본문으로 건너뛰기
cccskills
무료GitHub에 공개

team-dispatch

Turns a plan phase into a PhaseSpec v1 and dispatches it to a specialist subagent via the Agent tool. Subagent reads files itself, returns a structured result, the result is logged into progress.md. Markdown is the shared state; no in-memory shared store across agents.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

설치 방법 보기

포함된 파일(1)

  • SKILL.md5.4 KB

SKILL.md(원문)

설치하기 전에 에이전트에게 전달되는 지시 내용을 확인할 수 있습니다.

Team Dispatch

Subagent dispatch arm. The PhaseSpec v1 contract is the only handoff format used between parent and child agents.

PhaseSpec v1 (the contract)

phase_id: <stable id, used as filename prefix and progress.md anchor>
parent_plan_ref: <pointer to task_plan.md#phase-N>
goal: <one sentence>
done_when:                       # testable exit criteria, no goalpost moving
  - <criterion 1>
  - <criterion 2>
inputs:
  files: [<read-only file refs>]
  memory_keys: [<keys for memory-layer recall>]
tools_allowed: [<list>]
tools_denied: [<list>]
budget:
  max_tool_calls: <int>
  max_wall_seconds: <int>
hitl_checkpoints: [<named checkpoint strings the child surfaces>]
return_contract:
  format: markdown+frontmatter
  fields: [summary, artifacts, open_questions, next_phase_hint]

Six load-bearing properties:

  1. parent_plan_ref is a pointer (file path with anchor), not a copy of the phase text. The child reads the file itself.
  2. done_when is testable. Without it, completion is hallucinatable.
  3. tools_allowed and tools_denied map onto the Claude Code permission model and any other framework's permission story.
  4. hitl_checkpoints are named strings the child surfaces by printing HITL_CHECKPOINT: <name> to stdout.
  5. return_contract forces structured output. The child returns markdown with frontmatter; the parent parses it deterministically.
  6. phase_id doubles as the filename prefix for any artifacts the child writes.

Available subagents

Each is defined as agents/<name>.md in this plugin.

SubagentWhen to pick
plannerA phase exists in task_plan.md but has no PhaseSpec yet. The planner drafts one.
executorDefault. A PhaseSpec exists; the work is straightforward; one specialist runs it end to end.
reviewerAuto-approve mode is on; we need a surrogate human to check the executor's return against done_when.
memory-keeperStop hook fired; we need to classify session output into scratch/WHERE/WHY.
design-archeologist/mddesign:harvest invoked; long-running codebase scan needed.

Operation: /mddesign:team dispatch <phase_id> [--agent <name>]

Step 1: Resolve the phase

  • Read task_plan.md.
  • Find the section header matching phase_id (case-insensitive).
  • If not found, refuse.

Step 2: Pick the subagent

  • If --agent <name> was passed, use it.
  • Else infer from the phase: phases with "design", "ui", "tokens" in the title default to design-archeologist if /mddesign:harvest-shaped, else executor. All others default to executor.

Step 3: Build the PhaseSpec

  • Pull goal from the phase title or first bullet.
  • Pull done_when from the phase's "Status / Done When / Acceptance" section if it exists, else infer from the bullets and ask the user to confirm.
  • Set inputs.files to [task_plan.md, findings.md] plus any files the phase mentions.
  • Set tools_allowed to a sensible subset based on the phase (default: Read Write Edit Bash Glob Grep).
  • Set budget defaults: max_tool_calls: 80, max_wall_seconds: 600.
  • Set hitl_checkpoints to [before_commit, before_destructive_edit] plus phase-specific ones.

Step 4: HITL gate before dispatch

  • Print the full PhaseSpec to the user.
  • The PreToolUse hook intercepts the next Agent tool invocation (named checkpoint: before_subagent_dispatch) and requires approval.

Step 5: Dispatch

  • Call the Agent tool with subagent_type: <chosen subagent> and a prompt that contains:
    • The full PhaseSpec YAML
    • Plain-English instruction: "You are running PhaseSpec phase-<id>. Read every file in inputs.files. Do the work. Return a markdown document with the four fields in return_contract. If you hit any hitl_checkpoint, print HITL_CHECKPOINT: <name> and stop."

Step 6: Receive the return, validate

  • Parse the subagent's final message.
  • Verify it has the four required fields (summary, artifacts, open_questions, next_phase_hint).
  • If invalid, log the failure and surface to user.

Step 7: Write to progress.md

Append under ### Phase <phase_id> Result:

### Phase <phase_id> Result

**Subagent:** <name>
**Status:** complete | partial | blocked
**Wall time:** <seconds>
**Tool calls used:** <int>

#### Summary
<from return.summary>

#### Artifacts
<from return.artifacts>

#### Open questions
<from return.open_questions>

#### Next phase hint
<from return.next_phase_hint>

Step 8: Optional review

  • If auto-approve mode is on, dispatch reviewer with the executor's return + the PhaseSpec.
  • If reviewer says approve, mark phase as complete in progress.md.
  • If reject, append the reason to progress.md and surface to user.
  • If escalate, halt and ask user.

Core rules

  • Subagents never edit task_plan.md directly. Only this skill writes to progress.md, and only under ### Phase <id> Result.
  • The PhaseSpec is the entire handoff. Subagents do not see the parent conversation.
  • Markdown on disk is the shared state. No in-memory shared store across agents.
  • Every dispatch goes through the before_subagent_dispatch HITL gate.

리뷰

아직 리뷰가 없습니다. 사용해 본 소감을 남겨 주세요.

댓글

로그인 없이 쓸 수 있습니다. 사용 후기, 질문, 작동하지 않은 환경 등을 편하게 남겨 주세요.

아직 댓글이 없습니다. 첫 댓글을 남겨 보세요.

보내면 바로 공개됩니다. 이메일 주소나 전화번호 같은 개인 정보는 쓰지 마세요. URL은 2개까지입니다.

0 / 1,000자

이 양식은 reCAPTCHA로 보호되며 Google 개인정보처리방침 및 서비스 약관이 적용됩니다.

같은 저장소의 스킬

Continuously bridges DESIGN.md into the active planning phase. When the user is working on a UI phase, reads the project's DESIGN.md, extracts the relevant token slice, and appends it to findings.md under "## Design Context" so the agent sees design constraints in its attention window. Never edits DESIGN.md, never edits task_plan.md. Idles when DESIGN.md is absent.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

OthmanAdi/MDDesign142026년 4월 25일 업데이트

Harvests a DESIGN.md from an existing codebase, critiques it against the actual code, and applies safe quick-win fixes. Three modes harvest / critique / fix. Shells out to `npx @google/design.md lint` for structural validation; never reimplements Google's linter. The on-ramp that makes DESIGN.md adoptable for every existing repo, not just greenfield projects.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

OthmanAdi/MDDesign142026년 4월 25일 업데이트

Three-tier memory layer for the orchestrator. Routes scratch (file-tier), WHERE (QMD), and WHY (MemPalace) through the existing code-memory-router slash command. Never edits planning files directly. Every WHY-tier promotion goes through the HITL gate.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

OthmanAdi/MDDesign142026년 4월 25일 업데이트

Makes sessions durable across `/clear`, machine reboots, and IDE switches. SessionStart catchup pulls scratch + last MemPalace recall + planning files into context. Stop promotion writes a handoff note and runs memory-keeper. Manual `/mddesign:handoff` re-runs either at any time.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

OthmanAdi/MDDesign142026년 4월 25일 업데이트

OthmanAdi의 스킬 모두 보기

이 스킬의 문제 신고