Feature Decompose

Feature Decompose

Decomposes one roadmap item into one or more feature files at project/features/<slug>.md per spec/project/feature/<canonical_language>.md. This skill is the only authoring surface in the feature lifecycle that creates feature files; later status transitions belong to sprint-execute and sprint-review.

Why this is a skill, not an agent

• Mid-flow agent dispatch is the contract — the spec mandates a consistency check via the feature-consistency-reviewer agent before draft → ready; the orchestrator that dispatches an agent is always a skill per spec/claude/skill-vs-agent/ §Hybrid pattern.
• Per-feature operator approval — title, description, acceptance criteria, test hooks, and the verifies_sprint_value choice each require user confirmation; an agent's structured-report shape can't carry those checkpoints.
• Persistent on-disk artefact under iterative drafting — feature files live next to the roadmap and sprint corpus and are mutated again later by other skills; the iterative drafting (consistency findings → resolutions → final write) flows back into the main conversation naturally.
• Counter-dimension considered: a narrower agent prompt could sharpen acceptance-criterion phrasing, but the load-bearing dimensions here are mid-flow agent dispatch and per-step operator approval — skill wins.

User-language policy

Detect the user's language and respond in it. Feature files themselves (frontmatter, body sections, identifiers) are written in the project's primary language as declared by the consuming repo's root convention file (typically CLAUDE.md); when no primary language is declared, English is the default. Frontmatter field names, identifiers (F-<n>, R-<n>, acceptance-<n>), and the closed-vocabulary status tokens (pending, passing, skipped, failing) stay English regardless of prose language.

Preconditions

Before any write, confirm:

• The current repository carries spec/project/feature/<canonical_language>.md. If absent, stop — the spec is the input to this skill, not optional.
• project/roadmap.md exists and the target roadmap_item (passed by the user as R-<n>) is present in it; resolve its outcomes, detail, target_sprint, and mvp fields so they can be quoted into per-feature context.
• project/goals.md exists and the roadmap item's outcomes list resolves; if not, stop and surface the broken outcome reference rather than guessing.
• An audience artefact exists when the project carries project/mission.md; missing audiences block feature authoring the same way they block outcome and mission authoring (per spec/project/roadmap/ and spec/project/mission/). Dispatch audience-identify first if needed.
• A requirement artefact under project/requirements/ covers the work being decomposed, with U_gate ≥ τ_high, per spec/project/requirements-elicitation/ § Consumer contract. When none exists or it is below threshold, dispatch requirements-elicit first, or record an explicit operator override; decomposing against unstated or weakly-understood requirements is the failure this gate prevents.
• The feature-consistency-reviewer agent is reachable in the current plugin runtime, or the operator explicitly acknowledges the manual-fallback path and provides a manual-pass identifier of the form manual-<YYYY-MM-DD> (per spec/project/feature/ §Consistency check). The fallback is permitted only until the agent ships; once the agent is reachable, refuse to fall back.
• The next free F-<n> ID is determinable by reading the highest existing ID under project/features/; never reuse a retired ID.

Operations

1. Decompose

The decomposition produces N ≥ 1 features for the given roadmap item; the operator decides N.

1. Restate the roadmap item. Read its YAML block from project/roadmap.md and replay its title, outcomes, detail, target_sprint, and mvp flag back to the operator. If detail is coarse or backlog and the item is targeted at the current or next sprint, stop and recommend running roadmap-refine first — decomposing a non-fine item violates the roadmap spec.
2. Propose a feature split. Suggest one to four feature titles based on the roadmap item's body checklist (when present) or its prose. The operator confirms the split, adds, removes, or renames before any file is written.
3. For each feature, gather:
   • Title — one-line summary in the project's primary language.
   • Slug — ASCII kebab-case, ≤ 6 words, derived from the title; confirm with the operator and verify the path project/features/<slug>.md is free.
   • Description — one to three paragraphs phrased from the end-user perspective; ask follow-ups until the description is concrete enough that a reader unfamiliar with the roadmap can grasp the change.
   • Acceptance criteria — three to seven testable bullets, atomic and user-visible (no workflow gates such as "PR merged"). Each bullet uses the pattern - [ ] **acceptance-<n>** <criterion>. One is suspicious; more than ten suggests splitting the feature.
   • Test hooks — one entry per acceptance criterion, in the format - **acceptance-<n>** — <mechanism> — <status>, where <mechanism> names a test path, skill name, CLI command, or manual procedure, and <status> is one token from the closed vocabulary pending (default at decomposition), passing, skipped, failing.
4. Identify the value-verifier. Across the N features, ask the operator whether one feature carries verifies_sprint_value. The field is non-null on at most one feature per sprint (per spec/project/feature/ §Frontmatter schema). When this decomposition produces the feature that closes the MVP per spec/project/mission/, that feature MUST carry verifies_sprint_value: acceptance-<n> naming the criterion that proves the sprint's value_statement. When the operator is unsure, leave all features with verifies_sprint_value: null — sprint-plan may reassign it later.
5. Run the consistency check (mandatory; see Operation 2). Block the write until the check completes and every overlap or duplication finding is resolved.
6. Write the feature file(s). For each feature, render the frontmatter with the nine fields in the declared order (id, title, status: draft, roadmap_item, sprint: null, created, ended: null, verifies_sprint_value, consistency_check) and the five required level-2 sections in the declared order (## Description, ## Acceptance criteria, ## Test hooks, ## Consistency notes, ## Risks). Optional sections (## Open questions, ## References) MAY follow the required five.
7. Confirm paths back to the operator in their language and remind them that draft → ready, ready → in_progress, and in_progress → done are owned by sprint-plan, sprint-execute, and sprint-review respectively — not by this skill.

2. Run the consistency check

Mandatory before the feature file is written with status: draft AND a populated consistency_check frontmatter object. The check is what allows a downstream skill to later transition draft → ready; without it, the feature is malformed.

1. Dispatch the feature-consistency-reviewer agent with the draft frontmatter and body assembled in step 3 of Operation 1, the resolved roadmap_item, the slug, and a short git revision identifier (git rev-parse --short HEAD from the dispatching skill — the agent has no shell access by design and can't compute this itself). The agent reviews the existing feature corpus under project/features/, the project's primary source roots, and the spec corpus under spec/, and returns a findings array. Confirm the working tree is a git repository (git rev-parse --is-inside-work-tree) before invoking the agent — the agent has no shell access, so this skill owns that precondition.
2. Manual fallback (deprecated, transitional only). The agent ships with this skill, so the fallback is reached only when a plugin runtime predating it can't resolve the agent at all: walk the same investigation surface yourself — read every existing feature's frontmatter and ## Description, scan the project's primary source roots for already-implemented behaviour, and grep spec/ for prior decisions that constrain the new feature. Capture findings in the same shape the agent would emit. Acceptance of this path MUST be explicit: ask the operator before proceeding, and surface the runtime version gap so they know the right long-term fix is upgrading the plugin runtime, not normalising the manual pass.
3. Persist findings on the feature in two places:
   • Frontmatter — populate the consistency_check object with performed_at: <ISO date>, agent_version: <agent-id> (or manual-<YYYY-MM-DD> for the fallback), and a non-empty findings list. Each finding carries kind (overlap, duplication, drift, prior-art, or clean), target (file path or feature ID), and resolution (merge-into <id>, supersede <id>, split-out <ids>, proceed, or revisit-after <event>). A clean run still records exactly one finding with kind: clean so the array is never empty.
   • Body — populate ## Consistency notes with one paragraph per finding restating the finding in human-readable prose plus the chosen resolution. Per spec/claude/claim-provenance/, each paragraph marks the claim it rests on — the overlap, prior art, or drift it asserts — as established, citing the file:line or command output that established it, or as unestablished, naming the observation that would settle it and stating it wasn't made; make a cheap observation instead of taking the unestablished exit.
4. Block the write when any finding has kind: overlap or kind: duplication and its resolution is proceed without an explicit one-paragraph rationale in ## Consistency notes. Surface it and ask for a non-proceed resolution or a written rationale first. A merge-into <id> resolution is a first-class outcome and may end the decomposition for that feature; record the choice and skip writing the redundant file.
5. Record the manual-pass author in ## Consistency notes when agent_version starts with manual-. The spec's §Consistency check requires naming the operator who performed the fallback; capture it as a dedicated line of the exact form Manual pass performed by: <name> at the top of ## Consistency notes, so an auditor can attribute the resolution decisions without parsing prose. Every manual pass on a repo whose plugin runtime can resolve the agent is a workflow-health finding per spec/project/feature/ §Consistency check; the right resolution is to upgrade the plugin runtime and rerun the check via the agent path.

3. Re-run the consistency check on a material change

The consistency check is not a one-shot at decomposition. Per spec/project/feature/ §Consistency check, the check MUST be re-run while the feature is in ready or in_progress whenever any of these material changes occur:

• the ## Description section changes by more than typo-level wording;
• an acceptance criterion is added, or its core wording (excluding the checkbox state) is altered;
• the roadmap_item or sprint frontmatter field is changed;
• a feature with overlapping scope is added or removed elsewhere under project/features/.

Cosmetic edits — typo fixes, formatting, link-target normalisation, and flipping the checkbox state of an existing acceptance criterion — MUST NOT trigger a re-run. When a re-run is required, repeat Operation 2 but append a new dated findings block (keyed by its own performed_at) to the existing consistency_check object; never overwrite the historical findings. The ## Consistency notes section gains a new dated paragraph per re-run finding, preserving the prior notes. Re-invoking this skill on a feature that already carries a consistency_check is the canonical entry point for that re-run; when a downstream skill (sprint-plan retargeting sprint, or sprint-execute adding the feature mid-sprint) makes one of the triggering edits, it routes back here rather than re-running the check itself.

Examples

• Read examples/01-single-feature-from-roadmap-item.md when decomposing a single roadmap item into a feature for the first time.
• Read examples/02-split-into-multiple-features.md when a roadmap item is large enough to warrant multiple features.
• Read examples/03-consistency-check-overlap.md when the consistency checker reports potential overlap with existing features.

Gotchas

• The feature-consistency-reviewer agent has no shell access (per agent-management §Tool access — read-only invariant). The dispatching call from this skill MUST confirm the working tree is a git repository (git rev-parse --is-inside-work-tree) and pass the short SHA (git rev-parse --short HEAD) as a dispatch argument; the agent uses the SHA to populate agent_version in its findings report. Skipping this step and dispatching anyway produces an agent_version: unknown field in every consistency-check result.
• The manual-fallback path is deprecated, not removed. When the agent dispatch in Operation 2 step 1 fails because the plugin runtime predates the agent's release, the operator can still walk the investigation surface manually — but the skill MUST ask explicit permission before falling back, and the resulting consistency_check block carries agent: manual-<YYYY-MM-DD> instead of an agent name. Auditing later that bypasses surface as if it were a regular agent run mis-attributes the resolution decisions.
• verifies_sprint_value is a feature-side invariant, not a sprint-side one. At most one feature per sprint carries a non-null verifies_sprint_value; setting it on two features in the same sprint is a hard violation per spec/project/feature/ §Frontmatter schema. The skill defaults the field to null on every new feature; the operator opts in explicitly when authoring or when sprint-plan reassigns the verifier.
• The R-<n> and F-<n> ID counters are monotonic across the project's lifetime, never reused even after deletion. Deriving the next ID from "max existing ID + 1" without checking the git history's deleted IDs would silently re-use a retired ID. The skill reads the highest existing ID under project/features/ plus the highest deleted ID from git log before assigning.

Resumability

Per spec/claude/resumable-work/, this skill is resumable: true. State is persisted to .resume/feature-decompose/<run-id>.yml after every successful user-approval gate and after each named phase boundary. On re-invocation, scan that directory for files with status: in_progress whose inputs: snapshot matches the current invocation; if one matches, prompt the operator with Resume run <run_id> from phase <phase> (last checkpoint <last_checkpoint_at>)? [resume / start-new / discard]. The state-file envelope (schema_version, run_id, inputs, phase, decisions[], status, ...) and the fail-closed semantics on schema or YAML errors are load-bearing in the spec; don't duplicate those rules here.

Hard rules

• Never transition a feature past status: draft. Only sprint-plan (draft → ready after a sprint value is assigned) and sprint-execute / sprint-review (later transitions) own status flips. This skill writes status: draft and stops.
• Never write a feature file without a populated consistency_check frontmatter object whose findings array is non-empty (a clean run still records kind: clean) and a populated ## Consistency notes section.
• Never set verifies_sprint_value on more than one feature per sprint within a single decomposition; the at-most-one-per-sprint invariant is feature-side per the spec.
• Never invent a roadmap item, an outcome, or an audience inline; missing references block the write rather than fabricating them.
• Never rename a slug after the feature leaves draft; the skill writes the slug once and treats it as immutable thereafter.
• Never overwrite an existing feature file's status with anything but its current on-disk value: this skill re-reads the file before a rewrite and refuses when the requested write would change status at all — status transitions live in sprint-plan / sprint-execute / sprint-review, each of which verifies the legal predecessor on disk. An operator-decided lifecycle deviation is recorded as a dated deviation note in ## Consistency notes before any status write (F-17/F-18 precedent, #498).
• Never carry a process-internal acceptance criterion ("PR approved", "merged to develop", "CI green"); criteria are user-visible behaviour, not workflow gates.
• Never include effort estimates, points, due dates, or assignment fields beyond the nine declared frontmatter fields; lints flag unknown keys.
• Never dispatch feature-consistency-reviewer from inside an agent context. This is a skill operation; the parent dispatcher of this skill is the user or another skill, never an agent.
• Never write the feature file when the operator declines to resolve an overlap or duplication finding; the partial write would silently break the draft → ready gate downstream.
• Never record a manual-fallback consistency pass without a Manual pass performed by: <name> line in ## Consistency notes; the spec requires the notes section to name the operator who performed it.
• Never overwrite historical consistency_check.findings on a re-run; append a new dated findings block and a new dated ## Consistency notes paragraph per spec/project/feature/ §Consistency check, and never trigger a re-run on a cosmetic-only edit.
• When spec/project/feature/ disagrees with this skill, the spec wins. Propose updating this skill rather than silently diverging.