Surface Pydantic ValidationError to LLM in tool arg validation

Tool-argument validation failures raised as pydantic ValidationError are
now returned to the model as a clean, actionable message built from each
error's loc and msg, instead of the generic redacted text. All other
exceptions stay fully redacted.

The ValidationError handler is scoped to only the ptype.model_validate(args)
call that deserializes the tool arguments, so a ValidationError raised from
within a handler body is not surfaced and stays redacted by the broad
fallback like any other exception.

Safe to surface: the invalid values are arguments the LLM itself supplied,
and validator messages are authored by tool developers. The user-facing
text is assembled from only loc + msg; the full str(exc) (including raw
input) is kept in the debug-only error field.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

Author: idryzhov

python/copilot/tools.py

@@ -13,7 +13,7 @@
 from dataclasses import dataclass, field
 from typing import Any, Literal, TypeVar, get_type_hints, overload
 
-from pydantic import BaseModel
+from pydantic import BaseModel, ValidationError
 
 ToolResultType = Literal["success", "failure", "rejected", "denied", "timeout"]
 
@@ -211,7 +211,21 @@ async def wrapped_handler(invocation: ToolInvocation) -> ToolResult:
                 if takes_params:
                     args = invocation.arguments or {}
                     if ptype is not None and _is_pydantic_model(ptype):
-                        call_args.append(ptype.model_validate(args))
+                        try:
+                            call_args.append(ptype.model_validate(args))
+                        except ValidationError as exc:
+                            # Highlight input validation problems to the LLM.
+                            parts = []
+                            for err in exc.errors():
+                                loc = ".".join(map(str, err["loc"]))
+                                msg = err["msg"]
+                                parts.append(f"{loc}: {msg}" if loc else msg)
+                            return ToolResult(
+                                text_result_for_llm="Invalid tool arguments:\n" + "\n".join(parts),
+                                result_type="failure",
+                                error=str(exc),
+                                tool_telemetry={},
+                            )
                     else:
                         call_args.append(args)
                 if takes_invocation:

python/test_tools.py

@@ -3,7 +3,7 @@
 import json
 
 import pytest
-from pydantic import BaseModel, Field
+from pydantic import BaseModel, ConfigDict, Field, field_validator
 
 from copilot import define_tool
 from copilot.tools import (
@@ -197,6 +197,91 @@ def failing_tool(params: Params, invocation: ToolInvocation) -> str:
         # But the actual error is stored internally
         assert result.error == "secret error message"
 
+    async def test_validation_error_is_surfaced_to_llm(self):
+        class Params(BaseModel):
+            username: str
+
+            @field_validator("username")
+            @classmethod
+            def check_username(cls, v: str) -> str:
+                if v == "admin":
+                    raise ValueError("username 'admin' is reserved")
+                return v
+
+        @define_tool("validate", description="A validating tool")
+        def validating_tool(params: Params) -> str:
+            return "ok"
+
+        invocation = ToolInvocation(
+            session_id="s1",
+            tool_call_id="c1",
+            tool_name="validate",
+            arguments={"username": "admin"},
+        )
+
+        result = await validating_tool.handler(invocation)
+
+        assert result.result_type == "failure"
+        assert result.text_result_for_llm.startswith("Invalid tool arguments:")
+        assert "username 'admin' is reserved" in result.text_result_for_llm
+        # Full detail is retained in the debug field.
+        assert result.error is not None
+
+    async def test_validation_error_extra_forbid_includes_field_name(self):
+        class Params(BaseModel):
+            model_config = ConfigDict(extra="forbid")
+
+            request: str
+
+        @define_tool("strict", description="A strict tool")
+        def strict_tool(params: Params) -> str:
+            return "ok"
+
+        invocation = ToolInvocation(
+            session_id="s1",
+            tool_call_id="c1",
+            tool_name="strict",
+            arguments={"request": "ok", "extra_field": "unexpected"},
+        )
+
+        result = await strict_tool.handler(invocation)
+
+        assert result.result_type == "failure"
+        assert result.text_result_for_llm.startswith("Invalid tool arguments:")
+        # The offending key name is carried in `loc` even though the generic
+        # message is "Extra inputs are not permitted".
+        assert "extra_field" in result.text_result_for_llm
+        assert result.error is not None
+
+    async def test_validation_error_from_handler_body_is_redacted(self):
+        class Params(BaseModel):
+            pass
+
+        class Internal(BaseModel):
+            count: int
+
+        @define_tool("body", description="A tool that validates internally")
+        def body_tool(params: Params) -> str:
+            Internal.model_validate({"count": "secret-not-an-int"})
+            return "ok"
+
+        invocation = ToolInvocation(
+            session_id="s1",
+            tool_call_id="c1",
+            tool_name="body",
+            arguments={},
+        )
+
+        result = await body_tool.handler(invocation)
+
+        assert result.result_type == "failure"
+        # A ValidationError from the handler body must not be surfaced as an
+        # argument-validation error; it stays redacted like any other exception.
+        assert not result.text_result_for_llm.startswith("Invalid tool arguments:")
+        assert "secret-not-an-int" not in result.text_result_for_llm
+        assert "error" in result.text_result_for_llm.lower()
+        assert result.error is not None
+
     async def test_function_style_api(self):
         class Params(BaseModel):
             value: str