跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
开发与测试

aidd-dev:00:sdlc

Pure orchestrator for the full AIDD development flow. Use when a human (or Gardener) needs to take a free-form request from idea to shipped code, end-to-end. Coordinates spec generation, planning, implementation, and review by composing other skills and agents. Holds no business logic of its own — every step is delegated.

ai-driven-dev · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add ai-driven-dev/framework --skill aidd-dev:00:sdlc

技能说明


name: aidd-dev:00:sdlc description: Pure orchestrator for the full AIDD development flow. Use when a human (or Gardener) needs to take a free-form request from idea to shipped code, end-to-end. Coordinates spec generation, planning, implementation, and review by composing other skills and agents. Holds no business logic of its own — every step is delegated.

Skill: sdlc

Pure orchestrator. Takes a free-form request and drives it through spec → plan → implementation → done by composing other skills and agents. This skill is a workflow descriptor: every step delegates to a skill or an agent. No business logic lives here.

Iron rule

You are the conductor, not a player.

You never write code. You never run tests. You never commit. You never load a skill into your own context.

Every action — including invoking another skill — is delegated to a sub-agent that runs in a fresh context. You only ever read structured outputs (YAML) returned by sub-agents. You never read the contents of files produced (spec.md, plan.md, source code) — those belong to the agents that wrote them.

If you find yourself about to use the Skill, Write, Edit, Bash tool yourself for anything other than reading a sub-agent output or routing it to the next phase, you have violated the protocol. Stop and re-delegate.

Skill invocation protocol

You never call Skill(...) directly. Every skill invocation is wrapped in an Agent spawn:

  1. Spawn the Agent tool with subagent_type: general-purpose.

  2. Prompt template:

    Invoke skill `<skill-id>` with these inputs:
    <yaml inputs>
    
    Run the skill in your fresh context. When done, return ONLY the structured output described by the skill's frontmatter, in YAML, with no surrounding prose. Do not include the file contents you produced.
    
  3. Parse the agent's YAML response. Use the values, not the produced content.

This applies to: aidd-pm:05:spec, aidd-vcs:01:commit, aidd-vcs:02:pull-request. Skills used by planner, implementer, or reviewer are loaded inside those agents' own contexts — that already satisfies the rule.

When to use

  • A human invokes /sdlc <request> to start a new run from a free-form description
  • A Gardener (async) decides a new run is needed and invokes this skill
  • A previous run was paused and needs to be resumed with refined input

When NOT to use

  • A spec already exists and is validated → invoke the planner agent directly
  • A milestone implementation is needed in isolation → invoke the implementer agent directly
  • A standalone review is needed → invoke the reviewer agent directly
  • A human just wants a draft spec without execution → invoke aidd-pm:05:spec directly

Inputs

At least one of request or prd_path is required. If both are provided, prd_path wins and the request becomes additional context.

request: <free-form human description of what to build, optional>
prd_path: <path to an existing PRD file, optional>
working_dir: <optional — defaults to current working directory>
options:
  quality_threshold: <0-100, default 90>
  retry_cap_per_step: <n, default 3>
  open_pr: <true | false, default false>

Commits are always produced — at every phase boundary and at every ticked checkbox during implementation. There is no auto_commit: false mode. The audit trail is mandatory.

Outputs

status: done | blocked | aborted
spec_path: <path to validated spec>
plan_summary: <count of milestones, count of decisions made/blocked>
commits:
  spec_validated: <sha>
  plan_validated: <sha>
  per_milestone: [<sha>, ...]
  final: <sha if status=done>
pr_url: <if open_pr and status=done>
duration: <hh:mm>
notes: <relevant observations>

Orchestration

Four phases. Each phase delegates. Defaults below are overridable per invocation via options.

Phase 1 — Spec generation and validation

  1. Spawn a sub-agent (per Skill invocation protocol) to invoke aidd-pm:05:spec with inputs {request, prd_path, working_dir}. The agent loads the skill in its fresh context, writes the draft spec at working_dir/spec.md, and returns {spec_path, status, notes}. You never load the skill yourself; you only read the YAML output.
  2. Spawn reviewer agent (Agent tool, subagent_type: reviewer) with:
    • artifact: the draft spec at the returned spec_path
    • validator: {{TOOLS}}/plugins/aidd-pm/skills/05-spec/assets/spec-validator.yml (filesystem path, resolved by Claude Code)
  3. Read the reviewer's output:
    • completion_score = 100 and quality_score >= quality_threshold → proceed to step 4
    • quality_score < quality_threshold → re-spawn the spec sub-agent with existing_spec and findings, retry
    • Beyond retry_cap_per_step retries → escalate to human, return with status: blocked
  4. Commit the validated spec. Spawn a sub-agent (per Skill invocation protocol) to invoke aidd-vcs:01:commit with {mode: auto, message: "spec: <feature-name> validated", files: [<spec_path>], push: false}. Record the returned commit_sha in commits.spec_validated.

Phase 2 — Planning

  1. Spawn planner agent with the validated spec.
  2. The planner uses aidd-dev:01:plan internally. The skill writes to aidd_docs/tasks/<yyyy_mm>/...md (simple plan) or to *-master.md plus one *-part-N.md per child plan (master plan). The planner returns plan_path and child_paths in its output — use those, never assume a hardcoded path.
  3. Read the planner's output:
    • plan_path empty → planner failed to produce a plan, retry once then escalate
    • decisions_blocked non-empty → ask the human, get answers, re-spawn the planner with the answers as input
    • plan_status: blocked → escalate, return with status: blocked
    • Beyond retry_cap_per_step retries → escalate to human
  4. Commit the validated plan. Spawn a sub-agent (per Skill invocation protocol) to invoke aidd-vcs:01:commit with `{mode: auto, message: "plan: val
阅读完整来源 ↗