Codex + GLM-5.2 Proxy

Make OpenAI Codex CLI work with GLM-5.2 — Zhipu AI's SOTA coding model. A lightweight local protocol proxy that translates Codex's Responses API into Zhipu's Chat Completions API. Zero dependencies — standard library only.

Why?

Codex CLI only supports OpenAI models natively. GLM-5.2 is a 753B-parameter open-weights model that beats GPT-5.5 on coding benchmarks at 1/6 the cost. This proxy bridges the two, letting you use Codex with GLM-5.2 through Zhipu's API.

Verified working: Codex successfully creates files, writes code, and executes commands through GLM-5.2 via this proxy.

Quick Start

1. Install Codex CLI

npm install -g @openai/codex

2. Get a Zhipu API Key

Register at open.bigmodel.cn → API Keys → Create.

3. Clone & Run the Proxy

git clone https://github.com/KevinSHH/codex-glm-proxy.git
cd codex-glm-proxy

# Linux / macOS
export GLM_API_KEY=your_key_here
python glm_proxy.py

# Windows (cmd)
set GLM_API_KEY=your_key_here
python glm_proxy.py

# Windows (PowerShell)
$env:GLM_API_KEY = "your_key_here"
python glm_proxy.py

4. Configure Codex

Add to ~/.codex/config.toml:

model = "glm-5.2"
model_provider = "glm-proxy"

[model_providers.glm-proxy]
name = "GLM Proxy"
wire_api = "responses"
base_url = "http://127.0.0.1:8787"
env_key = "DUMMY_API_KEY"

5. Start Coding!

codex --yolo exec "Write a FastAPI server with user authentication"

Architecture

┌──────────┐  Responses API   ┌───────────────┐  Chat API   ┌──────────────┐
│  Codex   │ ───────────────▶ │  glm_proxy.py │ ──────────▶ │  Zhipu AI    │
│   CLI    │ ◀─────────────── │    :8787      │ ◀────────── │  (GLM-5.2)   │
└──────────┘  SSE/JSON stream └───────────────┘  JSON resp  └──────────────┘

The proxy handles:

• Protocol translation: Responses API ↔ Chat Completions
• Tool conversion: Responses format → function-calling format
• Thinking control: Disables GLM-5.2's mandatory CoT (critical for tool calls)
• SSE streaming: Server-Sent Events with function_call events
• Model metadata: Full capability info for Codex compatibility

Auto-Start Wrapper

Drop one of these scripts in your PATH to automatically start the proxy before Codex:

Windows (codex-glm.bat):

@echo off
REM Start proxy if not already running
curl -s http://127.0.0.1:8787/health >nul 2>&1
if errorlevel 1 (
    echo Starting GLM Proxy...
    start /B pythonw "%~dp0\glm_proxy.py"
    timeout /t 3 /nobreak >nul
)
codex %*

Linux/macOS (codex-glm):

#!/bin/bash
# Start proxy if not already running
curl -s http://127.0.0.1:8787/health >/dev/null 2>&1 || {
    echo "Starting GLM Proxy..."
    python "$(dirname "$0")/glm_proxy.py" &
    sleep 2
}
exec codex "$@"

Then use codex-glm instead of codex:

codex-glm --yolo exec "Build a React dashboard"

Key Fixes

This proxy includes critical patches discovered through extensive debugging:

Issue	Symptom	Fix
Forced thinking	Reasoning consumes all tokens; tool calls never appear	enable_thinking: False
Missing reasoning field	Codex doesn't send reasoning param; fix never triggers	Fallback when effort is empty
SSE without function_call	Streaming only emits text events; tool calls lost	Added function_call event support in SSE stream
Proxy crash on disconnect	Codex reports "stream disconnected before completion"; proxy process dead	ThreadingHTTPServer + socket error handling in _sse_stream + handle_one_request override

Debug Logging (Optional)

Set GLM_PROXY_DEBUG to a file path to enable request/response logging:

# Windows
set GLM_PROXY_DEBUG=C:\Users\you\.codex\proxy_debug.log
python glm_proxy.py

# Linux/macOS
export GLM_PROXY_DEBUG=/tmp/proxy_debug.log
python glm_proxy.py

API Endpoints

Path	Method	Purpose
/health	GET	Health check ({"status":"ok","provider":"zhipu"})
/models	GET	Model list (Codex-compatible format)
/v1/chat/completions	POST	Direct Chat Completions proxy
/responses	POST	Responses API → Chat API conversion
/v1/responses	POST	Same with /v1/ prefix

Limitations

• Looping: Codex may loop on simple tasks — this is a Codex/model interaction issue, not a proxy bug. The file IS created correctly.
• Coding endpoint: The /api/coding/paas/v4 endpoint requires a separate Zhipu subscription plan. This proxy uses the general /api/paas/v4 endpoint.
• Reasoning levels: When thinking is enabled, GLM-5.2 uses significant tokens for reasoning before outputting content. The proxy disables this by default.

Changelog

v5.2 (Jun 2026)

• Crash-resistant: ThreadingHTTPServer + handle_one_request override + _sse_stream socket error handling
• Client disconnects mid-stream no longer kill the proxy process
• Debug logging made optional via GLM_PROXY_DEBUG env var
• Security: removed hardcoded API key placeholder

v5.1 (Jun 2026)

• Initial public release: Responses API → Chat API protocol translation
• enable_thinking: False fix for tool calling
• SSE streaming with function_call event support

Related Projects

• GodeX — Multi-provider Codex proxy (Go/TS)
• opencodex — Universal Codex proxy (TS)
• GLM-5.2 — Official GLM-5.2 model weights

License

MIT © 2026