Summary

After enabling diagnostics-otel and seeing Anthropic + openai-codex metrics flowing correctly, no metrics appear for models on the openai provider (e.g. openai/gpt-5.4).

Code investigation

The emitDiagnosticEvent instrumentation in the source is provider-agnostic — it fires based on providerUsed/modelUsed dynamically. The diagnostics-otel plugin registers its listener on globalThis.__openclawDiagnosticEventsState.

Hypothesis: Models invoked via sessions_spawn subagents may run in a subprocess that does not share the globalThis listener store, causing emitDiagnosticEvent calls to be dispatched to an empty listener set in the subprocess. This would explain why openai-codex metrics appear (from cron isolated sessions, which likely run embedded in the gateway process), but openai provider metrics do not (only tested via sessions_spawn).

A secondary possibility: the usage object from the OpenAI Chat Completions API (prompt_tokens, completion_tokens, total_tokens) may not be normalized before the hasNonzeroUsage check, causing the guard to return false and skip event emission.

To reproduce

1. Enable diagnostics-otel with a working OTLP endpoint
2. Run sessions_spawn with model: openai/gpt-5.4 (or any openai provider model)
3. Confirm the model responds successfully
4. Query Prometheus — no openclaw_provider="openai" series appear

Expected

All provider backends emit openclaw_tokens_total and related metrics regardless of invocation path.

Environment

• diagnostics-otel plugin enabled, OTLP/HTTP configured
• openai:default auth profile configured, calls succeed
• openai-codex metrics appear correctly from cron runs