common/hermes
2
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

fix(deepseek): preserve v4 reasoning_content on replay

Browse Source
This commit is contained in:
lsdsjy
2026-04-28 14:17:11 +08:00
committed by Teknium
parent 8fbc9d7d78
commit b9b9ee3e6c
5 changed files with 398 additions and 33 deletions
Download Patch File Download Diff File Expand all files Collapse all files

44
run_agent.py
View File

@@ -8501,6 +8501,7 @@ class AIAgent:
Handles reasoning extraction, reasoning_details, and optional tool_calls
so both the tool-call path and the final-response path share one builder.
"""
assistant_tool_calls = getattr(assistant_message, "tool_calls", None)
reasoning_text = self._extract_reasoning(assistant_message)
_from_structured = bool(reasoning_text)
@@ -8560,16 +8561,19 @@ class AIAgent:
"finish_reason": finish_reason,
}
if hasattr(assistant_message, "reasoning_content"):
raw_reasoning_content = getattr(assistant_message, "reasoning_content", None)
if raw_reasoning_content is not None:
msg["reasoning_content"] = _sanitize_surrogates(raw_reasoning_content)
elif msg.get("tool_calls") and self._needs_deepseek_tool_reasoning():
# DeepSeek thinking mode requires reasoning_content on every
# assistant tool-call message. Without it, replaying the
# persisted message causes HTTP 400. Include empty string
# as a defensive compatibility fallback (refs #15250).
msg["reasoning_content"] = ""
raw_reasoning_content = getattr(assistant_message, "reasoning_content", None)
if raw_reasoning_content is None and hasattr(assistant_message, "model_extra"):
model_extra = getattr(assistant_message, "model_extra", None) or {}
if isinstance(model_extra, dict) and "reasoning_content" in model_extra:
raw_reasoning_content = model_extra["reasoning_content"]
if raw_reasoning_content is not None:
msg["reasoning_content"] = _sanitize_surrogates(raw_reasoning_content)
elif assistant_tool_calls and self._needs_deepseek_tool_reasoning():
# DeepSeek thinking mode requires reasoning_content on every
# assistant tool-call message. Without it, replaying the
# persisted message causes HTTP 400. Include empty string
# only when no structured reasoning text was captured.
msg["reasoning_content"] = reasoning_text or ""
# Additive fallback (refs #16844, #16884). Streaming-only providers
# (glm, MiniMax, gpt-5.x via aigw, Anthropic via openai-compat shims)
@@ -8626,9 +8630,9 @@ class AIAgent:
if codex_message_items:
msg["codex_message_items"] = codex_message_items
if assistant_message.tool_calls:
if assistant_tool_calls:
tool_calls = []
for tool_call in assistant_message.tool_calls:
for tool_call in assistant_tool_calls:
raw_id = getattr(tool_call, "id", None)
call_id = getattr(tool_call, "call_id", None)
if not isinstance(call_id, str) or not call_id.strip():
@@ -8728,11 +8732,11 @@ class AIAgent:
# if the source turn has tool_calls AND a 'reasoning' field but no
# 'reasoning_content' key, the 'reasoning' text was written by a
# prior provider (e.g. MiniMax) — DeepSeek's own _build_assistant_message
# always pins reasoning_content="" at creation time for tool-call turns,
# so the shape (reasoning set, reasoning_content absent, tool_calls
# present) is unreachable from same-provider DeepSeek history. Inject
# "" to satisfy the API without leaking another provider's chain of
# thought to DeepSeek/Kimi.
# pins reasoning_content at creation time for tool-call turns, so the
# shape (reasoning set, reasoning_content absent, tool_calls present)
# is unreachable from same-provider DeepSeek history after this fix.
# Inject "" to satisfy the API without leaking another provider's
# chain of thought to DeepSeek/Kimi.
normalized_reasoning = source_msg.get("reasoning")
if (
needs_thinking_pad
@@ -8745,9 +8749,9 @@ class AIAgent:
# 3. Healthy session: promote 'reasoning' field to 'reasoning_content'
# for providers that use the internal 'reasoning' key.
# This must happen BEFORE the DeepSeek/Kimi tool-call check so that
# genuine reasoning content is not overwritten by the empty-string
# fallback (#15812 regression in PR #15478).
# This must happen before the unconditional empty-string fallback so
# genuine reasoning content is not overwritten (#15812 regression in
# PR #15478).
if isinstance(normalized_reasoning, str) and normalized_reasoning:
api_msg["reasoning_content"] = normalized_reasoning
return