brainstorming

dinopowers: brainstorming

Wrapper over superpowers:brainstorming. Ensures creative exploration starts from real git signals (hotspots, owners, legacy debt) of target area — not from assumptions about how code looks today.

Iron Rule

tea-rags enrichment MUST be called BEFORE Skill(superpowers:brainstorming) — whenever target area identifiable. Correct preset selection (hotspots, ownership, techDebt) + correct parameters + correct ordering = core value of this wrapper.

If no code area identifiable from user request (pure conceptual brainstorming): skip enrichment, invoke superpowers:brainstorming directly. State explicitly — do not fabricate an area.

Chaining rule: see CHAINING.md — every dinopowers:X redirects superpowers:X. NEVER bypass wrapper.

Index freshness: see FRESHNESS.md and tea-rags/rules/index-freshness.md. No background reindex hook — worktree-plan freshness explicit (clone + per-task reindex in dinopowers:executing-plans); run mcp__tea-rags__index_codebase manually to search code edited but not yet committed, BEFORE first tea-rags call.

Step 1 — Extract target area

From user request identify:

Element	Example
Area — path or domain	src/core/domains/ingest/**, auth/, "rerank presets"
Change type — what the brainstorm is about	"new feature", "refactor", "bug-fix approach"
Scope hint — narrower filter if mentioned	file name, class name, symbol

Compose:

• pathPattern for Step 2 (e.g. **/auth/**, src/core/domains/ingest/**/chunker/**)
• query — single sentence describing change intent

If no area mentioned or derivable: skip to Step 4 with empty enrichment. Report "no area identifiable — proceeding without tea-rags enrichment".

Step 2 — Three tea-rags enrichment calls (four with codegraph)

Run these three mcp__tea-rags__semantic_search calls in parallel (same tool call block), each with metaOnly: true (want signals, not content):

Call	rerank preset	Why
A — Hotspots	"hotspots"	Bug-prone zones the brainstorm must account for (high churn + recency + bugFix)
B — Ownership	"ownership"	Knowledge silos — single-owner files that need reviewer pairing
C — Tech debt	"techDebt"	Legacy zones where "just refactor it" hides cost (age + churn + bugFix)
D — Backbone	"architecturalHub"	Codegraph only — structural hubs (high fanIn / isHub) the design must respect; a change to a backbone has wide blast radius

Call D runs ONLY when codegraph active (prime ## Enrichment lists codegraph.symbols). When that line absent, omit Call D — run A/B/C only, note structural backbone not assessed (do NOT substitute a similarity-ranked list as "the hubs"). See search-cascade "Graph navigation".

Exact parameters per call:

project:     <alias from list_projects — RECOMMENDED, omit path when set>
path:        <current project path — fallback when no alias is registered>
query:       <intent sentence from Step 1>
pathPattern: <pathPattern from Step 1>
rerank:      "hotspots" | "ownership" | "techDebt" | "architecturalHub"
limit:       10
metaOnly:    true

("architecturalHub" = Call D, codegraph-gated per the note above.)

Do NOT substitute:

Wrong tool	Why wrong
mcp__tea-rags__hybrid_search	Preset-based reranking is tied to semantic_search; hybrid ignores risk presets
mcp__tea-rags__find_similar	Requires an existing symbolId; brainstorm targets an area, not a known symbol
mcp__tea-rags__find_symbol	Returns one symbol; brainstorming needs area-wide risk distribution
Running the three calls sequentially	Wastes time; they are independent — issue them in one parallel block

Do NOT pass:

• metaOnly: false — content not needed; larger payloads slow wrapper
• Custom rerank weights instead of three named presets — loses calibrated risk-assessment semantics
• filter on git.ageDays / git.commitCount — presets already encode these signals; adding filters shrinks result set, hides debt

If all three calls return 0 results (area new, untouched by git): report "no git-enriched results for pathPattern <X> — area is new or excluded from indexing" and skip to Step 4 without enrichment block. Do NOT fabricate signals.

Step 3 — Build enrichment block

From three result sets, compose compact block (max ~15 lines):

### tea-rags enrichment for <pathPattern>

**Hotspots (bug-prone):**
- <relativePath>:<startLine>-<endLine> — bugFixRate <N>, commitCount <N>, ageDays <N>
- ... (top 3)

**Knowledge silos (single live-line owner):**
- <relativePath> — <blameDominantAuthor> (<blameDominantAuthorPct>%), <commitCount> commits
- ... (top 3)

**Tech debt (legacy + churn):**
- <relativePath> — ageDays <N>, commitCount <N>, relativeChurn <X>
- ... (top 3)

Cap each section at 3 entries. If a file appears in multiple sections, note explicitly (→ also in Hotspots) — overlap is strong brainstorming signal. Do NOT paste raw JSON. Extract readable relativePath + 2-3 key signals per entry.

Step 4 — Invoke superpowers:brainstorming

Invoke Skill tool with superpowers:brainstorming. Prepend enrichment block from Step 3 as context. Phrase handoff as:

"Before exploring, note these risk signals in the target area: …… Use them to pressure-test ideas — especially around the hotspot files.

Chaining rule reminder: if your cycle would next invoke a superpowers:Y skill (writing-plans, test-driven-development, etc.), invoke the dinopowers:Y wrapper instead when one exists — see the Chaining rule section above."

Let superpowers:brainstorming run full exploration cycle — this wrapper does not replace it, only grounds it.

Red Flags — STOP and restart from Step 2

• "I already know which files are hotspots here" → run Step 2 anyway; memory stale
• "tea-rags is slow for 3 calls, let me do one" → three presets = different lenses; one call loses two
• Substituted hybrid_search / find_similar / find_symbol → redo with semantic_search + preset
• Ran the three calls sequentially → reissue as parallel (same tool block)
• Invoked superpowers:brainstorming first, tea-rags after as "validation" → wrong order, restart
• Set metaOnly: false to "see content" → unnecessary; restart with metaOnly: true
• Let superpowers:brainstorming chain into a raw superpowers:Y (e.g. superpowers:writing-plans) without redirecting to dinopowers:Y → intercept and invoke wrapper instead (see Chaining rule)

Common Mistakes

Mistake	Reality
Skip tea-rags because "brainstorming is abstract"	If the user named a code area, it's not abstract. Enrich it.
Use rerank: "relevance" instead of the three risk presets	Loses the point — brainstorming should see risk, not just relevance
Collapse three calls into one with custom weights	Custom weights ≠ calibrated presets; output is noisy and hard to interpret
Paste raw results[] JSON into Step 4 handoff	superpowers:brainstorming drowns in noise; extract 3 entries per lens
Fabricate an area when user didn't name one	Explicit "no area — skipping enrichment" is honest; faked pathPattern is a lie