Server API

The HTTP server mashbox run <model> starts on 127.0.0.1 (reach it from your machine with an SSH tunnel). It speaks the OpenAI Chat Completions, Completions and Responses APIs and the Anthropic Messages API, so OpenAI and Anthropic clients and coding agents work against it unchanged. On 127.0.0.1 no API key is checked and any key a client sends is ignored; with --api-key-file or MASHBOX_API_KEY, and always when --host is not loopback, /v1 routes, /metrics and /openapi.json need a key (Authorization: Bearer <key> or x-api-key: <key>; 401 without). POST bodies must be Content-Type: application/json (415 otherwise), and a request from a web page of another origin, or naming the server by a host name it does not answer to, gets 403. Fields a route does not list are accepted and ignored.

The same description as an OpenAPI 3.1 file: openapi.json in these docs, or GET /openapi.json on a running server.

Method	Path	What it does
GET	`/v1/health`	Health
GET	`/health`	Health (short path)
GET	`/`	Health, or the chat page
HEAD	`/v1/health`	Ready check without a body
GET	`/v1/models`	List models
GET	`/v1/models/{model}`	Get a model
POST	`/v1/chat/completions`	Chat completion
POST	`/v1/completions`	Text completion
POST	`/v1/messages`	Messages (Anthropic)
POST	`/v1/messages/count_tokens`	Count tokens
POST	`/v1/responses`	Responses (OpenAI)
GET	`/metrics`	Metrics (Prometheus text)
GET	`/openapi.json`	This description

GET /v1/health

200 when the model is loaded and the server takes requests, 503 while it loads. GET /health is the same.

Responses:

200: OK application/json Health
503: The model is still loading. application/json Health

GET /health

Same as GET /v1/health.

Responses:

200: OK application/json Health
503: The model is still loading. application/json Health

GET /

Same as GET /v1/health. With mashbox run <model> --ui it is the MASHBOX Chat page instead.

Responses:

200: OK application/json Health
503: The model is still loading. application/json Health

HEAD /v1/health

200 when ready, 503 while loading, no body. The server answers HEAD the same way on every path.

Responses:

200: Ready.
503: Loading.

GET /v1/models

The one model this server runs.

Responses:

200: OK application/json ModelList

GET /v1/models/{model}

The model, when {model} is its id; 404 otherwise.

Responses:

200: OK application/json Model
404: Another model id. application/json Error

POST /v1/chat/completions

OpenAI Chat Completions with function tools, thinking and streaming. Coding agents such as Pi use it.

Request body (JSON):

Field	Type	Default	Description
`model`	string		Accepted for compatibility; the server answers with the one model it runs.
`messages` (required)	array of ChatMessage		The conversation.
`tools`	array of object		Function tools the model may call.
`tool_choice`	string or object		auto (default), none, required, or {“type”: “function”, “function”: {“name”: …}}.
`stream`	boolean	false	Stream the answer as server-sent events (see the 200 text/event-stream response). A client that reads nothing from the stream for 30 s is disconnected and its answer stops.
`stream_options`	object		Fields: `include_usage`.
`chat_template_kwargs`	object		Chat template options. Fields: `enable_thinking`.
`enable_thinking`	boolean		Same as chat_template_kwargs.enable_thinking.
`merge`	boolean	true	Accepted and ignored: every request joins the running batch at the next device call, and no request field changes how other requests are scheduled.
`prompt_cache_key`	string		Tags the request’s session in the server log (as the X-Session-Id header does).
`store`	boolean		Accepted and ignored.
`logprobs`	boolean		Not supported (log probabilities): a request that sets it gets 400 naming the field.
`top_logprobs`	integer		Not supported (log probabilities): a request that sets it gets 400 naming the field.
`response_format`	object		Only {“type”: “text”}. Not supported: json_object and json_schema (structured output), answered with 400.
`max_tokens`	integer		Most new tokens (>= 1).
`max_completion_tokens`	integer		Most new tokens (>= 1); max_tokens is the same. Default: up to the end of the context (–ctx).
`temperature`	number	0.0	0 = greedy (default). Up to 4.
`top_p`	number	1.0	Nucleus sampling, in (0, 1].
`top_k`	integer	0	Keep the k most likely tokens; 0 = off.
`min_p`	number	0.0	Drop tokens below min_p times the top probability, in [0, 1).
`seed`	integer		Sampling seed; a random one when absent.
`stop`	string or array		A string or a list of non-empty strings; the text ends before the first one.
`stop_sequences`	array of string		Same as stop (Anthropic’s name), used when stop is absent.
`repetition_penalty`	number	1.0	> 0; 1 = off.
`frequency_penalty`	number	0.0	-2 to 2; 0 = off.
`presence_penalty`	number	0.0	-2 to 2; 0 = off.
`n`	integer		Must be 1.

curl -s http://127.0.0.1:8000/v1/chat/completions -H 'Content-Type: application/json' \
  -d '{"model": "qwen3-8-27b", "messages": [{"role": "user", "content": "Name three planets."}], "max_tokens": 64}'

Responses:

200: OK application/json ChatCompletion, text/event-stream ChatCompletionChunk
400: Bad request. application/json Error
404: A model id other than the one this server serves (GET /v1/models), code model_not_found. Leaving model out is accepted. application/json Error
429: Too many requests are waiting for a slot: 32 of yours (per tenant: keys the key file names with one tenant share it, a key without a name is its own tenant; without keys, 32 in all), or the server’s queue is full (256 waiting in all). Retry after the Retry-After header’s seconds. application/json Error
500: The request failed on the device: its one retry failed too, or it failed after its stream had sent output. A request whose client has received nothing is sent again (once, from the start): not streamed, until its response is written; streamed, until its first output is sent. A stream that has sent output is never retried by the server, since a second run could repeat what the client already has; it ends with the route’s error event instead, carrying the response id as request_id. application/json Error
503: The model is still loading. Retry after the Retry-After header’s seconds. application/json Error

POST /v1/completions

Continue a plain prompt, no chat template.

Request body (JSON):

Field	Type	Default	Description
`model`	string		Accepted for compatibility; the server answers with the one model it runs.
`prompt` (required)	string or array		The text to continue (a string, or a list of one string).
`stream`	boolean	false	Stream the answer as server-sent events (see the 200 text/event-stream response). A client that reads nothing from the stream for 30 s is disconnected and its answer stops.
`merge`	boolean	true	Accepted and ignored: every request joins the running batch at the next device call, and no request field changes how other requests are scheduled.
`prompt_cache_key`	string		Tags the request’s session in the server log (as the X-Session-Id header does).
`logprobs`	integer		Not supported (log probabilities): a request that sets it gets 400 naming the field.
`max_tokens`	integer		Most new tokens (>= 1).
`max_completion_tokens`	integer		Same as max_tokens (OpenAI’s newer name); wins when both are set.
`temperature`	number	0.0	0 = greedy (default). Up to 4.
`top_p`	number	1.0	Nucleus sampling, in (0, 1].
`top_k`	integer	0	Keep the k most likely tokens; 0 = off.
`min_p`	number	0.0	Drop tokens below min_p times the top probability, in [0, 1).
`seed`	integer		Sampling seed; a random one when absent.
`stop`	string or array		A string or a list of non-empty strings; the text ends before the first one.
`stop_sequences`	array of string		Same as stop (Anthropic’s name), used when stop is absent.
`repetition_penalty`	number	1.0	> 0; 1 = off.
`frequency_penalty`	number	0.0	-2 to 2; 0 = off.
`presence_penalty`	number	0.0	-2 to 2; 0 = off.
`n`	integer		Must be 1.

curl -s http://127.0.0.1:8000/v1/completions -H 'Content-Type: application/json' \
  -d '{"prompt": "The three largest planets are", "max_tokens": 32}'

Responses:

200: OK application/json Completion, text/event-stream Completion
400: Bad request. application/json Error
404: A model id other than the one this server serves (GET /v1/models), code model_not_found. Leaving model out is accepted. application/json Error
429: Too many requests are waiting for a slot: 32 of yours (per tenant: keys the key file names with one tenant share it, a key without a name is its own tenant; without keys, 32 in all), or the server’s queue is full (256 waiting in all). Retry after the Retry-After header’s seconds. application/json Error
500: The request failed on the device: its one retry failed too, or it failed after its stream had sent output. A request whose client has received nothing is sent again (once, from the start): not streamed, until its response is written; streamed, until its first output is sent. A stream that has sent output is never retried by the server, since a second run could repeat what the client already has; it ends with the route’s error event instead, carrying the response id as request_id. application/json Error
503: The model is still loading. Retry after the Retry-After header’s seconds. application/json Error

POST /v1/messages

Anthropic Messages with client tools, thinking and streaming, for Claude Code. The ?beta=true query Claude Code adds is accepted.

Request body (JSON):

Field	Type	Default	Description
`model`	string		Accepted for compatibility; the server answers with the one model it runs.
`messages` (required)	array of object		The conversation.
`system`	string or array		System prompt: a string or text blocks.
`tools`	array of object		Client tools the model may call.
`tool_choice`	object		{“type”: “auto” \| “any” \| “tool” \| “none”, “name”: …}.
`thinking`	object		Fields: `type`, `budget_tokens`.
`stream`	boolean	false	Stream the answer as server-sent events (see the 200 text/event-stream response). A client that reads nothing from the stream for 30 s is disconnected and its answer stops.
`prompt_cache_key`	string		Tags the request’s session in the server log (as the X-Session-Id header does).
`metadata`	object		Accepted and ignored.
`output_config`	object		effort is accepted and ignored. Not supported: format (structured output), answered with 400.
`max_tokens`	integer		Most new tokens; default: up to the end of the context (–ctx).
`max_completion_tokens`	integer		Same as max_tokens (OpenAI’s newer name); wins when both are set.
`temperature`	number	0.0	0 = greedy (default). Up to 4.
`top_p`	number	1.0	Nucleus sampling, in (0, 1].
`top_k`	integer	0	Keep the k most likely tokens; 0 = off.
`min_p`	number	0.0	Drop tokens below min_p times the top probability, in [0, 1).
`seed`	integer		Sampling seed; a random one when absent.
`stop`	string or array		A string or a list of non-empty strings; the text ends before the first one.
`stop_sequences`	array of string		Same as stop (Anthropic’s name), used when stop is absent.
`repetition_penalty`	number	1.0	> 0; 1 = off.
`frequency_penalty`	number	0.0	-2 to 2; 0 = off.
`presence_penalty`	number	0.0	-2 to 2; 0 = off.
`n`	integer		Must be 1.

curl -s http://127.0.0.1:8000/v1/messages -H 'Content-Type: application/json' \
  -d '{"model": "qwen3-8-27b", "max_tokens": 64, "messages": [{"role": "user", "content": "Name three planets."}]}'

Responses:

200: OK application/json Message, text/event-stream MessageStreamEvent
400: Bad request. application/json AnthropicError
429: Too many requests are waiting for a slot: 32 of yours (per tenant: keys the key file names with one tenant share it, a key without a name is its own tenant; without keys, 32 in all), or the server’s queue is full (256 waiting in all). Retry after the Retry-After header’s seconds. application/json AnthropicError
500: The request failed on the device: its one retry failed too, or it failed after its stream had sent output. A request whose client has received nothing is sent again (once, from the start): not streamed, until its response is written; streamed, until its first output is sent. A stream that has sent output is never retried by the server, since a second run could repeat what the client already has; it ends with the route’s error event instead, carrying the response id as request_id. application/json AnthropicError
529: The model is still loading, or its worker is starting again after a fault, or the server’s queue is full (256 waiting in all): overloaded_error. Retry after the Retry-After header’s seconds (x-should-retry: true while not ready). application/json AnthropicError

POST /v1/messages/count_tokens

Prompt tokens a Messages request would use, after the chat template.

Request body (JSON):

Field	Type	Description
`model`	string	Accepted for compatibility; the server answers with the one model it runs.
`messages` (required)	array of object
`system`	string or array	As in MessagesRequest.
`tools`	array of object
`tool_choice`	object	As in MessagesRequest.
`thinking`	object	As in MessagesRequest.

curl -s http://127.0.0.1:8000/v1/messages/count_tokens -H 'Content-Type: application/json' \
  -d '{"model": "qwen3-8-27b", "messages": [{"role": "user", "content": "Name three planets."}]}'

Responses:

200: OK application/json TokenCount
400: Bad request. application/json AnthropicError
529: The model is still loading, or its worker is starting again after a fault, or the server’s queue is full (256 waiting in all): overloaded_error. Retry after the Retry-After header’s seconds (x-should-retry: true while not ready). application/json AnthropicError

POST /v1/responses

OpenAI Responses with function, custom and namespace tools, reasoning and streaming, for Codex.

Request body (JSON):

Field	Type	Default	Description
`model`	string		Accepted for compatibility; the server answers with the one model it runs.
`input` (required)	string or array		A string, or items: message (system, developer, user, assistant), function_call, function_call_output, custom_tool_call, custom_tool_call_output, local_shell_call, local_shell_call_output, reasoning. Text parts only: image, audio and file parts are refused with 400 (text-only model).
`instructions`	string		System prompt.
`tools`	array of object
`tool_choice`	string or object		auto (default), none, required, or {“type”: “function”, “function”: {“name”: …}}.
`parallel_tool_calls`	boolean	false	Echoed in the response.
`reasoning`	object		Fields: `effort`.
`stream`	boolean	false	Stream the answer as server-sent events (see the 200 text/event-stream response). A client that reads nothing from the stream for 30 s is disconnected and its answer stops.
`prompt_cache_key`	string		Tags the request’s session in the server log (as the X-Session-Id header does).
`store`	boolean		Accepted and ignored.
`client_metadata`	object		Accepted and ignored.
`include`	array		Accepted and ignored, except an entry naming logprobs: 400 (log probabilities are not supported).
`top_logprobs`	integer		Not supported (log probabilities): a request that sets it gets 400 naming the field.
`text`	object		Only format {“type”: “text”}. Not supported: other formats (structured output), answered with 400.
`max_output_tokens`	integer		Most new tokens; default: up to the end of the context (–ctx).
`max_tokens`	integer		Most new tokens (>= 1).
`max_completion_tokens`	integer		Same as max_tokens (OpenAI’s newer name); wins when both are set.
`temperature`	number	0.0	0 = greedy (default). Up to 4.
`top_p`	number	1.0	Nucleus sampling, in (0, 1].
`top_k`	integer	0	Keep the k most likely tokens; 0 = off.
`min_p`	number	0.0	Drop tokens below min_p times the top probability, in [0, 1).
`seed`	integer		Sampling seed; a random one when absent.
`stop`	string or array		A string or a list of non-empty strings; the text ends before the first one.
`stop_sequences`	array of string		Same as stop (Anthropic’s name), used when stop is absent.
`repetition_penalty`	number	1.0	> 0; 1 = off.
`frequency_penalty`	number	0.0	-2 to 2; 0 = off.
`presence_penalty`	number	0.0	-2 to 2; 0 = off.
`n`	integer		Must be 1.

curl -s http://127.0.0.1:8000/v1/responses -H 'Content-Type: application/json' \
  -d '{"model": "qwen3-8-27b", "input": "Name three planets.", "max_output_tokens": 64}'

Responses:

200: OK application/json Response, text/event-stream ResponseStreamEvent
400: Bad request. application/json Error
404: A model id other than the one this server serves (GET /v1/models), code model_not_found. Leaving model out is accepted. application/json Error
429: Too many requests are waiting for a slot: 32 of yours (per tenant: keys the key file names with one tenant share it, a key without a name is its own tenant; without keys, 32 in all), or the server’s queue is full (256 waiting in all). Retry after the Retry-After header’s seconds. application/json Error
500: The request failed on the device: its one retry failed too, or it failed after its stream had sent output. A request whose client has received nothing is sent again (once, from the start): not streamed, until its response is written; streamed, until its first output is sent. A stream that has sent output is never retried by the server, since a second run could repeat what the client already has; it ends with the route’s error event instead, carrying the response id as request_id. application/json Error
503: The model is still loading. Retry after the Retry-After header’s seconds. application/json Error

GET /metrics

What the server does, in the Prometheus text format: requests running and waiting, KV cache use, the cache layout (one slot holds one whole context: block_size = the context length, num_blocks = the slots), requests answered by finish reason, prompt and generated tokens, prefix cache lookups and hits (tokens), preemptions (never), worker restarts, programs compiled after ready, and histograms of time to first token, queue time, end-to-end time and time per output token. Every name starts with mashbox:. No label names a user, an API key or a session.

Responses:

200: Prometheus text exposition format. text/plain

GET /openapi.json

The OpenAPI 3.1 description of this server, with the running model’s id.

Responses:

200: OpenAPI 3.1 document. application/json

Schemas

Health

Field	Type	Description
`status` (required)	string	ok when the model is loaded, loading before (and while its worker restarts).
`live_rows`	integer or null	Requests being answered now.
`waiting`	integer	Requests waiting for a slot.
`cache_bytes`	integer	Bytes the prompt cache holds in host memory: host_cache_bytes plus snapshot_bytes, each copy counted once.
`host_cache_bytes`	integer	Bytes of the host memory prompt cache, each copy counted once.
`snapshot_bytes`	integer	Bytes of saved slot state points that host_cache_bytes does not already count.
`host_cache_cap_bytes`	integer	Most bytes the host memory prompt cache holds (0 = off).
`host_cache_evictions`	integer	Prompt cache entries removed from host memory to stay under that cap.
`cache_namespace`	string or null	Id of what cached prompts are valid for: model, weights, tokenizer, chat template and cache policy.
`deployment_id`	string or null	Id of the compiled programs this server runs (null until set).
`component_set_sha`	string or null	Id of the model’s component set (null until set).
`certification`	string	Whether this server’s build passed the certification gate on its chip.
`compiles_after_ready`	integer or null	Programs compiled since the server reported ready, in every worker process it has run (0 when the start-up warm-up compiled every program the requests ran; null before ready or where this process does not count them).
`routed_experts`	object or null	Mixture-of-experts models: the path the routed experts run on, read where the engine runs. path: kernel (the kernel package’s routed-expert kernel at every call site), exact (the exact path) or mixed; kernel: that kernel’s id, version and package (null when the package has none); sites: the routed-op call sites traced on each path; exact_why: why any site ran exact. Null for a dense model and before the first program is traced. Fields: `path`, `kernel`, `sites`, `exact_why`.
`worker_pid`	integer	Process id of the model’s worker, when the engine runs in a worker process.
`worker_restarts`	integer	Times the worker was started again after a fault.

Model

Field	Type	Description
`id` (required)	string	Model id, as in models.json.
`object` (required)	string	model
`owned_by` (required)	string	mashbox-labs
`created` (required)	integer	Always 0.
`max_model_len` (required)	integer	Context length the server runs (–ctx).
`weights`	string	Weights file loaded.
`gb`	number	Weights in device memory, GB.
`bpw`	number	Bits per weight of the weights file.
`prefix_cache`	boolean	Whether the server reuses cached prompt starts.
`slots`	integer	Requests decoded together (–slots).
`cache_ram_mib`	integer	Host memory for saved prompt states (–cache-ram).

ModelList

Field	Type	Default	Description
`object` (required)	string		list
`data` (required)	array of Model

Error

OpenAI error shape (all routes except /v1/messages*).

Field	Type	Default	Description
`error` (required)	object		Fields: `message`, `type`.

AnthropicError

Anthropic error shape (/v1/messages and /v1/messages/count_tokens).

Field	Type	Default	Description
`type` (required)	string		error
`error` (required)	object		Fields: `type`, `message`.

ChatMessage

Field	Type	Description
`role` (required)	string	system and developer join the system prompt.
`content`	string or array or null	Text, or a list of text parts. Text parts only: image, audio and file parts are refused with 400 (text-only model).
`tool_calls`	array of object	assistant: the calls it made.
`tool_call_id`	string	tool: the call this answers.
`reasoning_content`	string	assistant: its earlier thinking.
`name`	string	Accepted and ignored.

ChatCompletion

Field	Type	Description
`id` (required)	string	chatcmpl-…
`object` (required)	string	chat.completion
`created` (required)	integer	Unix time.
`model` (required)	string	Model id.
`choices` (required)	array of object
`usage` (required)	ChatUsage
`timings`	ChatTimings

ChatTimings

Server timings, the fields of llama-server’s timings block.

Field	Type	Description
`prompt_n` (required)	integer	Prompt tokens prefilled (not served from the prefix cache).
`prompt_ms` (required)	number	Milliseconds from the start of the run to the first new token.
`prompt_per_second`	number or null	prompt_n per second of prompt_ms.
`predicted_n` (required)	integer	New tokens.
`predicted_ms` (required)	number	Milliseconds from the first to the last new token.
`predicted_per_second`	number or null	Decode speed: new tokens after the first, per second of predicted_ms.
`cache_n` (required)	integer	Prompt tokens served from the prefix cache.
`queue_ms` (required)	number	Milliseconds the request waited before it ran.

ChatUsage

Field	Type	Description
`prompt_tokens` (required)	integer	Prompt tokens.
`completion_tokens` (required)	integer	New tokens.
`total_tokens` (required)	integer	Sum.
`prompt_tokens_details`	object	Fields: `cached_tokens`.

ChatCompletionChunk

One data: line. The stream ends with data: [DONE]. With stream_options.include_usage the last chunk has empty choices and usage. The chunk with the finish_reason, and the usage chunk, carry timings.

Field	Type	Description
`id` (required)	string	Same id on every chunk.
`object` (required)	string	chat.completion.chunk
`created` (required)	integer	Unix time.
`model` (required)	string	Model id.
`choices` (required)	array of object
`usage`	ChatUsage
`timings`	ChatTimings

Completion

Whole answer, or one data: line of a stream (no usage); the stream ends with data: [DONE].

Field	Type	Description
`id` (required)	string	cmpl-…
`object` (required)	string	text_completion
`created` (required)	integer	Unix time.
`model` (required)	string	Model id.
`choices` (required)	array of object
`usage`	object	Fields: `prompt_tokens`, `completion_tokens`, `total_tokens`.
`system_fingerprint`	string	seed-.

Message

Field	Type	Description
`id` (required)	string	msg_…
`type` (required)	string	message
`role` (required)	string	assistant
`model` (required)	string	The model field of the request, else the model id.
`content` (required)	array of ContentBlock
`stop_reason` (required)	string or null	end_turn, max_tokens, stop_sequence or tool_use; null in message_start.
`stop_sequence` (required)	string or null	The stop sequence that ended the text.
`usage` (required)	MessageUsage

ContentBlock

Field	Type	Description
`type` (required)	string	text, thinking or tool_use.
`text`	string	text block.
`thinking`	string	thinking block.
`signature`	string	thinking block: mashbox-local.
`id`	string	tool_use: toolu_…
`name`	string	tool_use: tool name.
`input`	object	tool_use: arguments.

MessageUsage

Field	Type	Description
`input_tokens`	integer	Prompt tokens not served from the prefix cache.
`cache_read_input_tokens`	integer	Prompt tokens served from the prefix cache.
`cache_creation_input_tokens`	integer	Always 0.
`output_tokens`	integer	New tokens.

MessageStreamEvent

One event: + data: pair: message_start, ping, then per block content_block_start / _delta / _stop, then message_delta and message_stop.

Field	Type	Description
`type` (required)	string	Same as the SSE event name.
`message`	Message
`index`	integer	Content block number.
`content_block`	ContentBlock
`delta`	object	text_delta, thinking_delta, signature_delta, input_json_delta; on message_delta: stop_reason and stop_sequence.
`usage`	MessageUsage
`error`	object	On error: type and message.

TokenCount

Field	Type	Default	Description
`input_tokens` (required)	integer		Prompt tokens after the chat template.

Response

Field	Type	Description
`id` (required)	string	resp_…
`object` (required)	string	response
`created_at` (required)	integer	Unix time.
`status` (required)	string	completed, or incomplete when max_output_tokens ran out.
`model` (required)	string	The model field of the request, else the model id.
`output` (required)	array of ResponseItem
`parallel_tool_calls`	boolean	From the request.
`tool_choice`	string or object	From the request.
`tools`	array	From the request.
`usage` (required)	object or null	input_tokens, input_tokens_details.cached_tokens, output_tokens, output_tokens_details.reasoning_tokens, total_tokens; null while in progress.
`incomplete_details`	object	Fields: `reason`.

ResponseItem

Field	Type	Description
`type` (required)	string	message, reasoning, function_call or custom_tool_call.
`id` (required)	string	Item id.
`status`	string	in_progress or completed.
`role`	string	message: assistant.
`content`	array	message: output_text parts; reasoning: reasoning_text parts.
`summary`	array	reasoning: empty.
`call_id`	string	Call id to answer with an *_output item.
`name`	string	Tool name.
`arguments`	string	function_call: arguments, JSON.
`input`	string	custom_tool_call: raw input.
`namespace`	string	function_call: the namespace tool it came from.

ResponseStreamEvent

One event: + data: pair. An error after the stream started is sent as event error, then response.failed (the response with status failed and its error).

Field	Type	Description
`type` (required)	string	Same as the SSE event name.
`sequence_number` (required)	integer	1, 2, 3, …
`response`	Response
`item`	ResponseItem
`item_id`	string	Item the event belongs to.
`output_index`	integer	Item number.
`content_index`	integer	Part number.
`delta`	string	New text or arguments.
`text`	string	Whole text (done events).
`part`	object	Content part.
`arguments`	string	Whole arguments (done event).
`input`	string	Whole input (done event).