GraphSmith · Protocol v0.2

AI workflows
that survive.

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

Kill-testedRed-teamed by 3 rival modelsMITZero-dependencyNo telemetry
Every workflow leaves the forge with
Save pointsExactly-onceCitationsKill-tested

They fix what the AI knows. We fix whether the job finishes.

Knowledge graph → smarter about your data GraphSmith → survives the workflow that uses it
The problem — and the fix

Three ways every agent breaks. One fix each.

None show up in the demo. All of them show up in week two — so GraphSmith hardens each one at generation time.

Failure 01

Amnesia

Crashes at step 7 of 10 and restarts from step 1 — redoing paid API calls, re-scraping, re-generating.

→ Save points

Resumes where it stopped

Every step saves progress; a crash resumes at the exact step it stopped, never from zero.

Failure 02

Duplication

A retry after a half-finished step sends the same email twice, charges the same card twice.

→ Exactly-once

Effects run once

Recorded effects run once, even after power loss; an uncertain send halts loudly instead of re-firing.

Failure 03

Hallucination

Ask about your own codebase and it invents functions, files, and APIs that don't exist.

→ Citations

Grounded in real files

Every claim cites a real file via KnoSky, or is flagged as a guess. Your code never leaves your machine.

The stakes

The math nobody puts on the slide.

A workflow that's 85% reliable per step fails four times out of five across ten steps.

20%
survive to step 10.
0.85¹⁰ ≈ 0.20.
Survival probabilityper added step →
Four in five runs are already dead by step 10 — and each restart pays for the finished steps again.
73%

of enterprise AI agent deployments hit reliability failures in year one.

source ↗
63%

of developers spend more time debugging AI code than writing it.

JetBrains 2026 ↗
43%

of AI-generated changes needed post-deployment debugging.

source ↗
4/5

people building with AI tools have no technical background.

Lovable ↗
How GraphSmith works

Forge a workflow that survives.

A plain-English goal goes into the forge — a hardened, self-defending workflow comes out. Then watch it take a crash and finish anyway.

A

What it does

goal in → hardened workflow out
“Build me an agent that researches leads and drafts outreach.”
— you, in plain English
would forget where it crashed would send the email twice would invent files that don't exist
GraphSmith
Save pointsresumes where it stopped
Exactly-onceeffects run one time only
Citationsclaims cite real files
Engineering discipline, hammered in the moment the code is generated — so you don't have to know it.
so when it crashes in production…
B

What that gets you

the same 10-step agent, run twice, killed at step 7
Scroll in to run the simulation…
Un-forged
no save points · no guard
↺ restart from zero — every finished step paid for twice.
Forged by GraphSmith
checkpoint after every step · exactly-once guard
→ resumes exactly where it stopped — proven by a kill test.
Four steps

You approve before a line is written.

01

Describe it

Say the outcome in plain English. The skill activates on its own.

02

Approve the blueprint

One screen: workers, handoffs, save points, stop rules. Nothing is built until you approve.

03

It builds the project

A runnable, zero-dependency project. Runs immediately with no API keys.

04 · the part nobody ships

It proves it

A chaos test kills the run mid-flight and verifies it recovered — no duplicated work.

Proof, not promises

"Trust us" doesn't fly for anything that matters. So we don't ask you to.

Kill-tested

A chaos harness crashes it on purpose

It SIGKILLs the run mid-flight, restarts, and asserts it resumed from the last save point with zero duplicated effects. Verification is executable — not "the AI says it's fine."

Red-teamed

By three rival models

They held the release twice until every finding was fixed or disclosed. Full reports — dissents preserved — live in the repo.

GPT-5.1Gemini 2.5 ProDeepSeek R1
Five invariants

Enforced, not promised

Verified integrity · immutable, gated core · isolated evaluation · local-by-default · observable & killable. A deterministic check runs at every boundary.

graphsmith verify

Enforceable on every pull request

Capability profiles — resumable, effect-reconciled, integrity-verified, adversarially-tested — with linked evidence, via a GitHub Action or GitLab template.

What's in the box

Everything GraphSmith gives you.

One install. Nine capabilities across the full lifecycle — all local, all readable in minutes. Pick a stage:

Scaffolder

A runnable, zero-dependency project: deterministic manager, worker steps, save points, resume, capped retries, structured logs — runs immediately with no API keys.

Blueprint gate

A one-screen plan — workers, handoffs, save points, stop rules — that you approve before a single line is written.

Architecture linter

Scans existing JS / TS / Python for the classic failure patterns — unbounded loops, missing persistence, unsafe side effects — ranked by severity, self-tested against a bundled corpus.

Smallest-fix diagnosis

Maps "it forgets / duplicates / loops forever" to the exact broken rule, with file and line — and proposes the minimal fix without rewriting what works.

Chaos harness

Kills the run mid-flight and asserts recovery: kill test, double-run, power-loss probe, lock probes, and a loud safety-halt path.

graphsmith verify

Capability profiles (resumable, effect-reconciled, integrity-verified, adversarially-tested) enforceable on every PR via a GitHub Action or GitLab template.

KnoSky citations

A local pointer index your AI cites from — claims about your code carry a real file reference, or are flagged as guesses. A map, not an oracle. Nothing leaves your machine.

Session-fresh index

KnoSky re-indexes each session and degrades gracefully offline — pinned and content-hash verified, never a silent global install.

Multi-agent rules

Lanes (one writer each), task claims with leases, frozen contracts, no self-certification, and risk-tiered human gates — so parallel agents never collide.

11-Document Build System

Optional, for production teams: registry, PRD-to-task traceability, adversarial QA charter, release & rollback runbooks, and agent eval scorecards.

Self-healing

A broken workflow gets a diagnosis and a smallest-fix repair, staged with evidence and one-command rollback — never a silent rewrite of your logic.

Gated learning

Bounded, project-local improvements that must clear an executable gate and a human before adoption. It can't learn past its own safety rules.

Assurance testing

Architectural, unit, smoke, regression & adversarial batteries against your workflows, with tamper-evident evidence.

Graduation ladder

Tells you exactly when to move up to SQLite, a framework checkpointer, or a durable engine — and when not to.

Not sure where to start? Ask the assistant in the corner — it answers what GraphSmith can do, how to install it, and more.

The outcome

What changes the day you adopt it.

Without this skill
With it
Crash → restart from zero, paying for every step again
Resumes at the exact step it stopped — proven by the kill test
Retry → duplicate emails, charges, posts
Recorded effects run exactly once; an uncertain send halts loudly, even after power loss
AI invents APIs and files in your codebase
Every claim cited to a real file, or flagged as inference
Architecture decided by vibes, discovered in production
You approve a one-screen blueprint before a line is written
For teams that answer to someone

Conservative by design. Local by default.

v0.2.0 turns the discipline into a protocol — five invariants a deterministic check enforces at every boundary.

Integrity

A deterministic check runs at every boundary; injected text can't reach control flow.

Gated core

A constitutional core no automated process can edit; every change passes a four-gate pipeline.

Isolated eval

Every candidate change is tried first in a disposable, mocked-effect copy.

Local by default

Healing, learning, evolution stay on your machine. No telemetry, ever.

Observable & killable

Every run declares budgets, streams to a live view, and is safe to kill and resume.

For regulated teams
Use cases

The jobs people are already trusting it with.

Growth

Lead research → outreach

Researches new leads and drafts emails. A crash never re-pays for the enrichment calls it already made.

Finance

Invoice & payment runs

Exactly-once so a retry never double-charges — and an uncertain send halts for a human instead of guessing.

Data

Nightly pipelines & ETL

Resumes at the failed step instead of re-scraping and re-enriching the whole dataset from zero.

Engineering

Codebase refactor agents

KnoSky citations mean it never edits a file or API it hallucinated into existence.

Scale

Multi-agent build swarms

Lanes, task claims with leases, and frozen contracts — parallel speed without parallel chaos.

Regulated

Finance / health / identity flows

Risk-tiered human gates and evidence packets for the workflows you can't afford to get wrong.

Get started

One install. Every agent.

Works identically via the Agent Skills open standard — plain Node 18+, macOS, Linux, and Windows.

Claude CodeCodex CLIGemini CLICursorWindsurf / DevinHermesVS Code Copilot+ 25 others
npx skills add SathiaAI/graphsmith

Then, in any agent: “Build me an agent that researches new leads and drafts outreach emails.”  ·  Free. MIT. No API keys. No telemetry.