Skip to main content
backpressure-plan turns a goal and its acceptance criteria into a gate matrix: each criterion is mapped to a verification method (schema, unit/integration test, eval, review, approval, or trace) and an enforcement policy. Use it before building or running an automation so every “done” claim has to be proven by evidence instead of an agent declaring victory and moving on. 
bunx smithers-orchestrator workflow run backpressure-plan \
  --prompt "The /login endpoint returns valid JSON, rejects bad credentials with 401, and stays under 200ms p95"

​
Stages

  1. extract-criteria: pull the prompt apart into a flat list of atomic, verifiable acceptance criteria. A goal with no verifiable criteria fails the run loudly instead of producing an empty gate matrix.
  2. plan-gates: map each criterion to a verificationMethod, gateType (blocking / warning / informational), who checks it, the failure action, the evidence required, and whether a human approval is needed. Every blocking criterion must map to a real verification method.
  3. verify: deterministic (no agent) parity check of the matrix against the extracted criteria (one gate per criterion, same order, verbatim text), returning { match, criteriaCount, gateCount, missing, unverifiedBlocking, summary }.

​
Inputs

InputTypeDefault
promptstring"Describe the goal and its acceptance criteria in plain English."
Use this when you want a run that can apply real backpressure: every criterion has a check and an evidence requirement, so the workflow cannot silently skip the parts it could not finish.