[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-minimax-binder-agent-eval":3,"mdc--eu1zdd-key":37,"related-repo-minimax-binder-agent-eval":701,"related-org-minimax-binder-agent-eval":792},{"slug":4,"name":4,"fn":5,"description":6,"org":7,"tags":12,"stars":26,"repoUrl":27,"updatedAt":28,"license":29,"forks":30,"topics":31,"repo":32,"sourceUrl":35,"mdContent":36},"binder-agent-eval","evaluate LLM agent decision quality","当需要在相同的阶段输入上评测 LLM agent 的结构化决策、或对比多个 provider 的决策质量时使用；对应 openbinder evaluate-agent（--provider\u002F--compare\u002F--case）与 openbinder provider doctor。典型触发语：\"用 mock provider 跑一遍 decision case 评测\"\"对比一下几个 provider 在同一批 case 上的表现\"\"先 doctor 一下 minimax 配置通不通\"。同义场景：agent 评测、decision evaluation、visible\u002Fhidden 分层、泄露检查、provider 对比、冻结决策。",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},"minimax","MiniMax","https:\u002F\u002Fpexgzepcugksgbtrxkhf.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Forg-logos\u002Fminimax.jpg","MiniMax-AI",[13,17,20,23],{"name":14,"slug":15,"type":16},"Benchmarking","benchmarking","tag",{"name":18,"slug":19,"type":16},"LLM","llm",{"name":21,"slug":22,"type":16},"Evals","evals",{"name":24,"slug":25,"type":16},"Agents","agents",5,"https:\u002F\u002Fgithub.com\u002FMiniMax-AI\u002FMiniMax-Code-Plugins","2026-08-27T13:32:30.156635",null,2,[],{"repoUrl":27,"stars":26,"forks":30,"topics":33,"description":34},[],"Community registry and contribution toolkit for MiniMax Code plugins.","https:\u002F\u002Fgithub.com\u002FMiniMax-AI\u002FMiniMax-Code-Plugins\u002Ftree\u002FHEAD\u002Fplugins\u002FHylouis233\u002Fopen-protein-binder-design\u002Fskills\u002Fbinder-agent-eval","---\nname: binder-agent-eval\ndescription: >-\n  当需要在相同的阶段输入上评测 LLM agent 的结构化决策、或对比多个 provider 的决策质量时使用；对应 openbinder evaluate-agent（--provider\u002F--compare\u002F--case）与 openbinder provider doctor。典型触发语：\"用 mock provider 跑一遍 decision case 评测\"\"对比一下几个 provider 在同一批 case 上的表现\"\"先 doctor 一下 minimax 配置通不通\"。同义场景：agent 评测、decision evaluation、visible\u002Fhidden 分层、泄露检查、provider 对比、冻结决策。\nargument-hint: '[--provider mock|replay|anthropic|minimax|--compare a,b --case \u003Ccases.json> --out \u003Cdir>]'\nmetadata:\n  domains: [protein-binder, llm-evaluation, leakage-control]\n  last_reviewed: '2026-08-20'\n---\n\n# binder-agent-eval：agent 决策评测与 provider 对比\n\n## 目的\n\n让 MiniMax、Claude、ReplayProvider、MockProvider 在**完全相同的阶段输入**上产出结构化 JSON 决策，然后按协议遵守、证据使用、编造信号、不确定性处理、与 recorded decision 的字段级一致、token 与成本逐项评分。它回答的是\"研究决策能力\"的可比性，不是蛋白模型能力——本仓库永远不评测也不重跑任何蛋白模型。\n\n## 前置检查\n\n1. case 文件就绪（如 `evals\u002Fdecision-cases.json`）：每个 case 显式分 `visible`（可以进 prompt）与 `hidden`（仅评分侧：recorded decision、wet-lab 标签、final rank、每 case 唯一的 `TRAP-` 陷阱 token）。\n2. 选 provider：\n   - `mock` \u002F `replay`：**完全离线**，不需要任何环境变量；replay 的决策直接取自 case 文件 hidden 半的 recorded decision（评分通道，永不进 prompt）；\n   - `anthropic` \u002F `minimax`：需要 `OPENBINDER_LLM_*` 环境变量与网络；配置缺失时 CLI 打印结构化 JSON 错误并 exit 2——**不设默认 endpoint\u002Fmodel\u002Fkey**。MiniMax 的能力以 `provider doctor` 实测为准，不预设。\n3. 联网 provider 先体检：`openbinder provider doctor --provider minimax [--json]`（配置完整性、可达性、模型接受度、JSON 输出、长 prompt、usage 上报、超时与错误结构；任何 fail 或缺配置 exit 2）。\n4. 产物落盘：`output\u002Fbinder-agent-eval\u002F\u003Cslug>\u002F\u003Ctimestamp>\u002F`。\"evaluation plan complete\" 是五个人工 stage gate 之一：评测方案（case 集、provider 名单、评分维度）先经 stage-gate 批准再跑。stage-gate 技能来自 openscience 插件；未安装时直接向用户确认评测方案后再跑即可。\n\n## 1 · 泄露红线（先于一切）\n\n- `build_request` 在类型层面不接收 hidden 参数；泄露守卫在**任何模型调用之前**扫描组装好的 prompt：TRAP token、recorded action\u002Fprose、wet-lab\u002Frank JSON 片段，命中即 `LeakageError`，CLI **exit 3** 并打印泄露细节；\n- exit 3 是红线事件：停止、报告、排查 prompt 组装链路，**不得**删掉陷阱 token 重跑来\"通过\"。\n\n## 2 · 冻结先于评分\n\n每个决策在读入任何 recorded decision 之前，先冻结到 `frozen\u002F\u003Ccase_id>__\u003Cprovider>.json`（含 SHA-256 摘要）。评分永远针对冻结件，不接受\"评分时再补一口\"的决策。\n\n## 3 · 执行评测\n\n```bash\n# 离线单 provider\nopenbinder evaluate-agent --provider mock --case evals\u002Fdecision-cases.json --out output\u002Fbinder-agent-eval\u002F\u003Cslug>\u002F\u003Ctimestamp>\u002F\n# 多 provider 对比（逐provider顺序跑，合并一份报告）\nopenbinder evaluate-agent --compare replay,mock --case evals\u002Fdecision-cases.json --out \u003Cdir>\n```\n\n- 启发式评分覆盖：JSON 合法性、协议违反、证据覆盖、不确定性承认（token 集 Jaccard）、编造信号、与 recorded decision 的字段级一致；**不可计算的 case 排除并披露数量，永不猜测**；\n- 终止性非法输出记为 `invalid_output`，最多 2 次格式修复重试，不进入正式结果；越界 evidence ID 记为协议违反而非静默丢弃；\n- 产物：`agent-comparison-report.md` + `agent-eval-results.json`，含固定的 Method-and-limitations 节、无墙钟时间——相同输入 byte-identical。\n\n## 输出模板\n\n```markdown\n## 决策评测结果：\u003Ccase 文件> × \u003Cprovider 名单>\n\n- 每 provider：case 数、invalid_output 数、各评分维度要点\n- 冻结件：frozen\u002F 目录（SHA-256 已记录）\n- 泄露守卫：通过（未命中任何 TRAP\u002Frecorded 片段）\u002F exit 3 红线详情\n- 不可计算 case：\u003Cn 个，已排除并披露>\n- 局限：复述报告中的 Method-and-limitations 要点（离线对照 ≠ 真实模型能力）\n```\n\n## 本技能不做什么\n\n- 不评蛋白模型：评的是**研究决策**，不声称任何模型会设计 binder。\n- 不泄露 hidden：任何把 hidden 内容带进 prompt 的做法都是红线，无例外。\n- 不猜缺失配置：缺 env 就 exit 2，不造默认 key\u002Fendpoint。\n- 不把 mock\u002Freplay 的对照结果表述成真实模型的能力结论。\n\n## 收尾与下一步\n\n- 报告每 provider 的评分要点、冻结件位置与不可计算 case 数；联网 provider 结果标注模型标签与成本。\n- 下一步通常是 binder-ranking-audit 或 binder-report（把评测产物纳入最终报告束，内容归类为 `model-generated`）。\n",{"data":38,"body":46},{"name":4,"description":6,"argument-hint":39,"metadata":40},"[--provider mock|replay|anthropic|minimax|--compare a,b --case \u003Ccases.json> --out \u003Cdir>]",{"domains":41,"last_reviewed":45},[42,43,44],"protein-binder","llm-evaluation","leakage-control","2026-08-20",{"type":47,"children":48},"root",[49,58,64,78,83,227,233,281,287,300,306,486,535,540,634,639,669,674,695],{"type":50,"tag":51,"props":52,"children":54},"element","h1",{"id":53},"binder-agent-evalagent-决策评测与-provider-对比",[55],{"type":56,"value":57},"text","binder-agent-eval：agent 决策评测与 provider 对比",{"type":50,"tag":59,"props":60,"children":62},"h2",{"id":61},"目的",[63],{"type":56,"value":61},{"type":50,"tag":65,"props":66,"children":67},"p",{},[68,70,76],{"type":56,"value":69},"让 MiniMax、Claude、ReplayProvider、MockProvider 在",{"type":50,"tag":71,"props":72,"children":73},"strong",{},[74],{"type":56,"value":75},"完全相同的阶段输入",{"type":56,"value":77},"上产出结构化 JSON 决策，然后按协议遵守、证据使用、编造信号、不确定性处理、与 recorded decision 的字段级一致、token 与成本逐项评分。它回答的是\"研究决策能力\"的可比性，不是蛋白模型能力——本仓库永远不评测也不重跑任何蛋白模型。",{"type":50,"tag":59,"props":79,"children":81},{"id":80},"前置检查",[82],{"type":56,"value":80},{"type":50,"tag":84,"props":85,"children":86},"ol",{},[87,126,201,214],{"type":50,"tag":88,"props":89,"children":90},"li",{},[91,93,100,102,108,110,116,118,124],{"type":56,"value":92},"case 文件就绪（如 ",{"type":50,"tag":94,"props":95,"children":97},"code",{"className":96},[],[98],{"type":56,"value":99},"evals\u002Fdecision-cases.json",{"type":56,"value":101},"）：每个 case 显式分 ",{"type":50,"tag":94,"props":103,"children":105},{"className":104},[],[106],{"type":56,"value":107},"visible",{"type":56,"value":109},"（可以进 prompt）与 ",{"type":50,"tag":94,"props":111,"children":113},{"className":112},[],[114],{"type":56,"value":115},"hidden",{"type":56,"value":117},"（仅评分侧：recorded decision、wet-lab 标签、final rank、每 case 唯一的 ",{"type":50,"tag":94,"props":119,"children":121},{"className":120},[],[122],{"type":56,"value":123},"TRAP-",{"type":56,"value":125}," 陷阱 token）。",{"type":50,"tag":88,"props":127,"children":128},{},[129,131],{"type":56,"value":130},"选 provider：\n",{"type":50,"tag":132,"props":133,"children":134},"ul",{},[135,161],{"type":50,"tag":88,"props":136,"children":137},{},[138,144,146,152,154,159],{"type":50,"tag":94,"props":139,"children":141},{"className":140},[],[142],{"type":56,"value":143},"mock",{"type":56,"value":145}," \u002F ",{"type":50,"tag":94,"props":147,"children":149},{"className":148},[],[150],{"type":56,"value":151},"replay",{"type":56,"value":153},"：",{"type":50,"tag":71,"props":155,"children":156},{},[157],{"type":56,"value":158},"完全离线",{"type":56,"value":160},"，不需要任何环境变量；replay 的决策直接取自 case 文件 hidden 半的 recorded decision（评分通道，永不进 prompt）；",{"type":50,"tag":88,"props":162,"children":163},{},[164,170,171,176,178,184,186,191,193,199],{"type":50,"tag":94,"props":165,"children":167},{"className":166},[],[168],{"type":56,"value":169},"anthropic",{"type":56,"value":145},{"type":50,"tag":94,"props":172,"children":174},{"className":173},[],[175],{"type":56,"value":8},{"type":56,"value":177},"：需要 ",{"type":50,"tag":94,"props":179,"children":181},{"className":180},[],[182],{"type":56,"value":183},"OPENBINDER_LLM_*",{"type":56,"value":185}," 环境变量与网络；配置缺失时 CLI 打印结构化 JSON 错误并 exit 2——",{"type":50,"tag":71,"props":187,"children":188},{},[189],{"type":56,"value":190},"不设默认 endpoint\u002Fmodel\u002Fkey",{"type":56,"value":192},"。MiniMax 的能力以 ",{"type":50,"tag":94,"props":194,"children":196},{"className":195},[],[197],{"type":56,"value":198},"provider doctor",{"type":56,"value":200}," 实测为准，不预设。",{"type":50,"tag":88,"props":202,"children":203},{},[204,206,212],{"type":56,"value":205},"联网 provider 先体检：",{"type":50,"tag":94,"props":207,"children":209},{"className":208},[],[210],{"type":56,"value":211},"openbinder provider doctor --provider minimax [--json]",{"type":56,"value":213},"（配置完整性、可达性、模型接受度、JSON 输出、长 prompt、usage 上报、超时与错误结构；任何 fail 或缺配置 exit 2）。",{"type":50,"tag":88,"props":215,"children":216},{},[217,219,225],{"type":56,"value":218},"产物落盘：",{"type":50,"tag":94,"props":220,"children":222},{"className":221},[],[223],{"type":56,"value":224},"output\u002Fbinder-agent-eval\u002F\u003Cslug>\u002F\u003Ctimestamp>\u002F",{"type":56,"value":226},"。\"evaluation plan complete\" 是五个人工 stage gate 之一：评测方案（case 集、provider 名单、评分维度）先经 stage-gate 批准再跑。stage-gate 技能来自 openscience 插件；未安装时直接向用户确认评测方案后再跑即可。",{"type":50,"tag":59,"props":228,"children":230},{"id":229},"_1-泄露红线先于一切",[231],{"type":56,"value":232},"1 · 泄露红线（先于一切）",{"type":50,"tag":132,"props":234,"children":235},{},[236,269],{"type":50,"tag":88,"props":237,"children":238},{},[239,245,247,252,254,260,262,267],{"type":50,"tag":94,"props":240,"children":242},{"className":241},[],[243],{"type":56,"value":244},"build_request",{"type":56,"value":246}," 在类型层面不接收 hidden 参数；泄露守卫在",{"type":50,"tag":71,"props":248,"children":249},{},[250],{"type":56,"value":251},"任何模型调用之前",{"type":56,"value":253},"扫描组装好的 prompt：TRAP token、recorded action\u002Fprose、wet-lab\u002Frank JSON 片段，命中即 ",{"type":50,"tag":94,"props":255,"children":257},{"className":256},[],[258],{"type":56,"value":259},"LeakageError",{"type":56,"value":261},"，CLI ",{"type":50,"tag":71,"props":263,"children":264},{},[265],{"type":56,"value":266},"exit 3",{"type":56,"value":268}," 并打印泄露细节；",{"type":50,"tag":88,"props":270,"children":271},{},[272,274,279],{"type":56,"value":273},"exit 3 是红线事件：停止、报告、排查 prompt 组装链路，",{"type":50,"tag":71,"props":275,"children":276},{},[277],{"type":56,"value":278},"不得",{"type":56,"value":280},"删掉陷阱 token 重跑来\"通过\"。",{"type":50,"tag":59,"props":282,"children":284},{"id":283},"_2-冻结先于评分",[285],{"type":56,"value":286},"2 · 冻结先于评分",{"type":50,"tag":65,"props":288,"children":289},{},[290,292,298],{"type":56,"value":291},"每个决策在读入任何 recorded decision 之前，先冻结到 ",{"type":50,"tag":94,"props":293,"children":295},{"className":294},[],[296],{"type":56,"value":297},"frozen\u002F\u003Ccase_id>__\u003Cprovider>.json",{"type":56,"value":299},"（含 SHA-256 摘要）。评分永远针对冻结件，不接受\"评分时再补一口\"的决策。",{"type":50,"tag":59,"props":301,"children":303},{"id":302},"_3-执行评测",[304],{"type":56,"value":305},"3 · 执行评测",{"type":50,"tag":307,"props":308,"children":313},"pre",{"className":309,"code":310,"language":311,"meta":312,"style":312},"language-bash shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","# 离线单 provider\nopenbinder evaluate-agent --provider mock --case evals\u002Fdecision-cases.json --out output\u002Fbinder-agent-eval\u002F\u003Cslug>\u002F\u003Ctimestamp>\u002F\n# 多 provider 对比（逐provider顺序跑，合并一份报告）\nopenbinder evaluate-agent --compare replay,mock --case evals\u002Fdecision-cases.json --out \u003Cdir>\n","bash","",[314],{"type":50,"tag":94,"props":315,"children":316},{"__ignoreMap":312},[317,329,423,432],{"type":50,"tag":318,"props":319,"children":322},"span",{"class":320,"line":321},"line",1,[323],{"type":50,"tag":318,"props":324,"children":326},{"style":325},"--shiki-light:#90A4AE;--shiki-light-font-style:italic;--shiki-default:#546E7A;--shiki-default-font-style:italic;--shiki-dark:#676E95;--shiki-dark-font-style:italic",[327],{"type":56,"value":328},"# 离线单 provider\n",{"type":50,"tag":318,"props":330,"children":331},{"class":320,"line":30},[332,338,344,349,354,359,364,369,374,380,385,391,396,401,405,410,414,418],{"type":50,"tag":318,"props":333,"children":335},{"style":334},"--shiki-light:#E2931D;--shiki-default:#FFCB6B;--shiki-dark:#FFCB6B",[336],{"type":56,"value":337},"openbinder",{"type":50,"tag":318,"props":339,"children":341},{"style":340},"--shiki-light:#91B859;--shiki-default:#C3E88D;--shiki-dark:#C3E88D",[342],{"type":56,"value":343}," evaluate-agent",{"type":50,"tag":318,"props":345,"children":346},{"style":340},[347],{"type":56,"value":348}," --provider",{"type":50,"tag":318,"props":350,"children":351},{"style":340},[352],{"type":56,"value":353}," mock",{"type":50,"tag":318,"props":355,"children":356},{"style":340},[357],{"type":56,"value":358}," --case",{"type":50,"tag":318,"props":360,"children":361},{"style":340},[362],{"type":56,"value":363}," evals\u002Fdecision-cases.json",{"type":50,"tag":318,"props":365,"children":366},{"style":340},[367],{"type":56,"value":368}," --out",{"type":50,"tag":318,"props":370,"children":371},{"style":340},[372],{"type":56,"value":373}," output\u002Fbinder-agent-eval\u002F",{"type":50,"tag":318,"props":375,"children":377},{"style":376},"--shiki-light:#39ADB5;--shiki-default:#89DDFF;--shiki-dark:#89DDFF",[378],{"type":56,"value":379},"\u003C",{"type":50,"tag":318,"props":381,"children":382},{"style":340},[383],{"type":56,"value":384},"slu",{"type":50,"tag":318,"props":386,"children":388},{"style":387},"--shiki-light:#90A4AE;--shiki-default:#EEFFFF;--shiki-dark:#BABED8",[389],{"type":56,"value":390},"g",{"type":50,"tag":318,"props":392,"children":393},{"style":376},[394],{"type":56,"value":395},">",{"type":50,"tag":318,"props":397,"children":398},{"style":340},[399],{"type":56,"value":400},"\u002F",{"type":50,"tag":318,"props":402,"children":403},{"style":376},[404],{"type":56,"value":379},{"type":50,"tag":318,"props":406,"children":407},{"style":340},[408],{"type":56,"value":409},"timestam",{"type":50,"tag":318,"props":411,"children":412},{"style":387},[413],{"type":56,"value":65},{"type":50,"tag":318,"props":415,"children":416},{"style":376},[417],{"type":56,"value":395},{"type":50,"tag":318,"props":419,"children":420},{"style":340},[421],{"type":56,"value":422},"\u002F\n",{"type":50,"tag":318,"props":424,"children":426},{"class":320,"line":425},3,[427],{"type":50,"tag":318,"props":428,"children":429},{"style":325},[430],{"type":56,"value":431},"# 多 provider 对比（逐provider顺序跑，合并一份报告）\n",{"type":50,"tag":318,"props":433,"children":435},{"class":320,"line":434},4,[436,440,444,449,454,458,462,466,471,476,481],{"type":50,"tag":318,"props":437,"children":438},{"style":334},[439],{"type":56,"value":337},{"type":50,"tag":318,"props":441,"children":442},{"style":340},[443],{"type":56,"value":343},{"type":50,"tag":318,"props":445,"children":446},{"style":340},[447],{"type":56,"value":448}," --compare",{"type":50,"tag":318,"props":450,"children":451},{"style":340},[452],{"type":56,"value":453}," replay,mock",{"type":50,"tag":318,"props":455,"children":456},{"style":340},[457],{"type":56,"value":358},{"type":50,"tag":318,"props":459,"children":460},{"style":340},[461],{"type":56,"value":363},{"type":50,"tag":318,"props":463,"children":464},{"style":340},[465],{"type":56,"value":368},{"type":50,"tag":318,"props":467,"children":468},{"style":376},[469],{"type":56,"value":470}," \u003C",{"type":50,"tag":318,"props":472,"children":473},{"style":340},[474],{"type":56,"value":475},"di",{"type":50,"tag":318,"props":477,"children":478},{"style":387},[479],{"type":56,"value":480},"r",{"type":50,"tag":318,"props":482,"children":483},{"style":376},[484],{"type":56,"value":485},">\n",{"type":50,"tag":132,"props":487,"children":488},{},[489,501,514],{"type":50,"tag":88,"props":490,"children":491},{},[492,494,499],{"type":56,"value":493},"启发式评分覆盖：JSON 合法性、协议违反、证据覆盖、不确定性承认（token 集 Jaccard）、编造信号、与 recorded decision 的字段级一致；",{"type":50,"tag":71,"props":495,"children":496},{},[497],{"type":56,"value":498},"不可计算的 case 排除并披露数量，永不猜测",{"type":56,"value":500},"；",{"type":50,"tag":88,"props":502,"children":503},{},[504,506,512],{"type":56,"value":505},"终止性非法输出记为 ",{"type":50,"tag":94,"props":507,"children":509},{"className":508},[],[510],{"type":56,"value":511},"invalid_output",{"type":56,"value":513},"，最多 2 次格式修复重试，不进入正式结果；越界 evidence ID 记为协议违反而非静默丢弃；",{"type":50,"tag":88,"props":515,"children":516},{},[517,519,525,527,533],{"type":56,"value":518},"产物：",{"type":50,"tag":94,"props":520,"children":522},{"className":521},[],[523],{"type":56,"value":524},"agent-comparison-report.md",{"type":56,"value":526}," + ",{"type":50,"tag":94,"props":528,"children":530},{"className":529},[],[531],{"type":56,"value":532},"agent-eval-results.json",{"type":56,"value":534},"，含固定的 Method-and-limitations 节、无墙钟时间——相同输入 byte-identical。",{"type":50,"tag":59,"props":536,"children":538},{"id":537},"输出模板",[539],{"type":56,"value":537},{"type":50,"tag":307,"props":541,"children":545},{"className":542,"code":543,"language":544,"meta":312,"style":312},"language-markdown shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","## 决策评测结果：\u003Ccase 文件> × \u003Cprovider 名单>\n\n- 每 provider：case 数、invalid_output 数、各评分维度要点\n- 冻结件：frozen\u002F 目录（SHA-256 已记录）\n- 泄露守卫：通过（未命中任何 TRAP\u002Frecorded 片段）\u002F exit 3 红线详情\n- 不可计算 case：\u003Cn 个，已排除并披露>\n- 局限：复述报告中的 Method-and-limitations 要点（离线对照 ≠ 真实模型能力）\n","markdown",[546],{"type":50,"tag":94,"props":547,"children":548},{"__ignoreMap":312},[549,562,571,584,596,608,621],{"type":50,"tag":318,"props":550,"children":551},{"class":320,"line":321},[552,557],{"type":50,"tag":318,"props":553,"children":554},{"style":376},[555],{"type":56,"value":556},"## ",{"type":50,"tag":318,"props":558,"children":559},{"style":334},[560],{"type":56,"value":561},"决策评测结果：\u003Ccase 文件> × \u003Cprovider 名单>\n",{"type":50,"tag":318,"props":563,"children":564},{"class":320,"line":30},[565],{"type":50,"tag":318,"props":566,"children":568},{"emptyLinePlaceholder":567},true,[569],{"type":56,"value":570},"\n",{"type":50,"tag":318,"props":572,"children":573},{"class":320,"line":425},[574,579],{"type":50,"tag":318,"props":575,"children":576},{"style":376},[577],{"type":56,"value":578},"-",{"type":50,"tag":318,"props":580,"children":581},{"style":387},[582],{"type":56,"value":583}," 每 provider：case 数、invalid_output 数、各评分维度要点\n",{"type":50,"tag":318,"props":585,"children":586},{"class":320,"line":434},[587,591],{"type":50,"tag":318,"props":588,"children":589},{"style":376},[590],{"type":56,"value":578},{"type":50,"tag":318,"props":592,"children":593},{"style":387},[594],{"type":56,"value":595}," 冻结件：frozen\u002F 目录（SHA-256 已记录）\n",{"type":50,"tag":318,"props":597,"children":598},{"class":320,"line":26},[599,603],{"type":50,"tag":318,"props":600,"children":601},{"style":376},[602],{"type":56,"value":578},{"type":50,"tag":318,"props":604,"children":605},{"style":387},[606],{"type":56,"value":607}," 泄露守卫：通过（未命中任何 TRAP\u002Frecorded 片段）\u002F exit 3 红线详情\n",{"type":50,"tag":318,"props":609,"children":611},{"class":320,"line":610},6,[612,616],{"type":50,"tag":318,"props":613,"children":614},{"style":376},[615],{"type":56,"value":578},{"type":50,"tag":318,"props":617,"children":618},{"style":387},[619],{"type":56,"value":620}," 不可计算 case：\u003Cn 个，已排除并披露>\n",{"type":50,"tag":318,"props":622,"children":624},{"class":320,"line":623},7,[625,629],{"type":50,"tag":318,"props":626,"children":627},{"style":376},[628],{"type":56,"value":578},{"type":50,"tag":318,"props":630,"children":631},{"style":387},[632],{"type":56,"value":633}," 局限：复述报告中的 Method-and-limitations 要点（离线对照 ≠ 真实模型能力）\n",{"type":50,"tag":59,"props":635,"children":637},{"id":636},"本技能不做什么",[638],{"type":56,"value":636},{"type":50,"tag":132,"props":640,"children":641},{},[642,654,659,664],{"type":50,"tag":88,"props":643,"children":644},{},[645,647,652],{"type":56,"value":646},"不评蛋白模型：评的是",{"type":50,"tag":71,"props":648,"children":649},{},[650],{"type":56,"value":651},"研究决策",{"type":56,"value":653},"，不声称任何模型会设计 binder。",{"type":50,"tag":88,"props":655,"children":656},{},[657],{"type":56,"value":658},"不泄露 hidden：任何把 hidden 内容带进 prompt 的做法都是红线，无例外。",{"type":50,"tag":88,"props":660,"children":661},{},[662],{"type":56,"value":663},"不猜缺失配置：缺 env 就 exit 2，不造默认 key\u002Fendpoint。",{"type":50,"tag":88,"props":665,"children":666},{},[667],{"type":56,"value":668},"不把 mock\u002Freplay 的对照结果表述成真实模型的能力结论。",{"type":50,"tag":59,"props":670,"children":672},{"id":671},"收尾与下一步",[673],{"type":56,"value":671},{"type":50,"tag":132,"props":675,"children":676},{},[677,682],{"type":50,"tag":88,"props":678,"children":679},{},[680],{"type":56,"value":681},"报告每 provider 的评分要点、冻结件位置与不可计算 case 数；联网 provider 结果标注模型标签与成本。",{"type":50,"tag":88,"props":683,"children":684},{},[685,687,693],{"type":56,"value":686},"下一步通常是 binder-ranking-audit 或 binder-report（把评测产物纳入最终报告束，内容归类为 ",{"type":50,"tag":94,"props":688,"children":690},{"className":689},[],[691],{"type":56,"value":692},"model-generated",{"type":56,"value":694},"）。",{"type":50,"tag":696,"props":697,"children":698},"style",{},[699],{"type":56,"value":700},"html .light .shiki span {color: var(--shiki-light);background: var(--shiki-light-bg);font-style: var(--shiki-light-font-style);font-weight: var(--shiki-light-font-weight);text-decoration: var(--shiki-light-text-decoration);}html.light .shiki span {color: var(--shiki-light);background: var(--shiki-light-bg);font-style: var(--shiki-light-font-style);font-weight: var(--shiki-light-font-weight);text-decoration: var(--shiki-light-text-decoration);}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"items":702,"total":791},[703,719,726,739,750,763,772],{"slug":704,"name":704,"fn":705,"description":706,"org":707,"tags":708,"stars":26,"repoUrl":27,"updatedAt":718},"article2tasks","convert articles into Dida365 tasks","将技术周报（批量）或单篇文章整理为滴答清单待办事项，自动分类到已有清单并打标签，全程通过 MCP 写入，跨平台无需任何本地脚本。当用户提供周报内容、文章链接，或提及「周报」「文章汇总」「添加到滴答」时，使用此技能。",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[709,712,715],{"name":710,"slug":711,"type":16},"Content Creation","content-creation",{"name":713,"slug":714,"type":16},"MCP","mcp",{"name":716,"slug":717,"type":16},"Task Management","task-management","2026-08-18T03:49:38.751656",{"slug":4,"name":4,"fn":5,"description":6,"org":720,"tags":721,"stars":26,"repoUrl":27,"updatedAt":28},{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[722,723,724,725],{"name":24,"slug":25,"type":16},{"name":14,"slug":15,"type":16},{"name":21,"slug":22,"type":16},{"name":18,"slug":19,"type":16},{"slug":727,"name":727,"fn":728,"description":729,"org":730,"tags":731,"stars":26,"repoUrl":27,"updatedAt":738},"binder-handoff","export and validate binder task packages","当需要把 binder 设计算力任务打包交给外部平台（Claude Science 或自有服务器）、或校验\u002F导入外部返回的结果包时使用；对应 openbinder handoff export\u002Fvalidate\u002Fimport。典型触发语：\"把这个 target 打包成 handoff 任务包\"\"校验一下对方返回的 result package\"\"把外部算回来的结果导入进来\"。同义场景：external handoff、任务包导出、结果包校验、回传包导入、RUNBOOK。",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[732,735],{"name":733,"slug":734,"type":16},"Automation","automation",{"name":736,"slug":737,"type":16},"Data Engineering","data-engineering","2026-08-27T13:32:32.957864",{"slug":740,"name":740,"fn":741,"description":742,"org":743,"tags":744,"stars":26,"repoUrl":27,"updatedAt":749},"binder-protocol","compile and validate binder design protocols","当需要把 binder 设计协议草稿编译成结构化 protocol.yaml、校验协议、或按阶段切出 stage slice 时使用；对应 openbinder protocol validate\u002Fcompile\u002Fshow 三个子命令。典型触发语：\"把这份 protocol 草稿编译一下\"\"校验这个 protocol.yaml\"\"把 cofold_scoring 阶段的切片拿出来\"。同义场景：协议编译、protocol compiler、十阶段、stage slice、阶段上下文、protocol validate。",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[745,746],{"name":733,"slug":734,"type":16},{"name":747,"slug":748,"type":16},"YAML","yaml","2026-08-27T13:32:33.760978",{"slug":751,"name":751,"fn":752,"description":753,"org":754,"tags":755,"stars":26,"repoUrl":27,"updatedAt":762},"binder-ranking-audit","audit binder campaign ranking results","当需要重算并审计 binder campaign 的候选排序、核对 recorded 与 recomputed 是否一致时使用；对应 openbinder audit-ranking（--fixture\u002F--source）。典型触发语：\"审计一下这个 campaign 的排序\"\"recorded 和 recomputed 对得上吗\"\"重算 top-N 命中率和 Spearman\"。同义场景：ranking audit、排序审计、ensemble 重算、hit rate、缺失数据处理、missing data 报告。",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[756,759],{"name":757,"slug":758,"type":16},"Data Analysis","data-analysis",{"name":760,"slug":761,"type":16},"QA","qa","2026-08-27T13:32:34.155444",{"slug":764,"name":764,"fn":765,"description":766,"org":767,"tags":768,"stars":26,"repoUrl":27,"updatedAt":771},"binder-replay","replay binder design campaigns","当需要回放已记录的 binder 设计 campaign、重现每一步\"当时知道什么、决定了什么\"时使用；对应 openbinder replay（--fixture\u002F--events\u002F--source）。典型触发语：\"帮我回放一下 IL-7Ra 的 binder campaign\"\"回放那个失败的 campaign\"\"看看当时每一步都知道了什么\"。同义场景：campaign replay、事件流回放、replay report、gap 分析、时间线重建。",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[769,770],{"name":733,"slug":734,"type":16},{"name":757,"slug":758,"type":16},"2026-08-27T13:32:33.346398",{"slug":773,"name":773,"fn":774,"description":775,"org":776,"tags":777,"stars":26,"repoUrl":27,"updatedAt":790},"binder-report","generate final reports from agent artifacts","当需要把 output\u002F 下已产生的 replay、agent 评测、ranking 审计产物汇总成最终报告束时使用；收集既有 artifact、按六类证据分级标注、生成 limitations 并提交 stage-gate 审批。典型触发语：\"把这些产物汇总成最终报告\"\"出一份 binder 最终报告和 limitations\"\"把 replay\u002F评测\u002F审计结果打包成报告束\"。同义场景：最终报告、报告束、evidence grading、limitations、claim check、收尾报告。",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[778,781,784,787],{"name":779,"slug":780,"type":16},"Documentation","documentation",{"name":782,"slug":783,"type":16},"Operations","operations",{"name":785,"slug":786,"type":16},"Reporting","reporting",{"name":788,"slug":789,"type":16},"Summarization","summarization","2026-08-27T13:32:26.101404",93,{"items":793,"total":992},[794,818,832,851,866,886,904,922,937,953,972,982],{"slug":795,"name":795,"fn":796,"description":797,"org":798,"tags":799,"stars":815,"repoUrl":816,"updatedAt":817},"android-native-dev","develop Android native applications","Android native application development and UI design guide. Covers Material Design 3, Kotlin\u002FCompose development, project configuration, accessibility, and build troubleshooting. Read this before Android native application development.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[800,803,806,809,812],{"name":801,"slug":802,"type":16},"Accessibility","accessibility",{"name":804,"slug":805,"type":16},"Android","android",{"name":807,"slug":808,"type":16},"Kotlin","kotlin",{"name":810,"slug":811,"type":16},"Mobile","mobile",{"name":813,"slug":814,"type":16},"UI Components","ui-components",13030,"https:\u002F\u002Fgithub.com\u002FMiniMax-AI\u002Fskills","2026-07-13T06:16:54.247834",{"slug":819,"name":819,"fn":820,"description":821,"org":822,"tags":823,"stars":815,"repoUrl":816,"updatedAt":831},"buddy-sings","generate singing performances for AI companions","Use when user wants their Claude Code pet (\u002Fbuddy) to sing a song. Triggers on any request that combines the concept of their Claude Code buddy, pet, or companion with singing or music. Supports multilingual triggers — match equivalent phrases in any language.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[824,825,828],{"name":24,"slug":25,"type":16},{"name":826,"slug":827,"type":16},"Audio","audio",{"name":829,"slug":830,"type":16},"Creative","creative","2026-07-13T06:16:35.130644",{"slug":833,"name":833,"fn":834,"description":835,"org":836,"tags":837,"stars":815,"repoUrl":816,"updatedAt":850},"color-font-skill","select color palettes and font pairings","Choose presentation-ready color palettes and font pairings for PPT\u002Fdesign tasks. Use when users ask for visual theme choices, brand-safe palettes, or font recommendations. Triggers include: 配色, 色板, 字体, color palette, font, PPT配色, 字体搭配.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[838,841,844,847],{"name":839,"slug":840,"type":16},"Design","design",{"name":842,"slug":843,"type":16},"Presentations","presentations",{"name":845,"slug":846,"type":16},"Themes","themes",{"name":848,"slug":849,"type":16},"Typography","typography","2026-07-13T06:17:02.785587",{"slug":852,"name":852,"fn":853,"description":854,"org":855,"tags":856,"stars":815,"repoUrl":816,"updatedAt":865},"design-style-skill","select visual design systems for presentations","Select a consistent visual design system for PPT slides using radius\u002Fspacing style recipes. Use when users ask for overall style direction or component styling consistency. Includes Sharp\u002FSoft\u002FRounded\u002FPill recipes, component mappings, typography\u002Fspacing rules, and mixing guidance. Triggers: 风格, style, radius, spacing, 圆角, 间距, PPT风格, 视觉风格, design style, component style.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[857,858,861,864],{"name":839,"slug":840,"type":16},{"name":859,"slug":860,"type":16},"Design System","design-system",{"name":862,"slug":863,"type":16},"PowerPoint","powerpoint",{"name":842,"slug":843,"type":16},"2026-07-13T06:17:10.398389",{"slug":867,"name":867,"fn":868,"description":869,"org":870,"tags":871,"stars":815,"repoUrl":816,"updatedAt":885},"flutter-dev","build cross-platform apps with Flutter","Flutter cross-platform development guide covering widget patterns, Riverpod\u002FBloc state management, GoRouter navigation, performance optimization, and platform-specific implementations. Includes const optimization, responsive layouts, testing strategies, and DevTools profiling.\nUse when: building Flutter apps, implementing state management (Riverpod\u002FBloc), setting up GoRouter navigation, creating custom widgets, optimizing performance, writing widget tests, cross-platform development.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[872,875,878,879,882],{"name":873,"slug":874,"type":16},"Dart","dart",{"name":876,"slug":877,"type":16},"Flutter","flutter",{"name":810,"slug":811,"type":16},{"name":880,"slug":881,"type":16},"Performance","performance",{"name":883,"slug":884,"type":16},"State Management","state-management","2026-07-13T06:16:36.626679",{"slug":887,"name":887,"fn":888,"description":889,"org":890,"tags":891,"stars":815,"repoUrl":816,"updatedAt":903},"frontend-dev","build visually striking frontend web pages","Full-stack frontend development combining premium UI design, cinematic animations,\nAI-generated media assets, persuasive copywriting, and visual art. Builds complete,\nvisually striking web pages with real media, advanced motion, and compelling copy.\nUse when: building landing pages, marketing sites, product pages, dashboards,\ngenerating media assets (image\u002Fvideo\u002Faudio\u002Fmusic), writing conversion copy,\ncreating generative art, or implementing cinematic scroll animations.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[892,895,896,897,900],{"name":893,"slug":894,"type":16},"Animation","animation",{"name":829,"slug":830,"type":16},{"name":839,"slug":840,"type":16},{"name":898,"slug":899,"type":16},"Frontend","frontend",{"name":901,"slug":902,"type":16},"Web Development","web-development","2026-07-13T06:16:39.108827",{"slug":905,"name":905,"fn":906,"description":907,"org":908,"tags":909,"stars":815,"repoUrl":816,"updatedAt":921},"fullstack-dev","build full-stack web applications","Full-stack backend architecture and frontend-backend integration guide.\nTRIGGER when: building a full-stack app, creating REST API with frontend, scaffolding backend service,\nbuilding todo app, building CRUD app, building real-time app, building chat app,\nExpress + React, Next.js API, Node.js backend, Python backend, Go backend,\ndesigning service layers, implementing error handling, managing config\u002Fauth,\nsetting up API clients, implementing auth flows, handling file uploads,\nadding real-time features (SSE\u002FWebSocket), hardening for production.\nDO NOT TRIGGER when: pure frontend UI work, pure CSS\u002Fstyling, database schema only.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[910,913,914,917,920],{"name":911,"slug":912,"type":16},"Backend","backend",{"name":898,"slug":899,"type":16},{"name":915,"slug":916,"type":16},"Full-stack","full-stack",{"name":918,"slug":919,"type":16},"REST API","rest-api",{"name":901,"slug":902,"type":16},"2026-07-13T06:16:43.219005",{"slug":923,"name":923,"fn":924,"description":925,"org":926,"tags":927,"stars":815,"repoUrl":816,"updatedAt":936},"gif-sticker-maker","create animated GIF stickers from photos","Convert photos (people, pets, objects, logos) into 4 animated GIF stickers with captions.\nUse when: user wants to create cartoon stickers, GIF expressions, emoji packs, animated avatars,\nor convert photos to Funko Pop \u002F Pop Mart blind box style animations.\nTriggers: sticker, GIF, cartoon, emoji, expression pack, avatar animation.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[928,929,930,933],{"name":893,"slug":894,"type":16},{"name":829,"slug":830,"type":16},{"name":931,"slug":932,"type":16},"Images","images",{"name":934,"slug":935,"type":16},"Media","media","2026-07-13T06:16:51.74461",{"slug":938,"name":938,"fn":939,"description":940,"org":941,"tags":942,"stars":815,"repoUrl":816,"updatedAt":952},"ios-application-dev","develop iOS applications with SwiftUI and UIKit","iOS application development guide covering UIKit, SnapKit, and SwiftUI. Includes touch targets, safe areas, navigation patterns, Dynamic Type, Dark Mode, accessibility, collection views, common UI components, and SwiftUI design guidelines. For detailed references on specific topics, see the reference files.\nUse when: developing iOS apps, implementing UI, reviewing iOS code, working with UIKit\u002FSnapKit\u002FSwiftUI layouts, building iPhone interfaces, Swift mobile development, Apple HIG compliance, iOS accessibility implementation.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[943,944,947,948,951],{"name":801,"slug":802,"type":16},{"name":945,"slug":946,"type":16},"iOS","ios",{"name":810,"slug":811,"type":16},{"name":949,"slug":950,"type":16},"SwiftUI","swiftui",{"name":813,"slug":814,"type":16},"2026-07-13T06:16:55.686092",{"slug":954,"name":954,"fn":955,"description":956,"org":957,"tags":958,"stars":815,"repoUrl":816,"updatedAt":971},"minimax-docx","create and edit DOCX documents","Professional DOCX document creation, editing, and formatting using OpenXML SDK (.NET). Three pipelines: (A) create new documents from scratch, (B) fill\u002Fedit content in existing documents, (C) apply template formatting with XSD validation gate-check. MUST use this skill whenever the user wants to produce, modify, or format a Word document — including when they say \"write a report\", \"draft a proposal\", \"make a contract\", \"fill in this form\", \"reformat to match this template\", or any task whose final output is a .docx file. Even if the user doesn't mention \"docx\" explicitly, if the task implies a printable\u002Fformal document, use this skill.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[959,962,965,968],{"name":960,"slug":961,"type":16},"Documents","documents",{"name":963,"slug":964,"type":16},"DOCX","docx",{"name":966,"slug":967,"type":16},"Office","office",{"name":969,"slug":970,"type":16},"Templates","templates","2026-07-13T06:16:40.461868",{"slug":973,"name":973,"fn":974,"description":975,"org":976,"tags":977,"stars":815,"repoUrl":816,"updatedAt":981},"minimax-music-gen","generate music and audio tracks","Use when user wants to generate music, songs, or audio tracks. Triggers on any request involving music creation, song writing, lyrics generation, audio production, or covers. Also triggers when user provides lyrics and wants them turned into a song, or describes a mood\u002Fscene and wants background music. Supports multilingual triggers — match equivalent phrases in any language. Do NOT use for music playback of existing files, music theory questions, or music recommendation without generation.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[978,979,980],{"name":826,"slug":827,"type":16},{"name":829,"slug":830,"type":16},{"name":934,"slug":935,"type":16},"2026-07-13T06:16:50.381758",{"slug":983,"name":983,"fn":984,"description":985,"org":986,"tags":987,"stars":815,"repoUrl":816,"updatedAt":991},"minimax-music-playlist","generate personalized music playlists","Generate personalized music playlists by analyzing the user's music taste and generation feedback history. Triggers on any request involving playlist generation, music taste profiling, or personalized music recommendations. Supports multilingual triggers — match equivalent phrases in any language.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[988,989,990],{"name":826,"slug":827,"type":16},{"name":829,"slug":830,"type":16},{"name":934,"slug":935,"type":16},"2026-07-13T06:16:57.002997",141]