Fix double-pipe DSML variant and promote on Responses API.

Parse <｜｜DSML｜｜> markup, add DSML recovery to openai_responses transport, and avoid persisting raw DSML as plain assistant text when promotion fails.

Co-authored-by: Cursor <cursoragent@cursor.com>

svc/llm/transports/openai_chat_completions.py

@@ -18,8 +18,8 @@ from svc.llm.transports.base import (
 )
 from runtime.dsml_tool_parse import (
     DeepSeekTextFilter,
-    dsml_text_tools_enabled,
-    try_parse_dsml_tool_calls_from_fields,
+    promote_dsml_tool_calls_in_response,
+    should_recover_dsml_tool_calls,
 )
 from svc.persistence.assistant_store import get_assistant_store
 
@@ -461,13 +461,11 @@ def _should_recover_dsml_tool_calls(
     *,
     thinking_mode_enabled: bool = False,
 ) -> bool:
-    if dsml_text_tools_enabled(base_url=str(base_url or ""), model_id=str(model or "")):
-        return True
-    if thinking_mode_enabled:
-        mid = str(model or "").strip().lower()
-        if mid.startswith("deepseek-") or "deepseek" in mid:
-            return True
-    return False
+    return should_recover_dsml_tool_calls(
+        model_id=str(model or ""),
+        base_url=str(base_url or ""),
+        thinking_mode_enabled=thinking_mode_enabled,
+    )
 
 
 def _promote_dsml_in_llm_response(
@@ -475,15 +473,7 @@ def _promote_dsml_in_llm_response(
     reasoning: str,
     tool_calls: list[LLMToolCall],
 ) -> tuple[str, str, list[LLMToolCall]]:
-    if tool_calls:
-        return content, reasoning, tool_calls
-    parsed, clean_content, clean_reasoning = try_parse_dsml_tool_calls_from_fields(
-        content=content,
-        reasoning_content=reasoning,
-    )
-    if parsed is not None:
-        return clean_content, clean_reasoning, parsed
-    return content, reasoning, tool_calls
+    return promote_dsml_tool_calls_in_response(content, reasoning, tool_calls)
 
 
 class OpenAIChatModel(ChatModel):

svc/llm/transports/openai_responses.py

@@ -9,6 +9,7 @@ from typing import Any, Optional
 from collections.abc import Callable, Iterable
 
 from svc.llm.transports.base import ChatModel, LLMResponse, LLMToolCall, normalize_image_b64_payload
+from runtime.dsml_tool_parse import promote_dsml_tool_calls_in_response, should_recover_dsml_tool_calls
 
 logger = logging.getLogger(__name__)
 
@@ -340,6 +341,18 @@ class OpenAIResponsesModel(ChatModel):
             kw["base_url"] = self.base_url
         self._client = OpenAI(**kw)
 
+    def _finalize_response(self, content: str, tool_calls: list[LLMToolCall], reasoning: str) -> LLMResponse:
+        text = str(content or "")
+        reasoning_text = str(reasoning or "")
+        calls = list(tool_calls or [])
+        if should_recover_dsml_tool_calls(
+            model_id=str(self.model or ""),
+            base_url=str(self.base_url or ""),
+            thinking_mode_enabled=bool(getattr(self, "thinking_mode_enabled", False)),
+        ):
+            text, reasoning_text, calls = promote_dsml_tool_calls_in_response(text, reasoning_text, calls)
+        return LLMResponse(content=text, tool_calls=calls, reasoning_content=reasoning_text.strip())
+
     @staticmethod
     def _strip_leading_system_messages(messages: list[dict[str, Any]]) -> tuple[str, list[dict[str, Any]]]:
         """Separate leading ``role=system`` rows for ``instructions=` (Responses API expects user-shaped ``input.messages``)."""
@@ -692,7 +705,7 @@ class OpenAIResponsesModel(ChatModel):
                                 on_token(text)
                     if cand_label != input_candidates[0][0]:
                         logger.info("openai_responses: succeeded with input variant %s", cand_label)
-                    return LLMResponse(content=text, tool_calls=tool_calls, reasoning_content=reasoning_text)
+                    return self._finalize_response(text, tool_calls, reasoning_text)
                 except Exception as exc:
                     emsg = str(exc)
                     # Fallback: provider thinking-mode replay contract.
@@ -721,7 +734,7 @@ class OpenAIResponsesModel(ChatModel):
                                         on_token(text)
                             if cand_label != input_candidates[0][0]:
                                 logger.info("openai_responses: succeeded with input variant %s", cand_label)
-                            return LLMResponse(content=text, tool_calls=tool_calls, reasoning_content=reasoning_text)
+                            return self._finalize_response(text, tool_calls, reasoning_text)
                         except Exception:
                             pass
                     if _env_truthy("AIA_RESPONSES_LOG_API_ERROR_DETAIL"):
@@ -759,7 +772,7 @@ class OpenAIResponsesModel(ChatModel):
                         on_token(text)
                     if cand_label != input_candidates[0][0]:
                         logger.info("openai_responses non-stream: succeeded with input variant %s", cand_label)
-                    return LLMResponse(content=text, tool_calls=tool_calls, reasoning_content=reasoning_text)
+                    return self._finalize_response(text, tool_calls, reasoning_text)
                 except Exception as e2:
                     emsg2 = str(e2)
                     if "reasoning_content" in emsg2 and "thinking mode" in emsg2 and "must be passed back" in emsg2:
@@ -787,7 +800,7 @@ class OpenAIResponsesModel(ChatModel):
                                     "openai_responses non-stream: succeeded with input variant %s",
                                     cand_label,
                                 )
-                            return LLMResponse(content=text, tool_calls=tool_calls, reasoning_content=reasoning_text)
+                            return self._finalize_response(text, tool_calls, reasoning_text)
                         except Exception:
                             pass
                     if _env_truthy("AIA_RESPONSES_LOG_API_ERROR_DETAIL"):