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
- 参考文档code-quality-reviewer-prompt.md1.3 KB
- 参考文档implementer-prompt.md5.1 KB
- 参考文档spec-reviewer-prompt.md2.2 KB
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
WorkerTaskPacketjust in time before selected delegated work. - Dispatch
one-subagentwhen one safe lane is ready. - Dispatch
parallel-subagentswhen 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-debugmay 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-teamsonly when Codex work needs durable team state, explicit join-point tracking, result files, or lifecycle control beyond one in-session subagent burst.
Process
- Select the owning workflow: Route to canonical
sp-taskswhen a plan lacks an executable task graph. Route to canonicalsp-implementfor planned feature execution,sp-quickfor lightweight tracked work,sp-debugfor 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. - Select the route: Keep small/tightly coupled work leader-direct. For an
independent or parallel lane, compile a validated
WorkerTaskPacketfrom the current task, stable refs, and live code. - Dispatch in the current runtime: Use the
native-subagentssurface such as Codexspawn_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. Usemanaged-teamonly for durable team state or lifecycle control, and useleader-inlineonly as the owning workflow's selected mode. - 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.
- 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
WorkerTaskPacketor 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_pathsor changed paths. A subagent must not silently narrow assigned scope; if the set does not fit, the worker result returns top-levelacceptance=fail_gap, marks affected paths with path-levelcoverage[].outcome="overflow"orcoverage[].outcome="blocked", and includes the smallest safe split or recovery suggestion.overflowandblockedare 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 tosp-tasksfirst if tasks are not generated. If a manual next step is needed, tell the user to run$sp-implement, or$sp-taskswhen 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-fastmay 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-debugroute selectedsubagent-assistedexecution 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
launchdeck-agent
Codex host adapter contract fixture for Launchdeck.
launchdeck-agent
Claude host adapter contract fixture for 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.
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.
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.
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