nexustechnology nexustechnologyThe persisted record is the execution; the process is just the thing currently advancing it. Everything else falls out of that inversion.
A tool call whose result is already recorded is never executed twice. Approve one of three calls tomorrow, the two that already ran stay ran.
Status, full conversation, pending approvals, cost. Mutable, atomically replaced.
Append-only: every LLM call, tool execution, and decision with who made it. Never rewritten. This is the audit artifact.
Both have public-domain JSON Schemas, the Open Agent Run spec. Any tool can audit; any conforming runtime can resume; anyone can replay.
Or replay against a new model or prompt: recorded tool results become fixtures, and every divergence is reported at the exact call where behavior changed.
stdio or HTTP, zero-dep JSON-RPC. Every tool call: journaled · budget-metered · approval-gated · resumable · replayable. One line of policy. MCP's own roadmap lists durability as an open problem, this is the answer you can deploy today.
Exposes Nexus itself as an MCP server over stdio: run, approve, deny, respond, resume, and list durable runs as MCP tools, so Claude or Cursor can drive them.
On a pause a human can approve, deny, approve with edited arguments (authorize $50 instead of the proposed $250 before it runs), or respond to answer the call directly so the tool never runs. Human-in-the-loop is an architecture, not a yes/no gate.
attachOtel(runtime, { endpoint: OTEL_ENDPOINT }); // GenAI spans: invoke_agent → chat → execute_tool // token usage · prompt content off by default // → Datadog, Grafana, Langfuse, any OTLP backend await runtime.run('support', t, { tags: { customerId: 'acme-42' } }); // tags ride record, journal, webhooks & spans
"Which customer moved the bill" becomes listRuns({ tags }), not a forensics project. The exporter speaks OTLP, so runs flow into your SIEM or APM with no vendor SDK.
Providers stream tokens as an llm:token event, and cancel() aborts a live model call in flight, not just at the next checkpoint.
Zero-dep validateRun/validateStep/validateJournal in the box, plus a stdlib-only Python reader (python -m oar validate|timeline). Two independent implementations of a public-domain format is what makes it a real standard.
Memory, file (zero-dep default), built-in SQLite, and Postgres for multi-node, bring your own pg client so it stays zero-dependency. The interface is 5 methods.