LLM Request Logger — AI skill for Claude Code

llm-request-logger

[English](README.md) | [中文](README.zh-CN.md)

See exactly what your AI coding agent sends to the model on every turn — and what each piece of your own config costs you.

Formerly `claude-code-prompt-probe`. v1 was a fake endpoint for measuring Claude Code's system prompt. v2 adds a real forwarding proxy and two analysis commands that no plain logger has: **attribution** (which file, skill list or MCP server put those characters in your request) and **cache-prefix diff** (where exactly two requests stop sharing a cache).

Python standard library only. No install step.

Two ways to capture

  fake   claude ──▶ llmlog ──▶ (nothing)           free, no credentials, single turn
  proxy  claude ──▶ llmlog ──▶ api.anthropic.com   real sessions, replies, token usage
                       │
                       ▼
                 cap/*.json  ──▶  report   what is in one request, and where it came from
                             ──▶  diff     where two requests stop sharing a cache prefix
                             ──▶  compare  system prompt sections across models
                             ──▶  export   scrubbed copies you can publish

	fake	proxy
Request leaves your machine	no	yes, forwarded untouched
Costs quota	no	yes
Credentials	a dummy key is enough	your real login passes through, never written to disk
You get	the request	the request, the reply, real token usage incl. cache reads and writes
Use it for	measuring the vendor's prompt from a clean config, comparing models	auditing your real setup, debugging cache misses

Quick start

git clone https://github.com/libaoming/llm-request-logger.git && cd llm-request-logger

./llmlog.py proxy                      # terminal A
ANTHROPIC_BASE_URL=http://127.0.0.1:8787 ENABLE_TOOL_SEARCH=true claude    # terminal B

./llmlog.py report cap/.json     # read a capture

`llmlog.py fake` prints its own command,