SuTang-vain/dsh-self-harness-tools--plugins-prior-probe0

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.

包名
dsh-prior-probe
版本
1.7.0
许可证
MIT
最近更新
2026年8月16日

安装

$npx -p @deepseek-ai/dsh dsh plugin --profile web add github:SuTang-vain/dsh-self-harness-tools#112d109fef14894ae95d7cc7d572384136115ff4&path:plugins/prior-probe

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_listList the bundled frozen batteries (8 general probes, 6 MCP TypeScript SDK probes).
prior_probe_scoreDeterministic, offline regex scoring of response texts — audit the scoring or hand-score answers.
prior_probe_runRun 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。

Install / 安装

Official bundle install (recommended) / 官方组合包安装(推荐)

The plugin ships as an official DSH bundle (package.json declares dsh.bundle; the layer lives in cordis.patch.yml; the entry is index.js). Install it into a profile with any of the three official channels:

插件以官方 DSH 组合包形式发布(package.json 声明 dsh.bundle,层定义在 cordis.patch.yml,入口为 index.js)。三种官方渠道任选其一安装进 profile:

# 1. from GitHub (source, plain JS — no build step needed)
dsh plugin --profile demo add github:SuTang-vain/dsh-self-harness-tools#path:plugins/prior-probe

# 2. from a packed tarball (no network)
dsh plugin --profile demo add ./dsh-prior-probe-1.5.0.tgz

# 3. from npm
dsh plugin --profile demo add dsh-prior-probe

Then verify the layer and start:

验证层并启动:

dsh --profile demo --dump-config   # shows the "# == dsh-prior-probe" layer
dsh --profile demo                 # or dsh --profile demo "run a probe battery"

Dynamic variant / 动态插件变体

plugin.host.js is the same instrument as a dynamic plugin body. In a DSH session, paste the whole file into the code.host parameter of a cordis_define call (new Plugin, any 3–6 letter id prefix), then cordis_run the returned packageId. Note: dynamic plugins are session-scoped and do not survive a process restart — prefer the bundle for anything persistent.

plugin.host.js 是同一工具的动态插件函数体。在 DSH 会话中把整个文件粘贴进 cordis_definecode.host 参数(新插件,任意 3–6 字母 id 前缀),然后对 返回的 packageId 执行 cordis_run。注意:动态插件是会话级的,进程重启后失效—— 需要持久安装请用上面的 bundle 方式。

There is no client half, no external dependency, and no runtime data read — both batteries are embedded, so the plugin works with the data/ directory deleted; those files ship only for provenance audit (node verify.js byte-checks the embedded copies against them).

无 client 半部分、无外部依赖、无运行时数据读取——两套电池均内嵌,删除 data/ 目录插件也能工作;这些文件仅用于溯源审计(node verify.js 会做字节比对)。

Bundled data / 内置数据

FileContent
data/prior-guessability-probes-v1.json8 general probes (frozen 2026-08-14)
data/prior-guessability-probes-ts-mcp-v1.json6 MCP TypeScript SDK probes (frozen 2026-08-14)

Batteries are frozen before use: do not edit a pattern without a recorded reason — the tooling is only meaningful against the frozen regexes.

电池冻结后使用:没有记录理由不要修改正则——工具只有针对冻结正则才有意义。

Example output / 示例输出

data/example-score-run.json is a deterministic offline demo produced by the shipped scoring code (no model calls; the responses are synthetic). It shows the exact output shape of prior_probe_score:

FieldDemo value
battery / frozen_at / stale_afterdsh-v1 / 2026-08-14 / 2026-11-14
scored5/8
G0.8
flags3 low-direction probes matched → flagged "above expected (prior-covered?)"

The demo input deliberately mixes matching and non-matching answers so both the pass logic and the low-direction flag logic are visible. For a live run against a real endpoint, call prior_probe_run with your own base_url/model/api_key_file.

data/example-score-run.json 是确定性离线演示,由随包发布的评分代码生成 (无模型调用;回答为合成样例),展示 prior_probe_score 的精确输出形态: dsh-v1 电池 5/8 计分、G=0.8,并演示了低方向探针命中时的 "above expected (prior-covered?)" 告警。对真实端点做实测请用 prior_probe_run 并传入你自己的 base_url/model/api_key_file