Configuration

compound.yaml at the repo root is shared by both halves. compound-bench reads the providers, budget, and models sections; the trace pipeline reads the rest. bun run compound init writes a starter file and compound validate checks it.

Providers

providers:
  openrouter:
    base_url: https://openrouter.ai/api/v1
    api_key_env: OPENROUTER_API_KEY
    type: openai_compatible
  doubleword:
    base_url: https://api.doubleword.ai/v1
    api_key_env: DOUBLEWORD_API_KEY
    type: flex
  myhost:                                   # any OpenAI-compatible server
    base_url: http://localhost:8000/v1
    api_key_env: MYHOST_API_KEY
    type: openai_compatible
    cache_strategy: none                    # implicit | explicit_marker | none

A block named myhost is addressable as the provider token direct/myhost in any --providers list, and sits in the same report as the OpenRouter routes.

Budget

budget:
  paid_runs_enabled: true
  hard_limit_usd: 25.00
  smoke_cases_per_benchmark: 2

bfcl and ds1000 enforce this limit and the per-run --cap. tau2 bills whatever the episodes cost, so size the subset with a dry run first. Note the known gaps in how the limit is enforced under concurrency.

Models

models:
  candidates:
    - id: gpt-4o-mini
      provider: openai
      provider_ids:                # same weights, different id per host
        openrouter: openai/gpt-4o-mini

provider_ids lets one logical model keep a single identity in storage while each host receives the id it expects.

Environment variables