Agent Skills › chrisbanes/skills › implement-with-subagents

implement-with-subagents

GitHub

通过子代理实现或审查任务编排,确保原子性、依赖顺序及隔离执行。支持审查与实施模式,验证依赖图并处理提交合并冲突,保障多任务并发安全与正确集成。

skills/implement-with-subagents/SKILL.md chrisbanes/skills

Trigger Scenarios

需要按依赖关系并行执行多个开发任务 审查任务队列的原子性与依赖顺序 在多子代理环境下进行代码集成与验证

Install

npx skills add chrisbanes/skills --skill implement-with-subagents -g -y
More Options

Use without installing

npx skills use chrisbanes/skills@implement-with-subagents

指定 Agent (Claude Code)

npx skills add chrisbanes/skills --skill implement-with-subagents -a claude-code -g -y

安装 repo 全部 skill

npx skills add chrisbanes/skills --all -g -y

预览 repo 内 skill

npx skills add chrisbanes/skills --list

SKILL.md

Frontmatter
{
    "name": "implement-with-subagents",
    "description": "Use when implementing or reviewing the orchestration of supplied tickets or plan tasks through separate implementation subagents, including queue atomicity, task-scoped commit acceptance, and repair ownership.",
    "compatibility": "Review mode has no external skill dependency. Implementation mode requires Matt Pocock's separately installed `implement` workflow; its current contract also invokes `tdd` and `code-review`.",
    "disable-model-invocation": true
}

Implement with subagents

Keep orchestration and implementation ownership separate: the controller schedules, and one implementation subagent owns each work item through completion. Validate task dependencies before dispatch. Run only ready, independent tasks concurrently in isolated worktrees; integrate accepted commits in dependency order and recheck affected evidence at the integrated head. Use serial execution when safe isolation or capacity is unavailable. Return failed acceptance evidence to the same owner, never repair it in the controller.

Check the prerequisite

Review mode has no external dependency. Implementation mode requires the implement skill from Matt Pocock's skill set. It is not bundled here. Never install it implicitly; report npx skills add mattpocock/skills and require the user to select implement, tdd, and code-review when it is absent.

Select the mode

  • Use review only to assess supplied orchestration without running it.
  • Use implement only to execute supplied tickets or plan tasks.

Review procedure

  1. Inspect only permitted repository and orchestration state. Do not start an agent, edit, commit, or contact a remote service.
  2. Assess queue atomicity, dependency order, implementation ownership, task-scoped acceptance, repair ownership, and controller mutation boundaries. Acceptance requires a task-scoped commit, independent inspection of its full diff, and requested task-level validation evidence before advancing. Separate worktrees do not make tasks independent when they edit the same file or one relies on another's unmerged code; serialize those tasks. Before dispatching a dependent, integrate each accepted prerequisite, check the joined diff for textual and semantic conflicts, and rerun affected validation at that exact integration head. Do not proceed until those checks pass; an isolated prerequisite commit alone is insufficient. For example, dispatch independent API and DOC, integrate their accepted commits, validate their joined head, and only then dispatch WIRE. Return failed acceptance to the same owner; reuse passing checks only when their inputs and environment remain unchanged. An accepted item is complete, not assignable again.
  3. Report the next action (or no action), evidence, and any acceptance gap. Stop before implementation. For an invalid dependency graph, hold dispatch and return the specific defects to the plan owner for clarification; do not choose which dependency to remove, invent a missing task, or prescribe a corrected queue as if that choice were authorized. Separate task-commit acceptance from integrated-head acceptance: explicitly require a textual and semantic conflict check of the joined diff and rerun affected validation at that exact head before accepting the join or dispatching a dependent. If an integration changed a validated input, name that rerun as a required acceptance action even when an invalid dependency graph also blocks dispatch. Calling the earlier report stale does not state the action.

Implementation mode

Before delegation, read the implementation-mode procedure completely. It is mandatory for implementation mode and owns queue construction, runtime selection, owner packets, independent acceptance, integration, repairs, and final validation. Stop when its dependency, task-scoped commit, or capability gate cannot be satisfied; never implement an item in the controller.

Runtime mapping

Select an implementation-capable owner and apply the implementation procedure's capability checks:

Runtime Implementation owner
Codex worker
Claude Code general-purpose
OpenCode general subagent; build is primary
Pi No built-in role; inspect its delegation extension and agent definitions. A bare or non-resumable Pi cannot own an item: stop and report it.
Other runtimes An exposed implementation-capable subagent that passes the checks

Finish gate

In review, finish only with the non-mutating assessment, evidence, and any acceptance gap. In implement, finish only when every queued item has a task-scoped, reviewed, verified commit, the final worktree matches the recorded pre-existing state, and no owner-reported blocker remains. Report the item-to-commit mapping and final validation; otherwise finish blocked at the first incomplete gate.

Version History

  • 2026.9.25 Current 2026-09-27 20:23

    增强对任务ID和依赖边的显式支持,优化并发调度逻辑,细化依赖感知下的工作树隔离与集成验证流程。

  • 2026.9.21 2026-09-22 08:34

    实现子代理选择的跨运行时可移植性,适配 Codex、Claude Code 和 OpenCode 等环境的能力映射,并更新相关文档与测试用例。

  • 2026.9.2 2026-09-09 02:59

    明确implement为外部工作流并使其要求可操作;记录所有外部提供者;为implement-with-subagents添加缺失提供者的校准用例

  • 2026.9.2 2026-09-03 04:30

    将高级工作流技能标记为显式仅用,更新UI元数据和SKILL.md前置信息,调整自动评估排除规则。

  • 2026.8.27 2026-08-28 05:51

    恢复技能评估行为,增加遥测数据捕获(如令牌、工具调用、耗时),对比基线与自动效率,并同步结果表至根目录和 README。

  • 2026.8.24 2026-08-27 17:15

    减少技能指令的冗余度

  • 2026.8.16 2026-08-19 19:31

Dependencies

  • suggested mattpocock/skills

Same Skill Collection

skills/android-benchmark-comparison/SKILL.md
skills/compose-animations/SKILL.md
skills/compose-component-design/SKILL.md
skills/compose-focus-navigation/SKILL.md
skills/compose-modifier-and-layout-style/SKILL.md
skills/compose-performance/SKILL.md
skills/compose-recomposition-performance/SKILL.md
skills/compose-side-effects/SKILL.md
skills/compose-slot-api-pattern/SKILL.md
skills/compose-stability-diagnostics/SKILL.md
skills/compose-state-and-effects/SKILL.md
skills/compose-state-authoring/SKILL.md
skills/compose-state-deferred-reads/SKILL.md
skills/compose-state-hoisting/SKILL.md
skills/compose-state-holder-ui-split/SKILL.md
skills/compose-ui-testing-patterns/SKILL.md
skills/gradle-run/SKILL.md
skills/grounded-writing/SKILL.md
skills/implement-issue/SKILL.md
skills/kotlin-api-design/SKILL.md
skills/kotlin-concurrency-and-flow/SKILL.md
skills/kotlin-control-flow/SKILL.md
skills/kotlin-coroutines-structured-concurrency/SKILL.md
skills/kotlin-flow-state-event-modeling/SKILL.md
skills/kotlin-functions/SKILL.md
skills/kotlin-multiplatform-expect-actual/SKILL.md
skills/kotlin-types-value-class/SKILL.md
skills/release-kotlin-library/SKILL.md
skills/run-github-project/SKILL.md
skills/shepherd/SKILL.md
skills/to-plan/SKILL.md
skills/using-chrisbanes-skills/SKILL.md

Metadata

Files
0
Version
2026.9.25
Hash
4dbf948c
Indexed
2026-08-19 19:31

Главная - Вики-сайт
Copyright © 2011-2026 iteam. Current version is 2.155.2. UTC+08:00, 2026-09-30 15:56
浙ICP备14020137号-1