SuTang-vain/dsh-plugins--plugins-prior-probe ↗★ 0
dsh-prior-probe
Frozen probe-battery tools for DeepSeek Harness: list batteries, offline deterministic regex scoring, and no-harness single-shot probing via any OpenAI-compatible endpoint.
安装
npx -p @deepseek-ai/dsh dsh plugin --profile web add github:SuTang-vain/dsh-plugins#261b971407deb0a133dff28b693c7890a17bda40&path:plugins/prior-probe说明文档
阅读完整 README ↗dsh-prior-probe
A frozen probe-battery tool for the DeepSeek Harness, packaged as a self-contained Host plugin that registers three model tools.
冻结探针电池工具(DeepSeek Harness 用),打包为自包含 Host 插件,注册三个模型工具。
What it does / 功能
For a frozen battery of questions, the tool measures how much of the expected answer a model can produce from prior knowledge alone — a no-harness, single-shot completion scored with deterministic grader-equivalent regexes.
对一套冻结的问题电池,测量模型仅凭先验知识能给出多少预期答案——无 harness、 单发补全 + 确定性正则评分。
| Tool / 工具 | Purpose / 用途 |
|---|---|
prior_probe_list | List the bundled frozen batteries (8 general probes, 6 MCP TypeScript SDK probes). |
prior_probe_score | Deterministic, offline regex scoring of response texts — audit the scoring or hand-score answers. |
prior_probe_run | Run a battery as no-harness single-shot completions (temperature 0, no tools, no system prompt) against any OpenAI-compatible /chat/completions endpoint. |
| 工具 | 用途 |
|---|---|
prior_probe_list | 列出内嵌冻结电池(8 条通用探针、6 条 MCP TypeScript SDK 探针)。 |
prior_probe_score | 离线确定性正则评分——审计评分逻辑或手工评分。 |
prior_probe_run | 以无 harness 单发方式(温度 0、无工具、无系统提示)对任意 OpenAI 兼容端点跑一套电池。 |
prior_probe_run requires an endpoint and model id; the Bearer key comes
from api_key (recorded in the session log — use a disposable key) or from
api_key_file (the file's first line, keeping the key out of the log). It
calls the model through the host web service; in a DSH runtime without a
registered web fetch provider the tool reports transport_error per probe
instead of crashing.
prior_probe_run 需要端点与模型 id;Bearer 密钥来自 api_key(会记录进会话
日志——请用一次性密钥)或 api_key_file(文件首行,密钥不进日志)。它通过
host 的 web 服务调用模型;在没有注册 web fetch provider 的 DSH 运行时,
工具会按探针返回 transport_error 而不会崩溃。