dsh-plugin-rag

Working across many sessions in DeepSeek Harness quickly exposes a gap: the model has no memory of the decisions you made, the code you wrote, or the context you built up last time, so you end up re-explaining everything each round. dsh-plugin-rag turns every conversation into a searchable semantic memory. As you chat, it incrementally indexes new messages and automatically un-indexes content shadowed by compaction or pruning, so retrieval always reflects the current surface of your sessions rather than a stale snapshot.

The design is deliberately lightweight. Vectors live in a single local JSON file with atomic tmp+rename writes and per-session cursors for idempotent restarts. There is no database, no native module, no extra service. The plugin is non-destructive by construction: it only subscribes to events the session store already publishes, never patches the agent loop, and uninstalling restores the harness to its exact original state.

Embeddings are model-agnostic—plug in BGE-M3, OpenAI text-embedding-3-small or large, Ollama nomic-embed-text, or any OpenAI-compatible endpoint. Once active, the model gains a first-class rag_search tool that embeds a query and returns the most relevant past messages with role, session title, and snippet. It is built for developers who iterate on long-running projects in Harness, reuse code and decisions across sessions, or need the agent to recall prior context without re-explaining. All data stays on your machine; the only outbound traffic is the embedding request you configure, and API keys are resolved through the harness credential store, never written into the index.