off, plus a pre-send self-check.
Use when the user mentions 飞书 / Feishu / Lark workspace operations: docx (云文档) read/write, bitable (多维表格) records / views / dashboards, drive (云盘) upload/download, wiki (知识库) link resolution, approval (审批) instance create/cancel/query, calendar (日历) event create/list/update + attendees, contact (联系人) user/department lookup, hire (招聘) job/talent/application listing. Trigger on phrases like 'OKR 周报', '把这份文档发到飞书云盘', '给团队拉个评审会议', '查 [姓名] 的联系方式', '撤销那条审批', '/wiki 链接', or any request that mentions a feishu / lark URL / token (doxcn.../bascn.../wikcn.../boxcn.../om_...).
Hope Agent browser automation — the standard `status → tabs → snapshot → act` loop, stale-ref recovery rules, and what to do when login / 2FA / captcha / camera-prompt / dialog blocks progress. Load this skill whenever you reach for the `browser` tool. Trigger on: user asks the agent to open / control / click / scrape / log into / verify something in a web app ('open X and click Y', '打开 X 然后点击 Y', 'log into my Gmail', 'scrape this page', 'fill out the form on X'); user reports a flow that requires real browser context (cookies, JS-rendered content, OAuth).
Discover and install third-party skills from external registries when the user needs a capability that no currently-active skill covers. Trigger when: (1) the user explicitly asks 'find a skill for X', 'is there a skill that does X', 'install a skill to X', (2) the user requests a well-known integration (Slack, Notion, Trello, GitHub, Hue, Sonos, iMessage, weather, TTS, transcription …) that isn't in the active skill catalog, (3) you are about to hand-write ad-hoc shell / API code for a domain that almost certainly has a published skill. Do NOT trigger if an active skill already covers the need — scan the visible skill catalog first.
Self-service diagnostics — query Hope Agent's local SQLite databases (logs / sessions / background jobs) directly via the `exec` tool to investigate problems, analyze usage, and locate root causes. Trigger on: user reports something broken / failing / slow / stuck / not responding ('X 不工作', 'X 报错', 'X 卡住', '为什么 X 失败', 'why did X fail', 'show me the logs', 'check what happened'); ad-hoc data analysis ('this week's token usage', '最近调用最多的工具', 'how many subagent runs failed', 'tool error rate', 'find sessions where X happened'); verifying a fix ('did the error stop after I changed Y'). Use BEFORE asking the user to paste log snippets — the data is on disk, query it directly. Read-only — SELECT only, never UPDATE/DELETE/INSERT/DROP.
Hope Agent native macOS desktop control — the standard `mac_control` status / diagnostics / apps / dock / spaces / snapshot / visual / windows / menu / clipboard / dialog loop, target-first action rules, no-blind-coordinate policy, and recovery for stale AX/window/menu/dialog state. Load whenever using `mac_control`, or when the user asks to control local Mac apps, Dock, Spaces, click/type/menu/window/dialog/clipboard, automate Finder/TextEdit/System Settings, visually locate UI, or says 控制 Mac, macOS 自动化, 点按钮, 打开应用, Dock, Space, 关闭窗口, 菜单点击, 视觉定位.
Self-understanding and issue reporting for Hope Agent itself. Use when the user asks how Hope Agent works internally, asks about its own source code/docs/runtime behavior, reports a bug/failure/slowness/crash, asks to diagnose logs, or asks to create/submit a GitHub issue for a bug, feature request, or improvement (including when there is no bug). Chinese triggers: 自查, 了解自己, 自我诊断, 排查 Hope Agent, 提交 issue, 需求 issue, 功能改进.