Kontour Flow Agents

Survey evidence for paused Flow gates

Decision

Flow Agents provides a library adapter for a host that has a server-owned Survey review session and needs to continue an already-paused Flow gate. The host configures a review-session resolver capability separately from continuation requests. A request carries only an opaque review-session reference; the resolver returns the persisted record, append-only events, current server state, producer identity, and its host-owned Flow subject binding. The adapter derives the review result and delegates the complete ReviewItem plus result projection to Survey’s public buildCanonicalReviewedTrustInput API. It verifies that the Survey-owned projection matches the current gate, then builds Survey’s ordinary trust bundle.

The adapter does not accept caller-authored review state, events, projection, or decision as authority. It also does not define a new gate, evaluate a gate itself, or imply lifecycle authority from a review result. The caller supplies an exact Flow run head and an explicit resume authority; Flow performs the one atomic evidence attachment, evaluation, and pass-only resume transaction.

For the reverse direction, Flow Agents discovers review work only from a persisted blocking outcome at the current gate and an exact Flow Run head. The host must explicitly identify which missing expectation ids are handled by a Survey producer; the adapter does not infer human-review semantics from claim type names. A discovery request carries the expectation selector and immutable run, gate, and expectation correlation, but no candidates.

The configured producer creates the canonical Survey ReviewItem, including its candidates and provenance. Flow Agents validates every candidate’s claim target against the persisted expectation, adds adapter-owned correlation annotations, and can publish the result through a host-owned queue capability using a deterministic idempotency key. It never fabricates a current or proposed value from a gate description or explore_hint.

Consequences

  • A rejected, incomplete, stale, foreign, or mismatched review cannot change the Flow run or its evidence manifest.
  • Survey is the single implementation source for review status precedence, effective edited values, claim identity, and candidate/source/extraction provenance. Flow Agents does not retain a parallel field comparator.
  • A non-passing canonical review remains inspectable through the returned Survey decisions and Flow outcome, while Flow keeps the paused run unchanged.
  • Native or in-process hosts call the same provider-neutral library API. This decision does not select a model runtime or transport.
  • Queue storage, ReviewItem construction, and review-session persistence remain host or producer capabilities. Flow Agents owns only discovery, validation, workflow correlation, and composition with the continuation adapter.