benjaminstelzer/scoville-brainstorm

Scoville Brainstorm

Read-only divergence before a material choice. For YES only, execute once:

CORE -> READ -> FRAME -> FREEZE -> RUN -> COLLECT -> CONVERGE -> RENDER -> STOP

Process state is trace-owned:

profile_targets={Compact:<=3,Standard:<=5,Deep:6..8}; spawned=successful tool-observed spawn/delegation calls; landscape_mode=native|research-owned

Named roles, reasoning passes, or imagined agents never count as isolated generators, a separate landscape agent, or an independent critic. Do not self-report spawn topology or isolation in coordinator text or the decision artifact; a host or evaluator may attach those facts only after inspecting calls and branch prompts.

Dispatch

Before answering any explicit combined Research and Brainstorm request, the next operation after Core is one observable read of research-composition.md. Answering that request from Core alone is invalid even when the topology appears inferable. The loaded reference owns the exact explicit_combined mode and retrieved-data trust boundary.

• Route by decision shape, not domain: an unknown root cause plus a request for several materially different failure mechanisms is YES; debugging is NO only after one cause is established or the request asks for direct diagnosis, implementation, or review.
• NO: canonical or single answer; known root cause; selected implementation, review, wording, small reversible work, durable planning, or transfer. Return the ordinary answer immediately. Run no tool and read no task source or inert fixture.
• ASK: broad exploration versus one answer materially changes cost and intent is unclear. Ask one question; read no task source.
• YES: explicit brainstorming or several materially different, unusual, or underexplored directions. Explicit invocation bypasses only the cost question, never authority or safety. A host instruction to use this Skill fixes YES.
• COMBINED: after YES and before READ, when the user explicitly requests Scoville Brainstorm together with Scoville Research and Research is independently available and applicable, read the reference exactly once and set landscape_mode=research-owned. Otherwise keep landscape_mode=native and do not read the reference.

Machine

1. CORE: The catalog/discovery read that exposed this body is activation and the only Core read. Never read, stat, list, or inspect it again.
2. READ: Read only every user-named task source, one file per shell operation and exactly once, so each tool result has one unambiguous owner. Never batch files or infer README.md, AGENTS.md, or another default; inventory, probe, retry, dummy command, metadata check, and later file IO are forbidden. Retain every nonempty result by its command path. A failed path invalidates only itself.
3. FRAME: Build one in-memory brief: outcome, language, effort profile, supplied facts and owners, hard constraints and authority, challengeable and fixed assumptions, and permitted source scope. From each non-landscape brief block, first scan every literal colon-terminated ID label such as - D1:, then copy all IDs character for character and in source order into fixed_ids; never synthesize, rename, or take IDs from an output contract. Include the final label and authority, selection-only, stop, or no-mutation labels. The label count and fixed_ids count must match before interpreting content. Observation versus rule does not change membership. If any non-landscape source contains a labeled line, fixed_ids=[] blocks FREEZE and RENDER; populate it first from the ledger. Before ideation, emit one compact nonfinal trace checkpoint: LEDGER blocks=<count>; fixed_ids=<exact comma list>. It is working state, not part of the requested artifact; emit it as plain text, never through or for a tool. Resolve factual uncertainty only when it changes this frame; do not research solutions yet.
4. FREEZE: Set the profile target from Process state. Before any branch starts, freeze every full generator prompt plus stable ID and content hash from the same brief. For landscape_mode=native, also freeze the native landscape prompt. For landscape_mode=research-owned, freeze the Research handoff contract and never freeze or dispatch a second landscape prompt. Later waves see no earlier output. Cover distinct causal operators, including one load-bearing-assumption challenge and one strongest practical comparator.
5. RUN: When fresh isolated agents are available, launch one per generator and, only for landscape_mode=native, a separate landscape agent, in parallel up to capacity. In research-owned mode the combined protocol supplies one Research-owned lane after generator collection; do not launch a native Brainstorm landscape agent. Generators receive only the brief, one operator, and a request for mechanism, preserved constraints, benefit, load-bearing risk, and cheapest falsifier. They never see sibling output or landscape evidence. The landscape agent sees only fixed facts and permitted sources and reports close matches, failed approaches, scope, and unresolved evidence. When spawned=0, freeze one consolidated generator prompt before ideation, run it once in the coordinator, and apply Process state exactly.
6. COLLECT: Wait until every started branch is terminal. Keep raw outputs separate and count the distinct surviving ideas. A failed branch is missing evidence, never permission to invent it. Accept exactly one applicable landscape result: native in standalone mode or Research-owned in combined mode.
7. CONVERGE: Normalize candidates to mechanism, constraints, evidence relationship, benefit, risk, and falsifier; merge paraphrases; reject broken constraints and unsupported facts; identify traps. Use the one collected landscape result as the sole landscape input. When available, use an independent critic. Retain at most three distinct directions (Compact: two) and deepen only those. Reject every direction breaking a fixed_id before RENDER.
8. RENDER: Obey the user's exact schema, key order, language, and closed values. Schema types dominate defaults: emit a requested scalar or enum as that scalar, never an enriched object. Otherwise return, in order: Brief, Landscape, Idea map, Shortlist, Traps, Deepened directions, Decision point. Prefer the smallest complete artifact over a branch transcript. In structured output, set activation.activated=true only when the schema requests activation; seal brief.fixed_constraint_ids=fixed_ids and implementation/experiment/ durable-record flags false. external_search_performed reflects actual tool use; named fixtures alone mean false. constraints.violated contains only IDs from fixed_ids that the proposed directions actually break; never include a challenged assumption, current defect, process limitation, or free-text sentinel. When no fixed ID is broken, emit []; after CONVERGE, a completed artifact therefore emits constraints.violated=[]. For JSON, pretty-print with two-space indentation, one key or value per line, and each closer on its own opener-matched line. Check required top-level fields once.
9. STOP: Return the decision artifact. Do not select for the user, edit, install, run a falsifier, create a Plan or Decision, send, publish, or deploy.

Evidence labels and transfer

Use exactly one label per candidate: Established = close same mechanism; Adaptation = known mechanism transferred to a materially different context; Recombination = known mechanisms whose interaction creates the difference; Candidate-original = no close match in the documented bounded search; and Unresolved = insufficient evidence. None proves novelty or patentability.

Transfer only after human selection. Family standalone: discovery != installed|active|applicable|required; absent|inactive => ignore/no require|install|simulate|reimplement; active+applicable => owner concern only, self continues; opt-out local.

Other owners: scoville-code-anti-ai-slop engineering/proof; scoville-ui-anti-ai-slop interface/rendered proof; scoville-scribe-anti-ai-slop wording/fidelity; scoville-plan records/lifecycle; scoville-handoff transfer.