Tribunal CLI

tribunal gets a structured second opinion from advocate, skeptic, and analyst LLM perspectives, then synthesizes the result with a deliberator model.

Run with API key environment variables already present:

npm --workspace @clipboard-health/tribunal run dev -- "Should we migrate?" --verbose

For local defaults, copy the checked-in example to a gitignored root config:

cp tribunal.config.example.json tribunal.config.json

The CLI searches for tribunal.config.json from the current working directory upward. Config values are defaults only; explicit CLI flags still take precedence.

{
  "apiKeys": {
    "anthropic": "op://Private/ANTHROPIC_API_KEY/credential",
    "googleGenerativeAi": "op://Private/GOOGLE_GENERATIVE_AI_API_KEY/credential",
    "openai": "op://Private/OPENAI_API_KEY/credential"
  },
  "models": {
    "advocate": "anthropic:claude-opus-4-7",
    "skeptic": "openai:gpt-5.5",
    "analyst": "google:gemini-3.1-pro-preview",
    "deliberator": "openai:gpt-5.5"
  },
  "reasoning": {
    "advocate": "max",
    "skeptic": "xhigh",
    "analyst": "high",
    "deliberator": "xhigh"
  },
  "outputFormat": "text",
  "showPerspectives": false,
  "saveIntermediates": true
}

--verbose writes progress to stderr, including per-role start, finish, and failure lines, plus a dot every five seconds while model calls are in flight.

Each run writes durable intermediate snapshots to .tribunal/runs/<timestamp>.json by default. The file is updated as each role starts, finishes, or fails, so completed perspective outputs are preserved even when a later model call fails. Use --save-intermediates <path> to choose a specific file, or --no-save-intermediates to disable snapshots.

After a successful run, an HTML report is written to .tribunal/reports/<timestamp>.html and opened in your default browser. The report shows the deliberator's verdict, recommendation, and confidence at the top, with each perspective (advocate, skeptic, analyst) as a collapsible accordion below. Override the path with --html <path>, skip opening with --no-open, or disable the report entirely with --no-html.

Override model reasoning or thinking levels per role:

npm --workspace @clipboard-health/tribunal run dev -- "Should we migrate?" \
  --model advocate=anthropic:claude-opus-4-7 \
  --model skeptic=openai:gpt-5.5-pro \
  --model analyst=google:gemini-3.1-pro-preview \
  --deliberator openai:gpt-5.5-pro \
  --reasoning advocate=max \
  --reasoning skeptic=xhigh \
  --reasoning analyst=high \
  --reasoning deliberator=xhigh

Reasoning levels are mapped to each provider's AI SDK options. OpenAI supports none, minimal, low, medium, high, and xhigh; Google supports minimal, low, medium, and high; Anthropic supports low, medium, high, xhigh, and max.

Run through 1Password without writing secret values to disk:

npm --workspace @clipboard-health/tribunal run dev:op -- "Should we migrate?"

When apiKeys are present in tribunal.config.json, tribunal-op passes those 1Password references to op run, which resolves them in the child process environment at runtime.