xyrrrrr-r/dsh-eval-control ↗★ 1

dsh-eval-control

Host-side Cordis control plugin for aeval: one-shot config injector (experiment variables, gateway-lease budgets, fork lineage, bundle descriptor), with a standalone status mode, a Web status chip, and a self-check command. 适合 aeval 评测任务,用于注入实验变量、控制预算及自检。

Package
dsh-eval-control
Compatibility
Unverified
Harness peer range
>=0.1.7-alpha.1 <0.1.8-0 || >=0.1.8-alpha.1 <0.2.0-0 || >=0.2.0-0 <0.3.0-0
Cordis peer range
^4.0.3
Version
0.2.2
License
Apache-2.0
Last updated
Oct 10, 2026

Install

$npx -p @deepseek-ai/dsh dsh plugin --profile web add github:xyrrrrr-r/dsh-eval-control

配置(EvalControlConfig)

{
  "run": {                        // 运行绑定:四元组互相锁死
    "run_id": "run-tbench-offline-1",
    "job_config_hash": "…",
    "config_file_sha256": "…",
    "runtime_lock_digest": "…"
  },
  "trialId": "trial-0",
  "sessionId": "…",
  "sessionRoot": "/…",            // 绝对路径、无符号链接,越界即拒
  "provider": "…", "model": "…", "reasoningEffort": "…",
  "maxSteps": 32, "maxTokens": 8192,
  "tools": { "allow": ["…"], "deny": ["…"] },
  "lineage": { "parentSessionId": "…", "parentTrialId": "…", "forkStep": 3 },
  "bundlePath": "/…/bundle.json",
  "gatewayUrl": "http://127.0.0.1:…",   // 宿主 broker
  "jobTokenFile": "/…/job-token",
  "refuseAuxiliaryCalls": true,          // 辅助调用默认拒绝
  "auxiliaryPolicy": { "…": "allow" }    // 按用途显式放行(优先于总开关)
}

配置按原样(as-authored)参与摘要计算,宿主组合的配置与沙箱内解析的配置 逐字段比对——两边不一致即失败,不做静默兼容。

除内联配置外,行配置还接受两种等价形态(解析结果与内联逐字段一致, configDigest 不变):{ "controlConfigPath": "/abs/config.json" } 指向 owner 写出的配置文件;或整行为空 {} 且设置 AEVAL_CONTROL_CONFIG。 两者都没有时进入 standalone:不挂载任何东西、不重定向模型调用, 只发布 evalControlStatus 状态——安装插件不会破坏一个普通 profile。