061-storm: divergent exploration and idea generation

[ahoward]

Summary

brane storm — divergent brainstorming against accumulated knowledge. "Here's what I know — what's missing? What connections haven't I made? What questions should I be asking?"

Why

Brane is purely retrieval right now. It finds what's there but never proposes what isn't. An agent that can only recall is half an agent — it also needs to speculate, question assumptions, and surface blind spots. Bunny's storm does this against markdown files; we do it against the structured graph.

What

CLI

brane storm                                    # brainstorm on bare knowledge
brane storm "security implications"            # seeded brainstorm
brane storm --input threat-model.md            # brainstorm against a document
brane storm --rounds 3                         # multi-round deepening

Flags

• --rounds <n> — iterative rounds (reload context each round, default: 1)
• --input <path> — seed from file
• --dry-run / -n — show what would be created without writing
• --json / -j — JSON output

Flow

1. Load context — vector search for relevant concepts + episodes (seeded by topic or broad sample)
2. LLM generation — Claude thinks divergently: gaps, questions, connections, contradictions
3. Output — new concepts, edges, episodes, plus suggestions

LLM Response Schema

{
  concepts: { name: string, type: string }[],
  edges: { source: string, target: string, relation: string }[],
  episodes: { observation: string, tags: string[] }[],
  suggestions: { kind: "question" | "source" | "lens", value: string, reason: string }[],
  reasoning: string
}

Suggestions

Storm doesn't just create — it suggests next actions:

• question — "Have you considered X?" (becomes a future ask)
• source — "You should read about Y" (becomes a future digest)
• lens — "Try looking at this from Z perspective" (becomes a future lens)

Multi-round

Each round reloads context (seeing previous round's additions), goes deeper. Stops early if LLM signals convergence.

MCP Tool

storm: { seed?: string, rounds?: number, dry_run?: boolean }

Depends On

• 059-digest: consume URLs, files, directories, stdin into knowledge graph #82 (059-digest) — storm is most useful with a populated graph
• 060-ask: conversational synthesis over accumulated knowledge #83 (060-ask) — ask validates that storm-created knowledge is retrievable

Acceptance

• Generates new concepts, edges, episodes from existing knowledge
• Suggestions include questions, sources, and lens ideas
• Multi-round deepening works
• Seeded and unseeded modes
• Dry-run shows what would be created
• MCP tool exposed
• Spike test passes