arcade-mcp/examples/mcp_servers/tools_with_output_schema/pyproject.toml
Eric Gustin 3204201360
fix: TypedDict total=False output breaks validation (#816)
When a tool’s output TypedDict uses total=False, MCP clients reject the
response with:
```
MCP error -32602: Structured content does not match the tool's output schema
```
Note that the bug also exists for the Engine transport
(/worker/tools/execute), but since the engine doesn't validate the
output schema, the bug never surfaced. This PR addresses the problem
holistically (MCP and Engine) in preparation for a future where the
Engine transport validates output schemas.

Two bugs combined to cause this:
1. Schema: The outputSchema had no required array and declared all
fields as strict types (e.g. "type": "string"), making every field look
mandatory and non-null.
2. Serialization: model_dump() on TypedDict-derived Pydantic models
emitted None for absent optional fields. A tool returning {"name":
"hello"} produced {"name": "hello", "optional_field": null} which is a
value the schema forbids.


<!-- CURSOR_SUMMARY -->
---

> [!NOTE]
> **Medium Risk**
> Adjusts core schema generation and MCP JSON Schema conversion for
TypedDicts, affecting how tool input/output contracts are emitted and
validated across clients; mistakes could break compatibility or
validation behavior.
> 
> **Overview**
> Fixes MCP/engine validation failures for `TypedDict(total=False)`
outputs by ensuring absent optional keys are **omitted from serialized
output** and that emitted schemas correctly describe **required vs
optional** keys.
> 
> `arcade-core` now tracks `required_keys`/`inner_required_keys` and
per-field `nullable` in `ValueSchema`, derives required sets from
TypedDict `__required_keys__`, and unwraps `Optional[T]` to support
optional nested TypedDicts; TypedDict-derived Pydantic models now
`model_dump(exclude_unset=True)` to avoid leaking missing fields as
`null`.
> 
> `arcade-mcp-server` JSON Schema conversion now emits `required` arrays
(including for arrays of objects), supports `nullable` by generating
`type: [<type>, "null"]` (and `enum` including `None`), and treats
nullable top-level objects as valid unwrapped output schemas. Adds
focused unit/end-to-end tests plus an expanded example server
demonstrating total-false, mixed required/optional, nullable, and
optional-nested TypedDict outputs, and bumps package
versions/dependencies accordingly.
> 
> <sup>Reviewed by [Cursor Bugbot](https://cursor.com/bugbot) for commit
53fe8365f613053599130520b75f30b614b465ca. Bugbot is set up for automated
code reviews on this repo. Configure
[here](https://www.cursor.com/dashboard/bugbot).</sup>
<!-- /CURSOR_SUMMARY -->
2026-04-09 17:47:57 -07:00

44 lines
1.2 KiB
TOML

[project]
name = "tools_with_output_schema"
version = "0.1.0"
description = "MCP Server created with Arcade.dev"
requires-python = ">=3.10"
dependencies = [
"arcade-mcp-server>=1.19.1,<2.0.0",
]
[project.optional-dependencies]
dev = [
"arcade-mcp[all]>=1.12.1,<2.0.0",
"pytest>=7.0.0",
"pytest-asyncio>=0.21.0",
"mypy>=1.0.0",
"ruff>=0.1.0",
]
# Tell Arcade.dev that this package has Arcade tools
[project.entry-points.arcade_toolkits]
toolkit_name = "tools_with_output_schema"
[build-system]
requires = ["hatchling"]
build-backend = "hatchling.build"
[tool.hatch.build.targets.wheel]
packages = ["src/tools_with_output_schema"]
[tool.ruff]
line-length = 100
target-version = "py312"
[tool.mypy]
python_version = "3.12"
warn_unused_configs = true
disallow_untyped_defs = false
# Uncomment the following if you are developing inside of the arcade-mcp repo & want to use editable mode
# Otherwise, you will install the following packages from PyPI
# [tool.uv.sources]
# arcade-mcp = { path = "../../../", editable = true }
# arcade-serve = { path = "../../../libs/arcade-serve/", editable = true }
# arcade-mcp-server = { path = "../../../libs/arcade-mcp-server/", editable = true }