# Page through the messages a model call in a session's trace was sent, with each message's full content. Source: GET /v1/sessions/{session\_id}/trace/turns/{turn}/steps/{step}/request Select code samplecURL (shell:curl) ```js const url = 'https://app.everruns.com/api/v1/sessions/example/trace/turns/example/steps/example/request'; const options = {method: 'GET'}; try { const response = await fetch(url, options); const data = await response.json(); console.log(data); } catch (error) { console.error(error); } ``` ```sh curl --request GET \ --url https://app.everruns.com/api/v1/sessions/example/trace/turns/example/steps/example/request ``` * Production APIapp.everruns.com/api/v1/sessions/{session\_id}/trace/turns/{turn}/steps/{step}/request ## Parameters ### Path Parameters **session\_id** required string Prefixed public identifier **turn** required string Prefixed public identifier **step** required string Prefixed public identifier ### Query Parameters **role** string Only messages of this role: `system`, `user`, `assistant` or `tool`. **offset** integer format: int32 First message index. Defaults to 0. **limit** integer format: int32 Messages per page, 1 to 200. Defaults to 100. **full** boolean Return every message’s content whole, however large. **native** boolean Also return the provider request body as the driver sent it. ## Responses ### 200 Success Media typeapplication/json A page of a model request’s messages. object **has\_native\_request** required Whether the provider request body was recorded for this call. boolean **message\_count** required Messages sent in the call. integer format: int32 **messages** required Requested messages, in order. Array\ One message of a model request. object **content** The whole message, when asked for and within the limit. **index** required Position in the request. integer format: int32 **preview** required Start of the message text. string **role** required Role as handed to the provider driver: `system`, `user`, `assistant` or `tool`. Stored messages say `agent` and `tool_result`; each driver then maps these four onto its own wire format. string **size\_bytes** required Size of the serialized message. integer **model** Model that served the call. string | null **native\_request** The request body exactly as the driver sent it to the provider, with inline media and reasoning replay state replaced; only when asked for with `native`, on the first page. **new\_from** required Messages before this index were already in the previous call. integer format: int32 **provider** Provider that served the call. string | null **request\_options** Request options sent with the call (temperature, caching, …); on the first page only. **tools** required Tools offered in the call; on the first page only. Array\ A tool offered to a model call. object **description** Tool description the model sees. string | null **name** required Tool name the model sees. string **usage** One of: * object Token usage of the call, split by prompt cache. object **cache\_creation\_tokens** required Prompt tokens written to the provider’s cache. integer format: int64 **cache\_read\_tokens** required Prompt tokens read from the provider’s cache. integer format: int64 **cache\_reset** One of: * object Set when this call lost most of the prompt prefix the previous model call had cached. object **idle\_seconds** required Seconds between the end of the previous call and the start of this one. integer format: int64 **lost\_tokens** required How many of those this call did not read back. integer format: int64 **previous\_cached\_tokens** required Prompt tokens the previous call left in the cache (read plus written). integer format: int64 **reason** required Likely cause, the first difference found: `model_changed`, `tools_changed`, `system_changed`, `history_changed`, `expired` (idle longer than the cache retention), or `unknown`. string * null null **cached\_through** Estimated index of the last message inside the cached prefix: the cache-read share of the prompt laid over the messages by size. `-1` when only the tools fit; absent when nothing was read from cache. integer | null format: int32 **caching\_would\_help** A long prompt sent with caching off that neither read nor wrote the cache: turning on prompt caching would likely make calls like it cheaper. boolean **input\_tokens** required Prompt tokens neither read from nor written to the cache. integer format: int64 **output\_tokens** required Completion tokens. integer format: int64 **prompt\_cache** How prompt caching was requested, such as `auto` or `anthropic_cache_control`; absent when caching was off. string | null * null Set when this call lost most of the prompt prefix the previous model call had cached. object **idle\_seconds** required Seconds between the end of the previous call and the start of this one. integer format: int64 **lost\_tokens** required How many of those this call did not read back. integer format: int64 **previous\_cached\_tokens** required Prompt tokens the previous call left in the cache (read plus written). integer format: int64 **reason** required Likely cause, the first difference found: `model_changed`, `tools_changed`, `system_changed`, `history_changed`, `expired` (idle longer than the cache retention), or `unknown`. string * object null * null null ##### Example ```json { "has_native_request": true, "message_count": 14, "messages": [ { "index": 12, "role": "tool", "size_bytes": 244 } ], "new_from": 12, "provider": "openrouter", "tools": [ { "description": "Read a file from the workspace.", "name": "read_file" } ], "usage": { "cache_creation_tokens": 0, "cache_read_tokens": 12529, "cache_reset": { "idle_seconds": 42, "lost_tokens": 18420, "previous_cached_tokens": 18420, "reason": "tools_changed" }, "cached_through": 11, "input_tokens": 10295, "output_tokens": 197, "prompt_cache": "auto" } } ```