fix(openai-agents): capture response.instructions as system prompt in generation spans

[dvirski]

Fixes #3738.

Problem:
When an OpenAI Agent had instructions set, the system prompt was silently dropped from generation spans — only the conversation history was recorded, with no role: system message.

Fix:
Prepend response.instructions as a role: system message to the input messages array in _end_generation_span() before passing to _extract_prompt_attributes(). Consistent with how responses_wrappers.py handles it in the plain openai instrumentation.

Summary by CodeRabbit

• Bug Fixes

  • Enhanced tracing to properly capture agent instructions in trace data when trace content is enabled.

• Tests

  • Updated test assertions to verify correct message ordering in traced content.
  • Added new tests for instruction handling in trace generation.

[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.

---

Dvir Rezenman 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.}

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: 73aa64ca-c1fc-4711-bb06-944986884d48

📥 Commits

Reviewing files that changed from the base of the PR and between 6d3e696 and c2d4b1d.

⛔ Files ignored due to path filters (2)

• packages/opentelemetry-instrumentation-openai-agents/uv.lock is excluded by !**/*.lock
• packages/sample-app/uv.lock is excluded by !**/*.lock

📒 Files selected for processing (3)

• packages/opentelemetry-instrumentation-openai-agents/opentelemetry/instrumentation/openai_agents/_hooks.py
• packages/opentelemetry-instrumentation-openai-agents/tests/test_openai_agents.py
• packages/opentelemetry-instrumentation-openai-agents/tests/test_tracing_processor.py

---

📝 Walkthrough

Walkthrough

The PR fixes a bug where OpenAI Agents instrumentation did not capture response.instructions as a system message in traced prompts. The implementation prepends instructions as the first system message when content tracing is enabled, with corresponding unit and integration tests validating the behavior under both tracing conditions.

Changes

Capture response.instructions as system message

Layer / File(s)	Summary
System message injection in _end_generation_span opentelemetry/instrumentation/openai_agents/_hooks.py	_end_generation_span now extracts response early and conditionally prepends response.instructions as a system message to input_data when trace_content is enabled, then passes the modified list to _extract_prompt_attributes.
Unit tests for response.instructions handling tests/test_tracing_processor.py	Two new test methods validate that response.instructions is inserted as the first system message in GEN_AI_INPUT_MESSAGES when trace_content=True, and omitted when trace_content=False.
Integration test assertions for message ordering tests/test_openai_agents.py	Updated test_agent_spans to locate the system message at index 0 and find the user message by role, ensuring the new system message injection is reflected in end-to-end test assertions.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

Poem

🐰 A system message hops into place,
First in the trace, taking its space,
Instructions now heard, no longer lost,
Agent wisdom captured—the bug is crossed! 🌟

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The pull request title clearly and accurately summarizes the main change: capturing response.instructions as a system prompt in generation spans for OpenAI Agents instrumentation.
Linked Issues check	✅ Passed	The code changes fully implement the fix specified in issue #3738: prepending response.instructions as a system message to input_data before calling _extract_prompt_attributes when trace_content is enabled.
Out of Scope Changes check	✅ Passed	All changes are directly related to fixing issue #3738. The modifications to _hooks.py, test_openai_agents.py, and test_tracing_processor.py are all scoped to implementing the response.instructions capture feature.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch dr/fix(openai-agents)-capture-response.instructions-as-system-prompt-in-generation-spans

---

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.}

[doronkopit5]

Overall looks good — clean, targeted fix that mirrors the pattern in responses_wrappers.py. Two small suggestions before merge:

1. Defensive normalization of input_data — at _hooks.py:946, if input_data is ever a str (the Responses API allows input: str), list + str will TypeError. Today the agents SDK passes a list, but a cheap guard avoids a future footgun:

   existing = input_data if isinstance(input_data, list) else []
   input_data = [{"role": "system", "content": response.instructions}] + existing

   This also keeps the line under the lint length limit.

2. Two tiny edge-case tests worth adding alongside the existing ones:

   • input_data == [] with only instructions → asserts single system-message output.
   • instructions == "" → asserts the empty/falsy case is skipped (locks in the current getattr(..., None) truthy check).

Otherwise LGTM 👍