eval-code

ryu-eval-code

Core-side code evaluators (unified-evaluator plan, P4): run a user-supplied (input, output, expected, vars) -> {score} function in an isolated runtime to score one eval case, then merge the real scores into the Gateway's eval-run response.

Role in the decomposition

An extracted Core capability crate, lifted verbatim from apps/core/src/eval_code. Code execution is a Core capability, so this stays a library crate on the Core hot path, compiled in by default and carrying zero dependency on apps/core. It builds on the sibling sandbox primitives (ryu-tool-exec, ryu-sandbox) rather than reaching back into the kernel.

Key API (src/lib.rs)

• run_code_evaluator(spec, case) -> CodeEvalOutcome — score a single case.
• merge_code_evaluators(...) — merge real scores into the Gateway eval-run response.
• CodeEvaluatorSpec / CaseInput / CodeEvalOutcome / CodeEvalLang — the evaluator contract (JS / Python).
• env_scrub.rs — child-env scrubbing for the evaluator subprocess.

Swap seam / backends

• JS — runs via the deny-all Deno backend (ryu-tool-exec::run_eval_js), gated by the tool-exec-deno feature forwarded to ryu-tool-exec. Default ON so cargo test -p ryu-eval-code exercises the JS path; with it OFF, run_js returns an honest executed: false skip.
• Python — runs via the swappable ryu-sandbox backend, or a host fallback.

Consumed as

Compiled-into-Core library crate (default path dependency); default = ["tool-exec-deno"].