Velanir LLM Usage Reporter

OpenClaw plugin that reports per-call model usage from the OpenClaw llm_output hook to the Velanir runtime observability API.

The plugin requires Node >=22.19.0 and OpenClaw 2026.6.6 or newer. That OpenClaw version documents conversation-observation hooks for model_call_started, model_call_ended, llm_input, and llm_output; the llm_output event includes normalized usage counters.

Runtime Contract

The plugin posts batches to:

POST /v1/runtime/observability/llm-usage

Authentication uses the existing runtime-v2 DPoP flow with scope observability:write. The platform derives organization, coworker, runtime identity, gateway, and deployment target from the runtime token; the plugin does not send or own those identifiers.

Privacy

By default the plugin does not send prompts, assistant text, or raw message content. It sends token counts plus non-content metadata such as provider, model, run id, session id, surface, trigger, and model-call timing.

Set metadata.includeContent: true only for local debugging.

Config

{
  "plugins": {
    "entries": {
      "velanir-llm-usage-reporter": {
        "enabled": true,
        "config": {
          "platform": {
            "baseUrl": "${OCT8_API_URL}",
            "endpointPath": "/v1/runtime/observability/llm-usage"
          },
          "batch": {
            "maxBatchSize": 20,
            "maxQueueSize": 2000,
            "flushIntervalMs": 5000,
            "requestTimeoutMs": 10000
          }
        }
      }
    }
  }
}

platform.baseUrl defaults to OCT8_API_URL, which is already part of the runtime-v2 environment. The reporter fails open: ingest failures are logged and queued for retry, but they do not block model execution or user-visible delivery.