Open-Open-Reasoning

A self-contained reproduction of the Open Reasoning demo — both the analysis of what Claude's thinking signature really is, and a working server side that unlocks the hidden chain-of-thought from it.

对 Open Reasoning 演示的一次自包含复现：既记录了「Claude 的 thinking signature 到底是什么」 的分析结论，也提供了一个可运行的服务端，用来从 signature 中解锁隐藏的思维链（CoT）。

Based on / inspired by the original Open Reasoning project by s-JoL: https://github.com/s-JoL/open-reasoning. That repo publishes the demo and unlocked examples but not the recovery implementation; this repo independently reverse-engineers and reproduces the server side.
本项目基于并致敬 s-JoL 的原始 Open Reasoning 项目： https://github.com/s-JoL/open-reasoning。原仓库公开了演示与解锁样例，但未公开其还原实现； 本仓库独立逆向并复现了服务端。

• ANALYSIS.md — full findings: the signature is a base64 protobuf envelope wrapping an AEAD-encrypted copy of the model's private reasoning, bound to the model name. Includes the decoded wire format, entropy measurements, and empirical harvest/replay results (secrets with special/multi-byte characters recovered verbatim).
  完整结论：signature 是一段 base64 的 protobuf 封装，内部包裹着模型私有推理的 AEAD 加密副本，并与模型名绑定。文档中给出了解码后的字段结构、熵值测量，以及实测的 harvest/replay 结果（含特殊字符、多字节字符的秘密被逐字符还原）。
• server.py — Flask backend implementing the exact /api/* contract the frontend uses.
  Flask 后端，实现前端所需的完整 /api/* 接口。
• static/index.html — the vendored demo frontend.
  内置（vendored）的演示前端。
• tools/decode_signature.py — standalone protobuf decoder for any signature.
  独立的 signature protobuf 解码器。

How it works / 工作原理

Two provider calls implement the whole demo (see ANALYSIS.md §4):

整个演示由两次对上游 provider 的调用构成（详见 ANALYSIS.md 第 4 节）：

1. Harvest — send a normal request with extended thinking. The provider returns a thinking block containing an opaque signature (the sealed reasoning) plus a visible answer. The plaintext reasoning never leaves the provider.
   Harvest（采集）：发送一次开启扩展思考（extended thinking）的普通请求。provider 会返回一个 thinking 块，其中包含不透明的 signature（被封存的推理）以及可见答案。明文推理永远不会离开 provider。
2. Replay (unseal) — paste that signature back as an assistant.thinking block in a fresh request and ask the model to recite its prior private reasoning. The provider decrypts the sealed blob server-side and the model surfaces it.
   Replay / Unseal（回放 / 解封）：把这个 signature 作为 assistant.thinking 块重新塞进一个 全新的请求里，再让模型复述它此前的私有推理。provider 会在服务端解密被封存的数据，模型随即将其还原出来。

The server reads the bound model name straight out of the signature header (via tools/decode_signature.py) so replays always use the matching model and pass the AEAD auth check.

服务端会直接从 signature 的头部读取绑定的模型名（通过 tools/decode_signature.py）， 因此每次回放都会使用匹配的模型，从而通过 AEAD 的完整性校验。

How the signature moves through the API / signature 如何在 API 中传递

The signature is never a top-level field of a request or response — it always rides inside a thinking content block. It enters and leaves the server at three distinct points:

signature 从来不是请求或响应的顶层字段——它始终藏在某个 thinking 内容块里。它在三个不同的 环节进出服务端：

1. Harvest — read out of the upstream response. POST /api/chat/send calls the Messages API with "thinking": {"type": "adaptive", "display": "omitted"} (see server.py), which tells the provider to seal the CoT into the signature instead of returning plaintext. _extract() then walks data["content"] and pulls block["signature"] from the thinking (or redacted_thinking) block.
   Harvest（采集）——从上游响应里读出。 POST /api/chat/send 调用 Messages API 时带上 "thinking": {"type": "adaptive", "display": "omitted"}（见 server.py），告诉 provider 把 CoT 封存进 signature 而不是返回明文。随后 _extract() 遍历 data["content"]，从 thinking （或 redacted_thinking）块中取出 block["signature"]。

2. Server-side only — the raw signature is not sent to the browser. In Live Conversation the harvested signature is stored in in-memory session state (sess["turns"][i]["signature"]) and written to logs/signatures/. The /api/chat/send JSON reply only exposes a boolean has_signature; the client later references a turn by index, never by the signature string.
   仅服务端持有——原始 signature 不会发给浏览器。 在 Live Conversation 中，采集到的 signature 被存进内存会话状态（sess["turns"][i]["signature"]）并落盘到 logs/signatures/。 /api/chat/send 的 JSON 响应只暴露一个布尔字段 has_signature；客户端之后是按轮次下标（turn index）引用某一轮，而不是传回 signature 字符串。

3. Replay — written back into an assistant.thinking block. _replay_prompt() rebuilds a messages array whose assistant turn contains {"type": "thinking", "thinking": "", "signature": sig} followed by a text block, then a fresh user turn asking the model to recite its working. The signature reaches this function two ways:

   • POST /api/byok/unseal — the browser supplies it directly in the request body as {"signature": "..."} (BYOK / Prove It Yourself).
   • POST /api/chat/unseal — the browser sends only {"session_id", "turn"}; the server looks the signature up from its own session state and replays it.
     Replay（回放）——写回一个 assistant.thinking 块。 _replay_prompt() 重建一个 messages 数组，其中 assistant 轮包含 {"type": "thinking", "thinking": "", "signature": sig} 以及一个 text 块，随后再接一个全新的 user 轮，要求模型复述其推理。signature 通过两条路径抵达该函数：
   • POST /api/byok/unseal——浏览器在请求体里直接以 {"signature": "..."} 传入（BYOK / Prove It Yourself）。
   • POST /api/chat/unseal——浏览器只发送 {"session_id", "turn"}；服务端从自己的会话状态里查出对应 signature 并回放。

Quick start / 快速开始

cd open-open-reasoning
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

cp .env.example .env      # then edit UPSTREAM_KEY / MODEL as needed / 然后按需修改 UPSTREAM_KEY / MODEL
set -a; source .env; set +a

python server.py          # serves the demo on http://localhost:8000 / 在此地址提供演示

Open http://localhost:8000 and use either Prove It Yourself (paste a signature to recover its sealed secret) or Live Conversation (chat, then "Unlock reasoning view").

打开 http://localhost:8000，即可使用 Prove It Yourself（粘贴一个 signature 以还原其中封存的秘密） 或 Live Conversation（先对话，再点击「Unlock reasoning view」解锁推理视图）。

If this environment lacks ensurepip / venv (as on this machine), install deps with pip install --user --break-system-packages -r requirements.txt instead.
若当前环境缺少 ensurepip / venv（本机即如此），可改用 pip install --user --break-system-packages -r requirements.txt 安装依赖。

API contract / API 接口约定

Method & path / 方法与路径	Purpose / 用途
GET /api/status	Per-IP daily quota + availability + running cost. / 按 IP 的每日配额、可用性与累计费用。
GET /api/pricing	Per-model price table + total USD spent. / 各模型价格表与累计花费（USD）。
GET /api/byok/template	Memorize-prompt template + model/limits for the curl builder. / 用于构建 curl 命令的记忆提示词模板及模型/上限。
POST /api/byok/unseal	Decode a submitted signature's bound model, replay it, return the recovered secret. / 解码提交的 signature 得到其绑定模型，回放并返回还原出的秘密。
POST /api/chat/send	Multi-turn chat; returns answer, summary, hidden-token count, has_signature. / 多轮对话；返回答案、摘要、隐藏 token 数、has_signature。
POST /api/chat/unseal	Replay a turn's signature to reveal the deeper reasoning trace. / 回放某一轮的 signature，揭示更深层的推理轨迹。
POST /api/chat/reset	Clear a session. / 清空一个会话。

Decode a signature by hand / 手动解码一个 signature

python tools/decode_signature.py @../replay_out/signature.txt

Configuration / 配置

All via environment variables (see .env.example): UPSTREAM_BASE, UPSTREAM_KEY, MODEL, PORT, PROVE_DAILY, LIVE_DAILY.

全部通过环境变量设置（见 .env.example）：UPSTREAM_BASE、UPSTREAM_KEY、MODEL、 PORT、PROVE_DAILY、LIVE_DAILY。

Notes & caveats / 说明与注意事项

• Elicitation is stochastic — a replay may occasionally return a refusal ("no string was shared"); retrying with the same signature usually succeeds.
  诱导（elicitation）是随机的：某次回放偶尔会返回拒绝（例如「没有提供任何字符串」）； 用同一个 signature 重试通常即可成功。
• Calls go to the official Anthropic Messages API (https://api.anthropic.com/v1/messages) authenticated with x-api-key; any Anthropic-compatible endpoint that forwards thinking blocks should also work.
  调用直接走 Anthropic 官方 Messages API（https://api.anthropic.com/v1/messages）， 使用 x-api-key 鉴权；任何能转发 thinking 块的、兼容 Anthropic 的端点同样可用。
• Every call is logged under LOG_DIR (default logs/): events.jsonl (conversation, signatures, recovered content, cost), raw/ (raw request+response JSON), signatures/ (harvested signatures). Per-call and running cost is computed from token usage and also returned in each API response and at GET /api/pricing.
  每次调用都会记录到 LOG_DIR（默认 logs/）：events.jsonl（对话、signature、还原内容、 费用）、raw/（原始请求+响应 JSON）、signatures/（采集到的 signature）。单次与累计费用会 根据 token 用量计算，并在每个 API 响应及 GET /api/pricing 中返回。
• Quotas and sessions are in-memory (single process). Use a shared store for real deployment.
  配额与会话均保存在内存中（单进程）。若要正式部署，请改用共享存储。