[Gareth.Y (linjiaye929)]

Summary

Keep saved evaluation parameters and JSON-schema defaults independent of each validation result. Currently, mutating a returned default list or nested saved configuration changes later calls to validate_parameters. Applying nested defaults also mutates caller-owned input, even when validation ultimately raises.

Copy dict/list containers before applying defaults and when injecting each default. Preserve opaque leaf values rather than applying arbitrary __deepcopy__ methods; local Pydantic/model/prompt parameter validation remains unchanged. This does not change sharing of parameter values between rows within one evaluation run.

Regression coverage

Pure local tests through validate_parameters and RemoteEvalParameters cover:

• Mutable defaults reused across validations.
• Nested defaults in arrays without modifying runtime inputs.
• Returned values without modifying saved parameter data.
• Opaque values whose __deepcopy__ must not run.
• Invalid input without partial default insertion into caller data.

Initial validation

Python 3.14.7, existing SDK test environment:

• Before: test_parameters.py: 4 failed, 10 passed, 10 skipped.
• After: 14 passed, 10 skipped. Skips require optional Pydantic, not installed in this environment.
• All seven whole-repository pre-commit hooks passed. Because mise is unavailable, invoked the underlying pre-commit run --all-files command directly with pinned existing tools rather than claiming literal make fixup execution.
• Production-file Pylint errors-only passed. Test-file Pylint reports eight missing-Pydantic imports; the unmodified baseline has the same eight errors.
• Ruff check/format and git diff --check passed.

The initial validation used no live service/provider calls, new dependencies, or full-core run. Full-core, provider/version/type-check matrices are not claimed as passed.

Follow-up validation (2026-10-31)

In an isolated Python 3.14.7 environment with the repository-locked Pydantic 2.14.1 and pydantic-core 2.50.1, the complete test_parameters.py passed all 24 tests with no skips. Pylint errors-only passed for both changed files; supplemental shared-container alias and opaque-lock checks passed. This resolves the earlier optional-Pydantic skips/import errors. The tests used the existing external-network guard; no outbound attempts were observed. The full-core suite was not retried, and provider/version matrices, including real Pydantic v1, remain untested.

AI assistance

Prepared with AI assistance. Independent AI code and evidence reviews found no blocking issues in the scoped change and follow-up validation. Human review and hosted CI are still pending.

Compatibility scope: ordinary dict/list containers are isolated; opaque leaf objects remain shared, and dict/list subclasses are normalized to ordinary containers. This does not introduce general support for cyclic parameter inputs.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-31T06:54:30.592854Z	7bd9d06	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "Codex (@codex) review" or "Codex (@codex) security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.