> ## Documentation Index > Fetch the complete documentation index at: https://www.bolna.ai/docs/llms.txt > Use this file to discover all available pages before exploring further. # Retrieve voice AI execution debug bundle API > Fetch everything needed to debug one call in a single request via the Bolna API: the execution, per-turn latency with the structured transcript, and the pipeline logs. Returns what the dashboard's **Debugging** view shows for one call, in a single response. Use it when you want to find out why a call was slow, dropped or went off script without stitching together [Get execution](/docs/api-reference/executions/get_execution.md), the per-turn latency and [Get raw logs](/docs/api-reference/executions/get_execution_raw_logs.md) yourself. The [Bolna MCP server](/docs/mcp-server.md) uses it for its `get_execution_debug` tool. ## Sections | Section | What's in it | | - | - | | `execution` | The call's status and error, provider and numbers, timestamps, duration, cost and usage breakdowns, `telephony_data`, `latency_data`, `context_details`, extractions, `summary`, tool calls, transfers and retries | | `progression` | `call_start_at`, the latency analytics for each turn, and the structured transcript: user and agent messages, function calls and call events such as transfers and hangups | | `logs` | The pipeline logs for the call (transcriber, LLM, synthesizer and function calls), followed by one `webhook` entry with the call's webhook deliveries. Webhook deliveries are included for calls from the last 14 days | Pass `include` to fetch only some sections, for example `?include=execution,logs`. Leave it out to get all three. Any other value returns `400`. ## Reading LLM requests Every LLM request resends the full prompt and the conversation so far, so in `logs` each `llm` request after the first carries only the text it adds. Its `data_shared_prefix_chars` says how many leading characters it shares with the previous request. To rebuild the full text of each request: ```python theme={"system"} full_requests, previous = [], None for row in bundle["logs"] or []: if row.get("component") == "llm" and row.get("type") == "request": text = row["data"] if "data_shared_prefix_chars" in row: text = previous[: row["data_shared_prefix_chars"]] + text full_requests.append(text) previous = text ``` The first request is always whole: it is the exact prompt that ran, with variables filled in. The bundle is trimmed for size. For the full data, use [Get execution](/docs/api-reference/executions/get_execution.md) and [Get raw logs](/docs/api-reference/executions/get_execution_raw_logs.md). ## When a section is missing One section never fails the whole response. A section that doesn't exist for this call, or that fails to load, comes back `null`, and `unavailable` names it with the reason: ```json theme={"system"} { "execution_id": "4c06b4d1-4096-4561-919a-4f94539c8d4a", "unavailable": { "progression": "not_available" }, "execution": { "id": "4c06b4d1-4096-4561-919a-4f94539c8d4a", "status": "completed" }, "progression": null, "logs": [] } ``` | Reason | Meaning | | - | - | | `not_available` | There is no data for this section on this call | | `error` | The section failed to load. Retry, or fetch the same data from its own endpoint | ## OpenAPI ````yaml GET /executions/{execution_id}/debug openapi: 3.1.0 info: title: Bolna API description: >- Use and leverage Bolna Voice AI using APIs through HTTP requests from any language in your applications and workflows. license: name: MIT version: 1.0.0 servers: - url: https://api.bolna.ai description: Production server security: - bearerAuth: [] tags: - name: SIP trunks description: >- Bring-your-own-carrier SIP trunks, their gateways, carrier IPs and numbers. paths: /executions/{execution_id}/debug: get: description: >- Retrieve everything needed to debug one execution in a single response, trimmed to fit an AI assistant's tool result. Returns what the dashboard's Debugging view shows. parameters: - in: path name: execution_id required: true schema: type: string format: uuid description: The unique `execution_id` - in: query name: include required: false schema: type: string example: execution,logs description: >- Comma-separated sections to return, from `execution`, `progression` and `logs`. Omit to return all three. Any other value returns `400`. responses: '200': description: Debug bundle for the execution content: application/json: schema: $ref: '#/components/schemas/ExecutionDebugBundle' '400': description: >- `include` names a section other than `execution`, `progression` or `logs` content: application/json: example: message: >- include must be a comma-separated subset of execution, progression, logs '404': description: No execution with this `execution_id` that your account can read content: application/json: example: message: Agent execution not found components: schemas: ExecutionDebugBundle: type: object properties: execution_id: type: string format: uuid description: The execution this bundle is for unavailable: type: object description: >- Sections that came back `null`, each with the reason. `not_available` means there is no data for that section on this call; `error` means it failed to load. additionalProperties: type: string enum: - not_available - error example: progression: not_available execution: type: object nullable: true description: >- The execution, keeping only fields that help debug a call (status, `error_message`, provider and numbers, timestamps, `conversation_duration`, `total_cost`, `cost_breakdown`, `usage_breakdown`, `telephony_data`, `latency_data`, `context_details`, extractions, `summary`, `tool_call_logs`, `transfer_call_data`, retries). Empty fields are dropped and the prose `transcript` is left out. Present only when requested. additionalProperties: true progression: type: object nullable: true description: >- Per-turn latency and the structured transcript. Present only when requested. properties: call_start_at: type: string description: When the call started turns: type: array description: >- Latency analytics for each turn that has them, in milliseconds. Fields without a value are dropped. items: $ref: '#/components/schemas/ExecutionDebugTurn' transcript: type: array description: >- The call in order, as user and agent messages, function calls and call events items: $ref: '#/components/schemas/ExecutionDebugTranscriptItem' logs: type: array nullable: true description: >- The call's pipeline logs in order, followed by one `webhook` entry with its webhook deliveries (calls from the last 14 days only). Present only when requested. items: $ref: '#/components/schemas/ExecutionDebugLog' ExecutionDebugTurn: type: object properties: sequence_id: type: integer turn_id: type: integer asr_end_to_user_end_ms: type: number description: >- Silence after the caller stopped speaking before the transcript was final llm_start_to_asr_end_ms: type: number description: Time from the final transcript to the LLM starting llm_duration_ms: type: number description: How long the LLM streamed, first token to last tts_start_to_llm_start_ms: type: number description: Time from the LLM starting to text-to-speech starting tts_start_to_llm_end_ms: type: number description: >- Gap between the LLM finishing and text-to-speech starting. Negative when speech started before the LLM finished agent_speech_to_tts_start_ms: type: number description: Time from text-to-speech starting to the agent starting to speak agent_speech_to_tts_first_audio_ms: type: number description: Time from the first synthesized audio to the agent starting to speak ExecutionDebugTranscriptItem: type: object properties: kind: type: string enum: - message - function_call - event role: type: string enum: - user - agent description: Who spoke (`message` rows) text: type: string description: What was said (`message` rows) interrupted: type: boolean description: Whether the message was interrupted message_category: type: string description: Tag for scripted agent speech, such as a hangup or transfer message not_sent_to_agent: type: boolean description: >- Present and `true` when the caller's speech was transcribed but not passed to the LLM ts_exact: type: boolean description: >- Present and `false` when the row's time was estimated for ordering rather than measured tool_name: type: string description: The tool that was called (`function_call` rows) status: type: string description: Outcome of the tool call (`function_call` rows) response_status_code: type: integer description: HTTP status the tool's API returned (`function_call` rows) tool_latency_ms: type: number description: How long the tool call took (`function_call` rows) label: type: string description: What happened, such as a transfer or hangup (`event` rows) ts: type: string description: When it happened (`function_call` and `event` rows) ts_ms: type: number description: When it happened, in milliseconds (`function_call` and `event` rows) ExecutionDebugLog: type: object properties: created_at: type: string description: Time of the log entry component: type: string description: >- The pipeline component, such as `transcriber`, `llm`, `synthesizer` or `webhook` example: llm type: type: string enum: - request - response provider: type: string description: The provider or model data: description: >- The log content. For every `llm` request after the first that shares a prefix with the previous request, only the text after that prefix. For the `webhook` entry, the list of webhook deliveries. oneOf: - type: string - type: array items: type: object data_shared_prefix_chars: type: integer description: >- Set on delta-encoded `llm` requests. The full request is the previous request's full text up to this many characters, followed by `data`. latency: type: number description: Latency of this step, when recorded tool_name: type: string description: The tool, on function-call rows reasoning_content: type: string description: The model's reasoning summary, when the provider returns one rag_info: type: object description: Knowledge base retrieval details, when the agent used one graph_routing_info: type: object description: Graph agent routing details, when the agent is a graph agent status: type: string description: >- Webhook delivery status on the `webhook` entry (`success`, `no_data` or `not_configured`) securitySchemes: bearerAuth: type: http scheme: bearer ```` This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.