fix(auxiliary): route Codex Responses path through shared converter (#5709)

The auxiliary Codex adapter maintained its own chat->Responses conversion
loop that forwarded every non-system message's role verbatim into
Responses input[]. When flush_memories()/compression replayed session
history containing assistant tool_calls + role=tool results, those tool
messages leaked into the request and the Responses API rejected them with
HTTP 400: Invalid value: 'tool'.

Route _CodexCompletionsAdapter.create() through the same shared converter
the main agent transport uses (_chat_messages_to_responses_input), so tool
calls become function_call items and tool results become function_call_output
items with a valid call_id. Single conversion path means no future drift.

Also remove the now-dead _convert_content_for_responses() helper — its only
caller was the private conversion loop this change deletes.

Co-authored-by: ProgramCaiCai <techxacm@gmail.com>
This commit is contained in:
Teknium 2026-06-07 19:48:28 -07:00
parent 568e127612
commit 2789bf4e25
3 changed files with 124 additions and 57 deletions

View file

@ -637,54 +637,6 @@ def _pool_runtime_base_url(entry: Any, fallback: str = "") -> str:
# calls to the Codex Responses API so callers don't need any changes.
def _convert_content_for_responses(content: Any) -> Any:
"""Convert chat.completions content to Responses API format.
chat.completions uses:
{"type": "text", "text": "..."}
{"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}}
Responses API uses:
{"type": "input_text", "text": "..."}
{"type": "input_image", "image_url": "data:image/png;base64,..."}
If content is a plain string, it's returned as-is (the Responses API
accepts strings directly for text-only messages).
"""
if isinstance(content, str):
return content
if not isinstance(content, list):
return str(content) if content else ""
converted: List[Dict[str, Any]] = []
for part in content:
if not isinstance(part, dict):
continue
ptype = part.get("type", "")
if ptype == "text":
converted.append({"type": "input_text", "text": part.get("text", "")})
elif ptype == "image_url":
# chat.completions nests the URL: {"image_url": {"url": "..."}}
image_data = part.get("image_url", {})
url = image_data.get("url", "") if isinstance(image_data, dict) else str(image_data)
entry: Dict[str, Any] = {"type": "input_image", "image_url": url}
# Preserve detail if specified
detail = image_data.get("detail") if isinstance(image_data, dict) else None
if detail:
entry["detail"] = detail
converted.append(entry)
elif ptype in {"input_text", "input_image"}:
# Already in Responses format — pass through
converted.append(part)
else:
# Unknown content type — try to preserve as text
text = part.get("text", "")
if text:
converted.append({"type": "input_text", "text": text})
return converted or ""
class _CodexCompletionsAdapter:
"""Drop-in shim that accepts chat.completions.create() kwargs and
routes them through the Codex Responses streaming API."""
@ -697,26 +649,37 @@ class _CodexCompletionsAdapter:
messages = kwargs.get("messages", [])
model = kwargs.get("model", self._model)
# Separate system/instructions from conversation messages.
# Convert chat.completions multimodal content blocks to Responses
# API format (input_text / input_image instead of text / image_url).
# Separate system/instructions from replayable conversation messages,
# then route the rest through the SINGLE shared chat->Responses
# converter used by the main agent transport
# (agent/transports/codex.py). Maintaining a private conversion loop
# here let chat-style messages with role="tool" leak straight into
# Responses input[] — which the Responses API rejects with
# "Invalid value: 'tool'. Supported values are: 'assistant', 'system',
# 'developer', and 'user'." (issue #5709, hit hard by flush_memories()
# / compression replaying real session history that includes assistant
# tool_calls + role="tool" results). The shared converter encodes
# assistant tool calls as `function_call` items and tool results as
# `function_call_output` items with a valid call_id, so every
# Responses path normalizes tool history identically and cannot drift.
from agent.codex_responses_adapter import _chat_messages_to_responses_input
instructions = "You are a helpful assistant."
input_msgs: List[Dict[str, Any]] = []
replay_messages: List[Dict[str, Any]] = []
for msg in messages:
role = msg.get("role", "user")
content = msg.get("content") or ""
if role == "system":
instructions = content if isinstance(content, str) else str(content)
else:
input_msgs.append({
"role": role,
"content": _convert_content_for_responses(content),
})
replay_messages.append(msg)
input_items = _chat_messages_to_responses_input(replay_messages)
resp_kwargs: Dict[str, Any] = {
"model": model,
"instructions": instructions,
"input": input_msgs or [{"role": "user", "content": ""}],
"input": input_items or [{"role": "user", "content": ""}],
"store": False,
}