dsh-advisor-plugin

The same model playing both coder and verifier easily falls into self-validation blind spots, gradual drift, and premature completion in long-running tasks. dsh-advisor-plugin addresses this structural problem with a simple strategy: a cheap, fast execution model keeps moving along the main path, while at critical checkpoints the full session is handed to a stronger reviewer that returns a plan, a correction, or a stop signal before execution resumes.

Two complementary channels are provided. The explicit channel lets the execution model call advisor() with zero parameters; the entire session and tool list are forwarded to the reviewer automatically, so the model never has to assemble context by hand. The patrol channel samples the session on a step-and-wall-clock schedule, injecting a correction into the next request when drift is detected and demanding the execution model halt and report to the user when a STOP fires. The reviewer only outputs judgments, never touches code, never mutates state, and every failure, timeout, or context overflow is isolated so the executing turn never breaks.

Built for developers running lightweight execution models through DeepSeek Harness who want an on-demand expert review at critical moments without paying for expensive inference on every step. Setup is a simple pick-a-model exercise on the DSH settings page; with no reviewer configured the plugin registers nothing and adds zero prompt overhead.