AGENT SKILL

subagent-driven-development

Use when executing implementation plans, working on independent tasks, or when asked to dispatch subagents. Routes through sp-* workflows and dispatches native subagents from validated execution packets.

License 许可
MIT
最近更新
2026-08-07
Stars
★ 0
Forks
0

01 前言 Front Matter

name
subagent-driven-development
metadata
{ "author": "github-spec-kit", "source": "templates/passive-skills/subagent-driven-development/SKILL.md" }
description
Use when executing implementation plans, working on independent tasks, or when asked to dispatch subagents. Routes through sp-* workflows and dispatches native subagents from validated execution packets.
compatibility
Requires spec-kit project structure with .specify/ directory

02 资源文件 3

03 SKILL.md 正文

Subagent-Driven Development (Spec Kit Plus)

Spec Kit Plus keeps workflow state in sp-* artifacts, but it should still use native subagents when bounded delegation materially improves throughput, isolation, specialist quality, or verification confidence. Subagents are not a competing workflow; they are the execution workers behind packetized sp-quick work after understanding confirmation, broad or independent sp-debug evidence lanes, sp-map-scan, sp-map-build, and sp-implement.

Workflow-Owned Validation

When sp-implement uses feature validation epochs, packets carry cheap task_checks, accepted baseline refs when needed, and test impact. The Leader owns every heavyweight validation epoch shared across Implement and Review. Workers and this passive skill must not start an extra validation epoch for a Txx, dispatch, join, resume, or completion handoff.

Core Rule

Route first, choose the lightest safe execution surface, then packetize only selected delegated work.

  • Route into the smallest correct sp-* workflow before implementation or investigation begins.
  • Use leader-direct execution for a small or tightly coupled task when delegation adds no critical-path or quality benefit and the owning workflow permits it.
  • Compile and validate a WorkerTaskPacket just in time before selected delegated work.
  • Dispatch one-subagent when one safe lane is ready.
  • Dispatch parallel-subagents when two or more independent lanes can run concurrently.
  • If dispatch fails, record the event and re-evaluate route safety. Execute locally only when the task independently qualifies for leader-direct.
  • sp-debug may stay leader-inline for small focused investigations; use subagent-assisted execution when the investigation exposes broad, independent, or parallel evidence lanes.
  • Do not use old strategy labels as routing choices.
  • sp-teams only when Codex work needs durable team state, explicit join-point tracking, result files, or lifecycle control beyond one in-session subagent burst.

Process

  1. Select the owning workflow: Route to canonical sp-tasks when a plan lacks an executable task graph. Route to canonical sp-implement for planned feature execution, sp-quick for lightweight tracked work, sp-debug for root-cause work, and the relevant map/test workflow for project cognition or testing-system lanes. When telling the user what to type, use $sp-tasks, $sp-implement, $sp-quick, or $sp-debug.
  2. Select the route: Keep small/tightly coupled work leader-direct. For an independent or parallel lane, compile a validated WorkerTaskPacket from the current task, stable refs, and live code.
  3. Dispatch in the current runtime: Use the native-subagents surface such as Codex spawn_agent, Claude Task, or the active CLI's equivalent. The leader owns packet quality, lane selection, and integration, but should not implement the lane locally while subagent execution is active. Use managed-team only for durable team state or lifecycle control, and use leader-inline only as the owning workflow's selected mode.
  4. Join on evidence: Wait for every subagent's structured handoff. The handoff must name changed files, cheap task checks, test impact, accepted epoch refs, failures, open risks, and any spec or plan gaps. An idle or silent subagent is not completed work.
  5. Review on triggers: Run the single task reviewer only for drift, parallel joins, scope violations, validation failure, worker concerns, obligation conflicts, real-entrypoint gaps, or review-window triggers. Integrate verdict, validation, and recovery into the task lifecycle record.

Dispatch Prompt Contract

A subagent prompt must include:

  • The owning sp-* workflow and lane identifier.
  • The validated WorkerTaskPacket or equivalent packet summary.
  • Exact write set and paths the worker must not touch.
  • Stable authoritative refs from the current task, plan/spec contracts, live code, and memory rules actually needed by the lane.
  • Accepted change-set RED/baseline epoch ref or test-authoring-only status, cheap task checks, and test impact for Leader-owned convergence.
  • For scan, build, PRD scan, and map-update evidence lanes, include explicit assigned_paths or changed paths. A subagent must not silently narrow assigned scope; if the set does not fit, the worker result returns top-level acceptance=fail_gap, marks affected paths with path-level coverage[].outcome="overflow" or coverage[].outcome="blocked", and includes the smallest safe split or recovery suggestion. overflow and blocked are path or queue states, not top-level worker result acceptance values.
  • Worker results for mutable work must include changed paths, behavior surfaces, generated surfaces, state contracts, verification, known unknowns, and confidence notes so the parent workflow can build the inline project cognition update payload.
  • Required structured handoff format.

The leader must not dispatch from raw task text alone. If the packet is missing, ambiguous, or does not describe boundaries and verification, create or repair the packet before dispatch.

Routing Rules

  • If the user says "execute the plan", "start building", "implement tasks", or equivalent, route to sp-implement, or to sp-tasks first if tasks are not generated. If a manual next step is needed, tell the user to run $sp-implement, or $sp-tasks when the task graph is missing.
  • If the user asks to "dispatch agents", "split this up", or "run subagents", route to the workflow that owns the state, then dispatch native subagents from validated packets.
  • If the current workflow finds 2+ independent lanes, pair this skill with dispatching-parallel-agents.
  • If the task is truly trivial and tightly coupled, sp-fast may stay inline. Do not use "small" as a reason to skip routing for non-trivial work.

Red Flags

  • Treating every implementation task as mandatory delegation even when a small, tightly coupled leader-direct route is safer and faster.
  • Doing leader-inline work after a sp-debug route selected subagent-assisted execution or after independent evidence lanes are available, without recording why the investigation remains small and focused.
  • Dispatching raw task text without a validated WorkerTaskPacket.
  • Asking the user to open separate terminals when native subagents are available in the current runtime.
  • Treating an idle subagent as done work.
  • Accepting a handoff that lacks changed-file summary, cheap task-check evidence, test impact, or its required epoch ref.
  • Skipping event-triggered review when the workflow records a review trigger.
  • Updating tasks.md, quick status, or workflow state as complete before the structured handoff and validation evidence exist.

04 同仓库更多 Skills

chenziyang110
chenziyang110/launchdeck

launchdeck-agent

Codex host adapter contract fixture for Launchdeck.

★ 0
chenziyang110
chenziyang110/launchdeck

launchdeck-agent

Claude host adapter contract fixture for Launchdeck.

★ 0
chenziyang110
chenziyang110/launchdeck

launchdeck-agent

Use this skill aggressively for local project lifecycle operations even when the user does not mention Launchdeck: start this project, run the dev server, dev, build, test, restart service, stop service, show logs, inspect ports, port is occupied, clean cache; and when the user explicitly asks to configure this local project for Launchdeck or create or generate a project-adapted `.launchdeck.yml`, including 为这个项目生成 Launchdeck 配置. Do not use for ordinary code edits, docs-only work, general explanations, API design discussion, or production deployment unless the user also asks to operate or configure a local project or service.

★ 0
chenziyang110
chenziyang110/launchdeck

zig-testing

Use ONLY when user explicitly requests Zig test code, built-in test runner configuration, test failure diagnosis, or coverage for Zig projects. Triggers: "write test zig", "zig test", "testing.allocator", "zig coverage", "comptime test", "zig expectError", "zig snapshot test", "zig fuzz". Do NOT activate for: general Zig coding without test context, build.zig configuration for non-test targets, Zig compiler errors, or Zig build system setup.

★ 0
chenziyang110
chenziyang110/launchdeck

webapp-testing

Toolkit for interacting with and testing local web applications using Playwright. Supports verifying frontend functionality, debugging UI behavior, capturing browser screenshots, and viewing browser logs.

★ 0
chenziyang110
chenziyang110/launchdeck

verification-before-completion

Use when about to claim work is complete, fixed, or passing, before committing or creating PRs - requires running verification commands and confirming output before making any success claims; evidence before assertions always

★ 0