Seam Finder

Seam Finder — Where to Cut Safely

/seam-finder answers the question "given my goal, which existing module is the smallest, safest place to extend or split?" It applies the Fowler / Thoughtworks Uncovering Mainframe Seams scoring methodology to a deterministic dependency graph.

A "seam" is a junction point where program flow can be diverted without rewriting upstream or downstream code. Good seams:

1. Sit at an observable boundary (HTTP route, queue, table, public class)
2. Are a funnel — many modules converge on them
3. Have read/write asymmetry — pure readers and pure writers are safer than mixed-use nodes

This skill is the bridge between brownfield discovery and /change, /refactor, /spec. Run it before deciding where in the existing code to land a change.

Usage

/seam-finder "add team invites to user onboarding"
/seam-finder "split the billing module into a separate service"
/seam-finder "introduce idempotency to webhook delivery"

Always pass a concrete goal. Without one, the skill ranks structural seams generically — useful for a refactor scan but less precise.

Prerequisites

specs/brownfield/code-graph.json must exist. If it does not:

1. Suggest running /code-map first.
2. Stop. Do not try to score seams from grep results.

Staleness check: if .claude/state/graph-dirty.jsonl exists and is non-empty, the graph is behind recent edits — the graph-refresh Stop hook normally drains it, but if entries remain, patch first: python3 .claude/skills/code-map/scripts/code_index/code_index.py --root . --out specs/brownfield/code-graph.json --files <dirty paths>.

If specs/brownfield/coupling-report.md exists, prefer it for hub data; otherwise compute on the fly from code-graph.json.

Outputs

Each candidate includes:

• Module / file path
• observable_score (0–1) — how externally visible the boundary is
• funnel_score (0–1) — fan-in + fan-out, normalized
• asymmetry_score (0–1) — read vs write imbalance
• total_score (0–1) — weighted combination (defaults: 0.4 / 0.4 / 0.2)
• Evidence — which files import or call it, with file:line refs
• Recommended action — extend, wrap, split, introduce-adapter, avoid

Steps

Step 1 — Load Graph

test -f specs/brownfield/code-graph.json || { echo "Run /code-map first"; exit 1; }

Step 2 — Score Seams

node .claude/skills/seam-finder/scripts/score_seams.js \
  --graph specs/brownfield/code-graph.json \
  --goal "$GOAL" \
  --out specs/brownfield/seams-${SLUG}.md

By default the scorer excludes test and fixture paths (tests/, __tests__/, spec/, *_test.*, *.spec.*, fixtures/, mocks/, examples/). Pass --include-tests or --include-fixtures if you genuinely want those ranked (e.g. you're refactoring the test suite itself).

The scorer:

1. Computes per-node fan-in / fan-out from the graph.
2. Tags nodes with an observable kind by file path heuristics:
   • routes/, controllers/, api/, handlers/, views/ → HTTP boundary (1.0)
   • queue/, events/, consumers/, producers/, webhooks/ → message boundary (0.9)
   • db/, repository/, models/, migrations/ → data boundary (0.8)
   • adapters/, gateways/, clients/, integrations/, providers/ → integration boundary (0.7)
   • services/, domain/, usecases/, core/, workflows/ → service layer (0.6)
   • Public classes/exports in any other module → module boundary (0.5)
   • Private helpers, deeply nested utils → internal (0.1)
3. Scores read/write asymmetry from imports direction and (if present) reads/writes edges.
4. Filters by goal keywords: nodes whose path or symbols match goal terms get a goal-relevance bump (multiplier 1.5).
5. Ranks descending by total_score.

Step 3 — Recommend Action

For each top-N candidate, the skill labels a recommended action based on score profile:

Step 4 — Verify Goal Fit

After the script writes the candidate list, re-read the top 3 candidates' source files and confirm they actually cover the goal's domain. The score is structural; the verification is semantic.

If none of the top 3 candidates fit the goal:

• Re-run with a refined goal phrase.
• Or, if the goal genuinely has no good seam, recommend /spec to plan a new module rather than forcing a fit.

Step 4.5 — Phase Evaluation Gate

Spawn the evaluator agent (artifact mode) to validate seam candidates.

Agent invocation:

Spawn Agent with subagent_type="evaluator" and prompt:

• Phase: seam
• Artifact: specs/brownfield/seams-{goal-slug}.md
• Upstream: specs/brownfield/code-graph.json
• Rubric: Read .claude/templates/phase-eval-rubrics.json, key "seam"
• Iteration: 1 (increment on retry)
• Previous score: null (or previous iteration's weighted_average)
• Verify top 3 candidates reference files/functions that exist in the codebase (grep verification).
• Write result to specs/reviews/phase-seam-eval.json

Ratchet loop (max 2 iterations):

1. If verdict is PASS — proceed to Step 5 (Hand Off) with eval summary.
2. If verdict is FAIL — re-score or re-rank candidates based on findings. Re-run evaluator.
3. Ratchet rule: weighted_average must be >= previous iteration. Revert on regression.
4. After 2 iterations — present best version with findings.

Step 5 — Hand Off

Reference the chosen seam in the next step:

• /change "<goal>" — extend seam: <path>
• /refactor <path> (when the seam is the refactor target)
• /spec (when no seam fits and a new module is the right call)

Goal-Slug Convention

The output filename uses a short, lowercase, dash-separated slug derived from the first 3–5 meaningful words of the goal.

Gotchas

• No graph, no seams. Do not attempt to score from filenames alone. Run /code-map first.
• Observable score is heuristic. A routes/ directory might be private; a lib/ directory might be the public API. Always re-read the top candidates to confirm.
• Goal relevance is keyword-based. Synonyms may miss the bump. If a candidate scores high but the goal mentions different terms, override manually.
• Hubs are not always seams. A god-module with fan-in 50 may be a refactor target, not a place to extend. The recommended action of introduce-adapter exists for this case — extract a public seam first.
• Cycles muddy the waters. A module inside a cycle has unreliable fan-in/fan-out. The script flags cycle members so the recommendation does not silently land in tangled territory.
• Do not auto-execute the recommendation. Seam-finder produces a plan input, not a refactor. Always confirm with the user before acting on extend / split.