跳到正文
cccskills
免费在 GitHub 公开

plan-pipeline-execute

Carry out a validated development plan in an isolated Git worktree, a separate working directory. Coordinate parallel implementation, run quality checks, update documentation, and take the changes through pull request creation, merging, execution metrics, and cleanup.

中文简介正在准备中。当前显示英语简介。

原文查看原文说明

Execute a validated plan: worktree isolation, TDD scaffolding, level-based parallel agents, quality gate with smoke test, PR creation and merge. Handles everything through to merged PR.

集成工具GitHub
查看安装方法

适用场景

  • Implementing a validated development plan
  • Running dependent tasks in parallel stages
  • Preparing failing tests before implementation
  • Completing checks, documentation, and PR merging

包含的文件(1)

  • SKILL.md7.1 KB

SKILL.md(原文)

安装前可以查看交给 AI 代理的指令内容。

/plan-pipeline:execute: Execution to merged PR

Execute the validated plan in an isolated worktree. Spawn per-task agents, verify quality, create and merge the PR. Handles everything through to cleanup.

Run /clear before this command.


Prerequisite

A validated plan must exist at docs/plans/plan-{name}.md with all issues resolved (output of /plan-pipeline:validate).


Step 1: Worktree setup

Create an isolated git worktree:

git worktree add .worktrees/{plan-name} -b feature/{plan-name}

All execution happens inside the worktree. Main branch remains clean throughout.


Step 2: TDD scaffolding

Only for tasks marked as TDD in the plan.

For each TDD task, before any implementation:

  1. Write the failing test(s) that define the acceptance criteria
  2. Run tests to confirm they fail (red)
  3. Commit the failing tests
  4. Mark the test file in the task for the implementation agent to find

Do not write implementation code in this step.


Step 3: Level-Based parallel execution

Parse the task list from the plan. Group tasks by layer (Layer 1 = foundation, Layer 2 = depends on Layer 1, etc.).

For each layer:

  1. Identify all tasks in the layer
  2. Spawn one agent per task in parallel (Task tool, run_in_background: true)
  3. Each agent receives: its task description, files to modify, acceptance criteria, and relevant ADRs
  4. Monitor all agents by reading the output path reported by the agent via Read (TaskOutput was removed in v2.1.278)
  5. Each agent commits on task completion: git commit -m "feat: {task-description}"
  6. Wait for all tasks in the layer to complete before starting the next layer

Drift detection: after each layer, diff the actual changes against the plan spec. If implementation deviates significantly from the plan (new files not in plan, plan files not touched), flag and ask how to proceed. Do not silently continue on drift.

Agent instructions for each task:

You are implementing one task from a validated plan.
Task: {description}
Files to modify: {file list}
Acceptance criteria: {criteria}
Relevant ADRs: {adr list}

First principles:
- Build state-of-the-art. No workarounds, no legacy patterns.
- Fix at the correct architectural level, never with component-level hacks.
- If you discover that the plan is wrong or missing context, stop and report. Do not improvise architecture.

Commit your changes when complete with message: "feat: {task-description}"

Step 4: Quality gate

Run in parallel:

  • Linter
  • Type checker (if applicable)
  • Full test suite

If all pass: proceed to smoke test.

If any fail: spawn a quality-fixer debug agent with the failure output. It gets up to 3 auto-fix attempts. After each attempt, re-run the quality gate. If still failing after 3 attempts: stop, report the failure with the full error output, and wait for human intervention.

Integration smoke test (skip for pure frontend or docs-only plans):

Run the smoke commands defined in the plan's ## Integration Verification section. Additionally:

  • If GraphQL: run an introspection probe to verify schema is accessible
  • If Docker services: scan container logs for ERROR-level entries
  • If new API routes: verify each returns expected status codes

Smoke test failures are debugged by a quality-fixer-smoke agent with the same 3-attempt limit.


Step 5: Pre-PR documentation

In the worktree, before creating the PR.

PRD Reconciliation: compare the implemented behavior against the original PRD. Note any deviations or additions discovered during implementation. Update the PRD with actuals. These updates ship in the same PR as the feature.

Plan Archival: move docs/plans/plan-{name}.md to docs/plans/completed/plan-{name}.md. Update the status header.

Commit documentation updates: docs: reconcile PRD and archive plan for {feature-name}.


Step 6: Push and PR

Push the worktree branch and create the PR:

git push origin feature/{plan-name}
gh pr create \
  --title "{feature-name}: {one-line summary from plan}" \
  --body "$(cat .pr-body.md)"

PR body template:

## Summary
{plan summary paragraph}

## Changes
{auto-generated from task list: bullet per task with files affected}

## ADRs
{list of ADRs created during this plan}

## Test Plan
{from plan test plan section}

## Smoke Test Results
{output from integration verification}

Merge using squash:

gh pr merge --squash --delete-branch

Step 7: Post-Merge metrics

Switch back to develop/main. Update docs/plans/metrics/{name}.json with execution data:

  • Task count and per-layer breakdown
  • TDD task count
  • Diff stats (files changed, lines added/removed)
  • Quality gate results (pass/fail, fix attempts)
  • Smoke test results
  • Drift score (0-1, how closely implementation matched plan)
  • PR data (number, merge commit, timestamp)

Commit metrics update.


Step 8: Worktree cleanup

git worktree remove .worktrees/{plan-name}

Usage

/plan-pipeline:execute

Picks up the most recent validated plan. Or specify:

/plan-pipeline:execute plan-user-authentication

Output

Setting up worktree: .worktrees/user-authentication
Branch: feature/user-authentication

TDD scaffolding: 2 tasks marked TDD
  ✓ Written failing tests for: auth-token-validation
  ✓ Written failing tests for: refresh-token-rotation
  Committed: "test: failing tests for auth pipeline (TDD)"

Executing Layer 1 (3 tasks, parallel)...
  [agent-1] Implementing: JWT token generation service
  [agent-2] Implementing: User session model
  [agent-3] Implementing: Auth middleware
  ✓ Layer 1 complete. 3 commits.

Drift check: Layer 1... ✓ No drift detected.

Executing Layer 2 (2 tasks, parallel)...
  [agent-4] Implementing: Login endpoint
  [agent-5] Implementing: Refresh endpoint
  ✓ Layer 2 complete. 2 commits.

Quality gate...
  ✓ Lint passed
  ✓ Type check passed
  ✓ Tests: 47 passed, 0 failed

Smoke test...
  ✓ GraphQL introspection: OK
  ✓ POST /api/auth/login: 200
  ✓ POST /api/auth/refresh: 200

Pre-PR docs...
  ✓ PRD reconciled (1 minor deviation noted)
  ✓ Plan archived to docs/plans/completed/

PR created: #142 "user-authentication: JWT auth with refresh token rotation"
PR merged (squash). Branch deleted.

Metrics committed. Worktree cleaned.
✅ Feature complete.

When to use

After /plan-pipeline:validate confirms all issues are resolved. Never skip validation: executing an unvalidated plan skips the independent review that catches ~18 issues on average.

Pipeline position

/plan-pipeline:ceo-review    → product direction locked
/plan-pipeline:eng-review    → architecture locked
/plan-pipeline:start         → produce implementation plan
/plan-pipeline:validate      → validate before execution
/plan-pipeline:execute       → execute to merged PR          ← you are here

评价

还没有评价。欢迎分享你的使用感受。

评论

无需登录即可发表。欢迎分享使用感受、提问,或说明在哪些环境下无法运行。

还没有评论。来写第一条吧。

提交后会立即公开。请勿填写邮箱、电话等个人信息。URL 最多 2 个。

0 / 1000 字

此表单受 reCAPTCHA 保护,适用 Google 的隐私权政策和服务条款。

同一仓库的技能

Evaluate Claude Code agents, skills, and commands with weighted quality criteria. Get scores, readiness grades, template comparisons, and practical improvement suggestions in Markdown or JSON reports to guide reviews before release.

中文简介正在准备中。当前显示英语简介。

  • Review skills before release
  • Compare agent definitions with reference templates
  • Quickly check critical quality criteria
FlorianBruniaux/claude-code-ultimate-guide62042026年10月18日 更新

Assess a codebase across seven areas, including exposed credentials, security, dependencies, structure, and tests. Receive category scores, findings tied to file locations, and a prioritized improvement plan with smaller tasks that can be completed quickly.

中文简介正在准备中。当前显示英语简介。

  • Assessing the overall health of an inherited project
  • Reviewing exposed credentials and security risks
  • Identifying dependency and test coverage gaps
FlorianBruniaux/claude-code-ultimate-guide62042026年10月18日 更新

Measure code patterns such as casts to any or disabled lint rules, then run repeated, targeted edits. Keep changes that improve the chosen metric, revert unsuccessful changes, and track results for each improvement loop.

中文简介正在准备中。当前显示英语简介。

  • Finding code patterns to prioritize for cleanup
  • Preparing measurement files for an improvement loop
  • Repeating targeted edits to reduce casts to any
FlorianBruniaux/claude-code-ultimate-guide62042026年10月18日 更新

best-of-n

免费

Generate several separate solutions, compare them against criteria fixed in advance, and verify the selected result through executable checks or an independent reviewer. Keep a proof log documenting candidates, evidence, failures, and remaining limits.

中文简介正在准备中。当前显示英语简介。

  • Comparing implementation options before choosing one
  • Verifying a solution when a wrong choice would be costly
  • Keeping evaluation evidence for delivery or audit
FlorianBruniaux/claude-code-ultimate-guide62042026年10月18日 更新

canary

免费

Monitor a live application after deployment and compare it with a saved baseline, a record of its earlier state. Check page responses, loading times, and new errors, flag persistent changes, and produce a per-page health report.

中文简介正在准备中。当前显示英语简介。

  • Recording a healthy baseline before deployment
  • Monitoring production immediately after deployment
  • Checking key pages for slower responses or failures
FlorianBruniaux/claude-code-ultimate-guide62042026年10月18日 更新

catchup

免费

Rebuild project context after clearing a conversation or returning to work. Review recent Git history, uncommitted changes, and TODO markers to get a structured summary of project status and suggested next steps.

中文简介正在准备中。当前显示英语简介。

  • Restoring project context after /clear
  • Reviewing recent changes after a long break
  • Checking unfinished work before choosing the next task
FlorianBruniaux/claude-code-ultimate-guide62042026年10月18日 更新

查看 FlorianBruniaux 的全部技能

报告此技能的问题