brainstrust

Brains Trust

Part of the shipwright method — see the shipwright skill for the work-loop this move fits into and when to invoke it.

Get a grounded second opinion from leading models. The defining move versus the old version: the consulted model is an agentic pair reviewer — it has read-only repo tools and explores the code itself, rather than reasoning over a bundle you hand-fed it. You point it at a few starting files; it follows the trail.

The one routing rule (read this first)

Who you consult decides the transport — and the cost:

Consulting…	Use	Cost
A Claude model (Opus/Sonnet/Haiku/Fable)	a Task subagent with an explore-and-critique prompt	free — rides the Claude Code subscription
A non-Anthropic model (Gemini / GPT / Qwen / DeepSeek …)	the OpenRouter agent harness (src/consult.ts)	paid OpenRouter tokens

A Claude subagent already has Read/Grep/Glob and an agent loop — it is an agentic pair reviewer out of the box, for free. So never pay the API to ask Claude. Pay only for the diverse voices — the whole value of a brains trust is different blind spots, and that means non-Claude models.

For an important call, a mixed panel is ideal: one free Claude subagent that reads the real files, plus one or two non-Anthropic voices via the harness.

Correlated blind spots: three Claude subagents "agreeing" is not three independent votes — they share Claude's blind spots. Same-provider panels too. For genuine diversity, span providers.

Methodologies (built-in recipes)

Each bundles a pattern + model count + a system-prompt lens. Pick by the kind of work:

Key	For	Pattern
review	correctness/security review of target code	consensus ×2
architecture	design trade-offs, simpler approaches	consensus ×2
debug	stuck after attempts → ranked root-cause hypotheses	consensus ×2
security	threat-model, ranked by exploitability × blast radius	consensus ×2
devils-advocate	argue hard AGAINST the current plan	single
strategy	product/approach decision	consensus ×2
explore	open: what are we missing? blind spots?	consensus ×2
ideate	DIVERGE — wide range of options, don't converge	consensus ×3

(ideate is brainstorming: same tools, different question — the model reads enough to ground ideas, then generates many distinct directions instead of narrowing.)

Setup — the harness needs its npm deps

src/consult.ts imports npm packages and node_modules is not committed. So before the first consult on a machine, install the deps: run npm install in this plugin's directory (idempotent — harmless to repeat, and a plugin update may need it again). Run the harness from that same directory so the deps and tsx resolve. If a consult ever errors with a missing module, that's the signal to install.

Running a non-Anthropic consult

Run from the plugin's directory:

export OPENROUTER_API_KEY=...        # Claude consults need no key
npx tsx src/consult.ts \
  --methodology review \
  --repo /path/to/repo \
  --question "Is the auth middleware order correct, and can a gated role reach /admin?" \
  --paths "src/server/index.ts:middleware registration; src/server/middleware/auth.ts:the gate"

--repo is the target repo to consult on (the harness reads files relative to it, so where you run from doesn't change which code the model sees).

• --methodology (default explore) picks the recipe. --pattern / --count / --models override it.
• --models is auto-chosen as flagship non-Anthropic, one per provider from the live list (models.flared.au), with a stale fallback if it's unreachable — or pass explicit ids (--models openai/gpt-5.4,google/gemini-3.1-pro-preview).
• --paths are hints (path:why, ;-separated), NOT limits — the model roams from there.
• --max-cost (default $0.50/model) and --max-steps (25) are hard ceilings via the agent loop's stopWhen. Each consult prints its own token cost.
• Output: each model's view to stdout (+ .brainstrust/<ts>-<methodology>/), with a total cost line.

Running a Claude consult (free)

Don't use the harness. Spawn a Task subagent: give it the question, the methodology lens (copy the relevant row's framing), and the starting file paths, and tell it to explore the repo and return a prioritised, file:line-cited critique. It reads the code natively. For consensus, spawn 2-3 with different stances (skeptic / pragmatist / security-hawk) in parallel.

The discipline (what keeps it honest)

• The panel is INPUT, not verdict. You synthesise: note where they agree/disagree, add your own read, and say plainly when you disagree with all of them. Don't defer to "the models said X" — they are often confidently wrong in chorus.
• Grounded only. The harness's system prompt forbids claims about code the model hasn't read and demands file:line citations. Hold consulted output to that — discount ungrounded assertions.
• Span providers for real diversity. Redundant providers give correlated errors.
• Mind the spend. Non-Claude consults cost money; Claude subagents don't. Run a panel for decisions that matter, not every edit.

When to use / not

Use: before a major architectural change; stuck debugging after 2+ attempts; security-sensitive code; challenging your own plan; ideation when you want breadth; any genuine "what are we missing?". Skip: simple syntax, well-known answers, every small edit (slow + costs money).