Technical Reports on Todd W. Bucy — Research Blog

WeaverTools — Current-State Snapshot (2026-05-14)

Thu, 14 May 2026 00:00:00 +0000

Methodology: Static analysis only — no tests run, no services hit. Every claim cites a file path. The intent is to ground the accompanying weavertools-PRD.md (delivered in the same PR as this snapshot) in code reality, not in PRD-claim reality.

0. TL;DR#

WeaverTools is a Rust-native harness that runs local language models as security-isolated agents (one OS user per agent) backed by per-agent ArangoDB memory graphs and a shared GPU-resident inference server. At HEAD the system can provision OS users + databases, spawn agents driven by agents/*.yaml, and dispatch HeroBench benchmark tasks via weaver agent start <task.yaml>. The nap (context-pressure memory offload) path fires on KV budget exhaustion and writes summaries to belief_nodes with embed-on-write stamping (Pen + engine-nap both wired). The in-process Rust candle encoder (EmbedderClient) is the sole embedder backend; the Python gRPC service is retired. OpenInference JSONL traces are emitted per run with llm.sampling, llm.weights_blake3, encoder.weights_blake3, and agent.prompt_stack attributes. Stubs: kind: chat returns an error, kind: chatroom and kind: training-loop both hit UnimplementedHandler. The HNSW vector index on belief_nodes.embedding is defined in code but creation at collection-bootstrap time has not been verified by a passing integration test; the preseed materializer (batch embed-on-write at agent-load) is wired and tested. The decoder-side candle stack in weaver-spu is present but the live inference path in production still uses the llama.cpp GGUF backend; the pure-candle decoder is Phase 2 work.

WeaverTools — Current-State Snapshot (2026-05-24)

Sun, 24 May 2026 00:00:00 +0000

Predecessor snapshot: current-state-2026-05-14.

Methodology: Static analysis + targeted grep/file reads; two code-grounded subagent passes (apparatus + attribution instrument). No tests run, no services hit, no sweep executed. Every claim cites a file path or PR. This snapshot exists to refresh the weavertools-PRD.md anchor, whose “verified commitments” table is now significantly out of date, and to answer one question: how far are we from running proper experiments?

Scope caveat — the memory/GNN cluster is a moving target. A parallel session is actively pushing on the memory system; the inductive GNN build-out (GraphSAGE-class graphsage-v1, replacing the mean-pool-v0 scaffold) begins ~2026-05-25. Status claims about surfacing / GNN / working-memory / engagement-edge code are bracketed as in-flux throughout and should not be treated as stable.

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

Wed, 27 May 2026 00:00:00 +0000

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.