How a decision moves

A Revue decision begins only when an agent reaches a genuine human fork. The decision plane then compresses duplicate asks, records one human answer, and routes it back to every dependent session.

agent reaches a fork
  → search prior decisions
      → applicable answer: reuse it + suppression receipt
      → no answer: submit a structured ask
          → group near-duplicate asks
          → present one decision to the human
              → dismiss: close the decision without an answer
              → answer once: create one route per waiting workstream
                  → deliver → consume → apply → resolve
                  → store the answer for scoped future reuse
                  → issue answer and suppression receipts

1. Search before asking

Every asker searches decision memory first. An applicable answer suppresses the interruption and returns immediately with a receipt. A potentially related answer can be cited for confirmation instead of silently applied.

Memory is deliberately scoped. Semantic similarity is not proof that two decisions are interchangeable, and humans can inspect or correct a remembered answer with revue why and revue correct.

2. Submit one answerable ask

When memory does not resolve the fork, the agent submits one concise question with:

mutually exclusive options;
a recommended default;
the risk and cost of guessing wrong;
evidence a human can trust;
concepts and files that define the affected scope.

Routine implementation details, status updates, and questions answered by the repository should never become human decisions.

3. Group equivalent blockers

Revue compares the question and its scope with open asks from other workstreams. Near-duplicates join one decision while each contributing phrasing remains inspectable. The human sees one compressed question plus who is waiting and what the answer affects.

Grouping can be wrong in either direction: over-grouping may apply one answer too broadly; under-grouping creates repeated interruptions. Revue therefore preserves contributors and keeps automatic memory reuse conservative.

4. Answer or dismiss

A verified human can answer or dismiss a decision. Answering records who answered, what the human saw, and the exact waiting workstreams resolved at that moment. One route is created for each target.

Dismissal is terminal for that ask. The agent should continue with its recommendation or safest reversible path instead of rephrasing the same question to force another interruption.

5. Route the answer with proof

Routes advance only as far as an integration can prove:

State	What it proves
`pending`	An answer route exists for the target
`claimed`	One eligible integration leased the route for delivery
`delivered`	The integration appended the answer to the target session
`consumed`	The target accepted the delivered instruction
`applied`	Later agent lifecycle evidence shows work continued after consumption
`resolved`	The integration closed the route after its deepest available proof
`failed`	Delivery or follow-through failed with recorded evidence

These states prove transport and observed lifecycle, not that the agent interpreted the decision correctly or produced correct code. Repository tests and human review remain separate evidence.

6. Show the value with receipts

Revue emits two human-readable receipt types:

Answer receipt: one answer was routed to the pooled askers and is eligible for future matching questions.
Suppression receipt: an agent used a prior human answer instead of creating another interruption.

The rollup counts interruptions made and prevented. Receipts also identify the remembered decision so a human can inspect it with revue why <decisionId> or narrow, retire, or expire it with revue correct.

Next, follow the Quickstart to exercise this loop or read the exact revue-mcp result states.