sprint-plan — author a sprint manifest with per-feature done-criteria

sprint-plan turns an MVP vision into a sprint manifest: a prioritized list of features, each carrying the done-criteria that make "done" mechanically checkable. The manifest is the input contract for sprint-oracle, which fails a sprint until every feature meets its criteria. This is Slice A of the mvp-sprint-parallel-cycles epic — the completeness half of the parallel-sprint goal (a sprint can be fast yet provably not partial).

What the manifest is

A JSON document (template-manifest.json is the shape). Each feature declares three done-criteria dimensions so completeness is a grep, not a judgment call:

• done_record — a reference proving the feature is specified (a spec AC id, e.g. AC-012). Non-empty.
• edge_tests — names of tests covering error / empty / edge states. A non-empty array.
• wiring_test — the name of one integration test exercising the feature end-to-end.

{
  "sprint": "<sprint name>",
  "features": [
    {
      "id": "search",
      "priority": "P0",
      "done_record": "AC-012",
      "edge_tests": ["search_empty_query", "search_unicode"],
      "wiring_test": "search_end_to_end"
    }
  ]
}

The test-tag convention (how sprint-oracle resolves the names)

A named test counts toward a feature only when the test file tags it. Immediately above the test('<name>', …) call, place:

// @sprint-feature:<feature-id> @kind:edge|wiring|happy
test('<name>', () => { /* … */ });

sprint-oracle builds a map of testName → {feature, kind} from these tags and resolves each manifest reference against it. A test tagged @kind:happy does not satisfy an edge or wiring requirement — that is the deliberate, mechanical line (no script can prove a test is semantically an edge case; the tag is the author's crisp claim, and the oracle verifies the claim resolves).

How to use

1. Decompose the MVP vision into features. Decide priority (P0/P1/…) per feature — scoping, not implementation.

2. For each feature, name its done_record (the spec AC it traces to), the edge_tests it must carry, and its wiring_test.

3. Write the manifest (default location: .claude/state/sprint/<sprint>/manifest.json, gitignored runtime state).

4. Validate the shape before handing it to the oracle:

   node .claude/skills/sprint-plan/validate-manifest.mjs validate <manifest-path>   # wraps validate-manifest.mjs -> validateManifest
   

   validateManifest(obj) returns {valid, errors:[{feature, field, reason}]} — it flags missing required fields per feature and duplicate feature ids. It checks shape, not completeness; completeness is sprint-oracle's job.

Constraints

• The manifest is a scoping artifact, not a design. sprint-plan decides what features and which done-criteria, never how a feature is built.
• Decisions live in main context (Article II). This skill is invoked there; it delegates nothing to a subagent.
• Completeness is enforced by sprint-oracle, not here — keep the two concerns separate.