The Graphnosis Skill Subgraph is a procedural layer in which a trained Skill — a standard operating procedure (SOP) with eight goal categories (Trigger, Prerequisites, Requires, Produces, Success, Out of scope, On failure, On completion) and a numbered step body — is compiled into a typed-edge subgraph stored in the same owner-held, encrypted, indelible knowledge multi-graph that holds the agent's declarative memory (the substrate of the companion paper, The Un-Brain). The procedure is the subgraph: there is no separate skill data type. Five evidence-tagged edge classes (skill: seq, skill: loop, skill: branch, skill: ctx, skill: calls) carry control flow as graph structure; each step's @needs annotation routes it to the cheapest model whose capabilities cover it, with sensitive-engram steps hard-locked to local inference; sub-skill calls compose procedures with argument binding and return capture; and a vitality model drives a governed retrain loop instead of letting procedures silently decay. The measured headline property is borrowability in the intra-cortex sense: because sibling agents share one owner-held cortex, an agent borrows a sibling's single skill through an in-cortex @skill: call — resolved at training time through an encrypted side-table at no model-token cost — instead of spawning the whole sibling agent. We measure this borrow-vs-spawn separation at ≈39–45× on this harness (two datapoints): a live single-run cortex record at ≈475 vs ≈21. 3k tokens (≈45×, the headline datapoint), corroborated by a reproducible-record sandbox run of five cold spawns at ≈25. 0k against a deterministic 637-token borrow (39. 3×) ; against the ≈700-token modeled typical borrow the separation is ~30×. The absolute figure is harness-specific and the architectural claim is the constant sub-agent warm-up a borrow avoids. A distinct cross-owner property — portability, exporting a skill as a signed. gsk pack (AES-256-GCM payload, Ed25519 signature) designed to bind a recipient's own engrams at walk time — the Ed25519 signature provides integrity, while standalone pack confidentiality requires an optional passphrase or recipient-key encryption — follows by construction from the empty-engram train invariant but is not evaluated here; an independent import-and-walk is future work. The contribution is a synthesis, not any single first: each ingredient (SOPs, prompt-skill libraries, workflow DAGs, model-cascade routing, retrieval-augmented and procedural-memory stores) exists elsewhere, but a source-audited comparison (LangGraph, GraphRAG, LightRAG/LazyGraphRAG, model routing/cascades, Mem0, Zep/Graphiti, Letta/MemGPT — the closest system — and Anthropic Agent Skills; author-rated positioning, not a benchmark) finds none holds the conjunction — in particular co-location of an executable procedure with declarative memory together with vitality-driven refresh — in one local-first indelible store. Five by-construction properties are stated and, where they hold, formalized: deterministic plan compilation (golden-test-pinned) ; a bounded-loop plan contract for capped loops composed with a depth-3 recursion bound, now also enforced at runtime by the shipped reference walker — loop bodies re-execute with the cap enforced in code (authored max=M, else a walker default of three for uncapped loops), with parallel dispatch remaining contract-only; structure-preserving training (a token-extraction validator reverts any rewrite that drops a structural token — a set-presence guarantee) ; cross-skill-edge stability under retrain; and the empty-engram train invariant (no automatically-recalled personal node content enters a skill body at training time; hand-authored facts and engram names remain the owner's to review). These are guarantees enforced by construction, not deep results. As a case study, an AI coding agent (Claude Code) read the Graphnosis repository and the founder's plans and sessions, identified the need for Skills, and trained a 76-skill corpus spanning code, business/go-to-market, and meta families — later extended to 108 skills across twelve families as a representational domain-compatibility probe (not measured cross-user transfer) — that then helped build Graphnosis. The corpus-wide routing accounting (E2) reproduces from route-b-artifacts/ (the frozen fixture plus the pinned model catalog vendored in lib/). That reproducible routing result is a corpus-wide cost accounting over 105 of the 108 skills (809 steps): ~99. 6% saved under an adaptive local+cloud policy (a cost-accounting figure, not a quality-preserving one: the ~99. 6% is realized by routing the judgment steps — reasoning ~51%, writing ~11% of the corpus — to free local models, which is exactly where the E6 pilot finds local quality collapses; so realizing this saving without quality loss is not established, and keeping those judgment steps on cloud would lower the saving toward ~96. 6%. E6 is a directional K=10 pilot scored by a single LLM judge), ~96. 6% cloud-only under the cheapest-qualifying rule, and 73. 3% under a premium coverage-max policy (that 73. 3% is the complement, 1 − 0. 2667, of the Haiku/Sonnet price ratio — a property of the pinned catalog, Haiku 4. 5 pinned at 0. 80/4. 00 and Sonnet 4. 6 at 3/15 → 0. 2667; at current retail, Haiku 4. 5 1/5, the same identity yields 66. 7%, so the identity structure — saving = 1 − price-ratio — is the claim, not the constant). A single annotated nine-step skill saved 99. 5% at measurement time; under the current shipped catalog (which now serves the writing capability locally) that skill routes fully local, so the 99. 5% is retained as a measurement-time result. Lazy dispatch loads roughly 8% of the library per session (48, 780 tokens to inline all 76 bodies versus ≈3, 600–4, 100 for the trigger index plus the one or two matched skills, a 91. 5–92. 6% reduction). A governed retrain measurably restores freshness (per-skill vitality 92→98 and 94→99). A curated engineering subset is published as a signed. gsk pack (the trained-skills repository) ; the full 76/108-skill corpus measured here is not published. The deterministic substrate, the routing savings, and every read-and-walk operation are license-free; only the in-app training compile and signed-pack export are gated.
Nelu Lazar (Sun,) studied this question.