[ethan-scitix]

Type

• fix — bug fix or alignment correction

Summary

• Root cause: the configuration guide documents a model's extra as "top level or args", and the schema check validated both placements, but model binding read only args.extra. A top-level extra was dropped: Model.extra == {}, tasks rendered plain prompts, and the inference identity left it out. This affected the canonical path and both sglang_legacy paths (root and derived). Inline grader / extractor configs flatten args, so they already accepted both placements.
• Fix: model_config_extra (sieval/cli/resolution.py) reads either placement. Every binding path uses it, so check_extra, Model.extra, the inference identity and --dry-run (which plans through the same session setup) all see the top-level value.
• Precedence: setting both placements is refused. Named models let the top level win for other dual-placement keys and inline roles let args win, but none of those keys is identity; inline extra used to silently resolve to args.extra, so refusing both there is a behaviour change too (in the CHANGELOG). extra is part of the inference identity, so silently picking one of two values would record a capability set nobody chose. The refusal fires at schema validation and at named/inline normalization, before any engine starts. A derived model (base:) inherits its base's extra unless it sets its own. A task's infer_args.extra replaces the model's. Neither merges keys, as before.
• Identity: unchanged for configs that set args.extra. The model_request_args digests for a root and a derived model match main. A 0.12.0 result directory written from a top-level extra ran without it, so a resume now sees an inference change and is refused (noted in the CHANGELOG).
• Docs: the guide states the both-refused rule and the inheritance and replacement behaviour.

Related Issues

Closes #227

Test Plan

Automated

• Lint/format clean (ruff check && ruff format --check)
• Type check clean (ty check)
• Unit tests pass: pytest tests/unit tests/integration tests/acceptance -m "not stress and not benchmark" -n auto: 9488 passed, 1 failed. The failure is test_exec_quotebench_anchor, a known macOS-only failure. Coverage on unit + integration: 97.11% (gate 95%).
• pre-commit run -a, scripts/check_preflight.py, scripts/sync_package_stubs.py --check

New tests (tests/unit/cli/test_session.py::TestModelExtraPlacement, full session setup through binding):

• top-level and args bind the same Model.extra and model_request_args()["extra"], on a canonical and an sglang_legacy model;
• a derived base: inherits a top-level extra, and its own top-level extra replaces the base's (both paths);
• both placements on a root or a derived model are refused at prelaunch, before any client is built;
• a task's infer_args.extra replaces a top-level model extra on the task's model and leaves the named model unchanged;
• inline grader: either placement binds, and both are refused.

tests/unit/cli/test_resolution.py::TestModelConfigExtra: either placement, null on one side defers to the other, both refused.

tests/unit/cli/test_validation.py: the schema rejects both placements, and the --dry-run schema check catches a malformed top-level or args extra, rejects both, and passes a valid top-level one.

Ten of the new session tests fail on main. The four that pass are the inline-placement cases and the args baseline.

Manual

• Script check: model_request_args digests for an args.extra root and derived model are byte-identical before and after.

Checklist

Required (all PRs)

• PR title follows conventional format (type(scope): description)
• No internal paths, credentials, or personal info in committed files
• AI-generated code has AI-Generated Code - <model> (<provider>) in module docstring (no new modules)
• No new upper-layer dependencies added to core/
• Deleted code verified — no remaining call sites depend on it

If: Breaking Change

• Described what breaks and migration path in Summary: a config that sets extra at both placements is now refused; keep one.

🤖 Generated with Claude Code