v0.2.1

v0.2.1 — hotfix

Detail drill-down windows now extract assistant/message text blocks (block arrays) and filter by block type, so the exact original text of assistant replies appears in drill-down results. Found and verified live during a real recall drill-down (coarse → hit list → exact window).

v0.2.0

v0.2.0 — progressive disclosure + manual/automatic dual mode

• Two-stage recall: recall returns a light coarse recall by default (session titles + snippets, far fewer tokens); a new detail parameter drills into the original text — detail:{sessionId} lists a session's hits, detail:{sessionId, seq, before, after} opens the exact original-text window (official readEvent), detail:{sessionId, browse, fromSeq, pageSize} pages through the session (fuzzy index, official filterEvents fallback)
• Proactive recall: description rewritten so the agent recalls on its own when needed (after compaction, when details are missing) — no need for the user to ask; explicit user requests still work. Scope red line (current session by default) and invisible-presentation rules preserved
• Compaction anchor: on compaction/summary, one lightweight anchor (LLM-written summary + up to 3 key original fragments picked by pure-local heuristics) is injected into the compacted session's agent for the next 3 assemblies; exact text stays one detail drill-down away
• Verified: smoke suites green (coarse/detail/browse/anchor), real-data browse on the live index, and live verification on the running instance (coarse recall default, detail window, paged browsing)

v0.1.0

dsh-recall v0.1.0 — conversation history recall for DeepSeek Harness. AI never forgets what you told it.

What's inside

• Three-layer hybrid retrieval: official FTS5 literal / self-built trigram+bigram fuzzy / local bge-small-zh semantic (int8, bundled, fully offline)
• One-command install: ships a dsh.bundle.patch — dsh plugin --profile web add dsh-recall wires the plugin and the search index automatically
• Invisible by design: quiet "Recalling…" sweep → one-line "Recall complete"; results are woven into the agent's reply, never dumped
• Compaction-proof: the full original log stays searchable, including shadowed history
• Worker-thread inference: WASM embedding never blocks the host event loop
• Zero runtime npm dependencies: vendored onnxruntime-web + bundled tokenizer; the 23.9MB model ships as the optional dsh-recall-models dependency (--omit=optional for a lightweight build)
• Bilingual README + locale-aware UI (zh / en)

See PROJECT.md for the full engineering doc and roadmap (v2: browse mode / two-stage recall / time filters / compaction anchors; v3: topic clustering / memory distillation).