@aicrew/openclaw-otel-patch

Patches for OpenClaw's diagnostics-otel plugin to fix telemetry export and add GenAI semantic convention spans.

What It Fixes

1. Bundler Listener Isolation (Critical)

The tsdown bundler duplicates diagnostic-events.ts into 5+ separate JS chunks, each with its own listeners Set. The gateway emits events into one copy, but the OTEL plugin subscribes via a different copy — so events never reach the plugin.

Fix: Share listeners via globalThis.__openclaw_diag_listeners.

Related: openclaw/openclaw#5190, PR #11168

2. GenAI Semantic Convention Spans

Adds gen_ai.completion trace spans following the OpenTelemetry GenAI Semantic Conventions, including:

• gen_ai.system, gen_ai.response.model, gen_ai.request.model
• gen_ai.usage.input_tokens, gen_ai.usage.output_tokens
• gen_ai.usage.cache_read_tokens, gen_ai.usage.cache_write_tokens
• openclaw.cost.usd

Related: PR #11100

Compatibility

• Base: OpenClaw v2026.2.18 (fa4772b4ce)
• Node.js: 22+

Usage

# Install globally
npm install -g @aicrew/openclaw-otel-patch

# Apply to your OpenClaw installation
openclaw-otel-patch ~/openclaw

# Hard restart (required — SIGUSR1 won't reinitialize OTEL SDK)
openclaw gateway stop && openclaw gateway start

Required Config

Add to ~/.openclaw/openclaw.json:

{
  "diagnostics": {
    "enabled": true,
    "otel": {
      "enabled": true,
      "endpoint": "https://your-otlp-endpoint",
      "protocol": "http/protobuf",
      "serviceName": "your-service-name",
      "traces": true,
      "metrics": true,
      "logs": true,
      "flushIntervalMs": 5000
    }
  },
  "models": {
    "providers": {
      "anthropic": {
        "models": [
          { "id": "claude-opus-4-6", "cost": { "input": 15, "output": 75, "cacheRead": 1.5, "cacheWrite": 18.75 } },
          { "id": "claude-sonnet-4-6", "cost": { "input": 3, "output": 15, "cacheRead": 0.3, "cacheWrite": 3.75 } },
          { "id": "claude-haiku-3-5", "cost": { "input": 0.8, "output": 4, "cacheRead": 0.08, "cacheWrite": 1 } }
        ]
      }
    }
  }
}

Metrics Exported

| Metric | Description | |--------|-------------| | openclaw.tokens | Token counts by type (input/output/cache_read/cache_write/total) | | openclaw.cost.usd | Estimated USD cost per model call | | openclaw.run.duration_ms | Agent run duration | | openclaw.context.tokens | Context window usage | | openclaw.message.queued | Messages queued | | openclaw.message.processed | Messages processed by outcome | | openclaw.queue.depth | Queue depth | | openclaw.queue.lane.enqueue | Lane enqueue events | | openclaw.session.state | Session state transitions |

Traces Exported

| Span | Description | |------|-------------| | gen_ai.completion | GenAI semantic convention span with token/cost attributes | | openclaw.model.usage | Legacy span with session/token details | | openclaw.webhook.processed | Webhook processing spans | | openclaw.message.processed | Message processing spans |

⚠️ Notes

• Dist patches are fragile — re-run after any pnpm build or OpenClaw update
• The resource attributes (agent ID, org ID) are hardcoded — edit service.ts for your own values
• Once PR #11168 merges, the dist patch becomes unnecessary
• Once PR #11100 merges, the gen_ai spans will be built-in with full tool tracing

License

MIT