[Ramcharan49]

When a tool accepts requests: list[EmailRequest] | list[SMSRequest], the null-default resolver currently selects the first array schema without checking its elements. With defaults of 2 and 5 retries respectively, [{"kind":"sms","retries":null}] becomes an SMS request with 2 retries: Pydantic chooses the SMS model after the wrong default has already been inserted, so validation succeeds with silently changed behavior.

Check items and positional prefixItems recursively when selecting a default-injection variant. This selects the SMS default of 5 and also handles nested lists and tuple arguments. If no variant matches the item schemas strictly, retain the previous shape-only fallback so existing Pydantic-coercible input can still receive defaults.

This follows the null-sentinel contract from merged #6444.

Validation on base 18c07ba0482d90093b361cb566b573a27e389515 (Python 3.12.14, Linux; dependencies installed from uv.lock):

• Final new regression suite against unchanged main: 4 failed, 6 passed. Each failure demonstrates the SMS tool receiving 2 instead of 5.
• Patched regression and related tool/session/schema suites: 417 passed across 12 modules. The new tests execute the prepared tools, validate list-shaped examples against their advertised strict schema, preserve genuine nullable values and input payloads, and cover explicit/omitted defaults plus non-strict numeric coercion.
• Repository-wide Ruff check and format check: pass (ruff 0.15.22, 1,207 files).
• Focused strict mypy: python -m mypy --follow-imports=silent --platform linux -m livekit.agents.llm.utils passes (mypy 2.3.0). Python 3.10 syntax parsing and git diff --check pass.
• The patch was reviewed in a separate pass and the checks above were rerun after the final edits.

CI on exact head 20974007bd6866e4090c36b3b6c69b18e346215b: Ruff, the release gate, and blockguard on Ubuntu/macOS/Windows pass. Unit-test collection and Python 3.13 all-plugin type checking fail on the pre-existing Thunderphone _shorten_call_id import after its upstream rename to _shorten_id; the Python 3.10 matrix job was cancelled. The relevant files are byte-identical to the tested base, and the import failure was reproduced in a clean worktree of that base. Existing upstream PR #8229 addresses it; this PR does not duplicate that fix.

Limits: the existing matcher remains a shape/literal heuristic, not a complete JSON Schema or Pydantic union validator. Ambiguous variants and variants requiring coercion retain the existing fallback behavior. No live providers or paid APIs were called. Full all-plugin type checking and the whole unit/audio suite were not run.

AI assistance: OpenAI Codex investigated, implemented and reviewed the patch, and executed the validation above. No personal human review is asserted. This PR is intentionally a draft; no CLA agreement has been accepted by this automated task.