Skip to content

Orchestration Release Acceptance

An agent change is releasable when role contracts, lifecycle authority, convergence, termination, trace, and final artifacts tell one coherent story. A persuasive final response is not evidence that the governed workflow ran correctly.

flowchart LR
    request[Typed request]
    lifecycle[Owned lifecycle]
    roles[Bounded role calls]
    decision[Judgment and convergence]
    outcome[Terminal outcome]
    trace[Complete trace]
    artifacts[Result and trace pair]

    request --> lifecycle --> roles --> decision --> outcome --> trace --> artifacts

Acceptance record

Changed surface Required evidence Release-blocking result
role input or output strict-model, unknown-field, failure, metadata, and serialization cases malformed or ambiguous role data reaches orchestration
lifecycle transition allowed/forbidden transition, passive-role, abort, and terminal-state evidence a role or provider can override lifecycle policy
merge or shard handling lineage, per-input status, conflict, and failure-assembly cases one successful shard hides another shard's failure
judgment or verification issues, action plan, decision, confidence, veto, and validation findings terminal content loses the decision that admitted or rejected it
convergence strategy, window, snapshot, hash, oscillation, maximum-iteration, and non-convergence cases a stopped loop is relabeled as converged success
trace contract mandatory fields, order, schema/runtime versions, deterministic exclusions, and reconstruction the terminal outcome cannot be rebuilt from the trace
provider adapter deterministic adapter failures plus opt-in live connectivity evidence a live response substitutes for lifecycle or failure tests
CLI or HTTP boundary typed parity, schema, dry-run, failure, and artifact-path evidence an adapter changes status or trace semantics
result custody final_result.json, named trace, and reconstruction comparison the two files are missing, divergent, or from different attempts

Terminal outcomes are not interchangeable

Acceptance evidence must preserve why orchestration stopped and which work is usable:

Terminal condition Required record Prohibited interpretation
converged success criterion, observation window, snapshots or hashes, admitted result, and complete trace the content is factually correct or runtime-accepted
completed without convergence maximum-iteration or stopping rule, last candidate, and non-convergence status successful convergence because the loop ended normally
partial result completed and failed inputs, lineage, usable artifacts, and explicit partial verdict whole-batch success based on one successful shard
veto or validation refusal source role, issues, action plan, targeted artifact, and final disposition infrastructure failure or missing output
provider or role failure call identity, stable error class, retry/fallback decisions, and lifecycle response empty content or a skipped role
interrupted or aborted last causal event, lifecycle state, incomplete work, and recovery boundary finalized trace or replayable completion
fatal orchestration failure terminal reason, preserved partial evidence, and non-success exit/response an ordinary role-level refusal

The result schema, trace schema, lifecycle, and public adapter must agree on the condition. A release is blocked if one surface reports success while another retains failure, partial completion, veto, or non-convergence.

Acceptance fixture set

The release candidate needs more than a golden success trace. Retain fixtures for an allowed lifecycle, a forbidden transition, a passive role, a failed provider call, a merge with one failed shard, a veto, oscillation or maximum iterations, an interrupted run, trace reconstruction, and a mismatched result/trace pair. Add CLI and HTTP cases when the changed contract crosses those boundaries.

The invariant, integration, API, end-to-end, snapshot, trace, and provider adapter suites carry different authority. Live-provider evidence may confirm connectivity, but it cannot replace deterministic lifecycle, failure, and trace fixtures. Record provider/model identity and isolate that evidence from the package-owned orchestration verdict.

Replay classification

A replayable trace carries complete input, configuration, model, version, and convergence identity with temperature zero. This classification supports deterministic reconstruction of recorded fields; it does not recreate a provider's historical model-serving environment.

Use change validation for focused routing and known limitations to keep provider, credential, and hosting claims bounded.