README.md

sprint

Orchestrate one large milestone as a series of brainstorm → plan → execute stages, tracked in a living sprint doc, with implementation delegated to native subagents on Claude Code or Oh My Pi (OMP), or to mimo/codex on Claude Code, while the conductor stays in charge.

You are the foreman. The executor digs.

What it's for

A milestone too big for a single spec-and-plan — a long, multistage effort that spans sessions and needs several rounds of brainstorming and planning. Instead of one giant plan, you decompose it into stages and drive them one at a time, resuming across sessions from a persistent doc.

Use when:

• A milestone needs multiple brainstorm/plan rounds, not one spec → done.
• Work spans sessions and you need to resume "what stage am I on".
• You want to hand the coding to an executor while steering design yourself.

Not for: a single-spec feature (use a standalone spec/plan workflow) or a one-off task (delegate directly — a plain subagent, mimo-code, or codex:rescue).

How it works

• Lean conductor. The main session holds only the sprint doc, current stage, decisions, and open questions. Every technical step runs in an isolated git worktree via a subagent, so diffs and build logs never flood it. Noisy investigation (debugging, repros, browser work) is likewise delegated to a named sprint-investigator subagent that returns a distilled finding.
• Per-stage lifecycle: brainstorm a spec in main → dispatch the named sprint-planner → isolate a worktree → executor implements → review in a dedicated sprint-reviewer subagent that loads the resolved review backend (the review named by the repository's instructions, or the runtime default), fans out its axes plus risk specialists (security, architecture, performance, test-quality), fixes supported findings, and re-reviews → verify → commit & land → update the doc.
• Executor engines. Availability is runtime-specific; the orchestration remains the same and only Execute differs:
  • native — a named sprint-stage-executor subagent on Claude Code or OMP; no external CLI, launcher, or model resolution. The recommended default; the executor model is asked per stage (or pinned).
  • mimo — via the mimo-code plugin on Claude Code only; provider/model resolved per stage (or pinned).
  • codex — via the official Codex plugin on Claude Code only; an optional runtime probe.
• Nesting. Each stage can run as one sprint-stage-runner subagent that nests the executor and the review gate, keeping the conductor's context clean. On Claude Code this is probed once per sprint (whether child agents get the Agent tool). On OMP it requires task.maxRecursionDepth >= 3 — main → sprint-stage-runner → sprint-reviewer → parallel review/fixer workers — and is otherwise a configuration fact, not a probe. Without nesting, the conductor drives the same steps flat, one subagent per step.
• Runtime-specific model dispatch. Claude Code passes model directly on Agent(...) and can inherit the session model when it's omitted. OMP's task has no per-call model field at all: every executor and review model must be resolved by the conductor up front and threaded explicitly through an eval agent(prompt, { agent, model }) bridge at each nested dispatch, including resumes.
• The engine (and any model/review pins) are recorded in the sprint-doc header, so cross-session resume keeps the same setup without re-asking.

Requirements

• Claude Code or Oh My Pi (OMP). The skill detects the runtime and loads the matching adapter (skills/sprint/runtime-claude.md or skills/sprint/runtime-omp.md); an unrecognized runtime stops before dispatch.

• mimo-code — hard plugin dependency, Claude Code engine only. sprint/.claude-plugin/plugin.json declares "dependencies": ["mimo-code"], so installing the Claude Code sprint plugin auto-installs it. OMP has no native mimo delegate and does not offer the engine.

• Codex — optional, Claude Code only. The official Codex plugin openai/codex-plugin-cc (codex@openai-codex) provides codex:rescue + the codex-companion runtime. It lives in an external marketplace, so it cannot be a hard dependency and remains a runtime probe. OMP does not offer it.

• OMP nesting depth. For the nested review gate, the OMP host must configure task.maxRecursionDepth: 3 or higher.

• Matt Pocock skills — strongly recommended, optional. Exact capability probes use grilling, codebase-design, and diagnosing-bugs. They improve main-thread brainstorming, child planning, and evidence-first diagnosis. Its code-review skill is one possible instruction-named review backend, never a default:

  claude plugin marketplace add mattpocock/skills
  claude plugin install mattpocock-skills@mattpocock

  A sprint-only install remains fully usable: direct main-thread brainstorming, the sprint-owned planner contract, evidence-first diagnosis, and runtime-native SDD all have built-in fallbacks. Missing Matt skills never block those paths. A code review named by the repository's own instructions (AGENTS.md/ CLAUDE.md) always wins as the review backend; otherwise the gate uses the runtime's built-in default — Claude Code's bundled /code-review skill, or OMP's bundled /review review (a fan-out of the built-in read-only reviewer agent).

• A git repository (stages use git worktree).

Usage

Command	Effect
/sprint	Ask among runtime-supported engines; native recommended. OMP currently offers native only.
/sprint native	Native engine; the executor model is asked per stage.
/sprint native <model>	Native, pinned to that model for the whole sprint.
/sprint mimo	mimo on Claude Code; provider/model + variant resolved and asked per stage. On OMP, stop with a missing-integration error.
/sprint mimo <provider/model> [variant]	mimo on Claude Code, pinned to that model (+variant) for the whole sprint.
/sprint codex	codex on Claude Code. On OMP, stop with a missing-integration error.
/sprint … review <model>	Any engine; pins or explicitly repins review to that exact model. On OMP, future children switch in the same session. A flat main-owned active review is cancelled and rerun as a complete gate on the retained worktree/diff; a nested stage-runner-owned gate finishes on its resolved model and the new pin applies to later reviews.

Describe the milestone after the command. The skill probes capabilities, decomposes the milestone into stages with you, and drives them. Re-invoking /sprint with no description resumes at the first non-done stage.

Skill reference → · Scenario suite →