Skill: roadmap

The decomposition layer between the idea and the per-feature pipeline. SDD builds one feature at a time under docs/features/<slug>/; roadmap answers the question that comes before any feature: where are we going, how does the idea break into steps, what depends on what, and in which order — and in what parallel lanes — do we walk them. One living docs/roadmap.md, repo-level utility (like survey).

Scope is whatever you bring. A whole product, one epic, or a single feature — the unit of the input does not change the unit of the output. A one-feature request gets decomposed the same way; there are simply fewer steps, and «fewer» is a legitimate answer. The skill never refuses a request for being too small, and never inflates one to look like a product plan.

Five load-bearing properties, in priority order:

1. A destination — one sentence for what is true once the last step ships. Without it the steps are a list of work with nothing to be a decomposition of.
2. Steps — the idea decomposed into incremental, source-anchored slices (each row cites the brief/PRD section it comes from; size XS–XL per ../_shared/size-matrix.md).
3. A word for what you don't know yet. Size is XS…XL or fog, in the same cell, because they answer the same question and only one of them can be true: sizing something nobody has formulated yet is how an estimate becomes a lie. Fog lives in ## Not yet specified until a recon pass sharpens it, and then the row trades fog for a real size.
4. Dependencies — an explicit graph, and the graph is where edges live. Every edge has a one-line reason (data model, UI zone, auth precondition). No edge without a real blocker, and no second copy of the edges in a column that could disagree with it.
5. Execution path — waves that respect the graph: wave N only contains steps whose deps are in earlier waves, and steps inside one wave are conflict-safe in the codebase (different modules / UI zones — so they can run as parallel worktree lanes). Each wave row names the zone.

Not here: RICE or any scoring (order IS the prioritization), dates (a decomposition, not a promise), solution detail (lives in the feature's spec). Question phrasing → ../_shared/ask-style.md; prose follows artifact_language, table structure and the Status / Size values stay English → ../_shared/artifact-language.md.

Owner

Whoever owns product direction (PM / lead / the solo maintainer).

Inputs

• Required — the idea source: docs/idea-brief.md (written by interview) / a PRD / a vision note (ask which, if several). This is the only hard input.
• Optional — docs/architecture-map.md (written by survey): zones for conflict-safety. Its absence costs precision in one column, not the roadmap.
• Optional — docs/design-system.md.
• Optional — existing docs/features/*/ + existing docs/roadmap.md — current statuses.

Protocol

1. Locate sources. Find the idea source (docs/idea-brief.md first, then PRD candidates). None found → say so and STOP: a roadmap without a source is fiction, and the fix is one command — /sdd:interview writes that brief. That is the only hard stop.

   docs/architecture-map.md is optional and never a gate. Present → its module inventory supplies the zones in ## Execution path. Absent → the roadmap is written anyway, every zone that is not an existing path is marked (new), and the handoff says plainly that the waves are a first cut worth re-cutting after /sdd:survey fills the map. A missing map degrades one column; it does not block the decomposition.

   Greenfield (empty repo, no docs/features/* at all): the first step is scaffold. Every other step assumes a skeleton to build into — a repo with no module layout, no test harness and no migration tooling cannot receive step 1 as written. Its source anchor is the foundation section of docs/architecture-map.md when the map exists, otherwise the brief itself (the project has to exist before anything in the brief can be built).

   If docs/roadmap.md exists, this run updates it (statuses, new steps, re-waving) — never silently rebuilds. A file written before 2.2 carries a Depends on column the template no longer has; this run folds it away — the edges it held move into the dependency graph, which is now the one place edges live. Say so in the handoff, so the dropped column reads as a migration rather than lost data.

2. Decompose — yourself, with the owner. Cutting an idea into steps is a live exchange with the person who holds the intent; it is not delegable, and a subagent standing in for the owner's half of it produces a plausible decomposition of a misunderstood idea. Read the sources and draft the steps under these rules:

   1. Every step traces to a source. Each row cites the section that justifies it (idea-brief.md §2 Problem, prd.md §5). No anchor → the step does not exist. Never invent scope the sources don't support.
   2. Steps are increments, not layers. A step is a walkable slice that leaves the product demonstrably better (UI + API + data as needed) — never «backend», «frontend», «tests».
   3. Dependencies are edges, not vibes. An edge names two step ids and has a reason you could defend in one line. Edges live in ## Dependency graph and nowhere else — a step with no real blocker has no edge, and inflated dependencies serialize work that could run in parallel.
   4. Fog is a size value, not a second column. Ask the Pocock test of each step: can you state the question precisely right now? — not «can you answer it». No → the step's Size cell reads fog, and the substance goes to ## Not yet specified as one area. Because fog and XS–XL share one cell, «the unformulated thing got sized anyway» stops being possible to write down. Do NOT pre-chop fog into ticket-sized pieces; that's inventing structure for something you haven't looked at yet.
   5. Size, don't score. Size every non-fog step XS–XL per ../_shared/size-matrix.md. No RICE, no priority numbers, no dates — order IS the prioritization.
   6. Existing state is respected. A step whose docs/features/<slug>/ already exists keeps its real status (spec'd / building / shipped); never re-plan shipped work.
   7. Name the destination. One sentence in ## Destination for what is true about the product once the last step ships. Write it before the steps and check it after: a step that moves nothing toward that sentence is scope that snuck in.

3. Close the AFK questions in parallel. Decomposition throws off two kinds of question. The ones that need the owner's judgment («is this in scope», «which of these matters more») you ask in step 4 — nobody can answer them for them. The ones that are just lookup («where does X live», «what format does the exporter emit», «is there already a library for this», «do steps 3 and 5 actually touch the same file») are AFK work, and today nobody does them: they get guessed at, and the guess becomes a dependency edge. For each such question, spawn one subagent — subagent_type: "general-purpose" (or the host's equivalent per ../_shared/tool-adapters.md), tools Read, Grep, Glob, Bash, WebSearch, WebFetch — all in one message so they run concurrently, one agent per question. Wrap each with the worker preamble from ../_shared/agent-roster.md and demand a fixed shape back: one line of answer + one citation (file:line or a URL). Uncited → dropped, and the question moves to ## Open decisions with type research. Zero AFK questions → skip this step and say so; do not manufacture questions to have something to fan out.

4. Review with the owner. Present the draft in prose (destination + steps + fog + waves; the mermaid goes to the file, never dumped raw to the terminal — same rule as design), calling each step by its name rather than its id. Then one AskUserQuestion call: (a) steps to merge/split/drop (multiSelect over flagged candidates + the questions step 3 could not close), (b) confirm the wave layout or name what must move. Apply.

5. Ask whether the file is worth keeping. If the pass surfaced no fog — every step is sized, ## Not yet specified is empty, ## Open decisions is empty — then say so and put the choice to the owner: a three-row file restating an already-clear plan is one more thing to keep current, and the alternative is to build straight from /sdd:specify <slug>. It is their call, not a stop. They want the file as a shared reference — write it. Say the tradeoff once, plainly, and take the answer; never refuse to write a roadmap because the request was small or clear.

6. Write. Fill ./templates/roadmap.md → docs/roadmap.md; set updated_at.

7. Structural self-check — per ../_shared/self-check.md, re-read from disk, verify 6 items:

   1. Every step row carries a source anchor, and the anchor resolves: the named document exists and the cited section is greppable in it (grep -F "<section>" <doc>). Text in the column is not evidence — a source anchor nobody checked is how invented scope gets in.
   2. Every graph edge names two existing step ids, and no step sits in a wave ≤ any of its dependencies' waves. (There is no Depends on column to reconcile against: the graph is the only place edges live, so an edge is wrong only by naming an id that isn't there or by contradicting the waves.)
   3. ## Destination is present and is one sentence — not a bulleted list, not a paragraph of goals. One sentence, or the section has failed at the only job it has.
   4. Every zone named in the execution-path table exists as a real path (test -e) or is explicitly marked (new).
   5. Every closed enum holds. Size ∈ {XS, S, M, L, XL, fog} — and every fog row points into ## Not yet specified and carries no XS–XL anywhere. Status ∈ {idea, spec'd, building, shipped}, and every spec'd+ step links an existing docs/features/<slug>/ (test -d). In ## Open decisions, Type ∈ {research, prototype, grilling, task} and Owner ∈ {agent, human} — a role, never a person's name. The frontmatter carries exactly the template's keys, none invented.
   6. Zero dates outside ## Shipped (\b20\d\d- scan) and updated_at = today.

   Fix + re-check ≤2 cycles; surface the rest.

8. Commit + handoff. Propose commit roadmap: <what changed>. Then emit the stage-handoff block per ../_shared/handoff.md (utility variant) — What I did (incl. «self-check: 6/6 pass» and the fog count) + Review (docs/roadmap.md) + Run next: name the first unblocked step the way the file names it, then give the command on its own line, ready to run — /sdd:specify <slug> for that step, /sdd:scaffold when the first step is the greenfield skeleton, or a recon pass on the top ## Not yet specified area when the first thing in the way is fog rather than work. When the run had no architecture map, add one line: the zones are unverified, /sdd:survey then a re-run re-cuts the waves.

Sync hooks (delivery keeps it current — anti-drift)

• specify sets the step's Status to spec'd (and → building as implement starts) and links docs/features/<slug>/. A spec'd feature with no step row gets one appended (source anchor = its spec). A step cannot go spec'd while its Size reads fog — specifying it is what clears the fog, so the hook replaces fog with a real size in the same edit.
• ship sets Status: shipped and adds the date + PR link to Shipped.
• Neither hook re-waves the graph — structure changes only happen here, with the owner.

Definition of Done

• docs/roadmap.md exists: a one-sentence Destination, the steps table (every row source-anchored + sized XS–XL or fog + status), the fog vocabulary (Not yet specified · Out of scope · Open decisions), Decisions so far as gists-with-links, the dependency graph that owns every edge, and execution waves that respect the graph with a named zone per parallel step.
• No scores, no dates outside Shipped; no XS–XL on a fog row; updated_at current.

Anti-patterns

• A prioritization scorecard. No RICE/effort-value math here — decomposition and order are the deliverable. If the owner wants scoring, that's a separate conversation, not this file.
• A dated Gantt / roadmap-as-promise. Zero dates outside Shipped; the disclaimer stays.
• Horizontal steps («backend», «the UI», «testing») — steps are vertical increments.
• A roadmap with no destination. Twelve rows and no sentence saying what they add up to. The steps then can't be wrong, because there's nothing for them to be wrong about.
• Phantom dependencies. An edge you can't justify in one line serializes parallelizable work.
• Same-wave conflicts. Two steps touching one code zone in one wave = a guaranteed merge conflict between lanes; the zone column exists to catch exactly this.
• Sizing the fog. A one-line step marked L, spec'd, one edge — and underneath it nine undecided things. The size was invented, the status was wrong, and the roadmap's unit was too coarse to say so. Size: fog is the honest cell.
• Refusing the small request. «This is one feature, you don't need a roadmap» is not an answer — it is a two-step roadmap. Say the tradeoff, then do what the owner asks.
• Stopping on a missing architecture map. The map sharpens the zones; the decomposition does not depend on it. Mark the zones (new) and say so in the handoff.
• Restating a decision instead of linking it. ## Decisions so far is a gist plus a link to where the decision actually lives (the ADR, the spec, the thread). A paragraph of solution detail copied in here is a second source of truth that will drift.
• Delegating the decomposition. A subagent has the same files and less context than you, and it cannot have the exchange with the owner that the cut depends on. Fan out the lookups (step 3), keep the judgment.
• Re-planning shipped work / duplicating specs. Shipped rows are history; solution detail lives in docs/features/<slug>/.

References & template

• ./templates/roadmap.md — the decomposition scaffold.
• ../_shared/size-matrix.md — XS–XL sizing heuristics.
• ../_shared/agent-roster.md — the shared contract every fanned-out question-closer follows.