Module monitor 

speccoder.task.monitor: a reasoning-blind gaming monitor.

A judge-side call after a task passed the verifier. It sees the task, the diff and the tool calls — never the implementer’s prose — and can only add a suspicion, which stops the tick for a person. It can never turn a failure into a pass, and the implementer never sees its findings: a model optimised against a monitor learns to hide (arXiv:2503.11926).

Structs

Monitor

The coded skill.

Constants

CAPABILITY

The capability the skill provides.

MAX_PATCH

The diff the monitor reads, at most.

Functions

input

The run input: the task’s spec context, the diff (capped), the tool calls (tool, target, decision) and the frozen test files.

policy

What the monitor is told.

schema

The answer’s schema.