fix(core): strictify schemas inside tuple prefixItems

[XuQuanxin04]

Summary

ensure_strict_json_schema recursed into items, anyOf, oneOf, allOf, properties, and $defs, but it never visited prefixItems. pydantic emits prefixItems for fixed-length tuple type hints, so a non-strict subschema hidden inside a tuple element was passed through unchanged.

Concretely, a dict[str, int] parameter is rejected at the top level because its additionalProperties is a schema rather than false:

additionalProperties should not be set for object types. This could be because you're using an older version of Pydantic

But the same type wrapped in a tuple — def f(data: tuple[dict[str, int], str]) — silently produced:

"prefixItems": [
  { "type": "object", "additionalProperties": { "type": "integer" } },
  { "type": "string" }
]

That schema is not valid for OpenAI Structured Outputs, so the tool/structured-output call is rejected by the API at request time.

Fix

Recurse into each prefixItems entry exactly like array items, so tuple elements are strictified consistently. $defs are still strictified at the root, variable-length tuples (tuple[int, ...]) already use items and were already handled, and primitive/model tuples remain valid.

Test plan

Added two tests to tests/test_strict_schema.py:

• test_prefix_items_entries_are_strictified — an object entry inside prefixItems gets additionalProperties: false and a required list.
• test_open_object_inside_prefix_items_is_rejected_like_top_level — an additionalProperties schema inside a tuple element raises the same UserError as the top-level case.

Verified directly with function_schema:

• tuple[dict[str, int], str] now raises the same UserError (before: silently non-strict).

• tuple[Coord, str], tuple[str, int, bool], tuple[Pt, Pt], and tuple[int, ...] all produce valid strict schemas.

• pytest tests/test_strict_schema.py tests/test_function_schema.py tests/test_function_tool.py — 172 passed

• pytest tests/models/test_openai_responses.py tests/models/test_openai_chatcompletions.py tests/models/test_openai_chatcompletions_stream.py — 317 passed

• ruff format --check and ruff check clean on changed files.

No issue number; self-contained strict-schema correctness fix.

Checks

• I've added new tests, if relevant
• I've run .agents/skills/code-change-verification/scripts/run.sh — could not run the bash script end-to-end on this Windows/Git-Bash host (POSIX process-group management via setsid/kill -PID); I ran its constituent steps directly (ruff format, ruff check, and the relevant pytest suites above).
• I've confirmed all verification steps pass

[seratch]

The prefixItems recursion is the correct implementation direction. Please add one caller-visible regression that starts from an actual annotated function or Pydantic model, runs through function_schema(), and verifies that an object nested in a fixed tuple is strictified. The current hand-written schema tests prove the helper recursion but not the public schema-generation path that exposed the bug.

[seratch]

Thanks for the focused strict-schema fix. #4696 now contains the same prefixItems traversal together with the originating fixed-tuple function-schema behavior and public-boundary regression coverage.

We should merge only one source of truth for this change, so I am closing this PR as superseded by #4696.