molecule-ai/hermes-agent
34
0
Fork 0
Code Issues 2 Pull Requests 4 Actions 14 Packages Projects Releases Wiki Activity

fix(deepseek): use non-empty reasoning_content placeholder for V4 Pro thinking mode

Browse Source 
DeepSeek V4 Pro tightened thinking-mode validation and rejects empty-string
reasoning_content with HTTP 400:

    The reasoning content in the thinking mode must be passed back to the API.

run_agent.py injected "" at three fallback sites — the tool-call pad in
_build_assistant_message and both injection branches of
_copy_reasoning_content_for_api (cross-provider poison guard + unconditional
thinking pad). All three now emit " " (single space), which satisfies the
non-empty check on V4 Pro without leaking fabricated reasoning.

Also upgrades stale empty-string placeholders on replay: sessions persisted
before this change have reasoning_content="" pinned at creation time; when
the active provider enforces thinking-mode echo, the replay path now rewrites
"" -> " " so existing users don't 400 on their first V4 Pro turn after
updating. Non-thinking providers still round-trip "" verbatim.

Updates 9 existing assertions + adds 2 regression tests (stale-placeholder
upgrade, non-thinking verbatim preservation).

Refs #15250, #17400.
Closes #17341.
This commit is contained in:
IMHaoyan 2026-04-30 22:49:55 -07:00 committed by Teknium
parent f0dc919f92
commit bfb704684e
4 changed files with 98 additions and 37 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

46
run_agent.py
View File

@ -8603,9 +8603,13 @@ class AIAgent:
# message. Without it, replaying the persisted message causes
# HTTP 400 ("The reasoning_content in the thinking mode must
# be passed back to the API"). Include streamed reasoning
# text when captured; otherwise pad with empty string.
# Refs #15250, #17400.
msg["reasoning_content"] = reasoning_text or ""
# text when captured; otherwise pad with a single space —
# DeepSeek V4 Pro tightened validation and rejects empty
# string ("The reasoning content in the thinking mode must
# be passed back to the API"). A space satisfies non-empty
# checks everywhere without leaking fabricated reasoning.
# Refs #15250, #17400, #17341.
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)
@ -8760,11 +8764,20 @@ class AIAgent:
return
# 1. Explicit reasoning_content already set — preserve it verbatim
# (includes DeepSeek/Kimi's own empty-string placeholder written at
# creation time, and any valid reasoning content from the same provider).
# (includes DeepSeek/Kimi's own space-placeholder written at creation
# time, and any valid reasoning content from the same provider).
#
# Exception: sessions persisted BEFORE #17341 have empty-string
# placeholders pinned at creation time. DeepSeek V4 Pro rejects
# those with HTTP 400. When the active provider enforces the
# thinking-mode echo, upgrade "" → " " on replay so stale history
# doesn't 400 the user on the next turn.
existing = source_msg.get("reasoning_content")
if isinstance(existing, str):
api_msg["reasoning_content"] = existing
if existing == "" and self._needs_thinking_reasoning_pad():
api_msg["reasoning_content"] = " "
else:
api_msg["reasoning_content"] = existing
return
needs_thinking_pad = self._needs_thinking_reasoning_pad()
@ -8776,8 +8789,10 @@ class AIAgent:
# 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.
# Inject a single space to satisfy the API without leaking another
# provider's chain of thought to DeepSeek/Kimi. Space (not "")
# because DeepSeek V4 Pro rejects empty-string reasoning_content
# in thinking mode (refs #17341).
normalized_reasoning = source_msg.get("reasoning")
if (
needs_thinking_pad
@ -8785,7 +8800,7 @@ class AIAgent:
and isinstance(normalized_reasoning, str)
and normalized_reasoning
):
api_msg["reasoning_content"] = ""
api_msg["reasoning_content"] = " "
return
# 3. Healthy session: promote 'reasoning' field to 'reasoning_content'
@ -8798,12 +8813,15 @@ class AIAgent:
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. Covers both
# tool-call turns (already-poisoned history with no reasoning at all)
# and plain text turns.
# reasoning_content. Inject a single space to satisfy the provider's
# requirement when no explicit reasoning content is present. Covers
# both tool-call turns (already-poisoned history with no reasoning
# at all) and plain text turns. Space (not "") because DeepSeek V4
# Pro tightened validation and rejects empty string with HTTP 400
# ("The reasoning content in the thinking mode must be passed back
# to the API"). Refs #17341.
if needs_thinking_pad:
api_msg["reasoning_content"] = ""
api_msg["reasoning_content"] = " "
return
# 5. reasoning_content was present but not a string (e.g. None after

1
scripts/release.py
View File

@ -50,6 +50,7 @@ AUTHOR_MAP = {
"rylen.anil@gmail.com": "rylena",
"godnanijatin@gmail.com": "jatingodnani",
"14046872+tmimmanuel@users.noreply.github.com": "tmimmanuel",
"657290301@qq.com": "IMHaoyan",
"revar@users.noreply.github.com": "revaraver",
# Matrix parity salvage batch (April 2026)
"sr@samirusani": "samrusani",

80
tests/run_agent/test_deepseek_reasoning_content_echo.py
View File

@ -10,15 +10,21 @@ field, DeepSeek rejects the next request with HTTP 400::
Fix covers three paths:
1. ``_build_assistant_message`` — new tool-call messages without raw
reasoning_content get ``""`` pinned at creation time so nothing gets
reasoning_content get ``" "`` pinned at creation time so nothing gets
persisted poisoned.
2. ``_copy_reasoning_content_for_api`` — already-poisoned history replays
with ``reasoning_content=""`` injected defensively.
with ``reasoning_content=" "`` injected defensively.
3. Detection covers three signals: ``provider == "deepseek"``,
``"deepseek" in model``, and ``api.deepseek.com`` host match. The third
catches custom-provider setups pointing at DeepSeek.
Refs #15250 / #15353.
The placeholder is a single space (not empty string) because DeepSeek V4 Pro
tightened validation and rejects empty-string reasoning_content with a
400 ("The reasoning content in the thinking mode must be passed back to
the API"). A space satisfies non-empty checks everywhere without leaking
fabricated reasoning.
Refs #15250 / #15353 / #17341.
"""
from __future__ import annotations
@ -105,8 +111,8 @@ class TestNeedsDeepSeekToolReasoning:
class TestCopyReasoningContentForApi:
"""_copy_reasoning_content_for_api pads reasoning_content for DeepSeek tool-calls."""
def test_deepseek_tool_call_poisoned_history_gets_empty_string(self) -> None:
"""Already-poisoned history (no reasoning_content, no reasoning) gets ''."""
def test_deepseek_tool_call_poisoned_history_gets_space_placeholder(self) -> None:
"""Already-poisoned history (no reasoning_content, no reasoning) gets ' '."""
agent = _make_agent(provider="deepseek", model="deepseek-v4-flash")
source = {
"role": "assistant",
@ -115,7 +121,7 @@ class TestCopyReasoningContentForApi:
}
api_msg: dict = {}
agent._copy_reasoning_content_for_api(source, api_msg)
assert api_msg.get("reasoning_content") == ""
assert api_msg.get("reasoning_content") == " "
def test_deepseek_assistant_no_tool_call_gets_padded(self) -> None:
"""DeepSeek thinking mode pads ALL assistant turns, even without tool_calls."""
@ -123,7 +129,7 @@ class TestCopyReasoningContentForApi:
source = {"role": "assistant", "content": "hello"}
api_msg: dict = {}
agent._copy_reasoning_content_for_api(source, api_msg)
assert api_msg.get("reasoning_content") == ""
assert api_msg.get("reasoning_content") == " "
def test_deepseek_explicit_reasoning_content_preserved(self) -> None:
"""When reasoning_content is already set, it's copied verbatim."""
@ -137,6 +143,42 @@ class TestCopyReasoningContentForApi:
agent._copy_reasoning_content_for_api(source, api_msg)
assert api_msg["reasoning_content"] == "<think>real chain of thought</think>"
def test_deepseek_stale_empty_placeholder_upgraded_to_space(self) -> None:
"""Sessions persisted before #17341 have ``reasoning_content=""`` pinned
at creation time. DeepSeek V4 Pro rejects "" with HTTP 400. When the
active provider enforces the thinking-mode echo, the replay path
upgrades "" → " " so stale history doesn't break the next turn.
"""
agent = _make_agent(provider="deepseek", model="deepseek-v4-pro")
source = {
"role": "assistant",
"content": "",
"reasoning_content": "",
"tool_calls": [{"id": "c1", "function": {"name": "terminal"}}],
}
api_msg: dict = {}
agent._copy_reasoning_content_for_api(source, api_msg)
assert api_msg["reasoning_content"] == " "
def test_non_thinking_provider_preserves_empty_reasoning_content_verbatim(self) -> None:
"""The stale-placeholder upgrade ONLY fires when the active provider
enforces thinking-mode echo. On non-thinking providers, an empty
reasoning_content must still round-trip verbatim.
"""
agent = _make_agent(
provider="openrouter",
model="anthropic/claude-sonnet-4.6",
base_url="https://openrouter.ai/api/v1",
)
source = {
"role": "assistant",
"content": "hi",
"reasoning_content": "",
}
api_msg: dict = {}
agent._copy_reasoning_content_for_api(source, api_msg)
assert api_msg["reasoning_content"] == ""
def test_deepseek_reasoning_field_promoted(self) -> None:
"""When only 'reasoning' is set, it gets promoted to reasoning_content."""
agent = _make_agent(provider="deepseek", model="deepseek-v4-flash")
@ -155,7 +197,7 @@ class TestCopyReasoningContentForApi:
If the source turn has tool_calls AND a 'reasoning' field but NO
'reasoning_content' key, it's from a prior provider (the DeepSeek
build path pins reasoning_content at creation). Inject "" instead
build path pins reasoning_content at creation). Inject " " instead
of forwarding the prior provider's chain of thought.
"""
agent = _make_agent(provider="deepseek", model="deepseek-v4-flash")
@ -167,7 +209,7 @@ class TestCopyReasoningContentForApi:
}
api_msg: dict = {}
agent._copy_reasoning_content_for_api(source, api_msg)
assert api_msg["reasoning_content"] == ""
assert api_msg["reasoning_content"] == " "
def test_kimi_poisoned_cross_provider_history_padded(self) -> None:
"""Kimi path of #15748 — same rule as DeepSeek."""
@ -180,7 +222,7 @@ class TestCopyReasoningContentForApi:
}
api_msg: dict = {}
agent._copy_reasoning_content_for_api(source, api_msg)
assert api_msg["reasoning_content"] == ""
assert api_msg["reasoning_content"] == " "
def test_kimi_path_still_works(self) -> None:
"""Existing Kimi detection still pads reasoning_content."""
@ -192,7 +234,7 @@ class TestCopyReasoningContentForApi:
}
api_msg: dict = {}
agent._copy_reasoning_content_for_api(source, api_msg)
assert api_msg.get("reasoning_content") == ""
assert api_msg.get("reasoning_content") == " "
def test_kimi_moonshot_base_url(self) -> None:
agent = _make_agent(
@ -205,7 +247,7 @@ class TestCopyReasoningContentForApi:
}
api_msg: dict = {}
agent._copy_reasoning_content_for_api(source, api_msg)
assert api_msg.get("reasoning_content") == ""
assert api_msg.get("reasoning_content") == " "
def test_non_thinking_provider_not_padded(self) -> None:
"""Providers that don't require the echo are untouched."""
@ -237,7 +279,7 @@ class TestCopyReasoningContentForApi:
}
api_msg: dict = {}
agent._copy_reasoning_content_for_api(source, api_msg)
assert api_msg.get("reasoning_content") == ""
assert api_msg.get("reasoning_content") == " "
def test_non_assistant_role_ignored(self) -> None:
"""User/tool messages are left alone."""
@ -302,7 +344,7 @@ class TestBuildAssistantMessageDeepSeekReasoningContent:
assert msg["reasoning_content"] == "DeepSeek model_extra reasoning"
def test_deepseek_tool_call_without_raw_reasoning_content_gets_empty_string(self) -> None:
def test_deepseek_tool_call_without_raw_reasoning_content_gets_space_placeholder(self) -> None:
agent = _make_agent(provider="deepseek", model="deepseek-v4-flash")
assistant_message = SimpleNamespace(
content=None,
@ -324,7 +366,7 @@ class TestBuildAssistantMessageDeepSeekReasoningContent:
msg = agent._build_assistant_message(assistant_message, "tool_calls")
assert msg["reasoning_content"] == ""
assert msg["reasoning_content"] == " "
assert msg["tool_calls"][0]["id"] == "call_1"
@ -345,22 +387,22 @@ class TestBuildAssistantMessagePadsStrictProviders:
[
pytest.param(
"deepseek", "deepseek-v4-pro", "",
None, "",
None, " ",
id="deepseek-attr-none",
),
pytest.param(
"deepseek", "deepseek-v4-pro", "",
_ATTR_ABSENT, "",
_ATTR_ABSENT, " ",
id="deepseek-attr-absent",
),
pytest.param(
"kimi-coding", "kimi-k2.6", "",
None, "",
None, " ",
id="kimi-attr-none",
),
pytest.param(
"custom", "kimi-k2", "https://api.moonshot.ai/v1",
_ATTR_ABSENT, "",
_ATTR_ABSENT, " ",
id="moonshot-base-url",
),
pytest.param(

8
tests/run_agent/test_run_agent.py
View File

@ -1465,8 +1465,8 @@ class TestBuildAssistantMessage:
This preserves ``_copy_reasoning_content_for_api``'s downstream
tiers at replay time — cross-provider leak guard (#15748),
promote-from-``reasoning``, and DeepSeek/Kimi ""-pad — which
would all be bypassed if we eagerly wrote ``reasoning_content=""``
promote-from-``reasoning``, and DeepSeek/Kimi " "-pad — which
would all be bypassed if we eagerly wrote ``reasoning_content=" "``
on every assistant turn regardless of provider.
"""
msg = _mock_assistant_msg(content="plain answer")
@ -4617,7 +4617,7 @@ class TestReasoningReplayForStrictProviders:
agent.compression_enabled = False
agent.save_trajectories = False
def test_kimi_tool_replay_includes_empty_reasoning_content(self, agent):
def test_kimi_tool_replay_includes_space_reasoning_content(self, agent):
self._setup_agent(agent)
agent.base_url = "https://api.kimi.com/coding/v1"
agent._base_url_lower = agent.base_url.lower()
@ -4654,7 +4654,7 @@ class TestReasoningReplayForStrictProviders:
assert replayed_assistant["role"] == "assistant"
assert replayed_assistant["tool_calls"][0]["function"]["name"] == "terminal"
assert "reasoning_content" in replayed_assistant
assert replayed_assistant["reasoning_content"] == ""
assert replayed_assistant["reasoning_content"] == " "
def test_explicit_reasoning_content_beats_normalized_reasoning_on_replay(self, agent):
self._setup_agent(agent)