dsh-session-recall

You have been chatting with your AI agent in DeepSeek Harness for months, but it has no memory of the bug you fixed last week or the font you settled on for a resume. The official session-query backend ships without a model-facing tool, its full-text index is disabled by default, and when enabled it lives in memory, vanishing on every restart. dsh-session-recall fills that gap: a recall tool the model can call directly to full-text search its own past session transcripts, working retroactively from day one with zero pre-extraction.

Unlike memory plugins that first extract structured notes with an LLM (lossy, token-costly), recall searches the original transcript text. The SQLite FTS5 index is built lazily on first search and persisted to disk, with incremental updates thereafter; warm queries land in 0.1 to 1.5 milliseconds on a 31-session corpus of over 8,000 indexed events. When the unicode61 tokenizer fails to match a short CJK phrase, the tool transparently falls back to an exact substring scan so Chinese queries still hit results. Search scope is cwd-scoped by default; cross-project access requires an explicit opt-in flagged by the deployment.

This plugin is for developers who run long-lived sessions in DeepSeek Harness and frequently need the agent to surface earlier context. Whether you are hunting down a configuration value the AI suggested three weeks ago or revisiting a design decision buried in a long chat, recall lets the model find the original text on its own instead of relying on you to scroll through history.