Graphsmith

Describe the job. Approve the plan. Get an AI workflow that's tested to survive failure — before you ever trust it.

Works identically in Claude Code, Codex CLI, Gemini CLI, Cursor, Windsurf / Devin Desktop, Hermes, and VS Code Copilot via the Agent Skills open standard. One install, every agent.

Reproduce it · Get started · What's inside · FAQ · Report a bug

• The problem
• The solution
• The outcome
• 60-second start
• What's in the box
• Detailed guide
• Warnings & disclaimers
• What this skill does NOT do
• Roadmap
• FAQ

The problem

AI makes it easy to build multi-step automations — and easy to build them wrong in the same three ways, every time:

1. Amnesia. The workflow crashes at step 7 of 10 and restarts from step 1 — redoing paid API calls, re-scraping, re-generating. 73% of enterprise AI agent deployments hit reliability failures in their first year, and the math is unforgiving: a workflow that's 85% reliable per step fails 4 out of 5 times across 10 steps (0.85¹⁰ ≈ 20% survival).

2. Duplication. A retry after a half-finished step sends the same email twice, charges the same card twice, posts the same message twice. Audits of AI-generated automation code find the same four omissions over and over: error handling, safe re-runs, retries, and logging (Sonar, State of Code survey, Jan 2026).

3. Hallucination. Ask the AI about your own codebase and it confidently invents functions, files, and APIs that don't exist — then builds on top of them.

None of these show up in the demo. All of them show up in week two — and the cost is measurable: 63% of developers have spent more time debugging AI-generated code than writing it themselves (JetBrains, 2026), 43% of AI-generated changes needed post-deployment debugging, and the documented incident list already includes an AI agent wiping a production database during an explicit code freeze.

And the people hitting these failures are increasingly not engineers — 4 out of 5 people building with AI tools have no technical background (Lovable, Build Economy report), while 37% of US developers are actively vibe coding (Stack Overflow, 2025). Founders, PMs, and operators are building real automations with no way to know these failure modes exist until they're living them.

The solution

GraphSmith is a skill your AI agent follows — it injects the engineering discipline at the moment code is generated, so you don't have to know it:

• A blueprint before any code. One screen in plain English: the workers (one job each), what they hand each other, where progress is saved, and exactly when everything stops. Nothing is built until you approve it.
• Architecture generated correctly by default. A plain, predictable manager controls the flow; AI works only inside individual steps. Every step saves its progress, is safe to re-run, retries a capped number of times, and writes one log line.
• Grounded in your actual code. Via KnoSky (local-first, auto-updated, your code never leaves your machine), every claim the AI makes about your existing codebase carries a citation to the real file — or is flagged as a guess.
• Proof, not promises. A chaos harness kills your workflow mid-run, restarts it, and asserts it resumed without redoing finished work — and that a crash inside a side-effect window halts for a human instead of guessing. Verification is executable, not "the AI says it's fine."

Already have a broken automation? The linter diagnoses it — mapping "it keeps forgetting / duplicating / looping" to the specific violation, with file and line.

The outcome

Reproduce it yourself — ~30s, public repo, zero dependencies, no API keys:

git clone https://github.com/SathiaAI/graphsmith
cd graphsmith
node scripts/scaffold.js demo && node scripts/chaos.js demo

A safety-halt on an uncertain send also counts as a pass — that is the guard doing its job.

Net effect: the reliability engineering that separates a demo from a production workflow, applied automatically, verified mechanically, at zero added cost.

For product and engineering leaders evaluating this: time-to-first-value is under a minute (the scaffold runs with no API keys); the risk profile is deliberately conservative (zero dependencies, read-only grounding, nothing leaves the machine); and there's no lock-in — references/graduation.md defines exactly when a workload outgrows this and should move to SQLite, a framework checkpointer, or a durable execution engine, with the migration seam already isolated in the generated code. It's the on-ramp to serious orchestration, not a competitor to it.

60-second start

git clone https://github.com/SathiaAI/graphsmith
node graphsmith/scripts/install.js

Or via the skills CLI (installs into whichever agents it detects):

npx skills add SathiaAI/graphsmith

The installer detects every AI coding agent on your machine and installs into each. Then open your agent and say:

"Build me an agent that researches new leads and drafts outreach emails."

The skill takes over: blueprint → your approval → working project → chaos-tested handover. (Devin CLI: invoke with /graphsmith.) On Claude.ai, just open the .skill file and click Save skill.

What's in the box

Detailed guide

Path A — "I want to build something new"

1. Say what you want in plain English; the skill activates on its own.
2. Approve the one-screen blueprint (workers, handoffs, save points, stop rules).
3. It generates the project. Run it: cd your-project && node manager.js
4. Replace the stub workers in workers/ with your real logic — AI/API calls live here, and each stub's comments say exactly what to preserve and why.
5. Prove it: node scripts/chaos.js your-project — green verifies crash recovery and once-and-replay-verified recorded effects (a safety-halt on an uncertain send also counts as a pass — that is the guard working).

Path B — "My automation keeps breaking"

Tell your agent the symptom (forgets where it left off / sent it three times / loops forever). The skill runs the linter, maps each finding to the broken rule with file and line, and proposes the smallest fix — it will not rewrite what works. With a repo present, KnoSky grounds every diagnostic claim in your actual files.

Path C — "I'm an engineer; skip the hand-holding"

Talk shop and the skill matches you: terse output, full vocabulary, diffs over explanations, graduation-ladder placement for your workload, and the chaos harness as a CI-able gate. The scripts are dependency-free Node — read them in five minutes, vendor them freely.

Path D — "I'm running several agents at once"

The moment two or more agents (or sessions) build against the same repo, individual reliability stops being enough — agents overwrite each other's files, redo each other's work, and rubber-stamp their own output. The skill applies a coordination layer distilled from production multi-agent delivery systems:

• Lanes — each agent owns a module boundary; one writer per lane, ever. Merge conflicts get treated as planning failures to fix in the plan, not battles to win in the diff.
• Claims with leases — work lives in a shared task file; an agent locks a task by claiming it atomically, and a dead agent's expired lease frees the task automatically. No task is ever worked twice or orphaned silently.
• Frozen contracts — the interfaces between lanes freeze before dependent work starts, which makes whole waves of tasks parallel-safe by construction instead of by luck.
• No self-certification — the agent that built something never reviews or merges it, and the checker ideally runs on a different model family, because a fresh context of the same model has the same blind spots.
• Risk-tiered autonomy — routine changes flow automatically; anything irreversible or touching money, identity, or private data halts at a human gate with evidence attached. When the tier is ambiguous, it promotes — always.
• Cite or verify — "should work" is not a status. Every claim an agent makes carries a resolvable reference or run evidence, with KnoSky as the lookup-before-inventing layer.

What you get: parallel speed without parallel chaos — no duplicated work, no clobbered files, no agent approving its own mistakes, and a clear record of what shipped, why, and on whose authority. It scales down honestly, too: a solo builder running two sessions needs only three of the six rules and a plain tasks file.

Minimum by default, full system if you want it. The skill ships with the distilled essentials — the seven build rules and six coordination rules cover most projects. For teams running many agents against production software (real users, money, or regulated data), the repo also includes the complete 11-Document Build System (references/full-build-system.md): document registry and manifest, PRD-to-task traceability, adversarial QA charter, release and rollback runbooks, agent operating protocol with risk-tiered human gates, and eval scorecards for the agents themselves. It's a great-to-have, not a prerequisite — adopt it when your project earns it, not before.

KnoSky (the grounding layer)

KnoSky turns your repo into a local map plus a read-only connector your AI cites from. It indexes pointers (titles, headings, short excerpts), never file bodies; nothing is uploaded anywhere. The skill keeps it updated each session and requires citations for any claim about existing code. Free — try it standalone with npx [email protected] . (the pinned, integrity-checked version this release ships against — the pin lives in scripts/knosky-sync.js)

Warnings & disclaimers

Read these before trusting the output with anything that matters:

• Templates and guardrails are starting points, not substitutes for product discovery, security review, legal advice, or experienced engineering judgment. Increase depth — and keep human gates wide — for anything handling money, health, identity, personal data, or regulated activity.
• The linter is heuristic — and measured. It is project-aware (import-graph based) but AST-free by design; node scripts/graphlint.js --selftest runs it against a bundled corpus of known-bad and known-clean files (including every probe from our published adversarial review) and fails on any precision/recall regression. Absence of findings is still not proof of correctness; known blind spots are documented in the script header. The chaos harness provides evidence only for the two properties it tests (crash recovery, once-and-replay-verified recorded effects) on projects following its conventions.
• Human gates are load-bearing. The skill defaults to halting for approval on irreversible or sensitive changes. Removing those gates is your decision and your risk — autonomy should be earned by track record, never assumed, and should narrow again when quality slips.
• Verification by the author is not verification. If you scale to multiple agents, keep the no-self-certification rule intact; an agent reviewing its own work is the single most common way bad changes reach production.
• Never put real production data in test environments, and never paste live secrets into documents, prompts, or indexes — agents read secrets from the environment, never from files they can echo back.
• AI-generated tests skew toward the happy path. Treat green tests written by the same agent that wrote the code as a floor, not a ceiling.

What this skill does NOT do

Equally important to state explicitly (for the GSA attestation layer, the full honesty boundary — what it verifies, and its replay, external-side-effect, generated-skill, per-skill-capability, and signer-trust limitations — is consolidated in KNOWN-LIMITATIONS.md, with a worked offline-verifier transcript):

• It does not replace orchestration frameworks or durable execution engines. LangGraph, Temporal, Inngest and peers solve distributed, long-running, at-scale delivery-guarantee workloads. GraphSmith is the disciplined on-ramp; the graduation ladder tells you when to move up, and the checkpoint seam in generated code is built for that migration.
• It does not run or host anything. No cloud service, no runtime, no daemon. Everything is local files and local scripts you can read in minutes.
• It does not write your business logic. The scaffold gives you the reliable skeleton; the workers' actual jobs — your API calls, your model calls, your rules — are yours to fill in.
• It does not show your logic is correct — or true single-delivery to external systems. The chaos harness tests exactly two properties: crash recovery, and single, replay-verified execution of recorded effects (with loud halts when an external send's outcome is uncertain). Single-delivery to an external system requires an idempotency key that system honors — the graduation ladder tells you when and how. A workflow can pass both and still compute the wrong answer — that's what your acceptance criteria and tests are for.
• It is not a security scanner or supply-chain audit. The linter finds architecture violations, not vulnerabilities. Use dedicated security tooling for that; the full build system (references/full-build-system.md) shows where such gates belong.
• It does not manage secrets or credentials. Agents read secrets from the environment — never from files, prompts, or indexes — and this skill never asks for them.
• It does not phone home. No telemetry, no analytics. The one network exception is explicit and inspectable: knosky-sync.js contacts the npm registry to verify the pinned KnoSky version against a baked-in content hash (refusing on mismatch), and npx downloads that package on first use — GRAPHSMITH_OFFLINE=1 disables all of it. Nothing else touches the network; your code and indexes stay on your machine (note: running knosky writes its index to a local .knosky/ folder in your repo).
• It does not remove humans from consequential decisions. Irreversible changes and anything touching money, identity, or private data are designed to stop at a human gate. That is a feature; the skill will not help you delete it quietly.
• Three boundaries are deliberate, not backlog. (1) No log rotation at this tier — a run's write-ahead logs grow with its step count, which the graduation ladder already bounds (rung 1 = low thousands of steps; beyond that, move to SQLite, rung 2). (2) No cross-file data-flow analysis in the linter — same-file clock/random value tracking is done; going further needs an AST, which would break the zero-dependency, read-it-in-five-minutes auditability the tool is built on (each blind spot is named in the linter's header). (3) No air-gapped supply-chain verification at the KnoSky-sync tier — the baked content hash defeats label reassignment but still queries the registry; offline, out-of-band-signed verification is delivered separately in the v0.3.0 regulated register (register Lane D), with its own adversarial review. Each exclusion states its trigger for revisiting; none is silent.
• It does not answer deep questions about your code by itself. KnoSky routes agents to the right files with citations; the agent still reads the live source. A map, not an oracle.

Roadmap

Driven by real user questions — the best way to influence it is to open an issue.

Shipped — v0.2.0: the GraphSmith Protocol. GraphSmith is no longer only a skill: v0.2.0 extracted its integrity and evolution behaviour into a published, versioned specification — the GraphSmith Protocol (v0.1-draft). It defines five invariants (I1–I5), evidence-carrying capability profiles (R/E/B/T/G/Q/X emitted by graphsmith verify --profiles), and a four-gate promotion pipeline, backed by JSON schemas, a conformance lab, a GitHub Action, and an evidence-carrying badge. Self-healing, gated learning (staged-only), and an assurance/red-team harness ship alongside. Reviewed by multiple non-Anthropic model families with dissents preserved in docs/reviews/.

Shipped — v0.3.0: GraphSmith Attestation (GSA). MCP connects your agents; GSA proves what they did. v0.3.0 turns each run into a signed, replayable, tamper-evident attestation bundle any CI can verify and any regulator can read — evidence, never a certification. It delivers:

• The GSA bundle + verifier — a portable, signed record (goal, plan, compiled-and-signed graph, skill provenance, capabilities, execution trace, outputs) with a conformance suite, extending the protocol's evidence-carrying profiles. plan → sign → apply → verify → replay.
• The regulated register — the compliance layer, delivered in v0.3.0 (previously scoped as the v0.2.0 regulated-industries extension, #1): approver-identity attestation, an obligations→controls coverage map that reports honestly and fail-closed, evidence retention (append-only, hash-chained, WORM-exportable), and air-gapped verification (an out-of-band maintainer signature so an offline site verifies with no network and no trust in the registry).
• Universal adoption — a thin MCP attestation shim so any MCP-speaking agent can emit a bundle, an open conformance kit + evidence-carrying badge, and a cross-agent property matrix.

Every security-tier component was built by one model family and adversarially tested by ≥2 different, non-Anthropic families — 16 real defects found and fixed (1 critical, 4 high), zero by rubber-stamp review — then closed out by a 282-scenario gauntlet (all HOLD / 0 BREAK), independently reproduced by a non-Anthropic harness. The MCP shim attests any MCP-speaking agent with no agent changes (Claude Code, Cursor, LangGraph, Devin, Windsurf verified). Full record in docs/reviews/.

In progress — v0.4.0: hardening the trust surface. Deferred review gaps become enforced controls, each built by one model family and adversarially reviewed by ≥2 independent non-Anthropic families (no self-certification):

• Capability conformance (R1) — per-skill filesystem / model / subprocess / network grants; requested ⊆ granted, attested only when the class is actually enforced.
• Side-effect receipts (R2) — recorded external effects reconcile against adapter receipts; evidence-less "success" is not vouched for.
• Signer lifecycle + recall (R3) — revoked signers and recalled attestations fail closed; a rotated key resolves only to a live successor.
• Secret / PII redaction (R4) — an exported trace carrying an unredacted secret or PII fails closed, against a declared-format recall gate.
• SBOM + build provenance (R5) — a hash-pinned bill of materials and provenance recomputed against actual artifact hashes; npm publish --provenance.
• Policy-as-code (R8) — versioned enterprise-safe profiles; a required control counts only when enforced.
• GSA §9.11 — the verifier recomputes all five run-time controls from a bundle's own evidence and fails closed on any control-lie; a v0.3.0 bundle verifies unchanged.
• Independent Assurance (R7) — a citable program formalizing the ≥2-non-Anthropic posture (docs/ASSURANCE.md).

A graphsmith verify <bundle> CLI ships from the graphsmith-skill package for CI and enterprise verifiers.

Also in design — System Blueprint & Architecture Review Gate (#2) Information architecture for multi-piece systems: a system blueprint (piece inventory, single-owner data map, frozen contract cards, blast-radius statements), lightweight decision records, a blocking architecture review gate with a concrete rubric and mandatory triggers, and chaos testing extended to the seams between pieces — kill one piece mid-handoff, and confirm the others' state survives.

Track record — review findings published Every release goes through multi-pass, cross-model adversarial review with dissents preserved. v0.1.1 was held by the council over a lead reviewer's dissent until the flagship guarantees survived power loss; v0.2.0 closed through a multi-model gauntlet (150+ orchestrator tests plus an independent execution pass, zero breaks); v0.3.0 through a 282-scenario gauntlet (all HOLD / 0 BREAK) independently reproduced by a non-Anthropic harness, after 16 real defects were found and fixed. Twice now the review process has caught our own claims outrunning our code — and both times we published the report and fixed the code. Same medicine the skill prescribes.

FAQ

Do I need to know how to code? No. Describe the outcome, approve the blueprint. You'll eventually edit worker files, but each is small, single-purpose, and commented in plain English.

Do I need API keys to try it? No. The scaffolded project runs immediately with stub workers; you wire in your model of choice when you replace them.

Which AI agents does this work with? Anything implementing the Agent Skills standard: Claude Code, Codex CLI, Gemini CLI, Cursor, Windsurf / Devin Desktop, Hermes, VS Code Copilot, and 25+ others. The scripts are plain Node 18+ and run identically everywhere, including Windows.

Does my code get uploaded anywhere? No. Scripts run locally; KnoSky is local-first by design and indexes pointers on your machine only.

How is this different from LangGraph / CrewAI / AutoGen? Those are frameworks you learn. This is the discipline, enforced at generation time, with a zero-dependency starting point — and an explicit graduation path to those frameworks when your workload earns it. On-ramp, not competitor.

What if the chaos test fails? That's the tool working: the message names the broken rule and the fix pattern. It caught a real bug in our own scaffold during development — a step killed after its side effect but before its save point re-executed on resume. The fix (guard every side effect with a check-before-write) is now modeled in every worker stub.

Why did it refuse my one-giant-agent design? Because that's the most documented failure pattern in the space. It proposes the split version and explains the tradeoff; you can overrule it, but it won't hide the risk.

Can multiple agents work on the same project at once? Yes — that's Path D. The skill applies lanes, single-commit task claims with leases, and frozen interface contracts so parallel agents never edit the same files or build against moving targets. See references/multi-agent-coordination.md.

Is there a full methodology for production teams? Yes — references/full-build-system.md contains the complete 11-Document Build System: registry, PRD, technical design, task graph, adversarial QA, release operations, agent operating protocol, and evals. It's included as an optional deep resource; the skill's defaults are the distilled minimum most projects need.

How do updates work? Re-run the installer with --force after pulling. KnoSky self-updates each session via knosky-sync.js, degrading gracefully offline.

Is it free? Yes. The skill is open source; KnoSky is free under FSL.

Something's broken — where do I report it? Open an issue with your agent name/version, OS, and the failing script's output. The scripts are deliberately tiny; most fixes are one-liners.

Acknowledgments

Born from building Sathia - where "just say it" becomes action.