qp-exa-dynamic

The first-party Exa search provider in DeepSeek Harness has a structural gap: its request body hardcodes highlightsPerUrl and never sends the Exa-Beta header, so Dynamic Highlights—a feature where Exa runs a single forward pass and allocates a shared context budget across all results—simply cannot be reached. Exa responds with HTTP 400. qp-exa-dynamic fixes both issues, ships with Dynamic Highlights on by default, and replaces the inert highlightsPerUrl knob with maxCharacters, which actually works when dynamic mode is off.

The plugin exposes all eight Exa retrieval types, from 464 ms keyword to 18 s deep-reasoning, and provides a composer-level /exa command to toggle highlights, switch retrieval type, and adjust result count at runtime—no restart needed. It also registers a second, model-facing tool called exa_search that owns its own per-call maxResults argument, letting the model request up to 50 sources in a single query without forking an agent preset to raise the dsh-tool-web ceiling that clamps web_search at 8.

It is built for developers already searching the web through DeepSeek Harness who find the default context quality too coarse, the source count too rigid at 8, or the retrieval type locked to a stale three-option subset. If you want to dial between keyword speed and deep reasoning mid-conversation, or let the model itself negotiate how many sources to pull per question, this plugin is exactly that control surface.