v1.26.0
v1.26.0 — B4 NLP Cost-Reduction + B5 Auto-Feedback + D3 Heartbeat Patrol
Three competitive-borrow landings (per docs/analysis/COMPETITIVE-REFERENCE.md): B4 (FastGraphRAG) + B5 (Cognee) + D3 (Letta).
Added
- B4
extract_noun_phrases(text)+saw_nlp_keywords(FastGraphRAG-inspired) — cheap NLP noun-phrase extraction (jieba CJK + regex Latin, no LLM/network) as a cost-reduction pre-index tier / OFFLINE-tier degrade path. - B5
saw_record_feedback(claim_uuid, helpful)(Cognee-inspired auto-feedback) — per-turn feedback bumps/lowers a claim's 4-level confidence (reusesclaims_repo.update_confidence); retrieval drifts toward claims the agent community found useful — no re-ingest. - D3
HeartbeatScheduler+saw_heartbeat_status(Letta-inspired heartbeat) — apscheduler background patrol runs Governor freshness + Detector unresolved-contradiction scans on an interval (SAW_HEARTBEAT_INTERVALenv), surfacing stale claims / open contradictions proactively.
Release Gate
pytest 2376 passed / 7 skipped / 0 failed (+13: B4×4 + B5×4 + D3×5) · coverage 68.28% · ruff 0 · vitest 64 pass · zero regressions.
Installation
pipx install smart-agent-wiki --force # or: pip install smart-agent-wiki==1.26.0
saw --versionDifferentiation
- FastGraphRAG uses NLP noun-phrases for cheap indexing — SAW lands it jieba-backed (CJK-ready) as a no-LLM degrade path for the OFFLINE tier, complementing (not replacing) the LLM claim extraction.
- Cognee auto-feedback tunes retrieval — SAW adjusts the 4-level confidence directly (so the B3 TRUE/FALSE/SUSPECTED status + Ed25519-receiptable provenance reflect the feedback), not a hidden retrieval score.
- Letta heartbeat runs the agent brain proactively — SAW runs Governor + Detector patrols (governance, not just memory), so stale/contradicted claims surface without a user trigger.
Deferred (gated)
AUDIT-F-05 activity persist (PRD rule 6) · AUDIT-F-06 banner SPEC drift · AUDIT-F-07 desktop signing / vLLM CI / Playwright (infra) · remaining competitive-borrow candidates → v1.27.0+.
Full changelog: CHANGELOG.md. PRD: docs/prd/PRD-nlp-feedback-heartbeat-v1.26.0.md.