Skip to content

Codebase intelligence for AI and humans: code health scores, auto-generated docs, git analytics, dead code detection, and architectural decisions via MCP.

fix(deepseek): correct thinking controls and effort mapping (#2239)

Selene29f044dd553ec9
diff --git a/packages/core/src/repowise/core/providers/llm/deepseek.py b/packages/core/src/repowise/core/providers/llm/deepseek.py
index 11ed96d52..9b7351045 100644
--- a/packages/core/src/repowise/core/providers/llm/deepseek.py
+++ b/packages/core/src/repowise/core/providers/llm/deepseek.py
@@ -52,16 +52,26 @@ _DEFAULT_BASE_URL = "https://api.deepseek.com"
_DEEPSEEK_REASONING_MODES: tuple[ReasoningMode, ...] = (
"off",
"none",
+ "minimal",
"low",
"medium",
"high",
"xhigh",
"max",
)
+_DEEPSEEK_REASONING_MODELS = frozenset(
+ {
+ "deepseek-flash",
+ "deepseek-v4-pro",
+ # Retired aliases remain accepted and are routed to V4.1 Flash.
+ "deepseek-v4-flash",
+ "deepseek-v4-flash-vision-exp",
+ }
+)
def _deepseek_supported_reasoning_modes(model: str) -> tuple[ReasoningMode, ...]:
- if model.startswith("deepseek-v4-"):
+ if model in _DEEPSEEK_REASONING_MODELS:
return _DEEPSEEK_REASONING_MODES
return ()
@@ -78,7 +88,7 @@ def _resolve_deepseek_reasoning_mode(
_deepseek_supported_reasoning_modes(model),
detail=(
"DeepSeek /models lists IDs only; reasoning controls are enabled "
- "for the documented V4 model family."
+ "for the documented DeepSeek Flash and V4 model families."
),
)
@@ -89,8 +99,11 @@ def _deepseek_reasoning_kwargs(reasoning: ReasoningMode) -> dict[str, Any]:
return {}
if mode in ("off", "none"):
return {"extra_body": {"thinking": {"type": "disabled"}}}
- effort = "max" if mode in ("xhigh", "max") else "high"
- return {"extra_body": {"thinking": {"type": "enabled", "reasoning_effort": effort}}}
+ effort = "low" if mode in ("minimal", "low") else "max" if mode == "max" else "high"
+ return {
+ "reasoning_effort": effort,
+ "extra_body": {"thinking": {"type": "enabled"}},
+ }
def _deepseek_model_options(
@@ -132,7 +145,7 @@ def _deepseek_model_options(
recommended=model_id == fallback_model,
source="api",
notes=(
- "reasoning controls documented for DeepSeek V4"
+ "thinking toggle and graded effort documented by DeepSeek"
if len(reasoning_modes) > 1
else ""
),
@@ -334,6 +347,12 @@ class DeepSeekProvider(BaseProvider):
"temperature": temperature,
"messages": full_messages,
"stream": True,
+ # Thinking tool calls require reasoning_content from every prior
+ # assistant turn. Repowise's shared chat history intentionally
+ # does not retain hidden reasoning, so use DeepSeek's documented
+ # non-thinking mode for this protocol instead of producing an
+ # invalid follow-up request after the first tool call.
+ "extra_body": {"thinking": {"type": "disabled"}},
}
if tools:
kwargs["tools"] = tools
diff --git a/tests/unit/test_providers/test_deepseek_provider.py b/tests/unit/test_providers/test_deepseek_provider.py
index e7e6003f9..d3797af7b 100644
--- a/tests/unit/test_providers/test_deepseek_provider.py
+++ b/tests/unit/test_providers/test_deepseek_provider.py
@@ -59,6 +59,7 @@ def test_available_model_options_uses_models_endpoint(monkeypatch):
def json(self) -> dict:
return {
"data": [
+ {"id": "deepseek-flash"},
{"id": "deepseek-v4-flash"},
{"id": "deepseek-v4-pro"},
]
@@ -83,6 +84,7 @@ def test_available_model_options_uses_models_endpoint(monkeypatch):
"auto",
"off",
"none",
+ "minimal",
"low",
"medium",
"high",
@@ -90,6 +92,8 @@ def test_available_model_options_uses_models_endpoint(monkeypatch):
"max",
)
assert flash.recommended is True
+ current_flash = next(option for option in options if option.model == "deepseek-flash")
+ assert current_flash.reasoning_modes == flash.reasoning_modes
def _make_mock_chat_response(
@@ -192,6 +196,30 @@ async def test_generate_forwards_disabled_thinking():
assert kwargs["extra_body"] == {"thinking": {"type": "disabled"}}
+ ("reasoning", "effort"),
+ [
+ ("minimal", "low"),
+ ("low", "low"),
+ ("medium", "high"),
+ ("xhigh", "high"),
+ ("max", "max"),
+ ],
+)
+async def test_generate_forwards_top_level_reasoning_effort(reasoning, effort):
+ provider = DeepSeekProvider(api_key="sk-test")
+ mock_response = _make_mock_chat_response()
+
+ with patch("openai.AsyncOpenAI") as mock_client:
+ mock_client.return_value.chat.completions.create = AsyncMock(return_value=mock_response)
+ provider._client = mock_client.return_value
+ await provider.generate("system", "user", reasoning=reasoning)
+
+ kwargs = mock_client.return_value.chat.completions.create.call_args.kwargs
+ assert kwargs["reasoning_effort"] == effort
+ assert kwargs["extra_body"] == {"thinking": {"type": "enabled"}}
+
+
async def test_generate_rejects_reasoning_for_non_v4_model():
provider = DeepSeekProvider(api_key="sk-test", model="deepseek-chat")
@@ -203,6 +231,31 @@ async def test_generate_rejects_reasoning_for_non_v4_model():
mock_client.return_value.chat.completions.create.assert_not_called()
+async def test_unknown_v4_name_does_not_inherit_speculative_reasoning_support():
+ provider = DeepSeekProvider(api_key="sk-test", model="deepseek-v4-future")
+
+ with patch("openai.AsyncOpenAI") as mock_client:
+ provider._client = mock_client.return_value
+ with pytest.raises(ProviderError, match="reasoning='high' is not supported"):
+ await provider.generate("system", "user", reasoning="high")
+
+ mock_client.return_value.chat.completions.create.assert_not_called()
+
+
+async def test_current_flash_model_accepts_documented_reasoning_effort():
+ provider = DeepSeekProvider(api_key="sk-test", model="deepseek-flash")
+ mock_response = _make_mock_chat_response()
+
+ with patch("openai.AsyncOpenAI") as mock_client:
+ mock_client.return_value.chat.completions.create = AsyncMock(return_value=mock_response)
+ provider._client = mock_client.return_value
+ await provider.generate("system", "user", reasoning="high")
+
+ kwargs = mock_client.return_value.chat.completions.create.call_args.kwargs
+ assert kwargs["reasoning_effort"] == "high"
+ assert kwargs["extra_body"] == {"thinking": {"type": "enabled"}}
+
+
async def test_generate_rate_limit_retry():
from openai import RateLimitError as _OpenAIRateLimitError
@@ -299,3 +352,5 @@ async def test_stream_chat_emits_text_delta_and_stop():
assert text_deltas[1].text == "i"
assert len(stops) == 1
assert stops[0].stop_reason == "end_turn"
+ kwargs = mock_client.return_value.chat.completions.create.call_args.kwargs
+ assert kwargs["extra_body"] == {"thinking": {"type": "disabled"}}