audit_launch.
The call is free and read-only. Poll it every few seconds after launching.
Authorization
string
required
Your bearer token:
Bearer <token>. See Authentication.Request body
string
required
The run to read.
Request
Knowing when it is done
Sections appear as they finish, so an early poll returns a partly filled response. The audit is finished whenreport, read, and verdicts are all
non-null:
replies fills in shortly after those three, so read it last. A run typically
completes in a few minutes; if all three are still null after 30 minutes, treat
the run as failed and prepare a new one.
Response
string
The run’s id.
string | null
The agent under review.
null on a run that was prepared but never launched.object
The parsed transcript the audit ran on, as
{ "messages": [...] }. Each
message carries the id (m1, m2, … in order), speaker, and text. The
ids are what the report’s evidence cites.object | null
The reception report:
health_score, summary, findings, per_user,
interactions, and interaction_totals. Same shape as
analyze. null until that section finishes.object | null
What your agent could have known about the people in the conversation.Every field inside is independently optional — a section that could not be
produced is omitted rather than failing the audit.
array | null
One entry per turn your agent took, each with the
index of the turn’s first
message, a risk grade, a summary of what the turn risked, and
predicted_message — what the other person was likely to say next.array
Rewrites of your agent’s worst turns. Each has the
index of the turn, the
rewritten reply, messages (the reply split the way a person would send
it), and the risk grade of the rewrite. Empty until the rewrites finish.Errors
Next: Analyze a transcript directly when you already
have structured messages and want just the report.