brainstrust

brainstrust

A second opinion that's grounded in your actual codebase. The consulted model doesn't get hand-fed a bundle of files — it has read-only repo tools (read_file / grep / find / list_dir) and explores the code itself, the way a Claude Code subagent or a pair programmer does. You point it at a couple of starting files; it follows the trail.

Sibling of decisions — decisions asks a human for one clean call; brainstrust asks other models for a grounded read.

The routing rule (this is the whole cost story)

Consulting…	Transport	Cost
a Claude model	a Task subagent (native Read/Grep/Glob)	free — rides the Claude Code subscription
a non-Anthropic model	the OpenRouter agent loop (src/consult.ts)	paid OpenRouter tokens

A Claude subagent is already an agentic pair reviewer, for free — so never pay the API to ask Claude. Pay only for the diverse voices (different providers = different blind spots), which is the actual point of a brains trust. Best of both: a mixed panel — one free Claude subagent reading the real files, plus a non-Anthropic voice or two.

How the non-Anthropic path works

src/consult.ts uses the OpenRouter Agent SDK (@openrouter/agent): it hands the model the read-only repo tools and runs the tool-calling loop automatically — the model requests read_file("…") / grep(…), the SDK executes it against the repo, feeds the result back, and loops until the model answers. Safety model: there is deliberately no write/edit/bash tool and every path is clamped to the repo root, so a read-only consult needs no container — the absence of hands is the sandbox.

export OPENROUTER_API_KEY=...
npm install
npx tsx src/consult.ts \
  --methodology review \
  --repo /path/to/repo \
  --question "Can a gated role reach /admin given the middleware order?" \
  --paths "src/server/index.ts:middleware registration; src/server/middleware/auth.ts:the gate"

Flags: --methodology (recipe, default explore), --models (explicit ids, else flagship non-Anthropic one-per-provider from models.flared.au with a stale fallback), --pattern / --count, --max-cost (default $0.50/model) and --max-steps (25) as hard ceilings, --question-file, --paths (hints, not limits). Each consult prints its own token cost.

Methodologies

review · architecture · debug · security · devils-advocate · strategy · explore · ideate — each bundles a pattern, model count, and a system-prompt lens. ideate is the brainstorming mode: same tools, different question (diverge into many options instead of converging).

What's in here

Path	What
skills/brainstrust/SKILL.md	the operating method (routing, methodologies, the discipline)
src/consult.ts	CLI: picks models + methodology, runs the agent loop per model, prints + saves
src/tools.ts	the read-only, repo-clamped filesystem tools handed to the model
src/models.ts	flagship non-Anthropic model selection (live list + stale fallback)
src/methodologies.ts	the built-in recipes

Discipline

The panel is input, not verdict — you synthesise, and say when you disagree with all of them. The system prompt forbids claims about code the model hasn't read and demands file:line citations; hold the output to that. Span providers for genuine diversity (same-provider or all-Claude panels share blind spots). Mind the spend — non-Claude consults cost money; Claude subagents don't.

Requirements

Node 18+ (for fetch/AbortSignal.timeout), OPENROUTER_API_KEY for non-Anthropic consults. rg (ripgrep) is used for grep if present, falling back to system grep.