dsh-session-index

Multi-turn, multi-session agent workflows suffer from a persistent gap — the session/event stream does not replay history from resumed conversations, and the framework's built-in SQLite FTS5 (unicode61 tokenizer) treats a whole CJK string as one token, making Chinese recall nearly impossible. Agents therefore cannot remember what was said in a previous session, nor search back tool-call results using natural language.

dsh-session-index extracts user messages, assistant replies, tool calls/results, and session titles from every conversation into searchable documents and builds a cross-session inverted index exposed via two model-facing tools: session_search and session_index_stats. It supports real-time incremental indexing, watermark-based backfill of persisted sessions at startup, and hybrid retrieval that runs FTS5 in parallel with a bigram in-memory index (the engine field reports memory / fts5 / hybrid). Hits can be projected into per-session grouped, expandable cards in the Web UI. The index is appended as JSONL locally; restarts only patch the delta, and session disposal cleans up its documents synchronously.

Built for developers running multi-turn, multi-session DeepSeek Harness workflows who need agents to recall content from any past conversation. If you have hit the wall of poor Chinese FTS5 recall or lost search coverage in resumed sessions, this plugin fills both gaps.