| name | dx-refactor-discover |
| description | Hunt the codebase for refactor opportunities, present them inline, and promote the ones you pick into changes. |
| disable-model-invocation | true |
| argument-hint | ["area or path"] |
dx-refactor-discover
Start with no concrete target — "find me refactor opportunities." Scan the codebase for design problems worth fixing — shallow modules to turn deep, plus whatever the wider design lenses surface — present them inline as markdown, and promote whichever the user picks into a normal container. This skill discovers and hands off — it can sketch alternative interfaces for whichever candidates the user picks, but it writes no plan.md and edits no code.
Guard. If context/ isn't scaffolded (no changes/ or efforts/), stop and tell the user to run /dx-init. The findings this run produces are ephemeral — there is no debt register; anything not promoted or recorded as a lesson leaves no trace.
1 — Load the vocabulary
Invoke dx-references with topic module-design — its terms (deep vs shallow, seam, leverage and locality, adapter, dependency category, the deletion test) are how every finding is phrased, and its ## Rejected framings says which words stay out. Read foundation/glossary.md for naming (a one-line habit — no section) and foundation/lessons.md so you skip anything a prior run already rejected ("don't re-deepen X because Y").
2 — Scan
Optionally scoped by [area or path]; unscoped means the whole tree. Invoke dx-references with design-lenses — module depth is the primary lens, not the only one, and problems outside it go unfound if nothing else is looking. Fan out to built-in Explore subagents to walk it — explore for friction, don't run rigid heuristics.
Recency is the default prior. Walk back a stretch of git log --oneline and let the paths that keep showing up pull attention first — churn is where a deepening pays back soonest. If the log is scattered with no clear hot spot, widen the net instead.
Look for: understanding one concept that means bouncing between many small modules; shallow modules (interface nearly as wide as the implementation); pure functions extracted only for testability while the real bug hides in how they're called; seams that leak; and what the other lenses catch — knowledge duplicated across modules, one module changing for unrelated reasons, coupling that makes a single edit ripple. Apply the deletion test to each suspect: would deleting it concentrate complexity, or just move it? "Concentrates" is the signal.
Whichever lens found it, module-design is still how it's said. A layering violation reports as "this module reaches through two interfaces to the store; the seam belongs at X" — not as the principle it violates.
3 — Present findings inline
Markdown only — no HTML, no report file, no clipboard. A concise numbered list; each candidate one tight entry:
- What & where — the module and files.
- Why it's shallow / tangled — in
module-design terms.
- Shape: before → after — one line (
4 wrappers + handler → 1 module, 2 methods).
- Proposed deepening — the move, plus a strength tag (
Strong | Worth exploring | Speculative).
- Dependency category — in-process / local-substitutable / remote-but-owned / true-external. The strength tag says whether the deepening is worth doing; this says whether the result can be tested afterwards, which is where a refactor actually stalls.
Cash out the win. Name it in module-design terms — "locality: bugs concentrate in one module", "leverage: one interface, 12 call sites", "two adapters justify the seam: HTTP in prod, in-memory in tests". "Cleaner code" and "easier to maintain" give the reader nothing to check and nothing to carry into the change.
The scan stays at shape level. Shape: before → after is as far as a candidate goes here. Designing interfaces across N candidates spends the effort before the user has said which one matters; that happens after the pick, in §4.
Close with a top pick — one sentence: which candidate to tackle first and why. The strength tag ranks confidence, not sequence; a Strong finding in a file nobody touches is worth less than a Worth exploring one in a hot path. Then ask which the user wants to promote.
4 — Design it twice for the pick (opt-in)
Offer this once the user has picked, before promoting. Skipping it is fine and goes straight to §5 — but the scan only ranked candidates on the shape they have, and a candidate whose alternatives all look bad shouldn't be promoted. Exploring here is cheap: the context is freshest at the moment of the pick, and the winning sketch rides into the change so /dx-plan doesn't re-derive it.
One finding picked → ask the plain yes/no ("explore <finding> before promoting?") and, if yes, run steps 1–3 below once for it.
More than one finding picked → don't repeat the yes/no per finding. Ask one batched question first: "sketch any of these before promoting?" — options none / all / specific ones, recommending just the top pick (2–4 concrete options with a recommendation, the interview reference's shape). Then run steps 1–3 below, plus the stop-and-ask that closes this section, once per selected finding, sequentially — never spawn the next finding's sub-agents before the current finding's sketch is confirmed. This is the fan-out safeguard: at most one finding's batch (3–4 agents) is ever in flight, no matter how many findings total, with no arbitrary cap to invent or maintain.
-
Frame the problem space to the user — the constraints any new interface has to satisfy, the dependencies and their category, and a rough code sketch to make the constraints concrete (an illustration, not a proposal). Show it and start the sub-agents immediately: the user reads and thinks while the agents work, which is the whole point of doing this in parallel. Don't block on a reply.
-
Spawn 3–4 built-in Plan sub-agents in parallel, each under a different forcing constraint. State each constraint as where it puts the seam, not as a value to maximize — two values can share an optimum, and then two agents hand back the same interface:
- Collapse to 1–3 entry points, everything else pushed behind them.
- Move part of the contract out of the interface — into structure, config, or something the caller declares rather than calls.
- Split the seam so the common caller and the rare caller reach different entry points.
- Ports & adapters — only when the dependency category is remote-but-owned or true-external.
On a module with one dominant caller the first and third collapse — the smallest surface already is the best default case — so pick one and spend the freed agent elsewhere. Use Plan, not Explore: Explore is read-only search, which is what §2 fans out to; this is design work.
-
Write each agent its own technical brief. A sub-agent starts cold, so the brief carries file paths, the coupling detail, the dependency category, what sits behind the seam, and its one constraint — separate from the user-facing framing in (1). Write the structure in module-design terms, and the domain in foundation/glossary.md terms when the candidate has a domain — infrastructure and tooling often doesn't, and forcing glossary words onto a build script buys nothing. Designs that name things the same way are the only ones worth comparing side by side.
Ask each for the same five things, so the answers line up: interface (types, methods, params — and the invariants, ordering and error modes that are just as much interface) · usage example from the caller's side · what the implementation hides behind the seam · dependency strategy and adapters · trade-offs, naming where leverage is high and where it's thin.
When it goes wrong. An agent that returns nothing, errors, or comes back unusable is dropped, not retried — compare the survivors and say plainly that fewer than three came back, so the user knows how wide the comparison actually was. And if the user's reply to the framing contradicts the constraints the briefs were built from, discard the in-flight sketches and re-spawn. Reconciling designs built on a superseded framing produces a shape nobody chose.
Present the survivors sequentially, compare them on depth, locality, and seam placement — the module-design axes, so this stays inside the one vocabulary — and close with an opinionated pick or hybrid, the same discipline as the top pick in §3. Two designs that share an entry point and differ only in policy are one design: say the constraint didn't bite rather than presenting them as two, because a comparison padded to three is worth less than an honest two.
Stop here and ask which sketch to carry forward. The pick is a recommendation, not a decision — §5 writes files, and a file written on the wrong sketch is a redo, not an edit. Don't promote on your own ranking the way §3 permits when the user is AFK; wait for the reply. For a multi-select promote, this reply is what unblocks the next selected finding — don't spawn its sub-agents until it lands.
5 — Promote the pick
- One → create the change directly, the same way
dx-diagnose self-contains its own promotion: write context/changes/<slug>/change.md stamped type: refactor, with the finding captured as its seed research/<topic>.md (or frame.md if it reads more like a framing than a research write-up). If §4 ran, wait for the user's reply to the stop-and-ask before writing anything — the sketch that goes into the seed, with the trade-offs that decided it, is the one the user confirmed, not the opinionated pick on its own. That confirmed sketch is what stops the exploration being thrown away; /dx-plan reads it as upstream context. Invoke dx-references with change-md for the exact schema. This is this skill's own deliverable, not a chain into /dx-plan — that stays the printed next command.
- Many → decomposing into an effort + roadmap + several child changes is already a multi-step flow owned by other skills (
dx-new for the effort, dx-roadmap for the slices, dx-new again per slice) — print the commands and let the user drive it, don't fold all of that in here. But don't make the user re-type what they just picked: compose a seed summary, the full entry (what & where, why, shape, proposed deepening, strength tag, dependency category) for each promoted finding — plus the chosen sketch for any finding §4 explored — under a heading that names /dx-refactor-discover as the source. Print it as the literal argument to hand to /dx-new so the handoff carries the detail, not just a slug. Carry the top pick into it as a closing line — /dx-roadmap sequences the slices, and the read on which one goes first is the thing it can't re-derive.
- Rejected with a load-bearing reason → offer
/dx-lesson to record "don't re-deepen X because Y" so the next run skips it. Rejected ephemerally or selected → no durable trace.
Done when
Findings have been presented inline, the user has chosen, and — if they took the §4 offer — the alternatives have been compared and one picked. For a single promoted change, the container now exists — print what was created and the next command. For everything else, print the exact command and stop — never run it:
Promoted one: Change created: context/changes/<slug>/change.md (type: refactor, seeded with the finding)
Next: /dx-plan <slug>
Promote many: Next: /dx-new "<seed summary>" → /dx-roadmap <effort-id>
Where <seed summary> is:
## Refactor opportunities (from /dx-refactor-discover)
1. **<title>** — <module/files>
Why: <shallow/tangled reason + the cashed-out win, in module-design terms>
Shape: <before → after>
Proposed: <the move> (<Strong|Worth exploring|Speculative>, <dependency category>)
Sketch: <the chosen interface and why it won> (only if §4 explored this one)
2. ...
Start with: <which one first, and why>
Record a no: /dx-lesson (don't re-deepen X because Y)
Stop. Do not chain into another skill.
