[weiguangli-io]

Summary

Fixes #4868

The dict origin branch in _parse_schema_from_parameter previously returned Schema(type=OBJECT) without any value type information for parameterized dict types like dict[str, str]. This caused Gemini to receive a bare {"type": "object"} with no field constraints, leading to:

• Empty/blank tool arguments ([{}])
• Infinite retry loops when combined with output_schema

Changes

• _function_parameter_parse_util.py: When the dict type has 2 type args (key and value), parse the value type and set additional_properties on the schema. For example, dict[str, str] now emits {"type": "OBJECT", "additionalProperties": {"type": "STRING"}} instead of just {"type": "OBJECT"}.

• test_build_function_declaration.py: Added tests for dict[str, str], dict[str, int], and list[dict[str, str]] to verify additional_properties is correctly set. Updated existing test_list to also assert on additional_properties for the dict items.

How it works

The fix follows the same pattern used by the list branch, which recursively calls _parse_schema_from_parameter on the element type to build schema.items. Similarly, the dict branch now recursively calls _parse_schema_from_parameter on the value type (args[1]) to build schema.additional_properties.

Test plan

• Existing tests pass (40/40 in test_build_function_declaration.py)
• New tests added for dict[str, str], dict[str, int], and list[dict[str, str]]
• Manual verification with the reproduction script from the issue

[rohityan]

Hi @weiguangli-io , Thank you for your contribution! We appreciate you taking the time to submit this pull request.
The _sanitize_schema_formats_for_gemini would likely delete the additional_properties added to _parse_schema_from_parameter.
Can you upload some screenshots/logs verifying the end to end testing and it won't get deleted.

[weiguangli-io]

@rohityan Thanks for the thorough review! I've investigated and here's the analysis:

_sanitize_schema_formats_for_gemini does intentionally strip additional_properties (line 167 in _gemini_schema_util.py), but this operates on a separate code path from the changes in this PR:

1. _parse_schema_from_parameter (this PR's path) — constructs types.Schema objects directly. When build_function_declaration() is called, it returns types.Schema objects with additional_properties correctly set. This is the path used for automatic function calling.

2. _to_gemini_schema → _sanitize_schema_formats_for_gemini — operates on JSON dict representations and strips additional_properties before sending to Gemini API, because Gemini rejects it. This is intentional and there's even an existing test (test_gemini_schema_util.py line 554) verifying this behavior.

These are independent paths. This PR adds additional_properties in path (1), which is correct — the types.Schema object carries the full schema information for automatic function calling. The Gemini sanitization in path (2) is a separate concern that only applies when converting schemas to Gemini's wire format.

Test evidence:

• All 3 new tests in test_build_function_declaration.py verify additional_properties is correctly set on the types.Schema objects: dict[str, str] → STRING, dict[str, int] → INTEGER, list[dict[str, str]] → nested STRING.
• The existing test_gemini_schema_util.py tests continue to pass, confirming no interference.

The behavior is consistent with how _parse_schema_from_parameter already handles other types.Schema fields that may later be filtered in the Gemini path.

[adk-bot]

Thank you @weiguangli-io for your contribution! 🎉

Your changes have been successfully imported and merged via Copybara in commit 93f57f4.

Closing this PR as the changes are now in the main branch.