1.5.2
release-tag / release-image (push) Successful in 11m6s

This commit is contained in:
2026-08-27 09:34:33 +02:00
parent 07b5cc4fc4
commit 173afb9c54
13 changed files with 738 additions and 45 deletions
+11
View File
@@ -0,0 +1,11 @@
# Migration v1.5.1 → v1.5.2
v1.5.2 is a drop-in research-planning fix. Persistent volumes and existing secrets are retained.
1. Build/publish the v1.5.2 images through the normal Gitea pipeline.
2. Set `IMAGE_TAG=1.5.2`.
3. `docker compose --profile research pull`
4. `docker compose --profile research up -d --force-recreate neuroforge neuroforge-worker`
5. Re-run the FortiClient SSLVPN 7200 goal and inspect `/api/v1/goals/{id}/research/live`.
Expected behavior: `query.planned` starts with `FortiClient SSLVPN 7200`; if the configured category mix produces no relevant result, `search.fallback` is emitted and a `general` search is used. Relevant Fortinet hits should proceed to `download.started` rather than being rejected before fetch.
+18
View File
@@ -0,0 +1,18 @@
# GLPI NeuroForge Mega v1.5.2
v1.5.2 fixes a real autonomous-research failure observed with a narrow FortiClient SSLVPN 7200 goal. SearXNG itself returned relevant Fortinet sources for a focused query, while NeuroForge's deterministic planner sent the complete goal description and target as one oversized query. With a mixed `general,science,it` category configuration this allowed arXiv results to dominate the top-N; the strict relevance gate then correctly rejected every result, leaving the cycle with zero evidence.
## Fixes
- Deterministic goal research no longer sends `title + full description + target` to SearXNG when `autonomy.use_llm=false`.
- The first deterministic query is the compact goal title; an optional second query adds only a small number of non-generic description terms.
- Scheduler/process text remains excluded from all search queries.
- Goal-bound research retries once with SearXNG `category=general` when the configured category mix returns no goal-anchored result.
- The strict pre-fetch relevance gate remains fail-closed; it is not weakened to compensate for bad search ranking.
- Regression coverage reproduces the observed `general,science,it -> unrelated arXiv` result set and verifies that the general fallback returns and learns FortiClient 7200 evidence.
## Validation
All four Go modules pass `go test ./...`, `go vet ./...`, and `go build ./...`. Race checks pass for NeuroForge brain/store/http/research, Agent core/config/knowledge/learning/state, Knowledge staging/store/server, and Control.
The real Docker/SearXNG/Ollama host smoke remains a deployment-host gate.