dsh-verify-judge

Autonomous goal loops let an AI agent run many rounds on its own, but the moment it says "I'm done" has always been pure self-declaration — a model can stamp complete while tests are still red. dsh-verify-judge closes that trust gap: it turns marking a goal complete from a chat statement into a gated action with an objective pass/fail gate.

The mechanism is straightforward. When the agent calls update_goal(complete), the plugin first runs the project's test suite (auto-detected or explicitly configured). Red tests mean the completion is denied, and the failing command output is fed back to the agent as instructions to keep fixing. Version 0.2.0 adds a turn-end gate: even if the agent tries to skip the formal stamp and simply declare done in prose, the plugin re-verifies at the turn boundary and sends it back to work. When everything genuinely passes, the agent never notices the plugin exists; when it tries to cut corners, it gets bounced back — twice by default, then stopped.

It is built for developers running DeepSeek Harness in goal mode who let an agent autonomously complete multi-step tasks. If your agent occasionally turns in half-finished work early, this plugin adds an objective completion standard with minimum intrusion — no prompt changes, no workflow restructuring. Enable the row and the gate goes live; disable it and behavior returns to stock.