dsh-advisor

Multi-step reasoning agents occasionally drift during long tasks, and the ideal way to course-correct is an independent perspective examining the current progress. Automatic per-turn reviewers are costly and interrupt the primary model's rhythm; forgoing review entirely risks missing the chance to correct at critical junctures. dsh-advisor takes a more restrained approach: it returns the review decision to the primary model itself, which calls a zero-parameter advisor() only when it decides a second brain is needed in the moment.

The core mechanism is deliberately simple. Once the primary model invokes advisor(), the plugin packages the agent's full derived session history together with its tool inventory and forwards them to a separately configured reviewer model for a single tool-less side call. The free-text response—a plan, a correction, or a stop signal—arrives as a normal tool result within the same turn. The primary model reads the guidance and continues; the reviewer model holds no tools and never writes into the session transcript, keeping intrusion minimal. Configuration composes across three surfaces: the plugin-row config provides base values, the dsh web Plugin Configuration page offers a visual toggle with provider and model selects, and the /advisor command flips only a per-session override without touching persisted settings. A hard gate ensures that enabling the advisor without both a provider and a model never triggers a model call.

This plugin suits developers running multi-step reasoning agents in dsh who want an on-demand external review perspective without committing to a resident observer, a per-turn probe, or a complex routing runtime. Zero-parameter invocation, per-agent-scope registration, and error paths that return explanatory text instead of throwing exceptions are all designed to keep the plugin's footprint on the primary model as small as possible.