Harness Routing Skill

cotcodec / harness/routing

Purpose

Routing chooses an orchestration condition per visible message from explicit, auditable features rather than using one condition globally.

Mental model & key files

• features.py extracts message/state features available at decision time.
• policy.py maps those features to a condition and records the decision.
• Optimization and evaluation live outside the online policy so frozen policies remain reproducible.

Patterns to follow / invariants

• Features must be available strictly before the routed outcome.
• Record feature values, policy version/hash, selected arm, and fallback reason.
• Keep a fixed-condition baseline and charge routing overhead to the routed arm.
• Fail closed or use a preregistered fallback for unknown message types.

Common tasks → first action

• Feature change: prove no future-label leakage and update serialization tests.
• Policy change: freeze a new version and compare against fixed-arm baselines.

Gotchas

• A router trained and evaluated on the same tasks is not a valid routing result.
• Model self-reports are not free features; count their tokens, latency, and risk.