agent-team（Agent 团队）

agent-team

This skill is now a navigation shell. Route all requests to the dedicated first-level skills below.

Do not keep adding new behavior here.

Skill Reference

Controller / human orchestration

task-orchestrator

Write-capable task lifecycle entry.

• Audience: controller, human
• Triggers: create task, assign task, complete task, archive task, task flow
• CLI: agent-team task create · task list · task show · task assign · task done · task archive
• Note: Prefer task-inspector for read-only queries with no mutation intent.

task-inspector

Read-only task status and lookup.

• Audience: controller, worker, human
• Triggers: inspect task, task status, list tasks, show task
• CLI: agent-team task list · task show
• Note: Route to task-orchestrator if the request includes create, assign, done, or archive intent.

task-splitting

Decompose requirement or design documents into task drafts, then create agent-team task packages after explicit approval.

• Audience: controller, human
• Triggers: turn document into tasks, split tasks, validate task boundaries, prepare task files for agent-team task create
• Workflow: intake source → extract goals → draft by function boundary → validate → write one file per task → confirm → create
• Note: Invokes brainstorming automatically when the document goal is ambiguous or scope conflicts exist.

workflow-orchestrator

Governance-only workflow plan entry.

• Audience: controller, human
• Triggers: workflow plan, approve plan, activate plan, close plan
• CLI: agent-team workflow plan generate · approve · activate · close
• Note: Not for worker assignment, worker recovery, or worker status inspection.

worker-dispatch

Controller-side surface for opening workers, replying to workers, and checking targeted worker status.

• Audience: controller, human
• Triggers: open worker, dispatch worker, reply to worker, inspect current worker before replying
• CLI: agent-team worker open · worker status · agent-team reply
• Note: Use worker-inspector for read-only fleet inspection with no dispatch or reply intent.

worker-inspector

Read-only worker status inspection.

• Audience: controller, human
• Triggers: worker status, inspect worker, list workers
• CLI: agent-team worker status
• Note: Route to worker-dispatch if the prompt implies open, dispatch, or reply behavior.

project-bootstrap

Project initialization and migration entry.

• Audience: human, controller
• Triggers: init project, migrate project, set up agent-team in a new repo
• CLI: agent-team init · agent-team migrate

rules-maintenance

Sync and refresh built-in rule templates and generated rule artifacts.

• Audience: human, controller
• Triggers: sync rules, fix rule drift, update rules
• CLI: agent-team rules sync

skill-maintenance

Skill cache maintenance: check, update, clean installed skill artifacts.

• Audience: human, controller
• Triggers: check skills, update skills, clean skills, refresh skill cache
• CLI: agent-team skill check · skill update · skill clean

role-browser

Read-only local role listing.

• Audience: controller, worker, human
• Triggers: role list, browse roles, view local roles
• CLI: agent-team role list
• Note: Does not handle role create or role-repo management.

role-creator

Create or update role skill packages with deterministic output files.

• Audience: human, controller
• Triggers: create role, new role, edit role, update role, add role skill, 修改角色配置
• Output: SKILL.md + references/role.yaml + system.md under skills/<role>/ or .agent-team/teams/<role>/
• CLI: agent-team role create
• Workflow: normalize name → auto-generate role fields → select skills → execute CLI → validate output

role-repo-manager

Manage where roles come from (add, list, check, update, remove role repositories).

• Audience: human, controller
• Triggers: search role repo, add role repo, update role repo, remove role repo
• CLI: agent-team role-repo find · add · list · check · update · remove
• Note: Not for browsing local roles; that belongs to role-browser or catalog-browser.

catalog-browser

Read-only catalog browsing: search roles from configured remote catalogs.

• Audience: human, controller
• Triggers: search catalog, browse roles, show catalog role, catalog stats
• CLI: agent-team catalog search · show · list · repo · stats
• Note: Does not expose normalize, discover, or serve commands.

Worker-first skills

worker-recovery

Resume a worker's current assignment from file artifacts.

• Audience: worker
• Triggers: resume work, recover task, continue current assignment
• Required entry: MUST read worker.yaml first
• Recovery order: worker.yaml → task.yaml → context.md → referenced materials only if needed
• Note: Not for worker → main reporting; use worker-reply-main for that.

worker-reply-main

Send completion, blocker, or decision-needed messages back to controller.

• Audience: worker
• Triggers: reply main, report completion, blocked, request decision
• CLI: agent-team reply-main
• Required entry: MUST read worker.yaml first
• Note: Not for controller → worker replies; use worker-dispatch for that.

Shared / strategy skills

context-cleanup

Clean session context and re-anchor from file artifacts when the session is drifting.

• Audience: controller, worker
• Triggers: clean context, session drift, re-anchor context, resume after pause, phase transition
• Required entry: controller → .agent-team/rules/index.md; worker → worker.yaml
• Note: Use worker-recovery for routine worker task resume when the problem is simply recovering the current assignment.

brainstorming

Turn ideas into validated design documents through collaborative dialogue before any implementation.

• Audience: any
• Triggers: explicitly ask to brainstorm, shape requirements, compare approaches, produce a planning/design document
• Hard gate: do NOT write code or invoke any implementation skill until the design is presented and approved
• Note: Not for straightforward implementation requests with clearly specified changes.

tdd

Acceptance-first TDD workflow for solo implementation work.

• Audience: worker, controller
• Triggers: implement with TDD, define acceptance criteria, write failing tests first, verify fix before concluding
• Steps: read context → define acceptance criteria → write failing tests (when feasible) → implement → verify → conclude with passed / failed / skipped
• Note: Invoke qa-expert if the task needs dedicated test-case design or regression execution.

systematic-debugging

Root-cause-first debugging for any bug, test failure, or unexpected behavior.

• Audience: any
• Triggers: any bug, test failure, unexpected behavior, performance problem, build failure, integration issue
• Iron law: NO FIXES WITHOUT ROOT CAUSE INVESTIGATION FIRST
• Four phases: Root Cause Investigation → Fix Design → Implementation → Verification
• Note: Use especially when under time pressure or when previous fixes did not work.

find-skills

Discover and install agent skills from the open agent skills ecosystem.

• Audience: any
• Triggers: "how do I do X", "find a skill for X", "is there a skill that can...", extending agent capabilities
• CLI: agent-team skill install
• Note: Use when the user is looking for functionality that might exist as an installable skill.

qa-expert

Establish comprehensive QA testing processes using Google Testing Standards and OWASP best practices.

• Audience: any
• Triggers: set up QA, write test cases, execute test plan, track bugs, generate QA report, OWASP security testing, third-party QA handoff
• Capabilities: test case writing (AAA pattern), bug tracking (P0–P4), quality metrics, autonomous LLM-driven execution, 90% coverage targets
• Note: Invoke from tdd when the task needs dedicated test-case design or regression execution beyond unit verification.

prompt-engineering-patterns

Master advanced prompt engineering techniques for production LLM applications.

• Audience: any
• Triggers: optimize prompts, improve LLM outputs, design production prompt templates, chain-of-thought, few-shot learning, structured outputs
• Capabilities: chain-of-thought / tree-of-thought patterns, dynamic few-shot selection, variable interpolation templates, JSON structured outputs, system prompt design

Routing rules

• If the request implies a specific lifecycle action, route to the dedicated skill immediately.
• If the request is about context stabilization or resumed work, route to context-cleanup or worker-recovery.
• Keep this file as a compatibility surface for historical prompts only.