codex-ds-api

中文

Codex CLI DeepSeek adapter.

This project runs a local FastAPI server that accepts Codex CLI Responses API requests and translates them to DeepSeek Chat Completions requests. It then translates DeepSeek responses back into the Responses/SSE shape that Codex CLI expects.

It is not a full OpenAI Responses API implementation. The goal is to make Codex CLI usable with DeepSeek for normal coding-agent workflows.

Features

• POST /v1/responses endpoint for Codex CLI.
• Streaming SSE response events used by Codex CLI.
• Codex model name to DeepSeek model name mapping.
• previous_response_id state tracking with memory or SQLite backend.
• Function tool translation.
• Custom/freeform tool translation for tools such as apply_patch.
• Codex CLI compact item handling.
• DeepSeek reasoning_content round-trip support for tool-call workflows.
• Optional MCP-backed web search replacement for Codex web_search.
• Optional image input degradation to metadata or an external vision endpoint.

Requirements

• Python 3.12+
• DeepSeek API key
• Codex CLI

Install

python3 -m venv .venv
source .venv/bin/activate
pip install -e .

Configure

cp config.example.toml config.toml

Edit config.toml:

[deepseek]
api_key = "sk-your-deepseek-key"
base_url = "https://api.deepseek.com"
model = "deepseek-v4-pro"

For longer Codex sessions, SQLite state is recommended:

[state]
backend = "sqlite"
db_path = ".proxy-state/responses-state.sqlite3"
ttl_seconds = 604800
max_entries = 10000

Run

source .venv/bin/activate
uvicorn ds_responses_proxy.app:app --host 127.0.0.1 --port 8000

Or pass the DeepSeek key through the start script:

./start.sh sk-your-deepseek-key

Debug logs:

./start-debug.sh sk-your-deepseek-key

Health check:

curl http://127.0.0.1:8000/health

Codex CLI Provider

Configure Codex CLI to use this proxy as a Responses API provider:

model_provider = "ds-proxy"
model = "gpt-5.5"

[model_providers.ds-proxy]
name = "ds-proxy"
base_url = "http://127.0.0.1:8000/v1"
wire_api = "responses"
requires_openai_auth = true

Codex CLI still requires a non-empty OpenAI API key value, but this proxy uses config.toml or DEEPSEEK_API_KEY to call DeepSeek:

export OPENAI_API_KEY=not-used-by-proxy

Quick Test

OPENAI_API_KEY=not-used-by-proxy codex exec \
  -c 'model_provider="ds-proxy"' \
  -c 'model_providers.ds-proxy.name="ds-proxy"' \
  -c 'model_providers.ds-proxy.base_url="http://127.0.0.1:8000/v1"' \
  -c 'model_providers.ds-proxy.wire_api="responses"' \
  -c 'model_providers.ds-proxy.requires_openai_auth=true' \
  --model gpt-5.5 \
  "用一句话介绍你自己"

See user-guide.md for a short usage guide.