feat(openai-agents): capture tool call arguments + result on tool spans

[hansmire]

Summary

FunctionSpanData carries the tool's arguments (input) and return value (output) on every function-tool invocation, but the openai-agents instrumentor never read either — tool spans only carried the tool name and type. Braintrust's own native Agents-SDK processor (braintrust/wrappers/openai.py:_function_log_data) logs both as top-level input/output fields; OTel-based integrations were the outlier.

This PR captures both values and emits them under two attribute pairs so they show up both as semantic-convention metadata and as first-class Input/Output panels in trace UIs.

Before / After

Same get_customer_info.tool span, same Braintrust workspace, rendered side-by-side:

Before: Output panel renders the tool's JSON schema ({"tool": {"name": ..., "strict_json_schema": true, "type": "function"}}) — a tool definition leaked into the output slot, not the tool's return value.
After: Output panel renders the real tool call result (hasChatEmailed: false, mailingAddress: "123 Main St, …", partyId: "risk_1a", …).

Changes (_hooks.py)

1. Capture FunctionSpanData.input / .output in on_span_end and emit them as:
   • gen_ai.tool.call.arguments / gen_ai.tool.call.result — OTel GenAI semconv (for OTel-aware backends).
   • traceloop.entity.input / traceloop.entity.output — Traceloop convention used by the LangChain instrumentor, which trace backends (Braintrust etc.) already map to the span's top-level Input / Output panels.
2. Drop the three gen_ai.completion.tool.* attributes previously set on tool spans at start time. They were a tool definition blob (name / type / strict_json_schema) that some backends' ingestion logic treated as the tool span's Output, producing {"tool": {...}} in the Output panel instead of the actual return value. The equivalent information is already carried by gen_ai.tool.name / gen_ai.tool.type + TRACELOOP_SPAN_KIND.
3. Add _serialize_span_payload helper — serialises the FunctionSpanData input/output reliably:
   • string passthrough,
   • model_dump() for pydantic models (e.g. Agents SDK ToolOutputText) and for lists of pydantic models,
   • json.dumps(..., default=str) fallback,
   • str() as last resort.
4. Gated on trace_content = should_send_prompts() so content-sensitive deployments that opt out of prompt capture also get the tool payload suppressed.

Tests

Extended the existing VCR-backed test_agent_with_function_tool_spans to assert:

• GEN_AI_TOOL_CALL_ARGUMENTS / GEN_AI_TOOL_CALL_RESULT are set and non-empty.
• TRACELOOP_ENTITY_INPUT / TRACELOOP_ENTITY_OUTPUT are set and match the GenAI pair.
• Tool arguments reference the query's city name (end-to-end sanity).

All 10 tests in tests/test_openai_agents.py pass locally. uv run ruff check clean.

Notes

Part of a small series of openai-agents parity fixes (#4061 cached_tokens + reasoning_tokens, #4062 tool span type + duration). Each stands alone off main and can be merged in any order.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: fc8df7b5-bbfe-4f10-872c-2ec2a0d29b66

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

Max Hansmire seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}