fix(llm): DeepSeek strict tools via function.strict, not extra_body

Match DeepSeek Tool Calls docs: set strict:true on each tools[].function.
Remove non-documented strict_tool_mode extra_body. Keep AIA_DEEPSEEK_STRICT_TOOL_MODE=0 to opt out.

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
oliver 2026-05-13 14:59:24 +08:00
parent 420abac9f1
commit 6ba3234a87
2 changed files with 60 additions and 74 deletions

View file

@ -109,11 +109,13 @@ def _messages_contain_list_with_image(messages: list[dict[str, Any]]) -> bool:
return False return False
def _deepseek_strict_tool_mode_enabled(*, base_url: str | None, model: str | None) -> bool: def _deepseek_strict_tools_enabled(*, base_url: str | None, model: str | None) -> bool:
"""DeepSeek OpenAI-compat: optional ``strict_tool_mode`` in request body (via ``extra_body``). """Whether to set ``strict: true`` on each OpenAI-style ``tools[].function`` for DeepSeek.
Upstream documents this as reducing DSML-in-text tool output in favor of native ``tool_calls``. DeepSeek documents strict tool output in
Opt out with ``AIA_DEEPSEEK_STRICT_TOOL_MODE=0`` when a proxy rejects unknown fields. https://api-docs.deepseek.com/zh-cn/guides/tool_calls :
each ``function`` should include ``\"strict\": true`` (Beta; also needs ``/beta`` base URL
and JSON Schema rules on the provider side). Opt out with ``AIA_DEEPSEEK_STRICT_TOOL_MODE=0``.
""" """
raw = str(os.getenv("AIA_DEEPSEEK_STRICT_TOOL_MODE") or "").strip().lower() raw = str(os.getenv("AIA_DEEPSEEK_STRICT_TOOL_MODE") or "").strip().lower()
if raw in ("0", "false", "no", "off"): if raw in ("0", "false", "no", "off"):
@ -122,19 +124,28 @@ def _deepseek_strict_tool_mode_enabled(*, base_url: str | None, model: str | Non
return "deepseek" in hay return "deepseek" in hay
def _merge_deepseek_strict_tool_mode_extra_body( def _apply_deepseek_strict_tools(tools: list[dict[str, Any]] | None) -> list[dict[str, Any]] | None:
kwargs: dict[str, Any], """Return a shallow-copied tools list with ``function.strict: true`` on each function tool."""
*, if not tools:
base_url: str | None, return tools
model: str | None, out: list[dict[str, Any]] = []
use_tools: bool, for raw in tools:
) -> None: if not isinstance(raw, dict):
if not use_tools or not _deepseek_strict_tool_mode_enabled(base_url=base_url, model=model): out.append(raw) # type: ignore[arg-type]
return continue
extra_body = kwargs.get("extra_body") if isinstance(kwargs.get("extra_body"), dict) else {} if str(raw.get("type") or "").strip().lower() != "function":
extra_body = dict(extra_body) out.append(dict(raw))
extra_body["strict_tool_mode"] = True continue
kwargs["extra_body"] = extra_body t = dict(raw)
fn = t.get("function")
if isinstance(fn, dict):
fn2 = dict(fn)
fn2["strict"] = True
t["function"] = fn2
else:
t["function"] = {"strict": True}
out.append(t)
return out
def _should_proactively_downgrade_multimodal_messages( def _should_proactively_downgrade_multimodal_messages(
@ -556,9 +567,10 @@ class OpenAIChatModel(ChatModel):
kwargs["tools"] = plan.tools_wired kwargs["tools"] = plan.tools_wired
except Exception: except Exception:
kwargs["tools"] = complete_openai_tools_wire_parameters(tools) kwargs["tools"] = complete_openai_tools_wire_parameters(tools)
_merge_deepseek_strict_tool_mode_extra_body( if use_tools and _deepseek_strict_tools_enabled(base_url=self.base_url, model=self.model):
kwargs, base_url=self.base_url, model=self.model, use_tools=use_tools tw = kwargs.get("tools")
) if isinstance(tw, list):
kwargs["tools"] = _apply_deepseek_strict_tools(tw)
try: try:
return self._client.chat.completions.create(**kwargs) return self._client.chat.completions.create(**kwargs)
except Exception as exc: except Exception as exc:
@ -712,8 +724,8 @@ __all__ = [
"OpenAIChatModel", "OpenAIChatModel",
"_likely_gemini_openai_compat_base_url", "_likely_gemini_openai_compat_base_url",
"_model_id_suggests_gemini", "_model_id_suggests_gemini",
"_deepseek_strict_tool_mode_enabled", "_deepseek_strict_tools_enabled",
"_merge_deepseek_strict_tool_mode_extra_body", "_apply_deepseek_strict_tools",
] ]

View file

@ -1,4 +1,4 @@
"""DeepSeek OpenAI-compat: strict_tool_mode via extra_body.""" """DeepSeek: per-function ``strict: true`` on tools (official Tool Calls / strict Beta)."""
from __future__ import annotations from __future__ import annotations
@ -6,15 +6,15 @@ import os
import unittest import unittest
from svc.llm.transports.openai_chat_completions import ( from svc.llm.transports.openai_chat_completions import (
_deepseek_strict_tool_mode_enabled, _apply_deepseek_strict_tools,
_merge_deepseek_strict_tool_mode_extra_body, _deepseek_strict_tools_enabled,
) )
class DeepseekStrictToolModeTests(unittest.TestCase): class DeepseekStrictToolsTests(unittest.TestCase):
def test_enabled_when_deepseek_in_base_url(self) -> None: def test_enabled_when_deepseek_in_base_url(self) -> None:
self.assertTrue( self.assertTrue(
_deepseek_strict_tool_mode_enabled( _deepseek_strict_tools_enabled(
base_url="https://api.deepseek.com/v1", base_url="https://api.deepseek.com/v1",
model="deepseek-chat", model="deepseek-chat",
) )
@ -25,7 +25,7 @@ class DeepseekStrictToolModeTests(unittest.TestCase):
try: try:
os.environ["AIA_DEEPSEEK_STRICT_TOOL_MODE"] = "0" os.environ["AIA_DEEPSEEK_STRICT_TOOL_MODE"] = "0"
self.assertFalse( self.assertFalse(
_deepseek_strict_tool_mode_enabled( _deepseek_strict_tools_enabled(
base_url="https://api.deepseek.com/v1", base_url="https://api.deepseek.com/v1",
model="x", model="x",
) )
@ -36,54 +36,28 @@ class DeepseekStrictToolModeTests(unittest.TestCase):
else: else:
os.environ["AIA_DEEPSEEK_STRICT_TOOL_MODE"] = prev os.environ["AIA_DEEPSEEK_STRICT_TOOL_MODE"] = prev
def test_merge_sets_extra_body_strict_tool_mode(self) -> None: def test_apply_sets_strict_on_function_tools(self) -> None:
kwargs: dict = {"model": "m", "messages": [], "stream": False, "tools": [{"type": "function"}]} tools = [
_merge_deepseek_strict_tool_mode_extra_body( {"type": "function", "function": {"name": "get_weather", "parameters": {"type": "object", "properties": {}}}},
kwargs, {"type": "not_function", "x": 1},
base_url="https://api.deepseek.com/v1", ]
model="deepseek-chat", out = _apply_deepseek_strict_tools(tools)
use_tools=True, self.assertIsNotNone(out)
) assert out is not None
self.assertEqual(kwargs.get("extra_body"), {"strict_tool_mode": True}) self.assertTrue(out[0]["function"]["strict"])
self.assertEqual(out[1].get("type"), "not_function")
self.assertNotIn("strict", out[1])
# input not mutated
self.assertNotIn("strict", tools[0].get("function", {}))
def test_merge_preserves_existing_extra_body(self) -> None: def test_apply_skips_non_deepseek_not_tested_here(self) -> None:
kwargs: dict = { out = _apply_deepseek_strict_tools([])
"model": "m", self.assertEqual(out, [])
"messages": [],
"stream": False,
"tools": [{"type": "function"}],
"extra_body": {"thinking": {"type": "disabled"}},
}
_merge_deepseek_strict_tool_mode_extra_body(
kwargs,
base_url="https://api.deepseek.com",
model="deepseek-reasoner",
use_tools=True,
)
self.assertEqual(
kwargs["extra_body"],
{"thinking": {"type": "disabled"}, "strict_tool_mode": True},
)
def test_merge_skips_without_tools(self) -> None: def test_apply_handles_missing_function_dict(self) -> None:
kwargs: dict = {"model": "m", "messages": [], "stream": False} out = _apply_deepseek_strict_tools([{"type": "function"}])
_merge_deepseek_strict_tool_mode_extra_body( assert out is not None
kwargs, self.assertEqual(out[0]["function"], {"strict": True})
base_url="https://api.deepseek.com/v1",
model="deepseek-chat",
use_tools=False,
)
self.assertNotIn("extra_body", kwargs)
def test_merge_skips_non_deepseek(self) -> None:
kwargs: dict = {"model": "m", "messages": [], "stream": False, "tools": [{"type": "function"}]}
_merge_deepseek_strict_tool_mode_extra_body(
kwargs,
base_url="https://api.openai.com/v1",
model="gpt-4o-mini",
use_tools=True,
)
self.assertNotIn("extra_body", kwargs)
if __name__ == "__main__": if __name__ == "__main__":