返回 Skill 列表
extension
分类: AI Agent 能力无需 API Key

use-agent-teams

将子代理调度升级为代理团队,用于执行do-plan、do-execute和do-discuss阶段。在以标准/深度级别运行do-plan、do-execute或do-discuss时使用。

person作者: jakexiaohubgithub

Overlay that replaces Spine's parallel subagent dispatch with Agent Teams for 5 phases.

All teammate personas MUST be defined inline in the spawn prompt — custom .claude/agents/ files are silently ignored for team agents.

Lifecycle per phase: create team → spawn teammates → wait for outputs and peer exchange → synthesize → shut down team. One team per phase; shut down before creating the next.

Spawn type: Use the custom agent type matching the phase's dispatch table (@framer, @planner, @debater, @analyst, @inspector, @worker). If a custom agent type is unavailable, fall back to general-purpose. Team agents write output to .scratch/<session>/.

Use team_name matching the phase: discuss-explore, plan-planning, plan-challenge, exec-polish, exec-review.

do-discuss Phase 4: Explore

Create one team. Perspective-committed framers react to each other's outputs — advisory dialogue, not independent position papers.

| Teammate | Peer directive | |----------|---------------| | stakeholder-advocate | Surface user needs and adoption friction. When peers raise technical concerns, assess user impact. | | systems-thinker | Map dependencies and second-order effects. When peers surface needs, assess feasibility and coupling. | | skeptic | Challenge assumptions and demand evidence. When peers agree, ask "what if that assumption is wrong?" |

Lead reads all teammate outputs + peer messages. Irreconcilable positions become Key Decisions in the problem frame. Convergent findings become Known Facts.


do-plan Phase 3: Planning

Create one team. Lead copies each persona from do-plan's dispatch table and appends the peer directive below:

| Teammate | Peer directive | |----------|---------------| | conservative | Share concerns about risk and blast radius with peers as you find them. | | thorough | When a peer raises a concern, verify whether your plan addresses it. | | innovative | React to conservative concerns by showing how innovations mitigate or accept them. |

Lead merges teammate outputs + peer message history into canonical_plan.


do-plan Phase 4: Challenge Debate

Create one team. Socratic dialogue, not independent position papers — debaters react to each other's arguments mid-flight.

| Teammate | Peer directive | |----------|---------------| | thesis-champion | When the dissenter raises an objection, respond with specific evidence and steelmanned rebuttals. Do not concede without E2+ counter-evidence. | | counterpoint-dissenter | Read the champion's defenses and escalate with stronger evidence or concede explicitly when rebutted. | | tradeoff-analyst | Track which arguments have been resolved vs remain contested. Quantify costs and reversibility for each open point. |

Lead reads all teammate outputs + full peer exchange. Findings that survive (E2+ with no viable alternative) are incorporated.


do-execute Phase 3: Polish Advisory

Create one team. Lead copies each persona from do-execute's dispatch table and appends:

| Teammate | Peer directive | |----------|---------------| | conventions-advisor | Share findings with peers so they can check for related implications. | | complexity-advisor | When the conventions advisor flags a pattern, check whether it also has complexity implications. | | efficiency-advisor | When peers flag a pattern deviation or bloat, check whether it also has performance or reuse implications. |

Lead reads all teammate outputs + peer messages. Deduplicates, assigns E-levels.


do-execute Phase 4: Adversarial Review

Create one team:

| Teammate | Peer directive | |----------|---------------| | spec-reviewer | Share spec gaps with peers so they can probe deeper. | | correctness-reviewer | When the spec reviewer flags a gap, check whether it causes correctness issues. | | risk-reviewer | React to correctness findings by assessing their risk severity. |

Lead reads all teammate outputs + peer messages. Deduplicates, assigns final E-levels and severity buckets per do-review rules.

Excluded Phases

All other phases: use standard subagent dispatch from Spine skill files.

Anti-Patterns

  • Creating teams for main-thread-only or single-subagent phases
  • Using delegate mode — always use default mode (delegate breaks tool access)
  • Spawning teams when env var is unset
  • Defining teammate personas only via .claude/agents/ files (silently ignored — inline only)
  • Forgetting to shut down a team before creating the next phase's team