Model plans

Model plans choose how a coding task uses OpenVibe.AI models. A plan has a mode, a list of roles, and checks. Every role call goes through AI's existing provider routing and uses the task's run token. The saved plan changes task behavior only when the owner enables CODES_MODEL_PLANS=1.

Modes

single: one selected model builds and verifies the result.
pair: a builder and a reviewer; the reviewer must approve. Findings return to the builder once.
council: two or three planner proposals; a judge picks or merges them with reasons before building and review.
escalate: start at the first ladder model and move up on a failed check, model refusal or low confidence. The task records the reason.

The copyable presets are Fast, Balanced, Deep and Council. Read them at GET /api/v1/model-plans/presets. Each role names an AI registry model id or auto, an effort from none, low, medium, high, xhigh, max, and optionally 1–5 samples and a token, minute or dollar budget. Saving checks availability and price against AI's registry; an unavailable id uses the first available model in the plan's ladder and returns that reason.

Configuration

The account default is saved at PUT /api/v1/model-plans with { "plan": { ... } }. Add ?project=<id> for a saved project default. A repository may contain .openvibe/codes.yml:

models:
  mode: pair
  roles:
    - role: builder
      model: openvibe/auto
      effort: high
    - role: reviewer
      model: openvibe/auto
      effort: medium
  checks:
    tests: true
    reviewer: true

For a task, send plan or preset in POST /api/v1/tasks. The task override wins over repository config, which wins over the saved project plan, which wins over the account default. The repository file is read and validated on the machine after cloning. Invalid YAML or a malformed models: block ends the task with a path and line or field error.

GET /api/v1/tasks/:id returns the resolved model_plan and each role's model, effort, tokens, cost, duration and outcome. The event stream exposes one AG-UI step per role call with metadata.ov.model. The verifier reads the final diff and repeats the configured tests and lint before success. The current automatic check runner uses test and lint scripts from the repository's package.json; missing scripts fail the requested check explicitly.