WeaverTools — Current-State Snapshot (2026-05-27)

Predecessor: current-state-2026-05-24.

Methodology — different from the prior two snapshots. 05-14 and 05-24 were static analysis only (“no tests run, no services hit”). This one is static analysis + live verification. Over this session the daemon was rebuilt/reinstalled and restarted, a 4-fresh-agent cohort was run end-to-end, per-agent ArangoDB databases were queried directly, and the seed-deletion bug was isolated by controlled create→load→inspect. Claims tagged [live] were observed on the running system; [code] claims are file-grounded; [proposed] are PRD/decision items not yet built. This snapshot exists to ground an imminent direction decision — read §9 first if that’s why you’re here.

0. TL;DR#

The clone-on-create memory-substrate apparatus now runs end-to-end: weaver experiment cohort run bakes/clones a vanilla belief graph into N identical fresh agents, runs a HeroBench task on each, and cohort diff reads per-member graph-formation readouts. A 4-member cartographer-style run completed 4 ok, 0 failed [live]. Getting there fixed a chain of substrate bugs (#413/#420/#421/#422) and surfaced two findings that drive the pending decision:

1. The per-agent encoder only loads if the daemon is built with --features embedder-rust. The documented install recipe (inference,viewer) omits it, producing an encoder-less daemon that forms no memory (no embed-on-write possible). The installed binary now includes it [live]; the install recipe is a real gap to fix [code].
2. The runtime autonomic writer writes events into belief_nodes (autonomic/writer.rs → BELIEF_NODES_COLLECTION), deviating from the architecture’s “beliefs mutate only at sleep” principle [live, code]. Runtime should produce events; sleep reconciles them into beliefs. This is the seam the next direction decision sits on.

Two new PRDs landed on main (proposed) framing a possible pivot in how event-triggers are recognized — measure the decoder’s natural phrasing distribution rather than force a grammar (event-trigger-phrasing-profile-PRD), and develop that measurement into a standing evaluation tool (phrasing-profile-eval-tool-PRD).

1. What changed since 2026-05-24#

Grouped by theme; all merged to main.

Working-memory event substrate (§6.1):

• #410 register engagement_edge + conversation_nodes collections.
• #411/#412 DB write API + write engagement edges (conversation_node → belief_node) at surfacing time.

Associative surfacing:

• #407 deterministic structural-expansion axis (assoc-surfacing §10); #408 resolve structural target from action-node markers; #409 specific craft-target + objective injection.

Memory-substrate images (clone-on-create):

• #405 weaver harness image bake verb; #406 clone-on-create. Catalog at /opt/weaver/images/, keyed by (domain, encoder, gnn, dim). Bake order: provision → materialize seed → embed → create HNSW index last → structural-embed → arangodump snapshot [code].

Vector-index correctness:

• #413 build the belief vector index at the encoder’s real dim (resolved from embedder.info()), not a hardcoded constant. The JINA_DIM = 2048 hardcode flagged in the 05-14 snapshot is gone from weaver-database [verified — grep empty].

Secrets:

• #415 agent db_password resolves from ~/.weaver_env (minted at provision), not committed yaml; KnowledgeState::resolved_db_password(name).

Cohort experiment tool:

• #414 experiment spec + graph_formation_readout (T3); #416/#417/#418 run/diff/list/show/teardown; #419 restore after a history rewrite. weaver experiment cohort …; registry at /opt/weaver/cohorts/.

Embed-before-insert chain (this session, the substrate-correctness arc):

• #420 defer per-agent prompt materialization to load and embed-before-insert (the non-sparse vector index rejects null-embedding inserts).
• #421 spawn bootstraps the system prompt from prompts_dir files on first load (sequencing fix) + cohort tolerates already-loaded member + diff reads member DB with member credentials.
• #422 use per-key upsert not import overwrite=true (which truncates the collection — the seed-killer).

Docs:

• event-trigger-phrasing-profile-PRD + phrasing-profile-eval-tool-PRD filed (proposed, on main).

2. Crate structure (8 crates)#

Unchanged from 05-24 except as noted. weaver-core, weaver-interface (binary weaver), weaver-database, weaver-spu, weaver-trace, weaver-analysis, weaver-demo, weaver-experimental.

weaver-experimental houses the swappable-hypothesis impls: the memory-peer GNN (memory_peer/ — mean-pool-v0, relational-pool-v0, both deterministic, untrained aggregators) and the sleep stages (sleep/ — FreudianSleepV1, presleep + Stage A/B/C/D, live_contexts). The canonical weaver-spu/src/structural/ home for the GNN (sprint task #219) is not built; the GNN lives in weaver-experimental as the v0 hypothesis [code].

3. The substrate facts established live this session (load-bearing)#

These are new, verified, and bear directly on the decision:

• The belief_nodes HNSW vector index is non-sparse and total [live]. It rejects any insert whose embedding is missing/null (ArangoError 10: Expecting type Array); sparse:true does not relax it; index creation also requires every doc to carry a valid embedding. Consequence: the “land with embedding: null, backfill later” escape (embed_write.rs, belief-nodes-embedding-Spec) is only valid before the index exists. Once present (clone-on-create, or post-bench-run index), embed-before-insert is mandatory. Documented in belief-nodes-embedding-Spec §4.1.3.
• Per-agent encoder load is gated on --features embedder-rust [live]. Without it load_per_agent_embedder compiles to a no-op → handle.embedder() == None → prompt materialization, semantic backfill, and structural-embed all skip → no runtime memory forms. inference = ["weaver-spu/cuda","weaver-spu/gguf"] does not include embedder-rust; scripts/install.sh’s die-message recipe is inference,viewer — i.e. the documented build produces an encoder-less daemon. Gap to fix (have inference imply embedder-rust, or correct the recipe + docs).
• insert_documents(overwrite=true) truncates the whole collection (ArangoDB import semantics), not per-key replace [live]. This was the seed-deletion bug (load wiped all 627 cloned world_graph beliefs, left 6 prompts). Fixed in #422 via upsert_documents (onDuplicate=replace). Post-fix verified: a fresh agent loads to 633 = 627 seed + 6 prompt [live].
• Sleep reconciles from episodic_memory (sleep/presleep.rs:126 FOR e IN episodic_memory), plus hanging_chad/meta_chad. It does not read conversation_nodes. And episodic_memory is not written at runtime today (a cohort member showed episodic_memory=0 after a run) — so the runtime→event-store→sleep pipeline is currently unfed [live, code].
• The autonomic writer writes to belief_nodes (autonomic/writer.rs uses memory_write(BELIEF_NODES_COLLECTION, …)). In a live run a member’s belief_nodes accumulated decision/observation note-shaped docs (27 + 5) while the working-memory path correctly wrote conversation_nodes (42) + engagement_edge (78) [live]. This is the architectural deviation §9 addresses.

4. Apparatus status — cohort + images (VERIFIED live)#

• weaver harness image bake --domain --seed --encoder --gnn produces a vanilla graph image (627 belief_nodes + edges + embeddings + HNSW index + structural embeddings), manifest at /opt/weaver/images/<key>/manifest.json [live].
• weaver experiment cohort run --agent-yaml --task --members N clones the resolved image into N fresh agents, runs the task on each serially, retains all [live, 4 ok].
• weaver experiment cohort diff <name> reads each member’s graph_formation_readout (seed vs formed partition) — requires the member’s own DB credentials (fixed #421); reads cleanly now (seed_node_count: 627) [live].
• Bake idempotency caveat: a failed bake leaves a partial indexed image DB that breaks subsequent re-bakes (the same null-insert-into-indexed-collection failure). Workaround applied this session: drop the stale image DB before re-bake. Not yet hardened into the bake verb [live].

5. GNN reality#

• Registered impls: mean-pool-v0, relational-pool-v0 (weaver-experimental/src/memory_peer/registry.rs) — deterministic, untrained aggregators [code].
• herobench-fullmem.yaml binds relational-pool-v0; bake + load run a structural-embed pass producing structural_embedding on belief_nodes [live — image had 627 structural-embedded].
• No trained GraphSAGE; #219 (the weaver-spu/src/structural/ scaffolding) and #225 (training) are pending. Engagement-contrastive training is density-gated (threshold ~50); the deterministic typed-edge walk + structural pool are the pre-GNN path and work now.

6. Schema reality#

• belief_nodes: 4-field embedding stamp (embedding/embedding_model/embedding_dim/embedding_task); encoder is qwen3-embedding-0.6b-gguf-q8, dim 1024 (the 2048 hardcode is resolved). HNSW index embedding_hnsw created at the encoder’s real dim — at bench-run (ensure_belief_vector_index, #413) and baked into images.
• Event collections exist: conversation_nodes (working-memory §6.1; the proposed today_ rename target), engagement_edge (conversation_node → belief_node), episodic_memory (sleep’s event source), hanging_chad/meta_chad (curiosity/contradiction).
• prompt-kind belief_nodes are legitimate constitutive memory (materialized at load); runtime observation/decision belief_nodes are not legitimate (the §9 issue).

7. Decision context — the two new PRDs#

• event-trigger-phrasing-profile-PRD (methodology, proposed, may still shift): measure the decoder’s natural event-signaling phrasing distribution, frequency-gate a recognizer, score per-(decoder, kind) coverage × collision against a closed-world oracle — before forcing a grammar or SFT. Borrows the templating premise from frequency-gated-recall-PRD.
• phrasing-profile-eval-tool-PRD: a standalone CLI that runs the methodology on arbitrary HuggingFace models (download → per-model adapter → bundled ground-truth task → per-(model,kind) report), publishing a HuggingFace dataset + writeup.

8. Numbers#

• Crates: 8. Agent YAMLs in agents/: 39. Task YAMLs in tasks/: 26. services/: weaver-daemon.service, weaver-infer.service, weaver.target (+ report-charts/). No weaver-embedder.service (in-process embedder).
• Installed binary /opt/weaver/bin/weaver built --features inference,viewer,embedder-rust this session (the embedder-rust addition is the live fix; not yet reflected in install.sh’s recipe).
• Decoder GGUFs on disk include a size ladder: qwen2.5-0.5b, qwen2.5-1.5b, gemma-3-1b (small) + gemma-4-26B-A4B, qwen3-coder-30B-A3B, gemma4-31b (large). Mid sweet-spot (Qwen3-4B/Gemma-3-4B) not local.

9. Open decisions (why this snapshot exists)#

The pending direction call centers on runtime events vs beliefs, and possibly a methodology pivot. The decisions, in dependency order:

1. today_ event container + stop writing beliefs at runtime. Rename conversation_nodes → today_ (one container for all between-sleep events), redirect the autonomic writer there with an event_type discriminator, keep engagement_edge: today_node → belief. Beliefs mutate only at sleep. (Task #303.) Prerequisite plumbing — likely lands regardless of the methodology pivot.
2. How event-triggers are recognized (the possible pivot): force a structured grammar (OBSERVE:) + SFT, or measure the natural phrasing distribution and frequency-gate a recognizer (the phrasing-profile PRD). Measurement-first is the stated lean.
3. Smaller decoders. Strong rationale (latency, faster iteration, more templated → easier recognition). Gated by the instruction-following floor; the cohort/image apparatus already supports decoder-size sweeps against a shared image (decoder is yaml, not in the image key).
4. Sleep cycle as first real test. Feed episodic_memory/today_ from runtime; a cartographer corpus (verified observations vs the map) becomes sleep’s first reconciliation test case.
5. Eval-tool build: standalone binary vs weaver experiment verb, and which ground-truth task to bundle.

Essential reading map (delta from 2026-05-24)#

• docs/prds/proposed/event-trigger-phrasing-profile-PRD.md, phrasing-profile-eval-tool-PRD.md — the pending-direction PRDs.
• docs/specs/in-progress/belief-nodes-embedding-Spec.md §4.1.3 — embed-before-insert under a non-sparse index (the substrate constraint).
• crates/weaver-core/src/embed_write.rs — materialize_embedded_belief_nodes (embed-before-insert, upsert), embed_unembedded_belief_nodes (load backfill).
• crates/weaver-core/src/autonomic/writer.rs — the runtime→belief_nodes write (the §9.1 issue).
• crates/weaver-interface/src/server.rs — load_per_agent_embedder (the embedder-rust gate), load-time prompt materialization + spawn bootstrap.
• crates/weaver-interface/src/cohort.rs — cohort run/diff; member_read_pool (member-cred DB read).
• crates/weaver-core/src/spec/image_bake.rs — bake order (embed → index last) + clone_image_into_db.
• crates/weaver-experimental/src/{memory_peer,sleep}/ — untrained GNN aggregators + Freudian sleep stages.
• crates/weaver-experimental/src/sleep/presleep.rs — sleep’s event source (episodic_memory).
• crates/weaver-database/src/graph/schema.rs — CONVERSATION_NODES_COLLECTION (the today_ rename target), engagement-edge endpoints.
• scripts/install.sh — the inference,viewer recipe gap (missing embedder-rust).