Coach mode

Coach mode pairs a free local draft (Ollama) with an optional paid review (OpenAI, Anthropic, etc.). You choose which version becomes training data.

Open Dashboard → Coach (/dashboard/coach).

Setup

1. Ollama — required for the local draft (see Agent Lab).
2. Paid provider — optional. Connect API keys under Dashboard → Settings → Model provider connections.

If no paid provider is connected, Coach still runs local drafts only.

Workflow

1. Enter a prompt and select local + coach models.
2. Click Run coach session.
   • Local agent drafts first (coach_local_draft trace, $0).
   • If enabled, the coach critiques and rewrites (coach_curated trace, billed by provider).
3. Click Train on this (local) or Train on coach version.
4. Export curated examples from Training data with Coach picks only.

API

Run session — POST /api/coach/run

{
  "prompt": "Summarize our refund policy.",
  "localModel": "llama3.2",
  "coachProvider": "openai",
  "coachModel": "gpt-4o-mini",
  "runCoach": true
}

Mark training pick — POST /api/coach/select

{
  "sessionId": "coach_…",
  "choice": "coach",
  "prompt": "…",
  "localTraceId": "…",
  "coachTraceId": "…",
  "coachResponse": "…"
}

List connected coaches — GET /api/coach/providers

Cost control

• Local draft always runs first ($0).
• Coach review is optional (checkbox in the UI).
• Only traces you mark with trainingPreferred are included in curated exports.