GBrain: Adopt, Port, or Build — the Skill-Registry Verification Path for RDCO's Harness
The question
Should RDCO adopt Garry Tan's GBrain directly, port concepts to our existing skill registry, or build our own equivalent from scratch — what's the maintenance-burden + dependency-risk tradeoff across the three options, and which of the 8 missing checklist steps (unit tests, integration tests, LLM evals, resolver eval, check-resolvable, DRY audit, E2E smoke, brain filing rules) does each option ship fastest? Context: the skillify-it brief flags this as a "decision worth making explicitly"; RDCO sits at steps 1-2 of Tan's 10-step gating, and the path choice sets the next 4-8 weeks of L5-unhobbling harness work.
What we already know (from the vault)
- The [[2026-04-22-garry-tan-skillify-it-workflow]] brief is the origin of this exact fork: it lays out the 10-step gating checklist, scores RDCO as ✅ on steps 1-2 and ❌/⚠️ on steps 3-10, and explicitly frames "(a) adopt directly, (b) port concepts, (c) build our own" — noting adoption is "meaningfully faster but creates a dependency on Tan's project trajectory."
- [[2026-04-11-garry-tan-thin-harness-fat-skills]] establishes that RDCO already is the thin-harness/fat-skills architecture by name (skills dir, resolvers via CLAUDE.md/SOUL.md, diarization = the vault, latent-vs-deterministic gating). The missing piece Tan has that we don't is the verification substrate, not the architecture.
- RDCO already owns a skill-manufacturing line: the skill-dev-pipeline brigade ([[2026-06-30-anthropic-skill-guide-vs-brigade-comparison]]) ships blind test authoring, static critics, an execution-eval A/B station (lift ± stddev), and a kill/86 decision — i.e. we have build/port muscle plus
rdco-doctorand a/skillifyskill that already scaffoldstests/. Per MEMORY the registry topology is a privatebrigade-housedev center + generatedray-pluginspublic surface (distribute.py + leak gate + manifest allowlist). - [[2026-05-12-garry-tan-ai-agent-complexity-ratchet-90-test-coverage]] is the "why bother" case: tests+docs+evals loaded each turn create a forward-only quality ratchet — but it also reveals GStack/GBrain move at 14 PRs / 72h with breaking-change velocity, which is the dependency-risk tell.
What the web says
- GBrain is real, open-source (MIT), and has drifted: the current repo (v0.41.x, up from the v0.30 the skillify brief saw) is primarily a self-wiring memory/knowledge-graph system (PGLite/Postgres + pgvector, 146K+ pages, LongMemEval benchmark), not chiefly the skill-quality checklist engine the April brief described. [github.com/garrytan/gbrain]
- The repo markets itself as standalone and portable — MCP-compatible with Claude Code/Cursor/Codex, "install in 30 minutes." Its CLI ships
gbrain doctor(health),gbrain eval(LongMemEval),gbrain schema lint, and per searchcheck-resolvable(reachability / MECE / DRY on the skill tree) plusskillpack-check --quiet(CI exit codes). [github.com/garrytan/gbrain] - An independent review flatly contradicts the "portable" framing for our case: skill-packs are first-class only for OpenClaw / Hermes; "no first-party packages for Claude Code," so a Claude Code shop faces a "weeks-long custom integration" (deploy the MCP server, wire tool-calling, run Postgres, re-author skills to Hermes's execution model). It warns the Minions job-queue + 34-skill scaffold is "over-engineered for most teams" and flags "frequent breaking changes." [vectorize.io/articles/gbrain-review]
- A fourth path has appeared since the April brief: Anthropic's native skill-creator now ships LLM evals, blind A/B comparators, trigger/resolver tuning (false-pos/neg reduction), regression tests, and benchmarking — free, on the Claude Code runtime RDCO actually uses. It does not cover unit/integration/E2E/DRY/filing. [claude.com/blog/improving-skill-creator]
- GStack (Tan's Claude Code setup, ~90-108K stars) is separate from GBrain (the brain/quality layer, ~14K stars); the two are frequently conflated but solve different problems. [github.com/garrytan/gstack]
Convergences and contradictions
- Converge: vault (skillify brief's "dependency on Tan's trajectory" caveat) and web (vectorize's "severe narrowing to Tan's ecosystem," breaking-change velocity) agree direct adoption is a real dependency-risk bet, and that RDCO already has its own doctor + brigade +
/skillifybuild muscle. - Contradict: GBrain's repo sells itself as portable/standalone/Claude-Code-compatible, but the independent review says the skill-quality layer (the exact part mapping to check-resolvable/DRY/resolver-eval) is OpenClaw/Hermes-first and Claude Code is not a supported target — the portability holds for the memory layer, not the checklist layer we want.
- New signal: GBrain has scope-drifted from "skill checklist engine" (April framing) to "memory system" (July reality), and Anthropic's native skill-creator now absorbs the eval-family steps for free on our runtime — both cut against paying an integration tax to adopt an external, moving dependency.
Synthesis for RDCO
Recommendation: PORT (hybrid), don't adopt, don't fully build. The maintenance-burden × dependency-risk matrix is decisive. Adopt directly is the worst fit despite being "meaningfully faster" on paper: GBrain's checklist layer is coupled to OpenClaw/Hermes + a Postgres/Bun substrate RDCO doesn't run, so the "30-minute install" becomes weeks of MCP wiring, and you inherit a fast-moving, breaking-change dependency whose scope has already drifted away from skill-quality toward memory. High ongoing maintenance (track upstream breaks, run a DB) + high dependency risk (external roadmap + runtime mismatch) for a payoff most of which we can get elsewhere free. Build from scratch is the opposite failure: lowest dependency risk but highest upfront cost, and it reinvents what Anthropic's native skill-creator and RDCO's own brigade already give us — a classic not-invented-here tax on the L5 roadmap.
Port threads the needle: take GBrain's transferable concepts — check-resolvable's reachability/MECE test, the DRY audit, resolver-eval discipline (both the deterministic table-presence check and the LLM routing check) — and re-implement them on RDCO's existing Claude Code substrate (rdco-doctor + the skill-dev brigade + the graph.duckdb reingest). For the eval-family steps, wire the native skill-creator rather than either building or importing them. Net maintenance burden: medium (small surface, our runtime, no DB); dependency risk: low-medium (Anthropic-native path + our own thin code, no Hermes coupling).
Which of the 8 steps each option ships fastest:
- Adopt GBrain ships check-resolvable + DRY audit fastest in raw capability (they are gbrain-native skill-tree commands today) and brain filing rules (GBrain is a self-filing brain). But for RDCO's Claude Code runtime this is "fastest in theory, slowest in practice" — the OpenClaw/Hermes wiring gate blocks delivery of any of them for weeks.
- Port (recommended) ships LLM evals + resolver eval + integration/E2E fastest — native skill-creator delivers the eval/trigger/regression trio essentially immediately on our runtime, and the brigade's execution-eval + blind test-author cover integration/E2E-analog A/B. Check-resolvable + DRY audit follow fast as thin re-implementations against our ~40 skills.
- Build ships brain filing rules + DRY audit fastest of its own steps, because RDCO's vault folder conventions and the typed graph.duckdb already exist — enforcing them programmatically fits our idiosyncratic layout better than any external tool. It is slowest on the eval-heavy steps (unit/LLM/resolver), which is exactly where adopt-native beats build.
Sequencing for the 4-8 week window: (1) turn on native skill-creator evals across the current ~22-40 skills (steps 3/5/7 — LLM evals, integration, resolver eval), (2) extend rdco-doctor with a check-resolvable + DRY axis ported from gbrain's semantics (step 8), (3) add unit + E2E smoke via the brigade's test-author station (steps 3/9), (4) codify brain filing rules against vault conventions + graph reingest (step 10). GBrain becomes a reference to read, not a dependency to run.
Why this is in the vault
This brief resolves the explicit adopt/port/build fork the skillify-it brief left open, setting RDCO's L5-unhobbling harness roadmap for the next 4-8 weeks: it commits the skill-registry to a port-plus-native-eval path on the existing brigade-house/rdco-doctor substrate rather than importing an OpenClaw/Hermes-coupled external dependency.
Open follow-ups
- What is the concrete
rdco-doctorextension spec for a ported check-resolvable + DRY axis over the current skill set, and does the brigade's critic station already cover part of it? - Does native skill-creator's blind A/B eval satisfy step 5 (LLM evals) well enough to skip a custom LLM-judge, and what is its per-run cost/latency on our cron cadence?
- Run gbrain's
check-resolvablesemantics manually against RDCO's ~40 skills once — how many "dark"/unreachable skills and overlapping triggers do we actually have (Tan found 15% dark)? - Should brain filing rules be enforced via the existing graph-reingest pipeline (graph.duckdb) rather than a new subsystem, and what's the enforcement hook?
- Does the "SKILLIFY IT" verb + "fucking shit"/"wtf" transcript eval-mining belong wired into the
/improveskill as the resolver-eval test-case generator?
Related
- [[2026-04-22-garry-tan-skillify-it-workflow]] — the origin brief that names the adopt/port/build fork and the 10-step checklist
- [[2026-04-11-garry-tan-thin-harness-fat-skills]] — the architecture RDCO already embodies; the verification substrate is the gap
- [[2026-05-12-garry-tan-ai-agent-complexity-ratchet-90-test-coverage]] — the ratchet case for why the 8 steps matter, plus GBrain's breaking-change velocity signal
- [[2026-06-30-anthropic-skill-guide-vs-brigade-comparison]] — RDCO's existing skill-manufacturing brigade (execution-eval, kill/86), the build/port muscle
Sources
- Vault: rdco-vault/06-reference/2026-04-22-garry-tan-skillify-it-workflow.md
- Vault: rdco-vault/06-reference/2026-04-11-garry-tan-thin-harness-fat-skills.md
- Vault: rdco-vault/06-reference/2026-05-12-garry-tan-ai-agent-complexity-ratchet-90-test-coverage.md
- Vault: rdco-vault/08-tooling/2026-06-30-anthropic-skill-guide-vs-brigade-comparison.md
- Web: https://github.com/garrytan/gbrain
- Web: https://vectorize.io/articles/gbrain-review
- Web: https://claude.com/blog/improving-skill-creator-test-measure-and-refine-agent-skills
- Web: https://github.com/garrytan/gstack