AgentSkillsCN

he-implement

在构建过程中，利用生成的项目上下文文档作为参考，采用依赖感知的并行批次、隔离的任务作业以及集成检查点，高效执行各项行动计划。

SKILL.md
--- frontmatter
name: he-implement
description: Executes active plans using dependency-aware parallel batches, isolated task work, and integration checkpoints while using generated project context docs for reference. Use during build execution.
argument-hint: "[slug or docs/plans/active/<slug>.md]"

HE Implement

Execute plan tasks with DAG-batch parallelism.

Inputs

  • •docs/plans/active/<slug>.md (plan_mode: lightweight|execution)

Generated Context

Before starting execution, refresh generated context if stale:

  • •docs/generated/db-schema.md (if present)
  • •docs/generated/api-schema.md (if present)
  • •docs/generated/component-tree.md (if present)
  • •docs/generated/dependency-graph.md (if present)
  • •Other generated reference docs in docs/generated/ (if present)

Each generated file should include a last_updated timestamp.

Execution Model

  1. •Parse task DAG/checklist from active plan and load concrete execution details from Task Details.
    • •For execution mode: load files_to_change, tests_to_run, verify_commands.
    • •For lightweight mode: load files, steps, test_type, verify, done_when.
  2. •Build next ready batch (all dependencies satisfied).
  3. •Launch one subagent per task in the batch. Each subagent receives the task details, target files, and relevant generated context. Run all subagents in the batch concurrently.
  4. •Collect subagent results and integrate batch outputs.
  5. •Repeat until no remaining tasks.

Use subagents aggressively — every independent task in a batch should be its own subagent. Keep the main context clean by offloading implementation work to subagents and only handling integration and plan updates in the main thread.

Subagent Return Contract

Each subagent returns:

  • •changed files
  • •planned target files (files_to_change or files) and whether each was touched (or explicit no-change reason)
  • •tests run and results (must match task's test_type — unit or e2e, no mocks)
  • •unresolved risks with priority
  • •integration notes

Integration Rules

  • •Integrate one batch at a time.
  • •If conflicts occur, split tasks or sequence conflicting tasks.
  • •Re-run targeted tests for integrated batch.
  • •If a task is missing concrete file paths or verify commands in Task Details, send it back to planning before execution.

Branch and PR Convention

  • •Work on the initiative's branch (one branch per slug).
  • •Use worktree-per-task for parallel execution when blast radius is non-trivial.
  • •Create PR at the implement-to-review boundary.

Plan Progress Updates

Update docs/plans/active/<slug>.md after each batch:

  • •Update Task DAG statuses (todo|in_progress|blocked|done) — this is the single source of truth
  • •Check/uncheck step checkboxes (implementation_steps in execution mode, steps in lightweight mode) to reflect executed work
  • •Append Progress Log entry with evidence
  • •Append Decision Log entry when scope/approach changes

Exit Gate

  • •All planned tasks completed or explicitly deferred
  • •Plan progress is current
  • •Docs commit gate passes

Transition Options

Present 2-3 explicit next-step options with a recommended default. Use request_user_input (Codex) or AskUserQuestion (Claude Code) in Plan mode; otherwise ask in chat. Wait for user selection before proceeding.

At least one option must be Next step: he-review.