prime-silo: Skill for Claude Code

LONGVIEW pipeline — operational knowledge

LONGVIEW map-reduces months of agent sessions (memo-ray store) into per-session cards via a local model, then graph + themes + report/book/PDF/audiobook. Pipeline: inventory → extract → map(walk) → (model or graph) → code → weave → enrich → review → reduce(TIMELINE) → opus → pdf. Code: scripts/longview/. Guide: runtime/docs/operations/LONGVIEW_GUIDE.md.

Governance & observability (v1.16.x) — all deterministic, ledger-sourced:

• Memory teleport (memory.mjs + memory_graph.py): label→resolve→teleport sensitive sessions (CV/job) to a quarantine workspace (moves files + Neo4j nodes + Chroma chunks, journalled, reversible via restore); quarantine.json filters teleported sids from inventory forever; lib/leak_gate.mjs gates deliverables. Graph Source names use the 8-char sid prefix, not the 32-char card filename. UI: :8788/memory.html.
• Dashboard (scratch/longview_run/dashboard/, bash dash.sh → :8788): Mission Control (/) + Lineage/governance (/lineage.html, OpenLineage DAG + artifact explorer + step-through reusing lib/record.mjs + execution register) + /kindle.html (ES5). Every number derives from ledger.jsonl + disk — no hidden state.
• Arc-driven opus (lib/arcs.mjs): timeline-walked cross-project arcs feed sections; draft → gate → critique → revise. LONGVIEW_OPUS_DIR=iterations/<name> builds without clobbering a prior book.
• Coverage gotcha: .meta.json also ends in .json — always filter(f=>f.endsWith('.json') && !f.endsWith('.meta.json')) or card counts / coverage denominators double.
• Ingest wedge fix: server-side deep_synthesis ingest kept flooding the LM host after client death — POST /api/rag/ingest/cancel + GET /api/rag/ingest/active + a 3-strike connectivity circuit breaker. subprocessEnv() forwards repo .env to spawned python (else LM endpoints unseen). reasoning_effort:"none" + empty-content→reasoning_content fallback (post-restart LM Studio diverts the whole reply into reasoning).

Model is env/profile-driven (not hardcoded): BENNY_DEFAULT_MODEL / LONGVIEW_MODEL pick it, the manifest carries only a preconfigured default. Verified LAN setup: LM Studio on 192.168.68.125:1234 serving lmstudio/google/gemma-4-12b (a reasoning model — the per-call token budget must clear its ~500-1000-tok preamble, e.g. LONGVIEW_FRAGMENT_MAX_TOKENS=1800, or content comes back empty → blank cards; full org/model id required; response_format must be json_schema|text, never json_object). A wedged gemma engine returns HTTP 200 but zero tokens — reload the model on the LM host to reset it; embeddings on the same host still answering isolates it as generation-specific. Don't confuse the two empty-content causes (bit us again 2026-07-27, twice): a small-max_tokens probe (e.g. 16) on this reasoning model returns content:"" + finish_reason:"length" because the preamble ate the whole budget — that is NOT a wedge. A real wedge = zero completion tokens / no bytes with a GENEROUS budget. Before declaring a wedge, re-probe with reasoning_effort:"none" OR max_tokens≥600 and check reasoning_content; only zero-token-under-ample-budget is the engine wedge. Also: the 12B loads slowly on the eGPU (~35s, ~22 tok/s) — a slow first token is not a hang.

Graph build — two paths, same schema (Source/Concept/RELATES_TO/SOURCED_FROM):

• model = deep_synthesis: a second LLM pass re-extracts triples per card section. ~60-120s/card.
• graph = longview_v2, deterministic: builds triples directly from the card's own concepts[]/applications[]/capabilities[]/skills_observed[] — no model call, ~0.4s/card. (:Project)-[INVOLVES/USES/DEMONSTRATES/APPLIES]-> entities + an anchor CO_OCCURS_WITH star per card. The map phase already distilled the entities once, so re-extraction is redundant — graph is the speed lever. Then run enrich to merge duplicate concepts across cards into shared hubs. Code: lib/card_triples.mjs (pure buildCardTriples), server route POST /rag/graph-upsert (→ save_knowledge_triples, no clustering), live earned-ETA via lib/eta.mjs → <ws>/longview/progress.json. Sentence-shaped fields (decisions/outcomes/failures) are deliberately NOT nodes — they stay in the card md for vectors. Blank cards (no entities) are counted + skipped, never written as empty hubs.