본문으로 건너뛰기
cccskills
무료GitHub에 공개

tdmcp-pipeline

Orchestrates the tdmcp feature team end-to-end: design/wireframe → build → integrate → QA → release. Use whenever the user wants to build, implement, develop, ship, or add one or more tdmcp features/tools/effects/controls/AI-prompts for TouchDesigner, or to run them through a coordinated design→develop→QA→deploy pipeline. Also use for follow-ups: re-run, continue, update, fix, or improve a feature/batch, ship the next version, or rebuild just one part of a previous run. Single feature or a whole batch.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

설치 방법 보기

포함된 파일(1)

  • SKILL.md7.6 KB

SKILL.md(원문)

설치하기 전에 에이전트에게 전달되는 지시 내용을 확인할 수 있습니다.

tdmcp-pipeline — feature delivery orchestrator

Coordinate the five tdmcp specialists to take feature ideas from spec to a pushed release, following the project's proven workflow: parallel isolated builders + a single-writer integrator + live-validated QA + an autonomous release.

Execution mode: hybrid

StageModeWhy
Designsub-agent (fan-out if many features)the architect is an isolated one-shot producer; specs land in _workspace/
Buildsub-agent (fan-out, parallel)builders are fully isolated by design (new files only, no inter-comms) — the textbook sub-agent case
Integrate → QA → Releaseagent teamproducer↔reviewer feedback loops (QA ↔ builder-fixer/integrator) and the release gate need live messaging

Why this split: builders must NOT touch shared files so they can run concurrently; that isolation removes any need for build-time team comms. The integrate/QA/release trio, by contrast, lives on a tight fix-and-re-verify loop, which is exactly what an agent team is for.

Agent roster

AgentTypeSkillOutput
td-architectcustomtd-feature-design_workspace/01_design_<feature>.md
td-builder (×N)customtd-feature-buildnew tool + test files; _workspace/02_build_<feature>.md
td-integratorcustomtd-feature-integratewired index.ts/agent.ts; _workspace/03_integrate.md
td-qacustomtd-feature-qa_workspace/04_qa_<batch>.md (PASS/FAIL/UNVERIFIED)
td-releasercustomtd-feature-releaseCHANGELOG + version bump + commit/tag/push; _workspace/05_release.md

All Agent / TeamCreate calls use model: "opus".

Workflow

Phase 0 — context check (follow-up support)

  1. Check whether _workspace/ exists.
  2. Decide the run mode:
    • No _workspace/ → fresh run. Go to Phase 1.
    • _workspace/ exists + user asks to fix/continue/re-run part → partial re-run. Re-invoke only the affected agent(s), passing the prior artifact paths so they refine rather than rewrite.
    • _workspace/ exists + a new batch of ideas → new run. Move the old _workspace/ to _workspace_<YYYYMMDD_HHMMSS>/, then Phase 1.

Phase 1 — prepare

  1. Parse the request into a concrete feature list. If the user named a roadmap cluster (e.g. "the cue sequencer + stage dashboard"), pull scope from docs/ROADMAP.md.
  2. Create _workspace/ and write the feature list to _workspace/00_input/features.md.

Phase 2 — design (sub-agent fan-out)

Spawn one td-architect per feature (or one for the whole small batch) via Agent, subagent_type: "td-architect", model: "opus", run_in_background: true for parallelism. Each writes _workspace/01_design_<feature>.md. Wait for all, then read the specs and resolve any flagged overlaps/contention before building.

Phase 3 — build (sub-agent fan-out, parallel)

Spawn N td-builder sub-agents in a single message (run_in_background: true), one per spec. Each prompt includes its spec path and the hard rule: new files only, never edit shared registries/CLI/docs. Each returns green-in-isolation files (vitest + biome) and a build note. Size N to the batch; keep it to a manageable parallel set (~3–5 at a time for a large batch, then a second wave).

Phase 4 — integrate + QA + release (agent team)

Form the team and run the convergence loop:

  1. TeamCreate(team_name: "tdmcp-delivery", members: [ { name: "integrator", agent_type: "td-integrator", model: "opus" }, { name: "qa", agent_type: "td-qa", model: "opus" }, { name: "builder-fixer", agent_type: "td-builder", model: "opus" }, { name: "releaser", agent_type: "td-releaser", model: "opus" }]) (builder-fixer closes QA's fix requests on handler/schema/test bugs without re-spawning.)
  2. TaskCreate the pipeline with dependencies:
    • integrate all built features (integrator)
    • QA each feature incrementally as it integrates (qa, depends on integrate) — not one big pass at the end
    • fix defects (builder-fixer/integrator, depends on QA findings)
    • release (releaser, depends on QA = PASS for everything shipping)
  3. Team self-coordinates via SendMessage: QA sends file:line + fix to the owner the instant a defect is found; boundary bugs go to both sides; QA re-validates after each fix (cap ~2–3 rounds/feature).
  4. Leader monitors with TaskGet, intervenes if a member stalls.

Phase 5 — cleanup + report

  1. Confirm _workspace/05_release.md exists (or that release was intentionally held).
  2. TeamDelete. Preserve _workspace/ for audit.
  3. Report to the user: features shipped, the released version + tag + SHA, anything held back (with reason), and any live-validation marked UNVERIFIED because the bridge was offline.

Data flow

[leader] → architect(s) ─ specs → builders (parallel) ─ files+notes →
   ┌─ TeamCreate(integrator, qa, builder-fixer, releaser) ─┐
   │  integrator wires → qa validates (live if bridge up)  │
   │     ↑ fix requests (SendMessage) ↓                    │
   │  builder-fixer / integrator patch → qa re-checks      │
   │  qa PASS → releaser cuts + pushes                     │
   └───────────────────────────────────────────────────────┘
                         ↓
              _workspace/0*_*.md + pushed tag

Error handling

SituationStrategy
One builder fails in isolationShip the rest; route its spec to builder-fixer in Phase 4 (or re-spawn). Don't block the batch.
Build breaks on integrateIntegrator bisects, wires the rest, sends the offender a precise fix; team continues.
QA finds a boundary bugSendMessage to both producer and consumer; re-validate after fix; cap ~2–3 rounds, else report blocker.
Bridge offlineQA runs offline gates, marks live checks UNVERIFIED-pending; release may still ship (note it) — never fail the pipeline for a missing TD.
QA = FAIL on a featureReleaser holds it for next cycle; ships only PASS features; report what was held.
Concurrent agent's WIP breaks compile project-wideValidate the slice in isolation (vitest run <file>); wait for their tree to go green before the full build/release.

Test scenarios

Normal: user asks to build the cue-sequencer feature → Phase 1 lists 1 feature → architect writes the spec → 1 builder produces a green tool+test → team forms → integrator wires + builds green → QA validates live (preview + post-cook errors clean) = PASS → releaser bumps minor, writes CHANGELOG, commits/tags/pushes → report names the new version.

Error: in a 4-feature batch, builder-3's tool cooks to a TD error caught by QA (a Level TOP gain no-op) → QA SendMessages builder-fixer with file:line + use brightness1 → fix re-validates PASS → the other 3 already PASS → releaser ships all 4. If builder-3 can't be fixed in 3 rounds, releaser ships the 3 PASS features and the report holds builder-3 for next cycle.

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

Use when manage multiple local CLI agents via tmux sessions (start/stop/monitor/assign) with cron-friendly scheduling.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

hybridlabor-api/aos62026년 10월 16일 업데이트

Use when a hybrid memory system that provides persistent, searchable knowledge management for AI agents (Architecture, Patterns, Decisions).

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

hybridlabor-api/aos62026년 10월 16일 업데이트

Reference for how BDB structures autonomous software engineering work — the seven-node dispatcher graph (Architect, TechLead, UI/UX, Engineering, Media/EventTech, Reviewer, Shipping) that /startcycle-graph actually runs. Use when you need the high-level lifecycle framing without inventing your own process.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

hybridlabor-api/aos62026년 10월 16일 업데이트

Tools are how AI agents interact with the world. A well-designed tool is the difference between an agent that works and one that hallucinates, fails silently, or costs 10x more tokens than necessary. This skill covers tool design from schema to error handling.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

hybridlabor-api/aos62026년 10월 16일 업데이트

Harness patterns for coding agents — memory, permissions, context engineering, delegation, skills, hooks, bootstrap.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

hybridlabor-api/aos62026년 10월 16일 업데이트

agenttrail: live map of a multi-agent build in the browser: which plan component is being worked on, by which agent or harness, what is done and what is stuck. Use when a multi-agent pipeline starts (/startcycle, /startcycle-graph, /teamwork-preview) or after a plan-canvas approve, when the user asks to see what the agents are doing ("live map", "build board", "who is running now"), or by yourself whenever a multi-agent run is under way — no user prompt needed.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

hybridlabor-api/aos62026년 10월 16일 업데이트

hybridlabor-api의 스킬 모두 보기

이 스킬의 문제 신고