dsh-search-first

Large language models fail most often by confidently hallucinating an answer before consulting any source. dsh-search-first plugs that gap with a hard gate in the agent pre-step hook: when Search First is toggled on, the loop is intercepted before the LLM composes its reply. The pipeline runs query extraction, multi-source web search, page fetch and organization, cache and run persistence, and finally injects the structured live leads back into context so the model can ground its final answer.

The plugin speaks to three search backends—AnySearch, TinyFish, and Tavily—and can schedule them in sequential failover or parallel fan-out with merge. The extraction sub-model is independently configurable, keeping token spend separate from the primary session model. Results are ranked across sources, selectively fetched, and fed back with standardized citation formatting so every factual claim in the final answer is traceable. Cache TTL, maximum injection length, and failure policy (continue or block) are all tunable in the settings panel, and a per-session toggle lets you flip the behavior without touching the global default.

If you rely on dsh for research workflows, knowledge-base customer support, or long-form writing assistance and want the model to do its homework before it speaks, this plugin is a lightweight but enforceable guardrail between your prompt and the answer.