Skip to content

fix(agent): use full tool schema for DeepSeek V4 #7862

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
imwyvern wants to merge 1 commit into from
Closed

fix(agent): use full tool schema for DeepSeek V4 #7862

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
26 changes: 24 additions & 2 deletions astrbot/core/agent/runners/tool_loop_agent_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -289,10 +289,12 @@ async def reset(
# Light tool schema does not include tool parameters.
# This can reduce token usage when tools have large descriptions.
# See #4681
self.tool_schema_mode = tool_schema_mode
self.tool_schema_mode = self._normalize_tool_schema_mode(
tool_schema_mode, provider, request
)
self._tool_schema_param_set = None
self._skill_like_raw_tool_set = None
if tool_schema_mode == "skills_like":
if self.tool_schema_mode == "skills_like":
tool_set = self.req.func_tool
if not tool_set:
return
Expand Down Expand Up @@ -322,6 +324,26 @@ async def reset(
self.stats = AgentStats()
self.stats.start_time = time.time()

@staticmethod
def _normalize_tool_schema_mode(
tool_schema_mode: str | None,
provider: Provider,
request: ProviderRequest,
) -> str | None:
if tool_schema_mode != "skills_like":
return tool_schema_mode

model = (request.model or provider.get_model() or "").lower().strip()
model_name = model.rsplit("/", 1)[-1]
if model_name not in {"deepseek-v4-flash", "deepseek-v4-pro"}:
Copy link
Copy Markdown

@chatgpt-codex-connector chatgpt-codex-connector Bot Apr 28, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Badge Handle DeepSeek V4 model aliases in schema normalization

The override only triggers when model_name is exactly deepseek-v4-flash or deepseek-v4-pro, so DeepSeek V4 identifiers with valid suffixes/prefix variants (for example provider-qualified or tier-suffixed IDs like deepseek-v4-flash:free) will miss this branch and remain in skills_like mode. In that case the runner still sends light tool schemas and can hit the same function-calling rejection this fix is meant to prevent.

Useful? React with 👍 / 👎.

return tool_schema_mode
Comment on lines +338 to +339
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist Bot Apr 28, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

Instead of hardcoding specific model variants, consider using a prefix check for deepseek-v4. This makes the logic more robust against future model releases within the same family (e.g., deepseek-v4-chat or simply deepseek-v4) that likely share the same tool schema limitations.

Suggested change
if model_name not in {"deepseek-v4-flash", "deepseek-v4-pro"}:
return tool_schema_mode
if not model_name.startswith("deepseek-v4"):
return tool_schema_mode


logger.info(
"DeepSeek V4 does not support skills-like light tool schemas; "
"using full tool schemas for function calling."
)
return "full"

def _read_tool_hint(self) -> str:
if self.read_tool is not None:
return f"`{self.read_tool.name}`"
Expand Down
34 changes: 34 additions & 0 deletions tests/test_tool_loop_agent_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1273,6 +1273,40 @@ async def text_chat(self, **kwargs) -> LLMResponse:
assert parts[0].text == "<image_caption>一张猫的照片</image_caption>"


@pytest.mark.asyncio
async def test_deepseek_v4_uses_full_tool_schema_instead_of_skills_like():
provider = MockProvider()
tool = FunctionTool(
name="test_tool",
description="测试",
parameters={"type": "object", "properties": {"query": {"type": "string"}}},
handler=AsyncMock(),
)
tool_set = ToolSet(tools=[tool])
req = ProviderRequest(
prompt="test",
func_tool=tool_set,
contexts=[],
model="deepseek-v4-flash",
Comment on lines +1286 to +1290
Copy link
Copy Markdown
Contributor

@sourcery-ai sourcery-ai Bot Apr 28, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

suggestion (testing): Consider adding a test where the model name comes from the provider instead of the request.

This path supports models from both request.model and provider.get_model(), but this test only covers the former. Please also cover the case where request.model is None and MockProvider.get_model() returns a DeepSeek V4 name (either via a separate test or parametrization) to ensure normalization works in that scenario.

Suggested implementation:

from unittest.mock import AsyncMock, Mock
@pytest.mark.asyncio
async def test_deepseek_v4_uses_full_tool_schema_instead_of_skills_like():
    provider = MockProvider()
    tool = FunctionTool(
        name="test_tool",
        description="测试",
        parameters={"type": "object", "properties": {"query": {"type": "string"}}},
        handler=AsyncMock(),
    )
    tool_set = ToolSet(tools=[tool])
    req = ProviderRequest(
        prompt="test",
        func_tool=tool_set,
        contexts=[],
        model="deepseek-v4-flash",
    )
    runner = ToolLoopAgentRunner()

    await runner.reset(
        provider=provider,
        request=req,
        run_context=ContextWrapper(context=None),
        tool_executor=cast(Any, MockToolExecutor()),
        agent_hooks=MockHooks(),
        tool_schema_mode="skills_like",
    )


@pytest.mark.asyncio
async def test_deepseek_v4_uses_full_tool_schema_when_model_from_provider():
    provider = MockProvider()
    # Ensure provider.get_model returns a DeepSeek V4 model name when request.model is None
    provider.get_model = Mock(return_value="deepseek-v4-flash")

    tool = FunctionTool(
        name="test_tool",
        description="测试",
        parameters={"type": "object", "properties": {"query": {"type": "string"}}},
        handler=AsyncMock(),
    )
    tool_set = ToolSet(tools=[tool])
    req = ProviderRequest(
        prompt="test",
        func_tool=tool_set,
        contexts=[],
        model=None,
    )
    runner = ToolLoopAgentRunner()

    await runner.reset(
        provider=provider,
        request=req,
        run_context=ContextWrapper(context=None),
        tool_executor=cast(Any, MockToolExecutor()),
        agent_hooks=MockHooks(),
        tool_schema_mode="skills_like",
    )

)
runner = ToolLoopAgentRunner()

await runner.reset(
provider=provider,
request=req,
run_context=ContextWrapper(context=None),
tool_executor=cast(Any, MockToolExecutor()),
agent_hooks=MockHooks(),
tool_schema_mode="skills_like",
)

assert runner.tool_schema_mode == "full"
assert runner.req.func_tool is tool_set
assert runner.req.func_tool.tools[0].parameters == tool.parameters
assert runner.req.func_tool.tools[0].handler is tool.handler
assert runner._tool_schema_param_set is None


@pytest.mark.asyncio
async def test_follow_up_accepted_when_active_and_not_stopping(
runner, mock_provider, provider_request, mock_tool_executor, mock_hooks
Expand Down