Skip to main content

​
States

failed and abandoned are explicit terminal exits from any live state. workflow.status_and_next is the agent’s read of the current state, its gates, the allowed actions, and the next action. Agents call it first and after every transition.

​
Gates

An experiment moves forward only when the gate has real evidence behind it: Transitions seal the artifact set in the same database transaction as the state change. Editing a file in your checkout changes nothing until you submit it again. Limits: at most seven non-terminal experiments per project, and no new experiments while a project reflection is mandatory (see Reflection waves).

​
Attempts

A rejection back to planned starts a new attempt; the old plan, evidence, and reviews stay on record. A rejection back to running keeps the attempt — the plan stands, the execution or the write-up doesn’t.

​
How a review happens

Four workflow roles: design_reviewer, experiment_reviewer, reflection_reviewer, consolidation_reviewer (review.request also accepts human and automated_check outside the gates).
  1. The producer agent calls review.request.
  2. The brain pins the target snapshot and returns a short-lived capability once, with a reviewer handoff prompt.
  3. A distinct reviewer session calls review.start and receives the pinned evidence plus bounded context — not the producer’s conversation.
  4. The reviewer submits one verdict and synopsis through review.submit.
  5. Submission re-checks that the request and snapshot are still current before it routes a rejection or satisfies a gate.
Capabilities are stored as hashes and expire after an hour; re-requesting reissues one. Producer and reviewer session ids must differ — a workflow guarantee, not cryptographic proof. Per client: Reviewer handoff.

​
What reviewers look for

  • Design review: can this plan actually test its claim? Is the success criterion decided before the run? Is the scope bounded?
  • Experiment review: does the evidence support the conclusion? Was the approved plan followed? Does the report interpret the metrics it shows? Is the logic graph honest?
Reviewers are told to be adversarial. A rejection with a reason is the normal case, not a failure.