菜单
Task management for session continuity. Use when coordinating multi-step work, managing subagent assignments, or preserving intent across compaction. Triggers on track tasks, manage work, coordinate agents, or when complex work requires sequencing.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/outfitter-dev/outfitter --skill maintain-tasks命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Patterns for @outfitter/daemon including lifecycle management, IPC communication, health checks, and PID files. Use when building background services, daemons, or when "daemon", "IPC", "health check", "background service", or "@outfitter/daemon" are mentioned.
Creates handlers, CLI commands, MCP tools, and daemon services following Outfitter Dev Kit conventions. Use when adding new components to a project, scaffolding code, or when "create handler", "new command", "add tool", or "daemon service" are mentioned.
Deep patterns for @outfitter/mcp including tool registration, Zod schemas, resources, and server configuration. Use when building MCP servers, registering tools, defining resources, or when "MCP server", "MCP tool", "registerTool", or "@outfitter/mcp" are mentioned.
正在显示 SKILL.md
基于 SOC 职业分类
| name | maintain-tasks |
| description | Task management for session continuity. Use when coordinating multi-step work, managing subagent assignments, or preserving intent across compaction. Triggers on track tasks, manage work, coordinate agents, or when complex work requires sequencing. |
| user-invocable | false |
Tasks are your working memory that survives compaction. Use them to maintain intent, sequence, and coordination across the natural lifecycle of a session.
| Tool | Purpose | Scope |
|---|---|---|
| Tasks | Session/project working memory | Current effort, active coordination |
| Linear/GitHub | Team-visible project management | Cross-session, multi-person tracking |
Tasks are NOT a replacement for filing issues. They are your local execution state — what you're doing now, what comes next, who (which agent) is responsible.
Sync pattern:
When context compacts, conversation history gets summarized. But Tasks persist in full:
blockedBy, blocks) preservedpendingin_progresscompletedThis means after compaction, you can read TaskList and know exactly where you were, what's next, and who's responsible.
Always assign subagents explicitly in task subjects.
After compaction, nuanced instructions like "have the reviewer check this" can lose fidelity. Explicit assignment survives:
[engineer] Implement auth refresh endpoint
[reviewer] Review auth implementation for security
[tester] Validate auth flow end-to-end
The [agent-name] prefix is not decoration — it's recoverable intent. After compaction, you can scan TaskList and immediately know which agent handles each task.
# Explicit agent assignment
[engineer] Build the feature
[analyst] Research the approach
[reviewer] Check for issues
[tester] Validate behavior
# Background agents (include task ID for retrieval)
[reviewer:background] Security audit (task-id: abc123)
# Resumable agents (include agent ID)
[analyst] Continue research (resume: a7227ac)
Use blockedBy and blocks to encode sequence:
Task 1: [analyst] Research auth patterns
Task 2: [engineer] Implement auth flow (blockedBy: 1)
Task 3: [reviewer] Review implementation (blockedBy: 2)
Task 4: [tester] Validate auth (blockedBy: 3)
After compaction, dependencies tell you what's actionable (no blockers) vs what's waiting.
in_progress before beginningAs context fills, ensure Tasks reflect:
in_progress task)Subject: [agent] Imperative action description
Description:
- Context needed to execute
- Acceptance criteria
- Any constraints or considerations
ActiveForm: Present continuous for spinner ("Implementing auth flow")
Subject format: [agent-name] Verb the thing
[engineer] Implement JWT refresh logic[reviewer] Audit auth module for vulnerabilities[analyst] Research rate limiting approachesDescription includes:
blockedBy1. [analyst] Research → creates findings
2. [engineer] Implement → uses findings (blockedBy: 1)
3. [reviewer] Review → checks implementation (blockedBy: 2)
4. [tester] Validate → confirms behavior (blockedBy: 3)
1. [engineer] Implement feature A
2. [engineer] Implement feature B
3. [tester] Integration tests (blockedBy: 1, 2)
Tasks 1 and 2 can run in parallel; Task 3 waits for both.
1. [engineer] Initial implementation
2. [reviewer] Review round 1 (blockedBy: 1)
3. [engineer] Address feedback (blockedBy: 2)
4. [reviewer] Review round 2 (blockedBy: 3)
| Anti-Pattern | Why It Fails | Better Approach |
|---|---|---|
| No agent assignment | Lost after compaction | Always prefix with [agent] |
| Vague subjects | Can't recover intent | Specific, actionable subjects |
| Batching completions | State becomes stale | Mark done immediately |
| Skipping for "quick" tasks | Consistency matters | If 3+ steps, use Tasks |
| No dependencies | Unclear sequence | Encode with blockedBy |
ALWAYS:
[agent] Task descriptionin_progress before starting, completed when doneNEVER:
in_progress simultaneously