extract-project-patterns

extract-project-patterns

Internal recipe for code-gen skills. Find battle-tested template in project for code you about to write, via three-level locality cascade. Invoked by parent skills; not users directly.

Inputs

Caller passes:

Input	Required	Notes
positiveIds	one-of	Chunk IDs from prior cascade results
positiveCode	one-of	Raw code snippet(s) (embedded on the fly by find_similar)
behaviorQuery	one-of	NL query if no chunk/code is available
pathPatternL1	yes	From explore PG-OUTPUT (deepest subdomain target)
limit	no	Default 10

At least one of positiveIds / positiveCode / behaviorQuery MUST be present. Else return { templates: [], locality: "none", diagnostics: ["no input"] }.

Recipe — three-level locality cascade

L1 pathPattern = pathPatternL1                    (deepest subdomain)
L2 pathPattern = first 2 path segments of L1      (broader domain)
                 if L1 has ≤ 2 segments → L2 = L1, skip L2 step entirely
L3 pathPattern = null                             (project-wide)

L2 derivation example.

• L1 = **/domains/trajectory/git/rerank/derived-signals/** → segments [domains, trajectory, git, rerank, derived-signals] → L2 = **/domains/trajectory/**.
• L1 = **/chunker/hooks/** → segments [chunker, hooks] → L2 = L1, skip L2, jump to L3.

For each level in [L1, L2, L3]:

1. Call find_similar (or semantic_search / hybrid_search if only behaviorQuery available) with:
   • rerank: "proven"
   • pathPattern: <level> (omit for L3)
   • limit: <input limit, default 10>
   • inputs: positiveIds | positiveCode | query: behaviorQuery
2. Apply quality gate over result overlay labels:
   • ideal_count = chunks where
     • commitCount label is "low" or "typical", AND
     • ageDays label is "old" or "legacy", AND
     • bugFixRate label is "healthy"
   • If ideal_count ≥ 2 → return top result + locality annotation. Stop.
3. Apply reject filter (regardless of gate pass):
   • chunks where bugFixRate is "critical" OR (ageDays is "recent" AND commitCount is "low") excluded from returned top.
4. If no qualifying chunk → next level.

All three levels fail → return diagnostic "no proven templates for <input> in this project" so caller falls back (generate from scratch, ask user, etc.).

Output

Structured object for caller:

{
  templates: [
    {
      chunkId,
      path,
      level: "L1" | "L2" | "L3",
      labels: { commitCount, ageDays, bugFixRate, blameContributorCount, ... },
      blameDominantAuthor,
    },
    ...
  ],
  locality: "L1" | "L2" | "L3" | "none",
  diagnostics: [<per-level fail reasons>],
}

Caller reads templates[0] as reference; locality informs how to use template:

• L1 → matches subdomain exactly. Use template's blameDominantAuthor for style + review routing.
• L2 → template from sibling subdomain in same broader domain. blameDominantAuthor reviews technique, not exact code.
• L3 → template from project at large. blameDominantAuthor reviews technique only; verify architectural fit before adopting verbatim.
• none → no template found. Caller generates from scratch and surfaces this to user so they know to scrutinize result.

Skip clause

Return immediately empty templates if:

• None of positiveIds / positiveCode / behaviorQuery provided
• Project has no git enrichment indexed (no overlay labels → quality gate cannot run)

Invoked by

• tea-rags:data-driven-generation Step 2 (TEMPLATE)
• dinopowers:writing-plans (per code-generation / code-modification Task)
• dinopowers:executing-plans (per Task during execute)

Eval coverage

/optimize-skill extract-project-patterns runs baseline cases. Fixture file evals/cases.json added in follow-up PR (out of scope for initial recipe landing — see spec Component E).