The full capability set, the five protocol invariants, and the complete 11-Document Build System — for teams shipping agents into finance, health, identity, and other regulated work. Local-by-default. No telemetry. Evidence on every run.
v0.2.0 turns GraphSmith's discipline into a protocol. A deterministic check runs at every boundary — a workflow either passes graphsmith verify or it doesn't.
A deterministic check runs at every boundary; injected text in a log or tool output can't reach control flow.
A constitutional core no automated process can edit; every change passes a four-gate pipeline before it's real.
Every candidate change is tried first in a disposable, mocked-effect copy; nothing touches your live project until it's earned it.
Healing, learning, and evolution stay on your machine. Nothing is sent anywhere. No telemetry, ever.
Every run declares budgets, streams to a live view, and is safe to kill and resume; tripwires halt a run heading somewhere dangerous.
One install. Every capability across the lifecycle, laid out in full for evaluation. All local, all dependency-free, all readable in minutes.
A runnable, zero-dependency project: deterministic manager, worker steps, fsync-durable save points, resume, capped retries, structured logs — runs immediately with no API keys.
A one-screen plan — workers, handoffs, save points, stop rules — that you approve before a single line is written.
A lease + heartbeat so a second manager on the same run refuses while the first is alive; a dead holder's lock is stolen automatically, so a crashed run self-recovers.
Import-graph-aware scan of existing JS / TS / Python for the classic failure patterns — unbounded LLM loops, missing persistence, clock/randomness near control flow, unkeyed external writes — ranked by severity.
--selftest runs the linter against a bundled corpus (every published adversarial-review probe is a case) and fails on any precision/recall regression.
Maps "it forgets / duplicates / loops forever" to the exact broken rule, with file and line — and proposes the minimal fix without rewriting what works.
Kill test, double-run, power-loss probe, three lock probes (dead-holder, live-recycled-pid, lease-expiry), and a loud safety-halt path. A forged halt fails — the verdict comes from on-disk state, never a string.
Reports capability profiles — resumable, effect-reconciled, budget-enforced, integrity-verified, gated-learning, assurance-tested, adversarially-tested — with linked evidence.
Enforceable on every pull request through a GitHub Action or a GitLab template. A workflow passes the gate or it doesn't merge.
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.
KnoSky is pinned by version AND content hash (refuses on registry mismatch), never a silent global install. GRAPHSMITH_OFFLINE=1 skips all network.
KnoSky indexes pointers (titles, headings, short excerpts), never file bodies. Its index writes to a local .knosky/ folder in your repo.
Each agent owns a module boundary; one writer per lane, ever. Merge conflicts are planning failures to fix in the plan, not battles to win in the diff.
Work lives in a shared task file; an agent locks a task atomically, and a dead agent's expired lease frees it. No task is worked twice or orphaned silently.
Interfaces freeze before dependent work starts; the agent that built a thing never reviews it — the checker runs on a different model family.
A halted or broken workflow gets a diagnosis and a smallest-fix repair, staged with evidence and one-command rollback. Code repairs stage for a human — never a silent rewrite of your logic.
Bounded, project-local improvements that must clear an executable gate and a human before adoption. It cannot learn its way past its own safety rules.
Architectural, unit, smoke, regression & adversarial batteries against your own workflows, with tamper-evident evidence. A passing battery is a floor, not a certification.
For teams running many agents against production software — real users, money, or regulated data — GraphSmith ships the complete agent-first delivery pipeline: a document for every stage from PRD to evals, each machine-readable and traceable by ID. Adopt it when your project earns it, not before.
The differentiator isn't a bigger claim; it's a smaller, provable one. Honesty is the enterprise sell.
GRAPHSMITH_OFFLINE=1.docs/compliance/register.yaml, risk-tier gates key off your data classifications, evidence packets cite your obligation IDs, and claims about your policies require citations — all while GraphSmith stays one public skill and your register stays in your repo.GraphSmith is the disciplined on-ramp, not a competitor to durable-execution engines. The graduation ladder tells you exactly when a workload outgrows this tier — and the migration seam is already isolated in the generated code.
Correct for single-machine jobs up to roughly thousands of steps. Where every new project starts.
When step counts and log volume grow past the file tier. Same conventions, more durable store.
A LangGraph-class checkpointer when you need its ecosystem and state model.
Temporal / Inngest for distributed, long-running, exactly-once workloads at scale.
Free, MIT, zero-dependency, no API keys. Time-to-first-value is under a minute.
npx skills add SathiaAI/graphsmith