ling/hermes-agent
1
0
Fork 0
mirror of https://github.com/NousResearch/hermes-agent.git synced 2026-04-28 06:51:16 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix(agent): block cross-provider reasoning leak to DeepSeek/Kimi (#15748) (#16500)

Browse Source 
On provider switches mid-session (e.g. MiniMax -> DeepSeek), the source
assistant turn carries a 'reasoning' field written by the prior provider
but no 'reasoning_content' key. _copy_reasoning_content_for_api would
promote that foreign 'reasoning' to 'reasoning_content' on the outbound
DeepSeek request, leaking a cross-provider chain of thought and in
practice causing HTTP 400.

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 — it can only come from a prior provider.
Pad with '' in that case instead of promoting.

Healthy same-provider 'reasoning' promotion (no tool_calls, or on
providers that do not require the empty-string pin) is unchanged.
This commit is contained in:
Teknium
2026-04-27 04:06:23 -07:00
committed by GitHub
parent 65f648ee84
commit ee1a07f9e9
2 changed files with 73 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files

52
run_agent.py
View File

@@ -7890,39 +7890,45 @@ class AIAgent:
api_msg["reasoning_content"] = existing
return
# 2. Healthy session: promote 'reasoning' field to 'reasoning_content'
needs_thinking_pad = (
self._needs_kimi_tool_reasoning()
or self._needs_deepseek_tool_reasoning()
)
# 2. Cross-provider poisoned history (#15748): on DeepSeek/Kimi,
# 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.
normalized_reasoning = source_msg.get("reasoning")
if (
needs_thinking_pad
and source_msg.get("tool_calls")
and isinstance(normalized_reasoning, str)
and normalized_reasoning
):
api_msg["reasoning_content"] = ""
return
# 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).
normalized_reasoning = source_msg.get("reasoning")
if isinstance(normalized_reasoning, str) and normalized_reasoning:
api_msg["reasoning_content"] = normalized_reasoning
return
# 3. DeepSeek / Kimi thinking mode: tool-call turns that lack
# reasoning_content are "poisoned history" — a prior provider (MiniMax,
# etc.) left them empty. DeepSeek returns HTTP 400 if reasoning_content
# is absent on replay; inject "" to satisfy the provider's requirement
# without forwarding any cross-provider reasoning content.
needs_empty_reasoning = (
source_msg.get("tool_calls")
and (
self._needs_kimi_tool_reasoning()
or self._needs_deepseek_tool_reasoning()
)
)
if needs_empty_reasoning:
api_msg["reasoning_content"] = ""
return
# 4. DeepSeek / Kimi thinking mode: all assistant messages need
# reasoning_content. Inject "" to satisfy the provider's requirement
# when no explicit reasoning content is present.
if (
self._needs_kimi_tool_reasoning()
or self._needs_deepseek_tool_reasoning()
):
# when no explicit reasoning content is present. Covers both
# tool-call turns (already-poisoned history with no reasoning at all)
# and plain text turns.
if needs_thinking_pad:
api_msg["reasoning_content"] = ""
return