fix: reject prefixItems in strict tool schemas

[mikemikimike]

Summary

Reject prefixItems during strict JSON schema conversion so ordinary fixed-length tuple parameters fail during tool construction with guidance for supported homogeneous tuples and lists. Non-strict schemas retain their existing behavior.

Test plan

• uv run --frozen pytest tests/test_function_schema.py tests/test_strict_schema.py tests/test_strict_schema_oneof.py -q (130 passed)
• uv run --frozen ruff check src/agents/strict_schema.py tests/test_function_schema.py (passed)
• uv run --frozen ruff format --check src/agents/strict_schema.py tests/test_function_schema.py (passed)
• uv run --frozen pyright src/agents/strict_schema.py (0 errors)
• git diff --check (passed)

The full pytest suite and single-file mypy check did not complete in the local Windows environment; both remained in dependency/test execution without producing a result and were terminated. No production failure was observed.

Issue number

Closes #4752

Checks

• I've added new tests, if relevant
• I've run .agents/skills/code-change-verification/scripts/run.sh
• I've confirmed all verification steps pass
• If using Codex, I've run /review before submitting this PR

[XuQuanxin04]

Automated Codex review identified a compatibility edge case in the current approach: traversing every dict while excluding annotation keys can reject a valid object property literally named prefixItems (and can skip schemas beneath property names such as default/examples). I reproduced this with a strict schema containing properties: {"prefixItems": {"type": "string"}}.

The traversal should follow schema-valued positions and treat entries of properties, $defs, definitions, patternProperties, and dependentSchemas as subschemas regardless of their names. I prepared a focused regression test and corrected walker locally; commit 96cbb4a in my fork branch (push retry may be needed due transient network).

[mikemikimike]

Addressed the compatibility edge case from the review comment in commit 54b6e83c.

The strict-schema rejection now traverses only schema-valued JSON Schema positions, including properties, $defs, definitions, patternProperties, dependentSchemas, unions, and conditional keywords. Literal property names such as prefixItems, and annotation values under default/examples, are preserved without false positives, while an actual prefixItems schema is still rejected during construction.

Validation:

• uv run --frozen pytest tests/test_function_schema.py tests/test_strict_schema.py tests/test_strict_schema_oneof.py -q — 134 passed
• ruff format --check src/agents/strict_schema.py tests/test_strict_schema.py — passed
• ruff check src/agents/strict_schema.py tests/test_strict_schema.py — passed
• uv run --frozen pyright src/agents/strict_schema.py tests/test_strict_schema.py — 0 errors
• git diff --check — passed

The repository-level make typecheck is not runnable as written on Windows because its parallel shell recipe uses POSIX syntax; direct Pyright passed, while repository mypy remains blocked by the pre-existing missing numpy stubs in src/agents/voice/imports.py. The full test command was attempted with the Windows-compatible equivalent, but xdist hit a permissions error in the system pytest temp directory; the single-process focused and three-file suites pass.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 54b6e83cbb

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Reuse the existing strict-schema traversal

For the supported case introduced here, Pydantic emits prefixItems beneath properties, and _ensure_strict_json_schema() already visits that node recursively. This standalone walker and its three keyword tables therefore create a second JSON Schema interpreter solely to validate additional direct-helper shapes; the tables can already diverge from the repository's other schema traversal (for example, ToolOutputTrimmer recognizes legacy dependencies, while this walker does not). Keep the prefixItems check inside _ensure_strict_json_schema() and remove the traversal-only tables and tests so strict conversion retains one source of truth.

AGENTS.md reference: AGENTS.md:L94-L96

Useful? React with 👍 / 👎.

[mikemikimike]

Addressed in commit 359c446b: removed the standalone traversal tables and helper, and kept the prefixItems check inside _ensure_strict_json_schema(), which already follows the schema nodes emitted by Pydantic. The focused regression suite remains green: 130 passed.