fix(tools): carry element types and constraints into tool schemas

[planetf1]

Pull Request

Issue

Fixes #1693

Description

Tools taking lists, dicts or constrained values now get their arguments as declared, on every backend.

Before

• A list[int] or list[float] argument reached the tool as strings ([1, 2] became ['1', '2']), so sum(nums) raised a TypeError inside the user's own code.
• list[bool], list[list[int]] and lists of Pydantic models (discriminated unions included) failed validation on every call, and were passed to the tool unchecked with a warning.
• The model was never told what goes inside a list or dict. list[int], list[Point] and dict[str, int] went out as a bare array or object, and Field constraints were dropped.
• An explicit null for an optional parameter (limit: int | None = None) failed validation. The whole call then fell back to the raw arguments, so "count": "3" was no longer coerced to 3. (Found while working on fix(backends): parse Granite XML tool calls on the local HF path #1692.)
• A tool schema containing a boolean sub-schema (valid JSON Schema, e.g. properties: {"a": true} from an MCP server) made validation raise even in lenient mode. On HF the generation failed; on OpenAI-compatible backends the tool call was dropped or the error escaped.

After

• Element types arrive intact, and model or union elements are validated against their fields.
• The model sees items, additionalProperties and constraints such as minimum, maxLength and minItems. On Ollama only items gets through, because the ollama client drops the other keywords.
• Optional parameters accept null. A parameter with a real default (page_size: int = 10) still rejects it.
• A schema the validator can't build falls back to the model's arguments with a warning, as lenient mode documents. Strict mode still raises.

How: convert_function_to_ollama_tool() rebuilt simple parameters from four keys and discarded the rest of Pydantic's schema. It now carries a fixed set of JSON Schema keywords (_CARRIED_SCHEMA_KEYWORDS) and flattens any discriminated unions that exposes. validate_tool_arguments() reads an items schema with no type as list[Any] rather than list[str], treats a non-required field with no non-null default as nullable, and builds its model inside the existing try.

Not covered

• Unions with several non-null branches (int | list[int]) and tuple element types still reach the model without element types. Their validation is unaffected.
• limit: int | None = 5 still rejects null. The schema doesn't record nullability, so it is inferred from the default.
• The validator still doesn't enforce numeric or length bounds, or dict value types (unchanged).
• On Ollama, a Pydantic-model parameter still reaches the model without its fields. The ollama client strips properties, required and anyOf before the request (upstream Add self-referencing properties field to Tool.Property ollama/ollama-python#724, feat: citation requirement #725), so this needs an upstream fix and a dependency bump.

Testing

• Tests added to the respective file if code was changed

• New code has 100% coverage if code was added

• Ensure existing tests and github automation passes (a maintainer will kick off the github automation when the rest of the PR is populated)

• New test/backends/test_tool_collection_params_unit.py: 41 tests covering each case above through schema generation and validation. 35 of them fail on main. Validation tests use strict=True, so a pass can't come from the lenient fallback.

• test/backends/test_tool_validation_integration.py: test_optional_param_none now runs strict, plus three new null tests (three of the four fail on main) and six boolean sub-schema tests (the three lenient ones fail on main).

• uv run pytest test/ -m "not qualitative": 4739 passed, 21 skipped. ruff and mypy are clean.

Attribution

• AI coding assistants used

---

Adding a new component, requirement, sampling strategy, or tool?

If your PR adds or modifies one of the types below, check the matching box. A checklist of type-specific review items will be posted as a comment.

• Component
• Requirement
• Sampling Strategy
• Tool

NOTE: Please ensure you have an issue that has been acknowledged by a core contributor and routed you to open a pull request against this repository. Otherwise, please open an issue before continuing with this pull request.