Nightly OSS: Responses API: undocumented reasoning+message pairing constraint breaks multi-turn conversations

.codex-nightly-retry-1.md

@@ -0,0 +1,13 @@
+The previous attempt produced no source, test, docs, or example changes.
+
+Continue working. A successful nightly slot must produce a small, honest,
+maintainer-reviewable PR unless the repository is technically impossible
+to patch for this target.
+
+Requirements for this retry:
+- Search the local repository for a directly related patch surface.
+- Prefer a focused test, docs/example correction, typing/schema fix, CLI behavior fix, or small error-handling improvement.
+- If the exact upstream symptom is server-side, produce a repository-local improvement that prevents user confusion or locks existing behavior without falsely claiming to fix the server.
+- Do not stop with only NIGHTLY_REPORT.md.
+- Modify at least one tracked source, test, docs, or example file if a defensible PR exists.
+- Keep the diff small and explain the rationale in NIGHTLY_REPORT.md.

NIGHTLY_CODEX_FINAL_ATTEMPT_1.md

@@ -0,0 +1,5 @@
+Wrote [NIGHTLY_REPORT.md](/home/runner/work/oss-nightly-control/oss-nightly-control/target/NIGHTLY_REPORT.md) and did not change repository code.
+
+I stopped because the issue already has two active upstream PRs covering the maintainable SDK-side fixes: `#3013` for helper/validation behavior and `#3016` for documentation. A new code/doc PR from this checkout would likely duplicate ongoing review work.
+
+Tests were not run because no code changed. Current untracked files are `.codex-nightly-prompt.md`, `uv.lock`, and `NIGHTLY_REPORT.md`; I only added the report.

NIGHTLY_CODEX_FINAL_ATTEMPT_2.md

@@ -0,0 +1,11 @@
+Implemented a small, defensible docs/example patch for `openai-python#3009`.
+
+Changed:
+- [README.md](/home/runner/work/oss-nightly-control/oss-nightly-control/target/README.md:47): added a `Conversation state` section showing `previous_response_id` and warning not to filter manual histories down to only messages.
+- [examples/responses/conversation_state.py](/home/runner/work/oss-nightly-control/oss-nightly-control/target/examples/responses/conversation_state.py:1): added a minimal two-turn Responses example.
+- [NIGHTLY_REPORT.md](/home/runner/work/oss-nightly-control/oss-nightly-control/target/NIGHTLY_REPORT.md:1): updated with rationale, tests, and risks.
+
+Verified:
+- `python -m py_compile examples/responses/conversation_state.py`
+
+I also checked the current upstream context: issue [#3009](https://github.com/openai/openai-python/issues/3009) is still open, with active related PRs [#3013](https://github.com/openai/openai-python/pull/3013) and [#3016](https://github.com/openai/openai-python/pull/3016), so this patch avoids duplicating their helper/docstring surfaces.

README.md

@@ -44,6 +44,35 @@ response = client.responses.create(
 print(response.output_text)
 ```
 
+### Conversation state
+
+For multi-turn conversations with the Responses API, use `previous_response_id`
+to have the API retain context between turns.
+
+```python
+from openai import OpenAI
+
+client = OpenAI()
+
+response = client.responses.create(
+    model="gpt-5.2",
+    input="Write a haiku about recursion in programming.",
+)
+print(response.output_text)
+
+response = client.responses.create(
+    model="gpt-5.2",
+    input="Now explain it in plain English.",
+    previous_response_id=response.id,
+)
+print(response.output_text)
+```
+
+If you manually manage conversation history instead, preserve all items from
+`response.output` in their original order. Reasoning models may return reasoning
+items together with assistant messages, and filtering those items down to only
+messages can break subsequent requests.
+
 The previous standard (supported indefinitely) for generating text is the [Chat Completions API](https://platform.openai.com/docs/api-reference/chat). You can use that API to generate text from the model with the code below.
 
 ```python

examples/responses/conversation_state.py

@@ -0,0 +1,22 @@
+from openai import OpenAI
+
+
+client = OpenAI()
+
+response = client.responses.create(
+    model="gpt-5.2",
+    input="Write a haiku about recursion in programming.",
+)
+print(response.output_text)
+
+response = client.responses.create(
+    model="gpt-5.2",
+    input="Now explain it in plain English.",
+    previous_response_id=response.id,
+)
+print(response.output_text)
+
+# If you manually manage conversation history instead of using
+# previous_response_id, append response.output items in order. Reasoning models
+# may return reasoning items together with assistant messages, and filtering
+# those items down to only messages can break the next request.