NewOh My Claude Code: Jev Judgments
A Claude Code orchestration plugin compares typed Jev judgments with existing heuristics through opt-in shadow calls.
Added to Jevfast
NewHow it uses Jev
At the model-routing judgment point, OMC records a Jev Choice under the key model-tier, asking whether a delegated task should use haiku for quick lookups or lightweight mechanical work, sonnet for standard coding and orchestration, or opus for complex architecture and deep analysis. The recorder constructs a request state with tool_name, subagent_type, and task, populated from the called tool name and the pinned enforcement result's original subagent type and prompt. In shadow mode the pinned heuristic result is returned immediately; the Jev comparison runs asynchronously and is logged. The point is non-blocking. Calls are off unless TYPESAFE_API_KEY is present and OMC_JEV opts the point in; OMC_JEV=off disables calls. This inspected model-routing example records comparisons while the heuristic controls routing.
What Jev decides
Choose a model tier for a delegated taskExample answers · not a recorded Jev response · Source ↗
Question 1 · model-tier
YOUR APP
INSTRUCTION
Choose the model tier for this delegated task.
STATE
One delegated Task or Agent invocation with tool_name, subagent_type, and task (the original prompt); the pinned enforcement result supplies the subagent type and task.
JEV · CHOICE
- haiku
- sonnet
- opus
App workflow
- Resolve the existing tier
The delegation enforcer computes a pinned tier and passes the tool name, original subagent type and task prompt to the model-routing recorder.
- Ask the bounded tier question
With the API key and point enabled, the recorder sends tool_name, subagent_type and task with the model-tier Choice to Jev. Excerpts are bounded before the request.
- Keep the heuristic in shadow mode
The non-blocking shadow path returns the pinned result immediately and logs the Jev comparison asynchronously. Disabled points make no request; failed calls degrade to the heuristic.
Why it is interesting
It adds observable model-tier comparisons to an existing delegation heuristic, with a non-blocking shadow path that can collect evidence before configured active routing is used.