cdxiaodong/dsh-llm-inspector

Unified LLM request/response inspector: reasoning-effort tuning, external-think export, traffic & bundle analysis.

Bundle verified Unknown JavaScript v0.1.1
Bundle verified

Listed

3

Dev

Bundle verified

Versionv0.1.1
LanguageJavaScript
LicenseUnknown
View on GitHub

Preview

Preview 1 of 2: cdxiaodong/dsh-llm-inspector
Preview 2 of 2: cdxiaodong/dsh-llm-inspector

What it does

Unified LLM request/response inspector: reasoning-effort tuning, external-think export, traffic & bundle analysis.

Best for

  • DSH developers comparing LLM traffic, reasoning blocks, tool calls, and per-session behavior.
  • Workflows that need configurable reasoning effort or an inspectable external-think trace.
  • Trusted development environments that need JSONL session auditing and traffic statistics.

Not ideal for

  • Workflows requiring guaranteed external-think capture; prompt-based think invocation depends on model compliance.
  • Environments where model content, code, reasoning, or credentials must not be written to local audit files.
  • Users who only need ordinary model responses and have no reasoning, traffic-analysis, or audit workflow.

README

dsh-llm-inspector

统一 LLM 请求/响应检查器 —— DeepSeek Harness 插件

一个插件收编四类能力,全部通过配置开关:reasoning 控制、外部思考(think)导出、流量/包分析、会话落盘审计。

功能

功能 实现机制 配置项
reasoning 控制 挂 agent/request waterfall,改写 LlmCallConfig.reasoningEffort(关推理 / 调 budget) reasoningEffort
外部思考导出 注册 think 工具(schema 自动流入请求体)+ systemPrompt 注入”先调 think”指令,再从 llm/stream 抓回 thoughts 参数 externalThinking
流量/包分析 挂 llm/stream waterfall 拦截每个 chunk,按类型统计 token/字符、捕获 reasoning 与 tool-call 块 trafficAnalysis
会话落盘审计 订阅 session/event + 命中的流式块,写入 audit.jsonl persistSession
外部思考(external thinking)说明

对明文返回 reasoning 的模型(DeepSeek 等开源模型),本插件直接从 llm/stream 抓 reasoning 块,零侵入。

对加密推理的模型(OpenAI / Anthropic / Gemini),原生推理被厂商加密藏起。本插件复刻 oh-my-pi 的 externalThinking 思路:

  1. reasoningEffort: 'none' 关掉原生隐藏推理;
  2. 注册一个 think 工具(description 为 “private scratchpad; not shown to user”),其 schema 经 ctx.tools.register() 自动进入发往 LLM 的请求体;
  3. 通过 systemPrompt.section() 注入”回答任何问题前必须先调用 think”的指令;
  4. 模型把推理写进 think 的 thoughts 参数,本插件在 llm/stream 拦截该 tool-call 块并抓回、落盘。

⚠️ 与 omp 的差异:DSH 核心 LlmCallConfig 刻意裁剪了 tool_choice,因此本插件用 system prompt 强制 替代 omp 的 API 层 tool_choice 强制。这意味着 think 触发是”软强制”,触发率取决于模型遵循度,不保证 100%。这是上架前需实测验证的已知边界。

安装

dsh plugin --profile <你的profile> add github:cdxiaodong/dsh-llm-inspector

配置

plugins:
  dsh-llm-inspector:
    reasoningEffort: 'none'      # 可选:'none' 关推理,或 adapter 支持的 'low'/'medium'/'high'
    externalThinking: true       # 是否启用 think 导出
    trafficAnalysis: true        # 是否做流量/包分析
    persistSession: true         # 是否落盘审计
    logDir: '.dsh-inspector'     # 落盘目录
    # thinkPrompt: '...'         # 可选,覆盖默认强制提示

服务

插件对外提供两个 Service(可被其他插件 inject):

  • inspectorTraffic —— 流量统计:snapshot() 返回 { calls, chunks, reasoningChars, thinkCalls, thinkChars, byKind }
  • inspectorAudit —— 审计:log(tag, payload) / flush()

开发

npm install
npm run build    # tsc 产出 lib/(已提交进 git,DSH 可直接从 Git 安装)
npm test         # node --test,7 个零依赖测试

安全与信任声明

按 DSH 社区约定:GitHub topics 不是安全审核,也不是官方背书。本插件会拦截并(在开启 persistSession 时)落盘你的模型请求/响应内容,其中可能包含推理过程、代码、甚至凭证。请:

  • 仅在可信环境开启 persistSession;
  • 注意 audit.jsonl 的存储位置与访问权限;
  • 安装前自审源码(src/index.ts,约 200 行)。

License

MIT

Frequently Asked QuestionsFAQ

Use the verified command dsh plugin --profile default add github:cdxiaodong/dsh-llm-inspector in a DSH-enabled shell. The command resolves the public package metadata and keeps the plugin attached to the catalog identity shown on this page.