uf-hy GitHub avatar

cpa-plugin-codexcomp

uf-hy

A CPA plugin that detects and fixes reasoning truncation in streaming requests for specific AI models (e.g., gpt-5.5) by automatically continuing the truncated thought process using encrypted_content replay, transparent to clients.

Stars

75

7-day growth

No data

Forks

7

Open issues

2

License

MIT

Last updated

2026-07-19

AI repository intelligence
FR-AI / ANALYSIS

Why it is worth attention

Directly addresses the known 518n−2 truncation pattern causing intelligence degradation in agent scenarios, with benchmarks showing accuracy improvement from 40% to 100% on candy evaluation tests.

Who it is for

  • AI agent users (Codex, Claude Code)
  • CPA proxy administrators
  • Developers deploying gpt-5.5/luna/terra models
  • Performance engineers optimizing reasoning depth

Use cases

  • Preventing truncation-related answer errors in coding agents
  • Improving reasoning quality for complex, multi-step problems
  • Serving as a transparent middleware without client configuration changes
  • A/B testing reasoning continuation via configurable parameters

Strengths

  • Integrated as a Go C-ABI shared library, zero overhead from separate proxy processes
  • Supports OpenAI Responses/Chat Completions and Anthropic Messages APIs with automatic protocol translation
  • Low first-byte latency (<500ms for simple questions) using async streaming
  • Open source (MIT) with clear credits and well-documented configuration

Considerations

  • Relies on non-contractual model behavior (518n−2 truncation pattern and encrypted_content field); may cease to work if OpenAI changes these
  • Only activates for configured models (default gpt-5.5, gpt-5.6-luna, gpt-5.6-terra); Terra has additional non-standard truncations not handled
  • Continuation rounds consume real tokens; metadata visibility limited to openai-response protocol
  • Experimental min_reasoning_tokens feature has poor real-world results

README quick start

安装

Description

CLIProxyAPI plugin that detects and repairs gpt-5.5 reasoning truncation (518n−2 pattern) via encrypted_content replay and multi-round folding

Related repositories

Similar projects matched by category, topics, and programming language.

nethical6
Featured
nethical6 GitHub avatar

conversation-steganography

A tool that hides encrypted secret messages inside innocent-looking AI-generated chat text for private communication over any messaging platform.

Go
1,149
programmersd21
Featured
programmersd21 GitHub avatar

flow

flow is a minimalist terminal dashboard that displays real-time network throughput with smooth animations, multiple responsive modes, and zero configuration.

Developer ToolsCLI & Terminal
277
MobAI-App
MobAI-App GitHub avatar

simslim

simslim reduces iOS simulator memory usage by 4x by disabling unnecessary background daemons, allowing many more simulators to run on a single Mac.

Go
248