Analysis results, rerun, and cost

Conversation detail shows Analysis status and the latest completed result. The Analysis is independent of the conversation's ended status: a model or schema failure does not rewrite a successfully ended conversation as failed.

Structured Agent Testing conversations are excluded from post-conversation Analysis and remain skipped. Their Simulation or Next Reply evaluator, Tool assertion, parameter matcher, and failure bucketing are separate test evidence; inspect them under Past runs and results.

Status and result fields#

The lifecycle is pending, processing, done, skipped, or failed. A pending run has not been claimed. Processing means one worker owns it; a stale processing claim can return to pending after 15 minutes. Done stores a summary, aggregate call result, criterion results and rationales, typed data and rationales, optional sentiment, model, version, and conversation IDs read. Skipped means there was no transcript or no active Analysis setting. Failed records a generic failure event without exposing provider text.

The aggregate result is success, failure, or unknown. A single failed criterion wins; all criteria must succeed for success; no criteria or unresolved evidence produces unknown.

Run Analysis again#

Use Run analysis or Run analysis again in the Conversation Overview only for an ended conversation. An active conversation is rejected. A done, skipped, or failed run is reset and claimed again; an already processing run is not duplicated. Rerun uses the original agent-version snapshot, not the latest published criteria.

Developers can use Run conversation Analysis. The dashboard mutation requires its normal session and request protection; API keys can read Conversation detail but cannot trigger this rerun.

Cost, quota, and delivery#

Each actual Analysis model request records tokens and cost, adds that cost to the conversation total, and counts toward the workspace daily quota. One run uses one model request for all configured outputs. A rerun is billable only if it reaches the model. An empty transcript or no active Analysis setting returns to skipped without a model request, tokens, or added cost. An invalid model result may still have consumed model usage.

After a successful Analysis, an enabled Workspace post-call webhook receives the result. Configure it with the get post-call webhook, create post-call webhook, and update operations. Delivery retries do not change a done Analysis result. The analysis_completed record is also visible under Activity in Conversation detail.