@spendgraph/harness

Workflows: the shapes an LLM app takes, built on prompts, tools and graphs.

npm install @spendgraph/harness

import { refine } from "@spendgraph/harness";

const result = await refine({
  draft: (ctx) => write(ctx.input),
  judge: (draft) => grade(draft),
  maxRounds: 3,
});

What you get

• Seven shapes, one signature — refine, route, chain, parallel, loop, orchestrate, cascade.
• Every one of them can be watched with streamed(), tokens and progress both.
• Pause for a person and resume days later, with the store, the expiry and the idempotency yours to supply.
• Failure is a result, not an exception, so a run that fell over is still a run you can read.

Docs

| | | | --- | --- | | Overview | What a workflow is, and what they all owe you | | Choosing one | Seven shapes, and which fits | | refine | Draft, judge, revise | | route | Classify, then dispatch | | chain | Steps in sequence, with gates | | parallel | Sectioning, and voting | | orchestrate | A lead decomposes and delegates | | loop | Tools until it stops asking | | cascade | Cheap first, escalate on rejection | | Human in the loop | Pause for a person, resume days later | | Streaming a run | Tokens as they arrive |

License

Apache-2.0