Roadmap Refine

Roadmap Refine

Enforces the detail-level invariant declared in spec/project/roadmap/<canonical_language>.md §Detail-level convention and refinement rule: at any point in time, every roadmap item with target_sprint set to the current sprint or the next sprint carries detail: fine. Coarse and backlog items two or more sprints out are fine; promotion to fine is the trigger for this skill to pull each affected item into shape.

German trigger phrases

This skill also triggers on equivalent German-language requests, including:

• "Roadmap verfeinern"
• "Roadmap-Detailstufen prüfen"

Coherence gate

Before a batch of transitions or when drift against goals.md / mission.md / the sprint files is suspected, dispatch the read-only roadmap-coherence-reviewer agent and act on its findings first — it is the hybrid-pattern review counterpart of this skill and catches cross-document inconsistencies the per-item validations here don't see.

Why this is a skill, not an agent

• Mid-flow interactivity is the contract — fix proposals (promote coarse → fine, write the missing body paragraph, expand the feature checklist) are negotiated with the user one item at a time; an agent's structured-report shape would lose that per-item iteration.
• Output flows back into the main conversation — the violation list and the per-item drafts must be visible in the working context so the user can confirm each fix before it hits disk.
• Orchestrator role — fixes that exceed simple promotion (re-targeting a sprint, adding a feature checklist) are dispatched into roadmap-plan; the orchestrator is always a skill per skill-vs-agent.
• Counter-dimension considered: the violation-detection step alone (read project/sprints/, walk roadmap items, emit records) would suit an agent's isolated context, but the load-bearing dimension is the per-item user dialogue around remediation, not the detection mechanics — skill wins.

User-language policy

Detect the user's language and respond in it. Violation records, remediation hints, and any prose the skill writes back into project/roadmap.md use the project's primary language (the same language already used in goals.md and roadmap.md).

Preconditions

Before walking the roadmap:

• project/roadmap.md and project/goals.md exist. When either is missing, stop and direct the user to roadmap-init.
• project/sprints/ exists and contains at least one sprint file under the <NNNN>-<slug>.md shape declared by spec/project/sprint/. When the directory is missing or empty, the current and next sprint can't be resolved; stop and report the gap so the operator can scaffold sprints first.
• The roadmap items under audit parse: every item exposes a ### <id> heading immediately followed by a fenced ```yaml block carrying id, title, detail, outcomes, target_sprint, mvp (when project/mission.md exists), and status. When parsing fails on any item, surface the parse error and stop — partial enforcement on a partially-parsed file would leak false positives.

Operations

1. Resolve the current and next sprint

The spec defines the resolution exactly; this skill MUST follow it without local interpretation:

1. Current sprint — the sprint whose status is active per spec/project/sprint/. At most one sprint per project carries active. When no sprint is active, fall back to the lowest-numbered sprint whose status is planned. When no sprint is planned, fall back to the highest-numbered sprint whose status is closed. When the project has only cancelled sprints, treat the current sprint as undefined and report the gap; the invariant cannot be checked, so emit a single non-violation diagnostic and exit zero.
2. Next sprint — the lowest-numbered sprint whose status is planned and whose number is strictly greater than the current sprint's number. When no such sprint exists, treat the next sprint as undefined and only enforce the invariant against the current sprint.

When the skill runs, it echoes both resolved values back to the user before walking items so the operator sees the assumption the skill is auditing against:

Current sprint: 0007 (active)
Next sprint:    0008 (planned)

2. Walk the roadmap

For every roadmap item:

1. Read its YAML block.
2. Skip when target_sprint is null or does not match the current or next sprint number.
3. When target_sprint matches the current or the next sprint and detail is anything other than fine, emit a violation record (see step 3).
4. Independently of the detail invariant, surface but do not treat as a violation here: a target_sprint pointing at a closed or cancelled sprint. That is a separate lint owned by sprint-plan per the spec; report it as an informational note and let roadmap-plan pick it up.

3. Emit violation records

Each violation MUST carry:

• id — the item's R-<n>;
• target_sprint — the offending sprint number;
• current_detail — the item's current detail value (coarse or backlog);
• resolved_current_sprint — the sprint number this skill resolved as current;
• resolved_next_sprint — the sprint number this skill resolved as next, or null;
• remediation_hint — one short line of advice ("Promote to detail: fine and add a feature checklist", "Retarget to a later sprint via roadmap-plan", "Split into a fine slice and a coarse follow-up").

Write each record to stderr (or the skill's structured output channel) so downstream tooling can pick them up. Exit non-zero when at least one violation is open. Exit zero only when every audited item is compliant.

Suggested record shape (one record per line, machine-readable):

violation id=R-3 target_sprint=8 current_detail=coarse resolved_current_sprint=7 resolved_next_sprint=8 hint="Promote to detail: fine and add feature checklist"

4. Walk fix proposals per item

For each violation, in roadmap order, propose a fix to the user one item at a time:

1. Show the current item (heading, YAML block, body) and the violation record.
2. Offer the canonical remediation paths:
   • Promote to fine — flip the YAML field and draft the missing body shape (paragraph stating the user-visible change plus a feature checklist). The skill drafts both; the user confirms or edits.
   • Retarget the sprint — change target_sprint to a later sprint. Refuse to do this directly; dispatch roadmap-plan so cross-references (outcome resolution, sprint resolution) are validated end-to-end.
   • Drop the sprint anchor — set target_sprint: null so the item drops back to unscheduled. Direct edit is acceptable here because no cross-reference materially changes.
3. Apply the chosen fix in-place inside project/roadmap.md. Refuse partial writes that would leave the file in a half-fixed state — when the user changes their mind mid-walk, revert to the pre-walk file and exit non-zero.
4. On every individual-fix completion, the skill re-resolves current and next sprint numbers from disk in case the operator changed sprint state in parallel.

When the user wants to accept all violations and just record the audit without fixing, refuse: this skill is the canonical enforcement point and writing the queue back unchanged after a violation would defeat its contract. Direct the user to either fix the items or dispatch roadmap-plan for non-trivial restructuring.

5. Final report

On run completion, the skill reports:

• The count of violations found, fixed, deferred, and skipped.
• The final exit code (zero only when no violations remain after the walk).
• When violations remain (the user deferred a fix), the skill keeps the exit code non-zero so calling automation sees the failure.

Examples

• Read examples/01-clean-no-violations.md when all roadmap items are already spec-compliant and the skill produces a clean report.
• Read examples/02-coarse-near-sprint-violation.md when a near-sprint item is still flagged coarse and the skill surfaces the violation.
• Read examples/03-walk-fixes-promote-and-retarget.md when the walk phase fixes a coarse violation by promoting to fine and retargeting the item.

Gotchas

• Detail-level invariant applies only to current and next sprint items: items targeted at sprints two or more out are intentionally coarse or backlog — do not flag them as violations; applying the fine requirement beyond the two-sprint horizon is over-enforcement.
• Sprint state must be re-read after each per-item fix: roadmap-refine step 4.4 explicitly re-resolves current and next sprint numbers from disk after each fix, because the operator may change sprint state in parallel; using a cached sprint resolution across multiple fix rounds can produce stale violation records.
• A missing or empty project/sprints/ directory is a hard stop, not a skip: without a resolvable current sprint the invariant cannot be checked; stop and report the gap rather than falling back to wall-clock heuristics or guessing from roadmap content.

Resumability

Per spec/claude/resumable-work/, this skill is resumable: true. State is persisted to .resume/roadmap-refine/<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 resolve the current or next sprint by any rule other than the one declared in spec/project/roadmap/ §Detail-level convention and refinement rule. The fallback chain (active → lowest planned → highest closed; next is lowest planned strictly greater than current) is exact and not negotiable.
• Never treat the absence of a planned next sprint as a violation. The invariant only applies to sprints that exist; when the next sprint is undefined, audit only against the current one.
• Never retarget target_sprint directly inside this skill. Cross-reference validation belongs to roadmap-plan; this skill dispatches into it.
• Never suppress a violation on stderr to keep automation green. The exit code and the violation records are the contract.
• Never rewrite a roadmap item's body without showing the draft to the user first; per-item user confirmation is the contract.
• Never silently demote a fine item to coarse or backlog. The spec only governs the upward direction (promotion to fine for near-sprint items); demotion is owned by roadmap-plan and requires explicit user intent.
• Never treat a target_sprint pointing at a closed or cancelled sprint as this skill's violation. Surface it as an informational note and route it to roadmap-plan and sprint-plan.
• When spec/project/roadmap/ disagrees with this skill, the spec wins. Propose updating this skill rather than silently diverging.