Transcript and Activity

Conversation detail has three views: Overview, Transcript, and Activity. Together they explain the summarized outcome, what the customer and agent said, and which runtime operations produced that exchange.

Use Overview and Transcript#

Overview shows duration, turns, input and output tokens, cost, end reason, source, external customer identifier when present, agent version, conversation ID, and the latest Analysis. Treat the displayed version as the source of truth for that conversation.

Transcript is derived from the append-only user and agent message events in sequence order. It shows customer and agent text from the selected conversation record. A Telegram or other channel may create separate continuation segments, and Analysis can internally read the supported chain. The dashboard detail page and public conversation response currently do not expose a continuation link or relationship between segments. Find and inspect each visible record separately through Conversation history.

Every conversation uses a frozen agent version or a frozen draft snapshot captured when it opened. Publishing, merging, rebasing, or editing later does not rewrite its transcript, Workflow labels, Procedure definitions, or version identifier.

Read Activity evidence#

Activity presents the ordered event record. The dashboard table shows the sequence number, rendered event type, and safe details derived from the payload. Follow that order to connect an observed reply to runtime work; the table does not display an event timestamp. The most useful evidence appears in this order of investigation:

• tool_call records show validated arguments, outcome, duration, assigned scalar values, and a bounded result or safe error.
• workflow_node_entered, workflow_node_completed, workflow_edge_evaluated, and workflow_error show Workflow movement using labels from the frozen snapshot.
• procedure_started, procedure_step_waiting, and procedure_completed identify Procedure selection and progress.
• rag_retrieval records show whether Knowledge retrieval ran and how many chunks qualified.
• guardrail_triggered and Guardrail error records show safe type, mode, action, outcome, stage, model, or error type.
• llm_request records show model, purpose, tokens, latency, finish reason, and cost, including Analysis requests.

Internal prompts, provider explanations, complete Secret values, draft configuration, and unsafe Guardrail text are deliberately excluded. Activity is evidence for diagnosis, not a raw database or provider-log viewer.

Diagnose with frozen context#

Start at the final message and end reason, then follow Activity sequence numbers backward. Compare each event type and safe payload details. Confirm the agent version, the relevant Workflow or Procedure event, any Tool or RAG evidence, Guardrail decisions, and the associated llm_request. A missing event often means the feature was not enabled or the runtime never reached that step; an explicit error event means it was reached and failed.

When current editor labels differ from Activity, compare against the frozen version instead of the latest draft. Reproduce the intended change with Test chat, then inspect the new conversation separately. Continue with Analysis results or troubleshooting.