qmd-autosearch

When a model runs grep or glob inside a knowledge-base directory, it typically catches only literal matches while semantically relevant documents slip through. Manually triggering a QMD semantic search mid-workflow breaks the chain and adds an unnecessary manual step. qmd-autosearch fills exactly that gap: it silently complements the keyword search with a semantic retrieval and asynchronously injects the hits into the next-step context, requiring zero extra instructions and zero external dependencies.

The trigger is deliberately conservative, activating only when the search path falls within a configured QMD collection directory and the call is top-level. The query term prefers the most recent user message as the semantic primary query, falling back to the grep pattern when the message is too short or clearly off-topic. Retrieval goes through QMDs lex + vec hybrid channel with server-side LLM re-ranking, then passes minScore filtering and limit truncation before being queued asynchronously into agent.inbox.nextStep, never blocking the current tool call. A tool, path, and pattern triple deduplicates per agent, and if the QMD service is unreachable or returns no hits, the plugin skips silently with zero side effects.

Ideal for teams already running DeepSeek Harness agent workflows with a local or intranet QMD MCP service and at least one built-out collection, especially when the knowledge base is large enough that keyword search alone feels incomplete. Once configured, no prompt changes are needed: every time the model finishes a grep, the next-step context carries a short list of semantic hits ready for a read call to expand the full document.