Pydantic AI v3+ tool spans show blank input/output in Conversation: map gen_ai.tool.call.arguments/result

Bug Description

Summary

• Pydantic AI tool spans are successfully ingested by PostHog, but their input and output are blank in the Conversation tab.
• The span’s Raw tab contains the current OpenTelemetry GenAI tool attributes:
  • gen_ai.tool.call.arguments
  • gen_ai.tool.call.result

However, Conversation displays which affects tool spans only. Model-generation spans in the same trace displaytheir input and output normally.

• Input: No input
• Output: No output
  Trace link: https://eu.posthog.com/project/27465/ai-observability/traces/80ce134e484d6cdfbf8c44e25b5e960c?timestamp=2026-09-04T09%3A33%3A43Z&event=01a06bc9-1a86-71c2-93a8-38fb451d4deb

Expected behavior

For Pydantic AI v3+ tool spans, PostHog should normalize:

• gen_ai.tool.call.arguments → $ai_input_state
• gen_ai.tool.call.result → $ai_output_state

The existing Conversation view should then render the tool arguments and tool result.

Legacy Pydantic AI v2 fields should continue to work:

• tool_arguments
• tool_response

Actual behavior

PostHog receives the dotted Pydantic AI / OpenTelemetry fields and leaves them visible in Raw, but does not create $ai_input_state or $ai_output_state.

The Conversation view reads those normalized state properties for $ai_span events, so it shows No input and No output.

Why this appears to be a compatibility gap

Pydantic AI defaults to instrumentation format v5. v3+ emits the current
OpenTelemetry-style names:

• gen_ai.tool.call.arguments
• gen_ai.tool.call.result

The Pydantic AI OTel middleware currently handles only the older flat
Pydantic v2 keys:

• tool_arguments
• tool_response

Relevant PostHog source:

• https://github.com/PostHog/posthog/blob/master/nodejs/src/ingestion/pipelines/ai/otel/middleware/pydantic-ai.ts
• https://github.com/PostHog/posthog/blob/master/products/ai_observability/frontend/ConversationDisplay/ConversationDisplay.tsx

How to reproduce

1. Configure Pydantic AI instrumentation with a PostHog OpenTelemetry exporter. Do not set InstrumentationSettings.version, so Pydantic AI uses its default instrumentation format v5.
2. Run an agent that calls a function tool with content capture enabled.
3. Open the resulting trace in PostHog AI Observability.
4. Select the tool span.
5. In the Raw tab, observe both:
   • gen_ai.tool.call.arguments
   • gen_ai.tool.call.result
6. Switch to the Conversation tab.
7. Observe that the span shows:
   • Input: No input
   • Output: No output

Additional context

The raw attributes prove this is not an exporter or include_content problem:
the tool arguments and result have already reached PostHog.

Suggested mapper behavior:

const rawArgs =
    props['gen_ai.tool.call.arguments'] ?? props['tool_arguments']

const rawResult =
    props['gen_ai.tool.call.result'] ?? props['tool_response']

PostHog should preserve its existing JSON parsing and write the resolved values to $ai_input_state and $ai_output_state.

Suggested regression coverage:

1. Legacy Pydantic AI v2:
   tool_arguments / tool_response
2. Current Pydantic AI v3+ and default v5:
   gen_ai.tool.call.arguments / gen_ai.tool.call.result
3. JSON-string arguments, JSON-string results, and plain-text results.

I did not find an existing public PostHog issue for this exact tool-span mapping problem.

Debug info

- [x] PostHog Cloud, region and project ID: EU, project ID: <fill in your project ID>
- [ ] PostHog Hobby self-hosted with `docker compose`, version/commit: N/A
- [ ] You’re self-hosted with Kubernetes: N/A

Client environment:
- posthog-python: 7.45.1
- pydantic-ai: 2.36.0
- opentelemetry-sdk: 1.44.0
- Pydantic AI instrumentation format: v5 default
- Event transport: OpenTelemetry to PostHog AI endpoint
- Tool arguments/results are visible in the PostHog Raw tab

Start in nodejs/src/ingestion/pipelines/ai/otel/middleware/pydantic-ai.ts and inspect how tool_arguments and tool_response are mapped into normalized AI state properties. Add regression coverage for legacy and gen_ai.tool.call.arguments/result attributes, including JSON-string and plain-text values, then run the relevant middleware tests and confirm Conversation displays the tool input and output.