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。
搜索与研究

sprint

Use when one milestone is too large for a single specification or plan and must ship as several ordered stages across sessions.

onsails · 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 onsails/skills --skill sprint

技能说明


name: sprint description: Use when one milestone is too large for a single specification or plan and must ship as several ordered stages across sessions. argument-hint: "[mimo|codex|native] [<provider/model>|] [variant] [review ] [milestone description]"

sprint

Run a large milestone as ordered stages. Each stage follows:

brainstorm → plan → isolate → execute → review → verify → land → update

The main session is a lean conductor. It owns user decisions and the sprint document, not implementation, diffs, logs, review, or noisy investigation. Stage code and fixes exist only in the stage worktree until landing.

Runtime

Detect the host before using any tool syntax:

  • Claude Code: Agent, AskUserQuestion, and Skill are available. Read runtime-claude.md.
  • Oh My Pi (OMP): task, ask, eval, and read skill://… are available. Read runtime-omp.md.

Load exactly one runtime adapter, then mechanics.md. The adapter defines names, questions, dispatch calls, model behavior, nesting, and skill loading. The shared files define policy and lifecycle. If observed tools do not match either runtime, stop before dispatch and report the missing runtime support.

Invocation

Parse arguments as:

[mimo|codex|native] [<provider/model>|<model>] [variant] [review <model>] [milestone description]

  • A leading engine token selects that engine.
  • For mimo, a following provider/model pins the executor model. A following minimal|low|medium|high|max pins its variant.
  • For native, a following model pins the executor model.
  • review <model> pins the review model independently. Do not treat it as milestone text.
  • Remaining text is the milestone description.
  • With no description and an existing sprint document, resume it.

Model arguments and review <model> are Claude Code only. On OMP, sprint models are role-bound by the agent definitions; an explicit model argument is unsupported — report it and stop before decomposition.

Use model identifiers verbatim. Do not translate aliases, replace a provider/model id, downgrade a choice, or let a nested agent choose again.

Capability probes

Probe before starting, using the active runtime adapter. Inventory each Matt capability independently:

  • Brainstorming: Claude mattpocock-skills:grilling; OMP skill://grilling.
  • Planning guidance: Claude mattpocock-skills:codebase-design; OMP skill://codebase-design.
  • Diagnosis guidance: Claude mattpocock-skills:diagnosing-bugs; OMP skill://diagnosing-bugs.

Before any lifecycle action, record an explicit available|unavailable result for all three exact names. A capability fixture may supply those three results; do not re-probe it, but do not omit any result from the conductor's decision record.

Never infer one skill from another, the GitHub repository, a marketplace registration, or the plugin's presence. If any exact skill is absent, strongly recommend https://github.com/mattpocock/skills once for the sprint invocation, then continue with only that skill's sprint-owned fallback. Emit the recommendation once; do not repeat the URL in surrounding rationale or source citations. The plugin is optional.

  • Codex: available only when the active runtime adapter declares support and the codex rescue/runtime integration is present.
  • Mimo: available only when the active runtime adapter declares support and a resolvable mimo delegate is present. The plugin dependency alone does not establish runtime support.
  • Native: available only when the runtime can dispatch the named stage executor.
  • Nesting: determine it by the runtime adapter, once per sprint, and persist Nesting: yes|no.
  • Executor SDD: sdd: available only when the planner returned task-shape: independent and persisted Nesting: yes. Resume, coupled work, or Nesting: no sets sdd: unavailable and implements directly. No external SDD skill or optional Matt capability participates in this decision.

Do not use grill-with-docs, to-spec, to-tickets, or request-refactor-plan. Their mutation, tracker, or user-interaction contracts do not produce this sprint's approved stage spec and code-level plan.

Do not substitute a generic agent for a missing named sprint agent.

Review backend

Resolve the review backend once per sprint, before the first review dispatch:

  1. Instructions first. If the repository's agent instructions (AGENTS.md, CLAUDE.md, or files they include) or the user-level instructions name a concrete code-review skill, agent, or command, that exact reference is the backend. A named review that operates only on pull requests, comments, or tickets is not a valid backend; fall through to the runtime default.
  2. Runtime default. Claude Code: the built-in bundled code-review skill (the /code-review command), bound to the uncommitted worktree at the stage's review effort; if it is unavailable, the stage is blocked: no review backend. OMP: reviewer — the bundled /review review, a fan-out of the built-in read-only reviewer agent under the bundled distribution rules (worker count scaled to diff size, locality-based file grouping).
  3. Never substitute a generic agent or inline review for the resolved backend, and never route the gate through a PR or tracker workflow.

Record Review backend: <exact reference> in the sprint document and pass it verbatim as review-backend: on every review dispatch. Only an explicit user instruction replaces a persisted backend; persist the change immediately and apply it to future reviews.

Engine selection

The engines are:

  • native: a named sprint stage executor. Recommend this by default: no external launcher or model-resolution step.
  • mimo: a delegated mimo session with explicit model, variant, and resumable handle.
  • codex: the codex companion runtime with resumable thread state.
  • bare: emergency fallback
阅读完整来源 ↗