jflowers’ opencode-auto-question answers an autonomous coding agent’s multiple-choice questions using Jev or a language model. It can also consult project guidance through a tool-using subagent before calling either engine.
Route uncertainty and preserve human decisions
- Questions marked
HUMAN_GATE:and questions from configured human-gated agents are left for a person, before guidance or model confidence is considered. - If project guidance covers a question but the guidance agent is unsure, the plugin defers instead of letting a generic model override the policy.
- The first Jev pass uses the immediate conversation. A low-confidence result triggers a context-gathering subagent and a second decision.
- The project defaults to a confidence threshold of 0.6 on the first pass and 0.35 on a successful second pass. These are configuration choices, not measured guarantees. The second pass can also retry with immediate context alone.
- On failed or malformed decisions, the question remains available for a human. A visible session summary records the result or deferral.
The permission and quality boundaries
The bundled autonomy agent permits external-directory access; the README warns that unrestricted access reaches the whole filesystem and shows a project-scoped alternative. The plugin sends the question and selected context to the configured model provider. Jev’s typed answer and confidence do not establish that it reflects the user’s wishes; the lower second-pass threshold especially needs evaluation on the actual questions.
Verification: inspected the README, Jev adapter and orchestration code. The project and live model were not run; this was not a security audit.