|
|
69c971708a
|
feat: Day 4+5 — R7/R9 fixes + integration tests (R8)
Day 4:
- R7 fixed: overlap-density ranking. p06-firmware-interface now
passes (was the last memory-ranking failure). Harness 16/18→17/18.
- R9 fixed: LLM extractor checks project registry before trusting
model-supplied project. Hallucinated projects fall back to
interaction's known scope. Registry lookup via
load_project_registry(), matched by project_id. Host-side script
mirrors this via GET /projects at startup.
Day 5:
- R8 addressed: 5 integration tests in test_extraction_pipeline.py
covering the full LLM extract → persist as candidate → promote/
reject flow, project fallback, failure handling, and dedup
behavior. Uses mocked subprocess to avoid real claude -p calls.
Harness: 17/18 (only p06-tailscale remains — chunk bleed from
source content, not a memory/ranking issue).
Tests: 280 → 286 (+6).
Batch complete. Before/after for this batch:
R1: fixed (extraction pipeline operational on Dalidou)
R5: fixed (batch endpoint + host-side script)
R7: fixed (overlap-density ranking)
R9: fixed (project trust-preservation via registry check)
R8: addressed (5 integration tests)
Harness: 16/18 → 17/18
Active memories: 36 → 41
Nightly pipeline: backup → cleanup → rsync → extract → auto-triage
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
|
2026-04-12 14:44:02 -04:00 |
|
