oi-language
oi-language/oi
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/oi-language/oi --skill bench-oi命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Use when an Agent needs to statically validate an Oi module without compiling, rewriting, or executing it.
Use when an Agent needs deterministic typed receipt replay without invoking the target entry or effects.
Use when an Agent needs to check whether an Oi module can use one explicitly requested installed language version without changing files or executing the target.
基于 SOC 职业分类
正在显示 SKILL.md
| name | bench-oi |
| description | Use when an Agent needs to run a structured Oi conformance suite against an explicitly identified harness and model. |
Use the colocated oi.mod and main.oi through the local using-oi bootstrap. Load exact Oi 0.0.2 execution.md once, close the manifested graph, and never reread the base specification.
Bind inputs to func main(suite bench.Suite, identity bench.RunIdentity). Map CompileTarget(target,input,observe,forbid) to tool.compile: sealed compile, target code zero, evidence absent. Map unchanged InvokeTarget(target,input,observe,forbid) to agent.invoke: sealed fresh run, actual receipt/digest. Both set LanguageVersion from target's nearest exact oi.mod; CompileTarget cross-binds target/version/source only. Before InvokeTarget returns, reject the first mismatch in Target→LanguageVersion→Artifacts(manifest/source order)→Inputs→Mappings→Authorities→AuthoritiesIdentity→Policy→Entry→Terminal→ReplyOperationID→Reply.Digest as BENCH_EVIDENCE_MISMATCH|runtime|target:1:1|sealed invocation mismatch: <FieldPath>|[]. Map Reply to caller.reply.
Host-invoke main once and expose its one typed BenchmarkReport Reply unchanged. Do not Oi-call main, infer a result, encode mode in text or fixture names, reuse one host effect then discard evidence, accept expected prose as evidence, or copy benchmark/runtime algorithms. Supports 0.0.1/0.0.2.