Troubleshooting OpenClaw Phoenix Otel Configuration to Populate All Trace Columns

Muni · 2026-04-04T00:26:26.301Z

Hello, I have installed phoenix on my openclaw instance and have been implementing tracing using otel. This is my current setup with phoenix running locally and diagnostics.otel configuration. As you can see, I am able to capture traces however, I am unable to fill out all the columns(input, output, tool calls, etc.). Can someone guide me as to what I am missing in my otel config to populate everything. Thanks!

Troubleshooting OpenClaw Phoenix Otel Configuration to Populate All Trace Columns | Arize AI Community

4 comments

· Sorted by Oldest

RunLLM
APP
·
·
To populate columns like input, output, and tool calls in Phoenix, you need to use OpenInference semantic conventions and/or OpenInference instrumentors, not just generic OpenTelemetry tracing. Standard OTEL spans alone do not automatically capture LLM-specific attributes (inputs, outputs, tool calls, etc.) required by Phoenix’s UI. You should either use OpenInference instrumentor packages for your frameworks (e.g., OpenAI, LangChain, etc.) or manually set semantic attributes on your spans according to the OpenInference conventions <https://docs.arize.com/phoenix/tracing/how-to-tracing/manual-instrumentation/custom-spans|[Phoenix manual instrumentation guide>, <https://arize.com/docs/phoenix/sdk-api-reference#environment-variables|OpenInference docs>]. For Python, install the relevant OpenInference instrumentor (e.g., `pip install openinference-instrumentation-openai`) and use `phoenix.otel.register(auto_instrument=True)` to auto-instrument supported libraries, or use decorators/context managers like `@tracer.chain` and `span.set_input()/set_output()` for manual spans. For each span, ensure you set attributes such as `input.value`, `output.value`, and tool-specific attributes as described in the <https://docs.arize.com/phoenix/tracing/how-to-tracing/manual-instrumentation/custom-spans|manual instrumentation docs>. Without these, Phoenix cannot display the detailed columns in its UI. Would you like a code example for your specific framework or more detail on manual attribute setting? Sources: - <https://arize.com/docs/phoenix/sdk-api-reference#environment-variables|Phoenix SDK/API reference>
💡 Hint: Mention RunLLM in the thread for followups.
Muni
·
·
How would I set auto-instrument=true for openclaw? RunLLM
JR (.
·
·
Hi Muni — I'm new to Phoenix and looking for some research challenges (I don't use OpenClaw). I deleted my last response because it needed some corrections. . After taking a look at the OpenClaw repo, here is what I suspect is happening: . Two questions first:
How did you connect Phoenix? Are you using @arizeai/phoenix-otel, the diagnostics-otel plugin, or something else? (I'm assuming you used diagnostics-otel)
Click into one of those spans — what do you see in the Attributes tab?
. The problem: Looking at service.ts on main, the diagnostics-otel plugin doesn't emit the data Phoenix needs. . Two reasons:
1.
No content capture. The plugin does not include prompt or response text in traces. There should be a captureContent option but it doesn't exist on main yet. Input/output will always show --.
2.
Wrong semantic conventions. The plugin emits openclaw.* attributes. Phoenix expects OpenInference (input.value, output.value, openinference.span.kind). That should be why kind shows "unknown" and cost reads $0.
. PR #21290 addresses both — adds captureContent and aligns with GenAI semantic conventions — but even after it merges, you'd still need a translation step to get from gen_ai.* to OpenInference. Arize's @arizeai/openinference-genai package handles that conversion at export time. Phoenix docs here. . Mitigation . I did notice a community plugin for Phoenix. It doesn't appear to be managed by Arize - for that reason I'd rather not include links. . I hope that this provides a good starting point 🙂
JR (.
·
·
After some digging, here is why I suspect that cost shows $0. . It's not that cost data is missing. It's being exported through a different OTel signal. The plugin (diagnostics-otel) does capture cost.
When a model invocation completes, service.ts calls costCounter.add(evt.costUsd, attrs), exporting it as an OTel counter metric. But Phoenix reads span attributes, not metrics .
In recordModelUsage, only token counts are written to the span (openclaw.tokens.*), not cost.
So Phoenix queries the span, finds nothing, and reports $0.
👍1

RunLLM
APP
·
·
To populate columns like input, output, and tool calls in Phoenix, you need to use OpenInference semantic conventions and/or OpenInference instrumentors, not just generic OpenTelemetry tracing. Standard OTEL spans alone do not automatically capture LLM-specific attributes (inputs, outputs, tool calls, etc.) required by Phoenix’s UI. You should either use OpenInference instrumentor packages for your frameworks (e.g., OpenAI, LangChain, etc.) or manually set semantic attributes on your spans according to the OpenInference conventions <https://docs.arize.com/phoenix/tracing/how-to-tracing/manual-instrumentation/custom-spans|[Phoenix manual instrumentation guide>, <https://arize.com/docs/phoenix/sdk-api-reference#environment-variables|OpenInference docs>]. For Python, install the relevant OpenInference instrumentor (e.g., `pip install openinference-instrumentation-openai`) and use `phoenix.otel.register(auto_instrument=True)` to auto-instrument supported libraries, or use decorators/context managers like `@tracer.chain` and `span.set_input()/set_output()` for manual spans. For each span, ensure you set attributes such as `input.value`, `output.value`, and tool-specific attributes as described in the <https://docs.arize.com/phoenix/tracing/how-to-tracing/manual-instrumentation/custom-spans|manual instrumentation docs>. Without these, Phoenix cannot display the detailed columns in its UI. Would you like a code example for your specific framework or more detail on manual attribute setting? Sources: - <https://arize.com/docs/phoenix/sdk-api-reference#environment-variables|Phoenix SDK/API reference>
💡 Hint: Mention RunLLM in the thread for followups.
Muni
·
·
How would I set auto-instrument=true for openclaw? RunLLM
JR (.
·
·
Hi Muni — I'm new to Phoenix and looking for some research challenges (I don't use OpenClaw). I deleted my last response because it needed some corrections. . After taking a look at the OpenClaw repo, here is what I suspect is happening: . Two questions first:
How did you connect Phoenix? Are you using @arizeai/phoenix-otel, the diagnostics-otel plugin, or something else? (I'm assuming you used diagnostics-otel)
Click into one of those spans — what do you see in the Attributes tab?
. The problem: Looking at service.ts on main, the diagnostics-otel plugin doesn't emit the data Phoenix needs. . Two reasons:
1.
No content capture. The plugin does not include prompt or response text in traces. There should be a captureContent option but it doesn't exist on main yet. Input/output will always show --.
2.
Wrong semantic conventions. The plugin emits openclaw.* attributes. Phoenix expects OpenInference (input.value, output.value, openinference.span.kind). That should be why kind shows "unknown" and cost reads $0.
. PR #21290 addresses both — adds captureContent and aligns with GenAI semantic conventions — but even after it merges, you'd still need a translation step to get from gen_ai.* to OpenInference. Arize's @arizeai/openinference-genai package handles that conversion at export time. Phoenix docs here. . Mitigation . I did notice a community plugin for Phoenix. It doesn't appear to be managed by Arize - for that reason I'd rather not include links. . I hope that this provides a good starting point 🙂
JR (.
·
·
After some digging, here is why I suspect that cost shows $0. . It's not that cost data is missing. It's being exported through a different OTel signal. The plugin (diagnostics-otel) does capture cost.
When a model invocation completes, service.ts calls costCounter.add(evt.costUsd, attrs), exporting it as an OTel counter metric. But Phoenix reads span attributes, not metrics .
In recordModelUsage, only token counts are written to the span (openclaw.tokens.*), not cost.
So Phoenix queries the span, finds nothing, and reports $0.
👍1