Shapes ReqLLM request and response payloads into the GenAI content attributes — gen_ai.input.messages, gen_ai.system_instructions, gen_ai.tool.definitions, gen_ai.output.messages.

Used by both ReqLLM.OpenTelemetry and ReqLLM.Telemetry.OpenTelemetry when content capture is opted in (content: :attributes or :event — default is :none). The caller must also have raw payload telemetry on (telemetry: [payloads: :raw]) for the request payload to be available.

Span attribute content is JSON-encoded before being returned, so attribute values are [String.t()] (an array of primitives — what the OpenTelemetry attribute spec accepts and what GenAI-aware backends like Langfuse, Honeycomb, and Grafana decode for display). Event content remains structured for the gen_ai.client.inference.operation.details event. request_attributes/1 returns roughly:

%{
  "gen_ai.system_instructions" => [
    ~s({"type":"text","content":"You are a helpful assistant."})
  ],
  "gen_ai.input.messages" => [
    ~s({"role":"user","parts":[{"type":"text","content":"Hi"}]})
  ],
  "gen_ai.tool.definitions" => [
    ~s({"type":"function","name":"get_weather","parameters":{...}})
  ]
}

Reasoning text is excluded in every mode — ReqLLM's payload sanitizer redacts it before this module sees the messages, and content parts are filtered to spec-friendly types (text, image, image_url, video_url, file) so thinking output cannot leak through.