Repository navigation
Tool outputSchema is generated in validation mode, so serialization aliases and computed fields don't match structuredContent #3644
Description
On v2 (2118f14), a tool returning a model with a serialization_alias or a @computed_field advertises an outputSchema that its own structuredContent violates:
from pydantic import BaseModel, Field, computed_field
from mcp.server.mcpserver.utilities.func_metadata import func_metadata
class Out(BaseModel):
user_id: int = Field(serialization_alias="userId")
@computed_field
@property
def doubled(self) -> int:
return self.user_id * 2
def tool() -> Out:
return Out(user_id=3)
m = func_metadata(tool)
print(sorted(m.output_schema["properties"])) # ['user_id']
print(sorted(m.convert_result(Out(user_id=3)).structured_content)) # ['doubled', 'userId']Clients that validate structuredContent against outputSchema (as the spec says they should) reject these results. camelCase serialization aliases are common for JS-facing tools, so this hits anyone using them.
I had a fix for this in #3118 (closed in the v2 cleanup) and can port it to v2 with a regression test if you'd like to assign this to me.
Activity
Ben-Home commentedon Oct 5, 2026
The mismatch comes from generating the output schema in validation mode while the result is serialized in serialization mode, and aliases plus computed fields are exactly where the two diverge: user_id in the schema, userId and doubled in structuredContent.
For each mechanism:
Worth adding to the fix: a validation pass in the SDK that checks the produced structuredContent against the generated outputSchema before it leaves the server, enabled at least in the test suite. Clients that validate (as the spec tells them to) will reject a mismatch anyway; failing inside the SDK's own tests makes this class of bug impossible to ship silently.
Your #3118 work mapped exactly this area, so porting it with alias and computed-field cases as separate regression tests, plus one nested-model case where partial fixes usually show up, should close it cleanly.
jayzuccarelli commentedon Oct 6, 2026
@maxisbey since this is labeled for both v1 and v2: the fix I'd port is small. Generate the output schema with
mode="serialization"inFuncMetadata.model_post_init, and when a tool returns its ownCallToolResult, dump the validatedstructured_contentwithby_alias=Trueso it matches that schema. Regression test with a serialization alias and a computed field, on v2 first. Happy to take it if you assign it to me.Kludex commentedon Oct 10, 2026
This is the same validation-versus-serialization
outputSchemamismatch tracked in #3100, including serialization aliases and computed fields. Closing as a duplicate. AI-assisted triage; I reviewed both reports.