codex-app-transfer-proxy

# Codex App Transfer Proxy > Skill by [ara.so](https://ara.so) — Codex Skills collection. Codex App Transfer is a local desktop gateway written in Rust that proxies OpenAI Codex CLI/Desktop requests. It translates the Responses API protocol (`/responses`) into Chat Completions format, enabling Codex to work with third-party providers like DeepSeek, Kimi, Xiaomi MiMo, Zhipu GLM, Alibaba Bailian, MiniMax, Gemini, Claude, and Grok. The tool runs as a Tauri 2.x desktop app with a local HTTP server (default `127.0.0.1:18080`) that intercepts Codex requests, performs protocol translation (including multi-turn context, `previous_response_id` history replay, tool calls, and `apply_patch` diffs), and forwards to upstream providers. ## Installation **From Releases (Recommended)** Download the latest binary for your platform from [GitHub Releases](https://github.com/Cmochance/codex-app-transfer/releases/latest): - **Windows**: `Codex-App-Transfer-v<version>-Windows-x64-Setup.exe` (NSIS installer) - **macOS (Apple Silicon)**: `Codex-App-Transfer-v<version>-macOS-arm64.dmg` - **macOS (Intel)**: `Codex-App-Transfer-v<version>-macOS-x64.dmg` - **Linux (Debian/Ubuntu)**: `Codex-App-Transfer-v<version>-Linux-x86_64.deb` - **Linux (AppImage)**: `Codex-App-Transfer-v<version>-Linux-x86_64.AppImage` **Build from Source** ```bash git clone https://github.com/Cmochance/codex-app-transfer.git cd codex-app-transfer cargo tauri build --bundles app,dmg # macOS cargo tauri build --bundles nsis # Windows cargo tauri build --bundles deb,appimage # Linux ``` **Development Mode** ```bash cargo tauri dev # Hot-reload desktop app cargo test --workspace --lib # Run unit tests ``` ## Configuration ### Adding a Provider 1. Launch Codex App Transfer desktop app 2. Click **+** in the dashboard top-right corner 3. Select a preset (DeepSeek, Kimi, MiMo, etc.) or choose **Custom** 4. Fill in: - **API Base URL**: e.g., `https://api.deepseek.com/v1` - **API Key**: `$DEEPSEEK_API_KEY` (stored encrypted) - **API Format**: `chat_completions` (or `gemini_native`, `anthropic_messages`, etc.) 5. Click **Fetch Models** to populate available models 6. Map OpenAI model slots to real models: - `gpt-5.5` → `deepseek-v4-pro` - `gpt-5.4` → `deepseek-v4-mini` - `gpt-5.3-codex` → `kimi-k2.6` 7. Click **Apply** at the bottom of the page 8. Click **↻ Restart Codex** (top-right) to reload Codex Desktop/CLI ### Provider Configuration Examples **DeepSeek V4** ```toml [[providers]] name = "DeepSeek V4" api_base = "https://api.deepseek.com/v1" api_key = "${DEEPSEEK_API_KEY}" api_format = "chat_completions" enabled = true default_model = "deepseek-v4-pro" [providers.models] "gpt-5.5" = "deepseek-v4-pro" "gpt-5.4" = "deepseek-v4-mini" ``` **Kimi For Coding** ```toml [[providers]] name = "Kimi Code" api_base = "https://api.moonshot.cn/v1" api_key = "${KIMI_API_KEY}" api_format = "chat_completions" enabled = true default_model = "kimi-k2.6" [providers.models] "gpt-5.5" = "kimi-k2.6" "gpt-5.4" = "kimi-k1.5" ``` **Gemini Native (Google AI Studio)** ```toml [[providers]] name = "Gemini" api_base = "https://generativelanguage.googleapis.com" api_key = "${GOOGLE_AI_STUDIO_KEY}" api_format = "gemini_native" enabled = true default_model = "gemini-3-pro" [providers.models] "gpt-5.5" = "gemini-3-pro" "gpt-5.4" = "gemini-2.0-flash" ``` **Anthropic Messages (Claude)** ```toml [[providers]] name = "Claude" api_base = "https://api.anthropic.com" api_key = "${ANTHROPIC_API_KEY}" api_format = "anthropic_messages" enabled = true default_model = "claude-3-7-sonnet-20250219" [providers.models] "gpt-5.5" = "claude-3-7-sonnet-20250219" ``` ### Configuration File Location - **macOS/Linux**: `~/.codex-app-transfer/config.toml` - **Windows**: `%USERPROFILE%\.codex-app-transfer\config.toml` Config is auto-saved when you click **Apply** in the UI. Manual edits require app restart. ### Session Persistence Multi-turn history is stored in two layers: - **L1 (Memory)**: LRU cache for active sessions - **L2 (SQLite)**: `~/.codex-app-transfer/sessions.db` (30-day TTL) Sessions survive app restarts. To clear: ```bash rm ~/.codex-app-transfer/sessions.db ``` ## Using with Codex CLI/Desktop ### Starting the Proxy 1. Launch Codex App Transfer 2. Enable at least one provider (toggle switch on provider card) 3. Click **Apply** 4. Proxy server starts on `http://127.0.0.1:18080` Codex App Transfer auto-patches `~/.codex/config.toml` and `~/.codex/auth.json` to route through the local proxy. Original config is snapshotted before modification. ### Verifying Connection In Codex Desktop, the model picker should show: ``` DeepSeek V4 / deepseek-v4-pro Kimi Code / kimi-k2.6 ``` If you see OpenAI model names only (`gpt-5.5`, `gpt-5.4`), the proxy is not active — click **↻ Restart Codex**. ### Example Codex Interaction ```bash # In Codex CLI codex --model gpt-5.5 "Explain this function" # Routed to deepseek-v4-pro via local proxy ``` Codex Desktop chat will show tool calls (e.g., `read_file`, `apply_patch`) properly round-tripped through the proxy's Responses ↔ Chat Completions adapter. ## Protocol Translation Details ### Responses API → Chat Completions Codex sends: ```http POST /responses HTTP/1.1 Content-Type: application/json { "model": "gpt-5.5", "messages": [...], "stream": true, "thinking": {"type": "enabled", "budget": "high"} } ``` Proxy translates to: ```http POST https://api.deepseek.com/v1/chat/completions Content-Type: application/json Authorization: Bearer ${DEEPSEEK_API_KEY} { "model": "deepseek-v4-pro", "messages": [...], "stream": true, "reasoning_effort": "high" } ``` Upstream response is translated back to Responses format with: - `response_id` rewritten from `chatcmpl-...` → `resp_<base64>` - Tool calls bridged: `function_call` ↔ `custom_tool_call` - Thinking content injected as `reasoning_content` for Codex UI - `previous_response_id` maintained across turns for history replay ### Tool Call Handling (`apply_patch`) Codex's `apply_patch` tool (file edit diff UI) works on chat-completions providers via bidirectional adapter: **Codex → Upstream (Chat)** ```json { "type": "custom_tool_call", "name": "apply_patch", "arguments": "{\"file_path\":\"main.rs\",\"patch\":\"...\"}" } ``` becomes: ```json { "type": "function", "function": { "name": "apply_patch", "arguments": "{\"file_path\":\"main.rs\",\"patch\":\"...\"}" } } ``` **Upstream → Codex (Responses)** ```json { "tool_calls": [{ "type": "function", "function": {"name": "apply_patch", "arguments": "..."} }] } ``` becomes: ```json { "custom_tool_calls": [{ "type": "custom_tool_call", "name": "apply_patch", "arguments": "..." }] } ``` Patch format follows V4A standard: ``` *** Begin Patch @@ main.rs 10-15 - old line 1 - old line 2 + new line 1 + new line 2 *** End Patch ``` ### Thinking Mode (`reasoning_effort`) Codex thinking budget (`low` / `medium` / `high` / `xhigh`) maps to provider-specific fields: | Provider | `xhigh`/`max` | Other levels | |----------------|----------------------------|-------------------| | DeepSeek V4 | `reasoning_effort: "max"` | `"high"` | | Kimi/GLM/MiMo | (omit field) | (omit field) | | Custom Chat | clamp to `"high"` | pass through | Reasoning content (e.g., DeepSeek's `<think>` blocks) is extracted and injected as Codex `reasoning_content` field. ## Codex Desktop Theme Injection (Optional) Codex App Transfer can inject custom CSS + background images into Codex Desktop (Electron) via Chrome DevTools Protocol: 1. Navigate to **Theme** tab 2. Enable **Theme Injection** toggle 3. Select a preset (Changli, Azur Lane, Nailin, Zani, Carton) or **Add Custom** 4. For custom: upload JPG/PNG → crop in 1:1 modal → Apply 5. Theme auto-applies on next Codex Desktop launch **Custom Background Upload** ```rust // Internal: Theme::Custom stores base64-encoded image + crop rect #[derive(Serialize, Deserialize)] pub struct CustomTheme { pub image_data: String, // base64 JPEG/PNG pub crop_rect: CropRect, // {x, y, width, height} } ``` Injected CSS example: ```css body::before { content: ''; position: fixed; inset: 0; background: url(data:image/jpeg;base64,...) center/cover no-repeat; opacity: 0.3; z-index: -1; } .chat-panel { background: rgba(255, 255, 255, 0.85); backdrop-filter: blur(10px); } ``` Turn off **Theme Injection** toggle to restore native Codex UI. ## Codex Document Management Sidebar → **Codex** tab provides editors for Codex's AI-readable index files: ### Agents (`AGENTS.md`) - Reads/writes `AGENTS.md` in any directory (auto-detects nearest `.git/` for project-root vs. subdir classification) - UI: textarea editor + file picker - Save triggers Codex reload if running ```markdown # Agents ## Project Agent You are a Rust backend engineer working on codex-app-transfer... ## Frontend Agent You specialize in Tauri + Vue.js... ``` ### Memories - **Main index**: `~/.codex/memories/MEMORY.md` - **Summary**: `~/.codex/memories/memory_summary.md` - UI: two-tab editor for both files ```markdown # Memory ## Coding Preferences - Always use `anyhow::Result` for error handling - Prefer `tracing` over `log` ## Project Context - DeepSeek V4 requires `reasoning_effort: "max"` for xhigh budget ``` ### Skills - Scans `~/.codex/skills/<skill-name>/SKILL.md` - UI: list of skills with raw markdown editor per skill