OpenAgents Coder
A terminal coding agent uses Jev for typed turn, risk and progress judgments before its generative planning and tool loop.
A terminal coding agent uses Jev for typed turn, risk and progress judgments before its generative planning and tool loop.
Before generating a turn response, the agent sends a bounded task and conversation state to Jev with a typed `action` Choice. The answer routes the turn to respond, clarify, or end; a missing or unrecognized answer halts the route. A separate shell-round judgment can pass, retry, or stop the command loop.
The acceptance test in `test` was written before the task in `task` was solved, to check the requirements in `requirements`. Does every assertion in the test follow from what the task states, from its examples, or from its rules, with nothing the task doesn’t ask for, and does every input the test builds follow the input format the task describes?
A single acceptance test with the task instruction, the requirements it is meant to check, and the test ID, kind, purpose, and clipped source.
The agent combines the current task, bounded conversation transcript, and repository member names into the classification state.
It sends that state with the `action` Choice question to the configured Jev client.
The returned choice maps to respond, clarify, or end; unanswered or unlisted actions halt, while a failed or absent classifier leaves generation unrouted.
For a response or clarification route, the generation door produces the next turn; shell outcomes can receive a separate pass, retry, or stop judgment.
Jev sits as a first-stage router in the agent loop, and its typed answer is mapped to explicit application routes. The current question set keeps the turn decision focused on one action rather than treating several unused judgments as routing signals.