Model scoring

The orchestrator scores local models found in CCSwitch.

The score is a local heuristic. It is not a paid benchmark run.

What gets scored

Each discovered model receives scores for:

Key	Meaning
code	Coding and refactoring fit
long_context	Large repo and long prompt fit
reasoning	Planning and review fit
speed	Expected response speed
stability	Expected reliability
cost	Cost or quota friendliness
tool_use	Agent and tool workflow fit
multimodal	Image or mixed input fit

Then each role receives a weighted score.

For example, implementation weighs code and reasoning more heavily. Ops weighs stability, tools, speed, and cost more heavily.

Run scoring

python "$CC_ORCHESTRATOR_HOME/cc_orchestrator.py" score-models

Write score and routing reports:

python "$CC_ORCHESTRATOR_HOME/cc_orchestrator.py" write-reports

Auto routes

The default policy can use aliases like:

{
  "code_strong": "auto:implementation",
  "review_strong": "auto:review",
  "fast": "auto:testing"
}

This means the orchestrator picks the highest local role score from the models present on that machine.

How to improve routing

1. Add a stronger model to CCSwitch.
2. Rerun score-models.
3. Rerun write-auto-policy.
4. Rerun workflow-plan.

The orchestrator does not mutate global CCSwitch state. It only reads profiles and injects the selected provider settings into the worker process.