/rune:brainstorm — Creation Log

Problem Statement

Rune users who want to explore an idea must run /rune:devise — a 7-phase pipeline that spawns agent teams, runs research, forge, and review. This is overkill for casual ideation. Users cannot brainstorm casually without triggering heavy infrastructure, cannot save brainstorm output persistently (devise writes to tmp/), cannot stop after brainstorming without being forced into planning, and cannot resume brainstorming later or hand off to /rune:devise when ready.

The brainstorm logic was embedded in devise/references/brainstorm-phase.md (474 lines) as Phase 0 of the devise pipeline, making it inaccessible as a standalone workflow.

Alternatives Considered

Alternative	Why Rejected
Keep brainstorm inline in devise Phase 0	Users must commit to full 7-phase pipeline to brainstorm. No standalone access. No persistent output.
Duplicate brainstorm logic into a new skill	Code duplication between devise Phase 0 and standalone skill. Two sources of truth for brainstorm behavior.
Simple conversational skill (no agents)	Misses the multi-perspective value of advisors. Solo mode covers this case, but Roundtable mode is the differentiator.
Full research agents in brainstorm	Over-engineered — deep research belongs in devise Phase 1. Advisors do lightweight 30-second codebase scans instead.
Advisors communicating directly with each other	Creates unmoderated cross-talk. All communication through Lead ensures coherent discussion flow and prevents advisor echo chambers.

Key Design Decisions

• Extract-don't-duplicate: Brainstorm skill becomes the single source of truth. Devise Phase 0 delegates to the brainstorm protocol. This mirrors how /rune:elicit is both standalone AND used internally by devise/forge/review.

• Three modes (Solo/Roundtable/Deep): Provides a spectrum from zero-overhead conversation to full multi-agent analysis. Users choose their depth at startup, or use --quick/--deep flag shortcuts.

• Roundtable Advisors as Agent Team teammates: Advisors run in their own context windows (subagent isolation) with tool access for lightweight codebase research. This grounds their questions in codebase reality without needing separate research agents. 200-300 word limit per response prevents context bloat.

• Lead as moderator (not relay): All advisor communication flows through the Lead, who curates inputs into a coherent discussion. Advisors never communicate with each other. This prevents unmoderated cross-talk and ensures the user gets a synthesized perspective, not raw forwarding.

• Persistent output to docs/brainstorms/: Unlike devise's tmp/ output, brainstorm documents persist as project knowledge. This enables auto-detection by future devise runs and serves as institutional memory.

• 7-dimension quality gate: Scoring model determines handoff readiness. Score >= 0.70 triggers devise suggestion. Score < 0.70 suggests another round. Prevents premature handoff to planning with underspecified requirements.

• disable-model-invocation: true: Prevents Claude from auto-loading this skill when it sees brainstorm-related keywords. The skill creates agent teams and interactive sessions — user must always consent via explicit invocation.

• State Machine Pre-Validation (Phase 4.5): In Deep mode, when >= 5 phase indicators are detected, state-weaver pre-validation triggers automatically. Provides early PASS/CONCERN/BLOCK verdict before handoff to devise, preventing state machine complexity from being discovered late in the planning pipeline.

• Workspace at tmp/brainstorm-{timestamp}/: Preserves full context chain (advisor observations, codebase scans, round history, elicitation outputs) for devise consumption. Cleaned up by /rune:rest like other tmp/ artifacts.

Iteration History

Date	Version	Change	Trigger
2026-03-03	v1.0	Initial creation — 3 modes, Roundtable Advisors, 7-dimension quality gate, workspace structure	Standalone brainstorm feature (v1.130.0)