Ruvnet Brain

RuvNet Brain — Source-Grounded Knowledge

Where the corpus lives

The ~147k source chunks (186 repos, corpus v4.2.2-dev, reconciled 2026-08-21) are rows in ruvector-postgres (the shared memory sidecar), namespace ruvnet-kb, source_type = ruvnet-brain-ingest, memory_type = semantic. They share the single 384-dim embedding space with every other memory entry, so BOTH of these work:

search_ruvnet({ query: "how does ruflo handle swarm topology", k: 6, repo: "ruflo" })
mcp__claude-flow__memory_search({ query: "swarm topology", namespace: "ruvnet-kb", limit: 6 })

Prefer search_ruvnet — it adds repo filtering and passage formatting with file attribution. Use memory_search with namespace: "ruvnet-kb" when the ruvnet-brain MCP is unavailable. ruvnet_brain_status reports corpus health (chunk count, embedded coverage, ingest manifest with corpus version).

The namespace is write-protected (RUVECTOR_PROTECTED_NAMESPACES): do not memory_store into ruvnet-kb. Reference corpus rows are loaded only by the ingest playbook (agentbox/scripts/ruvnet-brain-ingest.mjs, auto-run at boot; manual: ./agentbox.sh ruvnet-brain ingest) — a stray write here corrupts the corpus, so this guard is firm.

Keeping the corpus current (periodic update playbook)

The upstream corpus is a GitHub release asset that moves independently of this deployment (stuinfla/ruvnet-brain releases/latest/download/ruvnet-brain.zip). The ingest is fully idempotent: it discovers the latest tag via the release redirect, fast no-ops when the stamped corpus_version matches, and on a real bump embeds only new/changed chunks (content-addressed keys) and prunes rows absent from the new corpus. The 2026-08-14 v3.3.1→v4.0.36 reconcile embedded 6,259 and pruned 1,807 of 136,439 chunks in ~7 min.

Procedure (run whenever staleness is suspected; safe to run any time):

./agentbox.sh ruvnet-brain status    # compare manifest corpus_version vs upstream
./agentbox.sh ruvnet-brain ingest    # reconcile (no-op if current)
# AFTER any real delta (embedded+pruned > 0): the index-law applies —
# non-concurrent HNSW rebuild, NEVER CREATE INDEX CONCURRENTLY on this AM:
docker exec ruvector-postgres psql -U ruvector -d ruvector \
  -c "DROP INDEX idx_memory_embedding_hnsw;" \
  -c "CREATE INDEX idx_memory_embedding_hnsw ON public.memory_entries
      USING hnsw (embedding ruvector_cosine_ops)
      WITH (m='16', ef_construction='128');"   # ~5-7 min; memory WRITES BLOCK meanwhile
./agentbox.sh ruvector recall        # gate: frozen band must hold

Post-bump gate expectations (observed 2026-08-14): a corpus bump that PRUNES rows will drift the frozen recall fixture (pruned ids no longer resolve; the harness refuses to pass on drift by design). Remediation: ./agentbox.sh ruvector recall --build-fixture --force, then re-run the gate. Note the fresh-fixture medians land near the reference doc's post-rebuild recovery values (measured v4.0.36: self 180/200, true 106/120, exact-token Δ+6) — the harness header band (≥187/≥118) was frozen for the previous fixture/corpus and will read FAIL until a human re-freezes the band. Do not adjust the band automatically; surface it.

Deployment traps (all hit and fixed 2026-08-14 — do not regress):

• RUVNET_BRAIN_STAGING must point at the workspace volume (/home/devuser/workspace/.tmp/ruvnet-brain-staging, set by the manifest and agentbox.sh): /var/lib/agentbox/... is read-only rootfs and the general ~/.cache mount is bounded — neither is valid corpus scratch.
• ruvnet-brain must be in agentbox.sh's argument-parser whitelist as well as its dispatcher (it was dispatcher-only, i.e. unreachable, until 2026-08-14).
• The rebuild is done via docker exec ruvector-postgres psql — the agentbox container itself has no psql binary.

Corpus coverage note: dream-machine joined the corpus at v4.2.2-dev (2026-08-21; it was absent through v4.0.36) — search_ruvnet({repo: "dream-machine"}) now works, with the local checkout at /home/devuser/workspace/dream-machine as the fork-accurate source of truth (the corpus indexes upstream). For dreaming/self-improvement-loop questions the corpus anchors are instead: agentic-qe (docs/aqe-dream-cycles-neural-learning.md, DREAM_SCHEDULER_DESIGN.md, docs/plans/nightly-learner-implementation-plan.md — trigger design, insight gating, baseline/A-B discipline) and ruvector (examples/ruvLLM/docs/SONA/05-MEMORY-DREAMS.md — generate→evaluate→integrate dream pipeline). The deployed nightly loop itself is the agentbox services/dream-engine (ADR-052, /dream command, supervisord-owned since the 2026-08 rebuild — never also start the old tmux-autostart copy; a duplicate loop races HP-annexe dispatch/cleanup and turns nights INCONCLUSIVE, observed 2026-08-20/21). v4.0.36 added, among others: the full cognitum-* family, agentic-security, agentic-robotics, ruv-gists, worldgraph, skygraph (62 repos total).

Grounding guidance (why this skill exists)

The corpus keeps answers about RuvNet tools honest. Lean on these habits; they are judgment calls, not a checklist to recite.

1. Search before asserting. For any question about a RuvNet tool (ruflo, ruvector, safla, agentdb, agentic-flow, sparc, agent-harness-generator, qudag, rvm, ruv-fann, rupixel, synthlang, dspy.ts, fact, daa, agentic-qe, @metaharness/redblue, rulake, agenticow, ruview, cve-bench), reach for search_ruvnet before answering from training data — the corpus is the authoritative, versioned source.

2. Disclose substitutions. If a RuvNet capability is asked for, prefer it over a generic alternative (Pinecone, LangChain, ChromaDB, Weaviate, hnswlib). Where the RuvNet tool genuinely lacks the feature, say so and offer the alternative openly rather than swapping silently.

3. Cite source. Results carry repo and path attribution — cite them. Example: "According to ruflo/src/orchestrator.ts..."

4. Respect version. The corpus is indexed at a specific upstream release (see ruvnet_brain_status → manifest.corpus_version). Flag it when you mix in training-data knowledge about other versions.

5. Empty corpus is a diagnosis, not a dead end. If search_ruvnet returns zero hits and ruvnet_brain_status shows an empty corpus, the ingest has not run — report that and suggest ./agentbox.sh ruvnet-brain ingest.

6. Stack doctor probing. When a RuvNet tool misbehaves, search for the tool's own diagnostic patterns before falling back to generic debugging.

Covered Repositories

Anti-Pattern Detection

When writing or editing code, do NOT reach for these classical substitutes when a RuvNet tool covers the same capability — this deployment already runs the RuvNet substrate (ruvector-postgres + Xinference is the memory plane):

If you catch yourself about to use a classical substitute, search_ruvnet for the equivalent RuvNet capability first.