docs: remove Amp CLI integration guides and update references

Merge pull request #549 from router-for-me/log
Improve Request Logging Efficiency and Standardize Error Responses
2026-02-10 00:10:51 +08:00 · 2025-12-15 23:50:56 +08:00 · 2025-12-15 20:43:12 +08:00 · 2025-12-15 20:35:52 +08:00 · 2025-12-15 18:00:32 +08:00 · 2025-12-15 17:45:16 +08:00
44 changed files with 1070 additions and 1356 deletions
--- a/README.md
+++ b/README.md
@@ -59,7 +59,7 @@ CLIProxyAPI includes integrated support for [Amp CLI](https://ampcode.com) and A
 - **Model mapping** to route unavailable models to alternatives (e.g., `claude-opus-4.5` → `claude-sonnet-4`)
 - Security-first design with localhost-only management endpoints

-**→ [Complete Amp CLI Integration Guide](docs/amp-cli-integration.md)**
+**→ [Complete Amp CLI Integration Guide](https://help.router-for.me/agent-client/amp-cli.html)**

 ## SDK Docs

--- a/README_CN.md
+++ b/README_CN.md
@@ -57,7 +57,7 @@ CLIProxyAPI 已内置对 [Amp CLI](https://ampcode.com) 和 Amp IDE 扩展的支
 - 智能模型回退与自动路由
 - 以安全为先的设计，管理端点仅限 localhost

-**→ [Amp CLI 完整集成指南](docs/amp-cli-integration_CN.md)**
+**→ [Amp CLI 完整集成指南](https://help.router-for.me/cn/agent-client/amp-cli.html)**

 ## SDK 文档

--- a/config.example.yaml
+++ b/config.example.yaml
@@ -136,8 +136,8 @@ ws-auth: false
 #   upstream-url: "https://ampcode.com"
 #   # Optional: Override API key for Amp upstream (otherwise uses env or file)
 #   upstream-api-key: ""
-#   # Restrict Amp management routes (/api/auth, /api/user, etc.) to localhost only (recommended)
-#   restrict-management-to-localhost: true
+#   # Restrict Amp management routes (/api/auth, /api/user, etc.) to localhost only (default: false)
+#   restrict-management-to-localhost: false
 #   # Force model mappings to run before checking local API keys (default: false)
 #   force-model-mappings: false
 #   # Amp Model Mappings
--- a/docs/amp-cli-integration.md
+++ b/docs/amp-cli-integration.md
@@ -1,443 +0,0 @@
-# Amp CLI Integration Guide
-
-This guide explains how to use CLIProxyAPI with Amp CLI and Amp IDE extensions, enabling you to use your existing Google/ChatGPT/Claude subscriptions (via OAuth) with Amp's CLI.
-
-## Table of Contents
-
- [Overview](#overview)
-  - [Which Providers Should You Authenticate?](#which-providers-should-you-authenticate)
- [Architecture](#architecture)
- [Configuration](#configuration)
-  - [Model Mapping Configuration](#model-mapping-configuration)
- [Setup](#setup)
- [Usage](#usage)
- [Troubleshooting](#troubleshooting)
-
-## Overview
-
-The Amp CLI integration adds specialized routing to support Amp's API patterns while maintaining full compatibility with all existing CLIProxyAPI features. This allows you to use both traditional CLIProxyAPI features and Amp CLI with the same proxy server.
-
-### Key Features
-
- **Provider route aliases**: Maps Amp's `/api/provider/{provider}/v1...` patterns to CLIProxyAPI handlers
- **Management proxy**: Forwards OAuth and account management requests to Amp's control plane
- **Smart fallback**: Automatically routes unconfigured models to ampcode.com
- **Model mapping**: Route unavailable models to alternatives you have access to (e.g., `claude-opus-4.5` → `claude-sonnet-4`)
- **Secret management**: Configurable precedence (config > env > file) with 5-minute caching
- **Security-first**: Management routes restricted to localhost by default
- **Automatic gzip handling**: Decompresses responses from Amp upstream
-
-### What You Can Do
-
- Use Amp CLI with your Google account (Gemini 3 Pro Preview, Gemini 2.5 Pro, Gemini 2.5 Flash)
- Use Amp CLI with your ChatGPT Plus/Pro subscription (GPT-5, GPT-5 Codex models)
- Use Amp CLI with your Claude Pro/Max subscription (Claude Sonnet 4.5, Opus 4.1)
- Use Amp IDE extensions (VS Code, Cursor, Windsurf, etc.) with the same proxy
- Run multiple CLI tools (Factory + Amp) through one proxy server
- Route unconfigured models automatically through ampcode.com
-
-### Which Providers Should You Authenticate?
-
-**Important**: The providers you need to authenticate depend on which models and features your installed version of Amp currently uses. Amp employs different providers for various agent modes and specialized subagents:
-
- **Smart mode**: Uses Google/Gemini models (Gemini 3 Pro)
- **Rush mode**: Uses Anthropic/Claude models (Claude Haiku 4.5)
- **Oracle subagent**: Uses OpenAI/GPT models (GPT-5 medium reasoning)
- **Librarian subagent**: Uses Anthropic/Claude models (Claude Sonnet 4.5)
- **Search subagent**: Uses Anthropic/Claude models (Claude Haiku 4.5)
- **Review feature**: Uses Google/Gemini models (Gemini 2.5 Flash-Lite)
-
-For the most current information about which models Amp uses, see the **[Amp Models Documentation](https://ampcode.com/models)**.
-
-#### Fallback Behavior
-
-CLIProxyAPI uses a smart fallback system:
-
-1. **Provider authenticated locally** (`--login`, `--codex-login`, `--claude-login`):
-   - Requests use **your OAuth subscription** (ChatGPT Plus/Pro, Claude Pro/Max, Google account)
-   - You benefit from your subscription's included usage quotas
-   - No Amp credits consumed
-
-2. **Provider NOT authenticated locally**:
-   - Requests automatically forward to **ampcode.com**
-   - Uses Amp's backend provider connections
-   - **Requires Amp credits** if the provider is paid (OpenAI, Anthropic paid tiers)
-   - May result in errors if Amp credit balance is insufficient
-
-**Recommendation**: Authenticate all providers you have subscriptions for to maximize value and minimize Amp credit usage. If you don't have subscriptions to all providers Amp uses, ensure you have sufficient Amp credits available for fallback requests.
-
-## Architecture
-
-### Request Flow
-
-```
-Amp CLI/IDE
-  ↓
-  ├─ Provider API requests (/api/provider/{provider}/v1/...)
-  │   ↓
-  │   ├─ Model configured locally?
-  │   │   YES → Use local OAuth tokens (OpenAI/Claude/Gemini handlers)
-  │   │   NO  ↓
-  │   │       ├─ Model mapping configured?
-  │   │       │   YES → Rewrite model → Use local handler (free)
-  │   │       │   NO  → Forward to ampcode.com (uses Amp credits)
-  │   ↓
-  │   Response
-  │
-  └─ Management requests (/api/auth, /api/user, /api/threads, ...)
-      ↓
-      ├─ Localhost check (security)
-      ↓
-      └─ Reverse proxy to ampcode.com
-          ↓
-          Response (auto-decompressed if gzipped)
-```
-
-### Components
-
-The Amp integration is implemented as a modular routing module (`internal/api/modules/amp/`) with these components:
-
-1. **Route Aliases** (`routes.go`): Maps Amp-style paths to standard handlers
-2. **Reverse Proxy** (`proxy.go`): Forwards management requests to ampcode.com
-3. **Fallback Handler** (`fallback_handlers.go`): Routes unconfigured models to ampcode.com
-4. **Secret Management** (`secret.go`): Multi-source API key resolution with caching
-5. **Main Module** (`amp.go`): Orchestrates registration and configuration
-
-## Configuration
-
-### Basic Configuration
-
-Add these fields to your `config.yaml`:
-
-```yaml
-# Amp upstream control plane (required for management routes)
-amp-upstream-url: "https://ampcode.com"
-
-# Optional: Override API key (otherwise uses env or file)
-# amp-upstream-api-key: "your-amp-api-key"
-
-# Security: restrict management routes to localhost (recommended)
-amp-restrict-management-to-localhost: true
-```
-
-### Model Mapping Configuration
-
-When Amp CLI requests a model that you don't have access to, you can configure mappings to route those requests to alternative models that you DO have available. This avoids consuming Amp credits for models you could handle locally.
-
-```yaml
-# Route unavailable models to alternatives
-amp-model-mappings:
-  # Example: Route Claude Opus 4.5 requests to Claude Sonnet 4
-  - from: "claude-opus-4.5"
-    to: "claude-sonnet-4"
-  
-  # Example: Route GPT-5 requests to Gemini 2.5 Pro
-  - from: "gpt-5"
-    to: "gemini-2.5-pro"
-  
-  # Example: Map older model names to newer versions
-  - from: "claude-3-opus-20240229"
-    to: "claude-3-5-sonnet-20241022"
-```
-
-**How it works:**
-
-1. Amp CLI requests a model (e.g., `claude-opus-4.5`)
-2. CLIProxyAPI checks if a local provider is available for that model
-3. If not available, it checks the model mappings
-4. If a mapping exists, the request is rewritten to use the target model
-5. The request is then handled locally (free, using your OAuth subscription)
-
-**Benefits:**
- **Save Amp credits**: Use your local subscriptions instead of forwarding to ampcode.com
- **Hot-reload**: Mappings can be updated without restarting the proxy
- **Structured logging**: Clear logs show when mappings are applied
-
-**Routing Decision Logs:**
-
-The proxy logs each routing decision with structured fields:
-
-```
-[AMP] Using local provider for model: gemini-2.5-pro          # Local provider (free)
-[AMP] Model mapped: claude-opus-4.5 -> claude-sonnet-4        # Mapping applied (free)
-[AMP] Forwarding to ampcode.com (uses Amp credits) - model_id: gpt-5  # Fallback (costs credits)
-```
-
-### Secret Resolution Precedence
-
-The Amp module resolves API keys using this precedence order:
-
-| Source | Key | Priority | Cache |
-|--------|-----|----------|-------|
-| Config file | `amp-upstream-api-key` | High | No |
-| Environment | `AMP_API_KEY` | Medium | No |
-| Amp secrets file | `~/.local/share/amp/secrets.json` | Low | 5 min |
-
-**Recommendation**: Use the Amp secrets file (lowest precedence) for normal usage. This file is automatically managed by `amp login`.
-
-### Security Settings
-
-**`amp-restrict-management-to-localhost`** (default: `true`)
-
-When enabled, management routes (`/api/auth`, `/api/user`, `/api/threads`, etc.) only accept connections from localhost (127.0.0.1, ::1). This prevents:
- Drive-by browser attacks
- Remote access to management endpoints
- CORS-based attacks
- Header spoofing attacks (e.g., `X-Forwarded-For: 127.0.0.1`)
-
-#### How It Works
-
-This restriction uses the **actual TCP connection address** (`RemoteAddr`), not HTTP headers like `X-Forwarded-For`. This prevents header spoofing attacks but has important implications:
-
- ✅ **Works for direct connections**: Running CLIProxyAPI directly on your machine or server
- ⚠️ **May not work behind reverse proxies**: If deploying behind nginx, Cloudflare, or other proxies, the connection will appear to come from the proxy's IP, not localhost
-
-#### Reverse Proxy Deployments
-
-If you need to run CLIProxyAPI behind a reverse proxy (nginx, Caddy, Cloudflare Tunnel, etc.):
-
-1. **Disable the localhost restriction**:
-   ```yaml
-   amp-restrict-management-to-localhost: false
-   ```
-
-2. **Use alternative security measures**:
-   - Firewall rules restricting access to management routes
-   - Proxy-level authentication (HTTP Basic Auth, OAuth)
-   - Network-level isolation (VPN, Tailscale, Cloudflare Access)
-   - Bind CLIProxyAPI to `127.0.0.1` only and access via SSH tunnel
-
-3. **Example nginx configuration** (blocks external access to management routes):
-   ```nginx
-   location /api/auth { deny all; }
-   location /api/user { deny all; }
-   location /api/threads { deny all; }
-   location /api/internal { deny all; }
-   ```
-
-**Important**: Only disable `amp-restrict-management-to-localhost` if you understand the security implications and have other protections in place.
-
-## Setup
-
-### 1. Configure CLIProxyAPI
-
-Create or edit `config.yaml`:
-
-```yaml
-port: 8317
-auth-dir: "~/.cli-proxy-api"
-
-# Amp integration
-amp-upstream-url: "https://ampcode.com"
-amp-restrict-management-to-localhost: true
-
-# Other standard settings...
-debug: false
-logging-to-file: true
-```
-
-### 2. Authenticate with Providers
-
-Run OAuth login for the providers you want to use:
-
-**Google Account (Gemini 2.5 Pro, Gemini 2.5 Flash, Gemini 3 Pro Preview):**
-```bash
-./cli-proxy-api --login
-```
-
-**ChatGPT Plus/Pro (GPT-5, GPT-5 Codex):**
-```bash
-./cli-proxy-api --codex-login
-```
-
-**Claude Pro/Max (Claude Sonnet 4.5, Opus 4.1):**
-```bash
-./cli-proxy-api --claude-login
-```
-
-Tokens are saved to:
- Gemini: `~/.cli-proxy-api/gemini-<email>.json`
- OpenAI Codex: `~/.cli-proxy-api/codex-<email>.json`
- Claude: `~/.cli-proxy-api/claude-<email>.json`
-
-### 3. Start the Proxy
-
-```bash
-./cli-proxy-api --config config.yaml
-```
-
-Or run in background with tmux (recommended for remote servers):
-
-```bash
-tmux new-session -d -s proxy "./cli-proxy-api --config config.yaml"
-```
-
-### 4. Configure Amp CLI
-
-#### Option A: Settings File
-
-Edit `~/.config/amp/settings.json`:
-
-```json
-{
-  "amp.url": "http://localhost:8317"
-}
-```
-
-#### Option B: Environment Variable
-
-```bash
-export AMP_URL=http://localhost:8317
-```
-
-### 5. Login and Use Amp
-
-Login through the proxy (proxied to ampcode.com):
-
-```bash
-amp login
-```
-
-Use Amp as normal:
-
-```bash
-amp "Write a hello world program in Python"
-```
-
-### 6. (Optional) Configure Amp IDE Extension
-
-The proxy also works with Amp IDE extensions for VS Code, Cursor, Windsurf, etc.
-
-1. Open Amp extension settings in your IDE
-2. Set **Amp URL** to `http://localhost:8317`
-3. Login with your Amp account
-4. Start using Amp in your IDE
-
-Both CLI and IDE can use the proxy simultaneously.
-
-## Usage
-
-### Supported Routes
-
-#### Provider Aliases (Always Available)
-
-These routes work even without `amp-upstream-url` configured:
-
- `/api/provider/openai/v1/chat/completions`
- `/api/provider/openai/v1/responses`
- `/api/provider/anthropic/v1/messages`
- `/api/provider/google/v1beta/models/:action`
-
-Amp CLI calls these routes with your OAuth-authenticated models configured in CLIProxyAPI.
-
-#### Management Routes (Require `amp-upstream-url`)
-
-These routes are proxied to ampcode.com:
-
- `/api/auth` - Authentication
- `/api/user` - User profile
- `/api/meta` - Metadata
- `/api/threads` - Conversation threads
- `/api/telemetry` - Usage telemetry
- `/api/internal` - Internal APIs
-
-**Security**: Restricted to localhost by default.
-
-### Model Fallback Behavior
-
-When Amp requests a model:
-
-1. **Check local configuration**: Does CLIProxyAPI have OAuth tokens for this model's provider?
-2. **If YES**: Route to local handler (use your OAuth subscription)
-3. **If NO**: Check if a model mapping exists
-4. **If mapping exists**: Rewrite request to mapped model → Route to local handler (free)
-5. **If no mapping**: Forward to ampcode.com (uses Amp credits)
-
-This enables seamless mixed usage:
- Models you've configured (Gemini, ChatGPT, Claude) → Your OAuth subscriptions
- Models with mappings configured → Routed to alternative local models (free)
- Models you haven't configured and have no mapping → Amp's default providers (uses credits)
-
-### Example API Calls
-
-**Chat completion with local OAuth:**
-```bash
-curl http://localhost:8317/api/provider/openai/v1/chat/completions \
-  -H "Content-Type: application/json" \
-  -d '{
-    "model": "gpt-5",
-    "messages": [{"role": "user", "content": "Hello"}]
-  }'
-```
-
-**Management endpoint (localhost only):**
-```bash
-curl http://localhost:8317/api/user
-```
-
-## Troubleshooting
-
-### Common Issues
-
-| Symptom | Likely Cause | Fix |
-|---------|--------------|-----|
-| 404 on `/api/provider/...` | Incorrect route path | Ensure exact path: `/api/provider/{provider}/v1...` |
-| 403 on `/api/user` | Non-localhost request | Run from same machine or disable `amp-restrict-management-to-localhost` (not recommended) |
-| 401/403 from provider | Missing/expired OAuth | Re-run `--codex-login` or `--claude-login` |
-| Amp gzip errors | Response decompression issue | Update to latest build; auto-decompression should handle this |
-| Models not using proxy | Wrong Amp URL | Verify `amp.url` setting or `AMP_URL` environment variable |
-| CORS errors | Protected management endpoint | Use CLI/terminal, not browser |
-
-### Diagnostics
-
-**Check proxy logs:**
-```bash
-# If logging-to-file: true
-tail -f logs/requests.log
-
-# If running in tmux
-tmux attach-session -t proxy
-```
-
-**Enable debug mode** (temporarily):
-```yaml
-debug: true
-```
-
-**Test basic connectivity:**
-```bash
-# Check if proxy is running
-curl http://localhost:8317/v1/models
-
-# Check Amp-specific route
-curl http://localhost:8317/api/provider/openai/v1/models
-```
-
-**Verify Amp configuration:**
-```bash
-# Check if Amp is using proxy
-amp config get amp.url
-
-# Or check environment
-echo $AMP_URL
-```
-
-### Security Checklist
-
- ✅ Keep `amp-restrict-management-to-localhost: true` (default)
- ✅ Don't expose proxy publicly (bind to localhost or use firewall/VPN)
- ✅ Use the Amp secrets file (`~/.local/share/amp/secrets.json`) managed by `amp login`
- ✅ Rotate OAuth tokens periodically by re-running login commands
- ✅ Store config and auth-dir on encrypted disk if handling sensitive data
- ✅ Keep proxy binary up to date for security fixes
-
-## Additional Resources
-
- [CLIProxyAPI Main Documentation](https://help.router-for.me/)
- [Amp CLI Official Manual](https://ampcode.com/manual)
- [Management API Reference](https://help.router-for.me/management/api)
- [SDK Documentation](sdk-usage.md)
-
-## Disclaimer
-
-This integration is for personal/educational use. Using reverse proxies or alternate API bases may violate provider Terms of Service. You are solely responsible for how you use this software. Accounts may be rate-limited, locked, or banned. No warranties. Use at your own risk.
--- a/docs/amp-cli-integration_CN.md
+++ b/docs/amp-cli-integration_CN.md
@@ -1,392 +0,0 @@
-# Amp CLI 集成指南
-
-本指南说明如何在 Amp CLI 和 Amp IDE 扩展中使用 CLIProxyAPI，通过 OAuth 让你能够把已有的 Google/ChatGPT/Claude 订阅与 Amp 的 CLI 一起使用。
-
-## 目录
-
- [概述](#概述)
-  - [应该认证哪些服务提供商？](#应该认证哪些服务提供商)
- [架构](#架构)
- [配置](#配置)
- [设置](#设置)
- [用法](#用法)
- [故障排查](#故障排查)
-
-## 概述
-
-Amp CLI 集成为 Amp 的 API 模式添加了专用路由，同时保持与现有 CLIProxyAPI 功能的完全兼容。这样你可以在同一个代理服务器上同时使用传统 CLIProxyAPI 功能和 Amp CLI。
-
-### 主要特性
-
- **提供者路由别名**：将 Amp 的 `/api/provider/{provider}/v1...` 路径映射到 CLIProxyAPI 处理器
- **管理代理**：将 OAuth 和账号管理请求转发到 Amp 控制平面
- **智能回退**：自动将未配置的模型路由到 ampcode.com
- **密钥管理**：可配置优先级（配置 > 环境变量 > 文件），缓存 5 分钟
- **安全优先**：管理路由默认限制为 localhost
- **自动 gzip 处理**：自动解压来自 Amp 上游的响应
-
-### 你可以做什么
-
- 使用 Amp CLI 搭配你的 Google 账号（Gemini 3 Pro Preview、Gemini 2.5 Pro、Gemini 2.5 Flash）
- 使用 Amp CLI 搭配你的 ChatGPT Plus/Pro 订阅（GPT-5、GPT-5 Codex 模型）
- 使用 Amp CLI 搭配你的 Claude Pro/Max 订阅（Claude Sonnet 4.5、Opus 4.1）
- 将 Amp IDE 扩展（VS Code、Cursor、Windsurf 等）与同一个代理一起使用
- 通过一个代理同时运行多个 CLI 工具（Factory + Amp）
- 将未配置的模型自动路由到 ampcode.com
-
-### 应该认证哪些服务提供商？
-
-**重要**：需要认证的提供商取决于你安装的 Amp 版本当前使用的模型和功能。Amp 的不同智能模式和子代理会使用不同的提供商：
-
- **Smart 模式**：使用 Google/Gemini 模型（Gemini 3 Pro）
- **Rush 模式**：使用 Anthropic/Claude 模型（Claude Haiku 4.5）
- **Oracle 子代理**：使用 OpenAI/GPT 模型（GPT-5 medium reasoning）
- **Librarian 子代理**：使用 Anthropic/Claude 模型（Claude Sonnet 4.5）
- **Search 子代理**：使用 Anthropic/Claude 模型（Claude Haiku 4.5）
- **Review 功能**：使用 Google/Gemini 模型（Gemini 2.5 Flash-Lite）
-
-有关 Amp 当前使用哪些模型的最新信息，请参阅 **[Amp 模型文档](https://ampcode.com/models)**。
-
-#### 回退行为
-
-CLIProxyAPI 采用智能回退机制：
-
-1. **本地已认证提供商**（`--login`、`--codex-login`、`--claude-login`）：
-   - 请求使用**你的 OAuth 订阅**（ChatGPT Plus/Pro、Claude Pro/Max、Google 账号）
-   - 享受订阅自带的额度
-   - 不消耗 Amp 额度
-
-2. **本地未认证提供商**：
-   - 请求自动转发到 **ampcode.com**
-   - 使用 Amp 的后端提供商连接
-   - 如果提供商是付费的（OpenAI、Anthropic 付费档），**需要消耗 Amp 额度**
-   - 若 Amp 额度不足，可能产生错误
-
-**建议**：对你有订阅的所有提供商都进行认证，以最大化价值并尽量减少 Amp 额度消耗。如果没有覆盖 Amp 使用的全部提供商，请确保为回退请求准备足够的 Amp 额度。
-
-## 架构
-
-### 请求流
-
-```
-Amp CLI/IDE
-  ↓
-  ├─ Provider API requests (/api/provider/{provider}/v1/...)
-  │   ↓
-  │   ├─ Model configured locally?
-  │   │   YES → Use local OAuth tokens (OpenAI/Claude/Gemini handlers)
-  │   │   NO  → Forward to ampcode.com (reverse proxy)
-  │   ↓
-  │   Response
-  │
-  └─ Management requests (/api/auth, /api/user, /api/threads, ...)
-      ↓
-      ├─ Localhost check (security)
-      ↓
-      └─ Reverse proxy to ampcode.com
-          ↓
-          Response (auto-decompressed if gzipped)
-```
-
-### 组件
-
-Amp 集成以模块化路由模块（`internal/api/modules/amp/`）实现，包含以下组件：
-
-1. **路由别名**（`routes.go`）：将 Amp 风格的路径映射到标准处理器
-2. **反向代理**（`proxy.go`）：将管理请求转发到 ampcode.com
-3. **回退处理器**（`fallback_handlers.go`）：将未配置的模型路由到 ampcode.com
-4. **密钥管理**（`secret.go`）：多来源 API 密钥解析并带缓存
-5. **主模块**（`amp.go`）：负责注册和配置
-
-## 配置
-
-### 基础配置
-
-在 `config.yaml` 中新增以下字段：
-
-```yaml
-# Amp 上游控制平面（管理路由必需）
-amp-upstream-url: "https://ampcode.com"
-
-# 可选：覆盖 API key（否则使用环境变量或文件）
-# amp-upstream-api-key: "your-amp-api-key"
-
-# 安全性：将管理路由限制为 localhost（推荐）
-amp-restrict-management-to-localhost: true
-```
-
-### 密钥解析优先级
-
-Amp 模块以如下优先级解析 API key：
-
-| 来源 | 键名 | 优先级 | 缓存 |
-|------|------|--------|------|
-| 配置文件 | `amp-upstream-api-key` | 高 | 无 |
-| 环境变量 | `AMP_API_KEY` | 中 | 无 |
-| Amp 密钥文件 | `~/.local/share/amp/secrets.json` | 低 | 5 分钟 |
-
-**建议**：日常使用时采用 Amp 密钥文件（最低优先级）。该文件由 `amp login` 自动管理。
-
-### 安全设置
-
-**`amp-restrict-management-to-localhost`**（默认：`true`）
-
-启用后，管理路由（`/api/auth`、`/api/user`、`/api/threads` 等）只接受来自 localhost（127.0.0.1、::1）的连接，可防止：
- 浏览器探测式攻击
- 对管理端点的远程访问
- 基于 CORS 的攻击
- 伪造头攻击（例如 `X-Forwarded-For: 127.0.0.1`）
-
-#### 工作原理
-
-此限制使用**实际的 TCP 连接地址**（`RemoteAddr`），而非 `X-Forwarded-For` 等 HTTP 头，能防止头部伪造，但有重要影响：
-
- ✅ **直接连接可用**：在本机或服务器直接运行 CLIProxyAPI 时适用
- ⚠️ **可能不适用于反向代理场景**：部署在 nginx、Cloudflare 等代理后，请求源会显示为代理 IP 而非 localhost
-
-#### 反向代理部署
-
-若需要在反向代理（nginx、Caddy、Cloudflare Tunnel 等）后运行 CLIProxyAPI：
-
-1. **关闭 localhost 限制**：
-   ```yaml
-   amp-restrict-management-to-localhost: false
-   ```
-
-2. **使用替代安全措施**：
-   - 防火墙规则限制管理路由访问
-   - 代理层认证（HTTP Basic Auth、OAuth）
-   - 网络隔离（VPN、Tailscale、Cloudflare Access）
-   - 将 CLIProxyAPI 仅绑定 `127.0.0.1`，并通过 SSH 隧道访问
-
-3. **nginx 示例配置**（阻止外部访问管理路由）：
-   ```nginx
-   location /api/auth { deny all; }
-   location /api/user { deny all; }
-   location /api/threads { deny all; }
-   location /api/internal { deny all; }
-   ```
-
-**重要**：只有在理解安全影响并已采取其他防护措施时，才关闭 `amp-restrict-management-to-localhost`。
-
-## 设置
-
-### 1. 配置 CLIProxyAPI
-
-创建或编辑 `config.yaml`：
-
-```yaml
-port: 8317
-auth-dir: "~/.cli-proxy-api"
-
-# Amp 集成
-amp-upstream-url: "https://ampcode.com"
-amp-restrict-management-to-localhost: true
-
-# 其他常规设置...
-debug: false
-logging-to-file: true
-```
-
-### 2. 认证提供商
-
-为要使用的提供商执行 OAuth 登录：
-
-**Google 账号（Gemini 2.5 Pro、Gemini 2.5 Flash、Gemini 3 Pro Preview）：**
-```bash
-./cli-proxy-api --login
-```
-
-**ChatGPT Plus/Pro（GPT-5、GPT-5 Codex）：**
-```bash
-./cli-proxy-api --codex-login
-```
-
-**Claude Pro/Max（Claude Sonnet 4.5、Opus 4.1）：**
-```bash
-./cli-proxy-api --claude-login
-```
-
-令牌会保存到：
- Gemini: `~/.cli-proxy-api/gemini-<email>.json`
- OpenAI Codex: `~/.cli-proxy-api/codex-<email>.json`
- Claude: `~/.cli-proxy-api/claude-<email>.json`
-
-### 3. 启动代理
-
-```bash
-./cli-proxy-api --config config.yaml
-```
-
-或使用 tmux 在后台运行（推荐用于远程服务器）：
-
-```bash
-tmux new-session -d -s proxy "./cli-proxy-api --config config.yaml"
-```
-
-### 4. 配置 Amp CLI
-
-#### 方案 A：配置文件
-
-编辑 `~/.config/amp/settings.json`：
-
-```json
-{
-  "amp.url": "http://localhost:8317"
-}
-```
-
-#### 方案 B：环境变量
-
-```bash
-export AMP_URL=http://localhost:8317
-```
-
-### 5. 登录并使用 Amp
-
-通过代理登录（请求会被代理到 ampcode.com）：
-
-```bash
-amp login
-```
-
-像平常一样使用 Amp：
-
-```bash
-amp "Write a hello world program in Python"
-```
-
-### 6. （可选）配置 Amp IDE 扩展
-
-该代理同样适用于 VS Code、Cursor、Windsurf 等 Amp IDE 扩展。
-
-1. 在 IDE 中打开 Amp 扩展设置
-2. 将 **Amp URL** 设置为 `http://localhost:8317`
-3. 用你的 Amp 账号登录
-4. 在 IDE 中开始使用 Amp
-
-CLI 和 IDE 可同时使用该代理。
-
-## 用法
-
-### 支持的路由
-
-#### 提供商别名（始终可用）
-
-这些路由即使未配置 `amp-upstream-url` 也可使用：
-
- `/api/provider/openai/v1/chat/completions`
- `/api/provider/openai/v1/responses`
- `/api/provider/anthropic/v1/messages`
- `/api/provider/google/v1beta/models/:action`
-
-Amp CLI 会使用你在 CLIProxyAPI 中通过 OAuth 认证的模型来调用这些路由。
-
-#### 管理路由（需要 `amp-upstream-url`）
-
-这些路由会被代理到 ampcode.com：
-
- `/api/auth` - 认证
- `/api/user` - 用户资料
- `/api/meta` - 元数据
- `/api/threads` - 会话线程
- `/api/telemetry` - 使用遥测
- `/api/internal` - 内部 API
-
-**安全性**：默认限制为 localhost。
-
-### 模型回退行为
-
-当 Amp 请求模型时：
-
-1. **检查本地配置**：CLIProxyAPI 是否有该模型提供商的 OAuth 令牌？
-2. **如果有**：路由到本地处理器（使用你的 OAuth 订阅）
-3. **如果没有**：转发到 ampcode.com（使用 Amp 的默认路由）
-
-这实现了无缝混用：
- 你已配置的模型（Gemini、ChatGPT、Claude）→ 你的 OAuth 订阅
- 未配置的模型 → Amp 的默认提供商
-
-### 示例 API 调用
-
-**使用本地 OAuth 的聊天补全：**
-```bash
-curl http://localhost:8317/api/provider/openai/v1/chat/completions \
-  -H "Content-Type: application/json" \
-  -d '{
-    "model": "gpt-5",
-    "messages": [{"role": "user", "content": "Hello"}]
-  }'
-```
-
-**管理端点（仅限 localhost）：**
-```bash
-curl http://localhost:8317/api/user
-```
-
-## 故障排查
-
-### 常见问题
-
-| 症状 | 可能原因 | 解决方案 |
-|------|----------|----------|
-| `/api/provider/...` 返回 404 | 路径错误 | 确保路径准确：`/api/provider/{provider}/v1...` |
-| `/api/user` 返回 403 | 非 localhost 请求 | 在同一机器上访问，或关闭 `amp-restrict-management-to-localhost`（不推荐） |
-| 提供商返回 401/403 | OAuth 缺失或过期 | 重新运行 `--codex-login` 或 `--claude-login` |
-| Amp gzip 错误 | 响应解压问题 | 更新到最新构建；自动解压应能处理 |
-| 模型未走代理 | Amp URL 设置错误 | 检查 `amp.url` 设置或 `AMP_URL` 环境变量 |
-| CORS 错误 | 受保护的管理端点 | 使用 CLI/终端而非浏览器 |
-
-### 诊断
-
-**查看代理日志：**
-```bash
-# 若 logging-to-file: true
-tail -f logs/requests.log
-
-# 若运行在 tmux 中
-tmux attach-session -t proxy
-```
-
-**临时开启调试模式：**
-```yaml
-debug: true
-```
-
-**测试基础连通性：**
-```bash
-# 检查代理是否运行
-curl http://localhost:8317/v1/models
-
-# 检查 Amp 特定路由
-curl http://localhost:8317/api/provider/openai/v1/models
-```
-
-**验证 Amp 配置：**
-```bash
-# 检查 Amp 是否使用代理
-amp config get amp.url
-
-# 或检查环境变量
-echo $AMP_URL
-```
-
-### 安全清单
-
- ✅ 保持 `amp-restrict-management-to-localhost: true`（默认）
- ✅ 不要将代理暴露到公共网络（绑定到 localhost 或使用防火墙/VPN）
- ✅ 使用 `amp login` 管理的 Amp 密钥文件（`~/.local/share/amp/secrets.json`）
- ✅ 定期重新登录轮换 OAuth 令牌
- ✅ 若处理敏感数据，使用加密磁盘存储配置和 auth-dir
- ✅ 保持代理二进制为最新版本以获取安全修复
-
-## 其他资源
-
- [CLIProxyAPI 主文档](https://help.router-for.me/)
- [Amp CLI 官方手册](https://ampcode.com/manual)
- [管理 API 参考](https://help.router-for.me/management/api)
- [SDK 文档](sdk-usage.md)
-
-## 免责声明
-
-此集成仅用于个人或教育用途。使用反向代理或替代 API 基址可能违反提供商的服务条款。你需要对自己的使用方式负责。账号可能会被限速、锁定或封禁。软件不附带任何保证，使用风险自负。
--- a/internal/api/handlers/management/auth_files.go
+++ b/internal/api/handlers/management/auth_files.go
@@ -26,6 +26,7 @@ import (
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/auth/qwen"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/interfaces"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	sdkAuth "github.com/router-for-me/CLIProxyAPI/v6/sdk/auth"
 	coreauth "github.com/router-for-me/CLIProxyAPI/v6/sdk/cliproxy/auth"
@@ -266,6 +267,54 @@ func (h *Handler) ListAuthFiles(c *gin.Context) {
 	c.JSON(200, gin.H{"files": files})
 }

+// GetAuthFileModels returns the models supported by a specific auth file
+func (h *Handler) GetAuthFileModels(c *gin.Context) {
+	name := c.Query("name")
+	if name == "" {
+		c.JSON(400, gin.H{"error": "name is required"})
+		return
+	}
+
+	// Try to find auth ID via authManager
+	var authID string
+	if h.authManager != nil {
+		auths := h.authManager.List()
+		for _, auth := range auths {
+			if auth.FileName == name || auth.ID == name {
+				authID = auth.ID
+				break
+			}
+		}
+	}
+
+	if authID == "" {
+		authID = name // fallback to filename as ID
+	}
+
+	// Get models from registry
+	reg := registry.GetGlobalRegistry()
+	models := reg.GetModelsForClient(authID)
+
+	result := make([]gin.H, 0, len(models))
+	for _, m := range models {
+		entry := gin.H{
+			"id": m.ID,
+		}
+		if m.DisplayName != "" {
+			entry["display_name"] = m.DisplayName
+		}
+		if m.Type != "" {
+			entry["type"] = m.Type
+		}
+		if m.OwnedBy != "" {
+			entry["owned_by"] = m.OwnedBy
+		}
+		result = append(result, entry)
+	}
+
+	c.JSON(200, gin.H{"models": result})
+}
+
 // List auth files from disk when the auth manager is unavailable.
 func (h *Handler) listAuthFilesFromDisk(c *gin.Context) {
 	entries, err := os.ReadDir(h.cfg.AuthDir)
--- a/internal/api/middleware/response_writer.go
+++ b/internal/api/middleware/response_writer.go
@@ -71,22 +71,64 @@ func (w *ResponseWriterWrapper) Write(data []byte) (int, error) {
 	n, err := w.ResponseWriter.Write(data)

 	// THEN: Handle logging based on response type
-	if w.isStreaming {
+	if w.isStreaming && w.chunkChannel != nil {
 		// For streaming responses: Send to async logging channel (non-blocking)
-		if w.chunkChannel != nil {
-			select {
-			case w.chunkChannel <- append([]byte(nil), data...): // Non-blocking send with copy
-			default: // Channel full, skip logging to avoid blocking
-			}
+		select {
+		case w.chunkChannel <- append([]byte(nil), data...): // Non-blocking send with copy
+		default: // Channel full, skip logging to avoid blocking
 		}
-	} else {
-		// For non-streaming responses: Buffer complete response
+		return n, err
+	}
+
+	if w.shouldBufferResponseBody() {
 		w.body.Write(data)
 	}

 	return n, err
 }

+func (w *ResponseWriterWrapper) shouldBufferResponseBody() bool {
+	if w.logger != nil && w.logger.IsEnabled() {
+		return true
+	}
+	if !w.logOnErrorOnly {
+		return false
+	}
+	status := w.statusCode
+	if status == 0 {
+		if statusWriter, ok := w.ResponseWriter.(interface{ Status() int }); ok && statusWriter != nil {
+			status = statusWriter.Status()
+		} else {
+			status = http.StatusOK
+		}
+	}
+	return status >= http.StatusBadRequest
+}
+
+// WriteString wraps the underlying ResponseWriter's WriteString method to capture response data.
+// Some handlers (and fmt/io helpers) write via io.StringWriter; without this override, those writes
+// bypass Write() and would be missing from request logs.
+func (w *ResponseWriterWrapper) WriteString(data string) (int, error) {
+	w.ensureHeadersCaptured()
+
+	// CRITICAL: Write to client first (zero latency)
+	n, err := w.ResponseWriter.WriteString(data)
+
+	// THEN: Capture for logging
+	if w.isStreaming && w.chunkChannel != nil {
+		select {
+		case w.chunkChannel <- []byte(data):
+		default:
+		}
+		return n, err
+	}
+
+	if w.shouldBufferResponseBody() {
+		w.body.WriteString(data)
+	}
+	return n, err
+}
+
 // WriteHeader wraps the underlying ResponseWriter's WriteHeader method.
 // It captures the status code, detects if the response is streaming based on the Content-Type header,
 // and initializes the appropriate logging mechanism (standard or streaming).
@@ -160,12 +202,16 @@ func (w *ResponseWriterWrapper) detectStreaming(contentType string) bool {
 		return true
 	}

-	// Check request body for streaming indicators
-	if w.requestInfo.Body != nil {
+	// If a concrete Content-Type is already set (e.g., application/json for error responses),
+	// treat it as non-streaming instead of inferring from the request payload.
+	if strings.TrimSpace(contentType) != "" {
+		return false
+	}
+
+	// Only fall back to request payload hints when Content-Type is not set yet.
+	if w.requestInfo != nil && len(w.requestInfo.Body) > 0 {
 		bodyStr := string(w.requestInfo.Body)
-		if strings.Contains(bodyStr, `"stream": true`) || strings.Contains(bodyStr, `"stream":true`) {
-			return true
-		}
+		return strings.Contains(bodyStr, `"stream": true`) || strings.Contains(bodyStr, `"stream":true`)
 	}

 	return false
@@ -221,7 +267,7 @@ func (w *ResponseWriterWrapper) Finalize(c *gin.Context) error {
 		return nil
 	}

-	if w.isStreaming {
+	if w.isStreaming && w.streamWriter != nil {
 		if w.chunkChannel != nil {
 			close(w.chunkChannel)
 			w.chunkChannel = nil
@@ -233,24 +279,19 @@ func (w *ResponseWriterWrapper) Finalize(c *gin.Context) error {
 		}

 		// Write API Request and Response to the streaming log before closing
-		if w.streamWriter != nil {
-			apiRequest := w.extractAPIRequest(c)
-			if len(apiRequest) > 0 {
-				_ = w.streamWriter.WriteAPIRequest(apiRequest)
-			}
-			apiResponse := w.extractAPIResponse(c)
-			if len(apiResponse) > 0 {
-				_ = w.streamWriter.WriteAPIResponse(apiResponse)
-			}
-			if err := w.streamWriter.Close(); err != nil {
-				w.streamWriter = nil
-				return err
-			}
+		apiRequest := w.extractAPIRequest(c)
+		if len(apiRequest) > 0 {
+			_ = w.streamWriter.WriteAPIRequest(apiRequest)
+		}
+		apiResponse := w.extractAPIResponse(c)
+		if len(apiResponse) > 0 {
+			_ = w.streamWriter.WriteAPIResponse(apiResponse)
+		}
+		if err := w.streamWriter.Close(); err != nil {
 			w.streamWriter = nil
+			return err
 		}
-		if forceLog {
-			return w.logRequest(finalStatusCode, w.cloneHeaders(), w.body.Bytes(), w.extractAPIRequest(c), w.extractAPIResponse(c), slicesAPIResponseError, forceLog)
-		}
+		w.streamWriter = nil
 		return nil
 	}

@@ -335,26 +376,3 @@ func (w *ResponseWriterWrapper) logRequest(statusCode int, headers map[string][]
 		apiResponseErrors,
 	)
 }
-
-// Status returns the HTTP response status code captured by the wrapper.
-// It defaults to 200 if WriteHeader has not been called.
-func (w *ResponseWriterWrapper) Status() int {
-	if w.statusCode == 0 {
-		return 200 // Default status code
-	}
-	return w.statusCode
-}
-
-// Size returns the size of the response body in bytes for non-streaming responses.
-// For streaming responses, it returns -1, as the total size is unknown.
-func (w *ResponseWriterWrapper) Size() int {
-	if w.isStreaming {
-		return -1 // Unknown size for streaming responses
-	}
-	return w.body.Len()
-}
-
-// Written returns true if the response header has been written (i.e., a status code has been set).
-func (w *ResponseWriterWrapper) Written() bool {
-	return w.statusCode != 0
-}
--- a/internal/api/modules/amp/amp.go
+++ b/internal/api/modules/amp/amp.go
@@ -137,7 +137,8 @@ func (m *AmpModule) Register(ctx modules.Context) error {
 		m.registerProviderAliases(ctx.Engine, ctx.BaseHandler, auth)

 		// Register management proxy routes once; middleware will gate access when upstream is unavailable.
-		m.registerManagementRoutes(ctx.Engine, ctx.BaseHandler)
+		// Pass auth middleware to require valid API key for all management routes.
+		m.registerManagementRoutes(ctx.Engine, ctx.BaseHandler, auth)

 		// If no upstream URL, skip proxy routes but provider aliases are still available
 		if upstreamURL == "" {
@@ -187,9 +188,6 @@ func (m *AmpModule) OnConfigUpdated(cfg *config.Config) error {

 	if oldSettings != nil && oldSettings.RestrictManagementToLocalhost != newSettings.RestrictManagementToLocalhost {
 		m.setRestrictToLocalhost(newSettings.RestrictManagementToLocalhost)
-		if !newSettings.RestrictManagementToLocalhost {
-			log.Warnf("amp management routes now accessible from any IP - this is insecure!")
-		}
 	}

 	newUpstreamURL := strings.TrimSpace(newSettings.UpstreamURL)
--- a/internal/api/modules/amp/fallback_handlers.go
+++ b/internal/api/modules/amp/fallback_handlers.go
@@ -64,7 +64,7 @@ func logAmpRouting(routeType AmpRouteType, requestedModel, resolvedModel, provid
 		fields["cost"] = "amp_credits"
 		fields["source"] = "ampcode.com"
 		fields["model_id"] = requestedModel // Explicit model_id for easy config reference
-		log.WithFields(fields).Warnf("forwarding to ampcode.com (uses amp credits) - model_id: %s | To use local proxy, add to config: amp-model-mappings: [{from: \"%s\", to: \"<your-local-model>\"}]", requestedModel, requestedModel)
+		log.WithFields(fields).Warnf("forwarding to ampcode.com (uses amp credits) - model_id: %s | To use local provider, add to config: ampcode.model-mappings: [{from: \"%s\", to: \"<your-local-model>\"}]", requestedModel, requestedModel)

 	case RouteTypeNoProvider:
 		fields["cost"] = "none"
--- a/internal/api/modules/amp/proxy.go
+++ b/internal/api/modules/amp/proxy.go
@@ -41,6 +41,11 @@ func createReverseProxy(upstreamURL string, secretSource SecretSource) (*httputi
 		originalDirector(req)
 		req.Host = parsed.Host

+		// Remove client's Authorization header - it was only used for CLI Proxy API authentication
+		// We will set our own Authorization using the configured upstream-api-key
+		req.Header.Del("Authorization")
+		req.Header.Del("X-Api-Key")
+
 		// Preserve correlation headers for debugging
 		if req.Header.Get("X-Request-ID") == "" {
 			// Could generate one here if needed
@@ -50,7 +55,7 @@ func createReverseProxy(upstreamURL string, secretSource SecretSource) (*httputi
 		// Users going through ampcode.com proxy are paying for the service and should get all features
 		// including 1M context window (context-1m-2025-08-07)

-		// Inject API key from secret source (precedence: config > env > file)
+		// Inject API key from secret source (only uses upstream-api-key from config)
 		if key, err := secretSource.Get(req.Context()); err == nil && key != "" {
 			req.Header.Set("X-Api-Key", key)
 			req.Header.Set("Authorization", fmt.Sprintf("Bearer %s", key))
--- a/internal/api/modules/amp/response_rewriter.go
+++ b/internal/api/modules/amp/response_rewriter.go
@@ -39,7 +39,13 @@ func (rw *ResponseRewriter) Write(data []byte) (int, error) {
 	}

 	if rw.isStreaming {
-		return rw.ResponseWriter.Write(rw.rewriteStreamChunk(data))
+		n, err := rw.ResponseWriter.Write(rw.rewriteStreamChunk(data))
+		if err == nil {
+			if flusher, ok := rw.ResponseWriter.(http.Flusher); ok {
+				flusher.Flush()
+			}
+		}
+		return n, err
 	}
 	return rw.body.Write(data)
 }
--- a/internal/api/modules/amp/routes.go
+++ b/internal/api/modules/amp/routes.go
@@ -98,7 +98,8 @@ func (m *AmpModule) managementAvailabilityMiddleware() gin.HandlerFunc {
 // registerManagementRoutes registers Amp management proxy routes
 // These routes proxy through to the Amp control plane for OAuth, user management, etc.
 // Uses dynamic middleware and proxy getter for hot-reload support.
-func (m *AmpModule) registerManagementRoutes(engine *gin.Engine, baseHandler *handlers.BaseAPIHandler) {
+// The auth middleware validates Authorization header against configured API keys.
+func (m *AmpModule) registerManagementRoutes(engine *gin.Engine, baseHandler *handlers.BaseAPIHandler, auth gin.HandlerFunc) {
 	ampAPI := engine.Group("/api")

 	// Always disable CORS for management routes to prevent browser-based attacks
@@ -107,8 +108,9 @@ func (m *AmpModule) registerManagementRoutes(engine *gin.Engine, baseHandler *ha
 	// Apply dynamic localhost-only restriction (hot-reloadable via m.IsRestrictedToLocalhost())
 	ampAPI.Use(m.localhostOnlyMiddleware())

-	if !m.IsRestrictedToLocalhost() {
-		log.Warn("amp management routes are NOT restricted to localhost - this is insecure!")
+	// Apply authentication middleware - requires valid API key in Authorization header
+	if auth != nil {
+		ampAPI.Use(auth)
 	}

 	// Dynamic proxy handler that uses m.getProxy() for hot-reload support
@@ -154,6 +156,9 @@ func (m *AmpModule) registerManagementRoutes(engine *gin.Engine, baseHandler *ha
 	// Root-level routes that AMP CLI expects without /api prefix
 	// These need the same security middleware as the /api/* routes (dynamic for hot-reload)
 	rootMiddleware := []gin.HandlerFunc{m.managementAvailabilityMiddleware(), noCORSMiddleware(), m.localhostOnlyMiddleware()}
+	if auth != nil {
+		rootMiddleware = append(rootMiddleware, auth)
+	}
 	engine.GET("/threads/*path", append(rootMiddleware, proxyHandler)...)
 	engine.GET("/threads.rss", append(rootMiddleware, proxyHandler)...)
 	engine.GET("/news.rss", append(rootMiddleware, proxyHandler)...)
--- a/internal/api/server.go
+++ b/internal/api/server.go
@@ -568,6 +568,7 @@ func (s *Server) registerManagementRoutes() {
 		mgmt.DELETE("/oauth-excluded-models", s.mgmt.DeleteOAuthExcludedModels)

 		mgmt.GET("/auth-files", s.mgmt.ListAuthFiles)
+		mgmt.GET("/auth-files/models", s.mgmt.GetAuthFileModels)
 		mgmt.GET("/auth-files/download", s.mgmt.DownloadAuthFile)
 		mgmt.POST("/auth-files", s.mgmt.UploadAuthFile)
 		mgmt.DELETE("/auth-files", s.mgmt.DeleteAuthFile)
--- a/internal/config/config.go
+++ b/internal/config/config.go
@@ -139,7 +139,7 @@ type AmpCode struct {

 	// RestrictManagementToLocalhost restricts Amp management routes (/api/user, /api/threads, etc.)
 	// to only accept connections from localhost (127.0.0.1, ::1). When true, prevents drive-by
-	// browser attacks and remote access to management endpoints. Default: true (recommended).
+	// browser attacks and remote access to management endpoints. Default: false (API key auth is sufficient).
 	RestrictManagementToLocalhost bool `yaml:"restrict-management-to-localhost" json:"restrict-management-to-localhost"`

 	// ModelMappings defines model name mappings for Amp CLI requests.
@@ -327,7 +327,7 @@ func LoadConfigOptional(configFile string, optional bool) (*Config, error) {
 	cfg.LoggingToFile = false
 	cfg.UsageStatisticsEnabled = false
 	cfg.DisableCooling = false
-	cfg.AmpCode.RestrictManagementToLocalhost = true // Default to secure: only localhost access
+	cfg.AmpCode.RestrictManagementToLocalhost = false // Default to false: API key auth is sufficient
 	if err = yaml.Unmarshal(data, &cfg); err != nil {
 		if optional {
 			// In cloud deploy mode, if YAML parsing fails, return empty config instead of error.
--- a/internal/registry/model_definitions.go
+++ b/internal/registry/model_definitions.go
@@ -648,10 +648,11 @@ func GetIFlowModels() []*ModelInfo {
 		{ID: "glm-4.6", DisplayName: "GLM-4.6", Description: "Zhipu GLM 4.6 general model", Created: 1759190400},
 		{ID: "kimi-k2", DisplayName: "Kimi-K2", Description: "Moonshot Kimi K2 general model", Created: 1752192000},
 		{ID: "kimi-k2-thinking", DisplayName: "Kimi-K2-Thinking", Description: "Moonshot Kimi K2 thinking model", Created: 1762387200, Thinking: &ThinkingSupport{Levels: []string{"low", "medium", "high"}}},
-		{ID: "deepseek-v3.2-chat", DisplayName: "DeepSeek-V3.2", Description: "DeepSeek V3.2", Created: 1764576000},
+		{ID: "deepseek-v3.2-chat", DisplayName: "DeepSeek-V3.2", Description: "DeepSeek V3.2 Chat", Created: 1764576000},
+		{ID: "deepseek-v3.2-reasoner", DisplayName: "DeepSeek-V3.2", Description: "DeepSeek V3.2 Reasoner", Created: 1764576000},
 		{ID: "deepseek-v3.2", DisplayName: "DeepSeek-V3.2-Exp", Description: "DeepSeek V3.2 experimental", Created: 1759104000},
 		{ID: "deepseek-v3.1", DisplayName: "DeepSeek-V3.1-Terminus", Description: "DeepSeek V3.1 Terminus", Created: 1756339200},
-		{ID: "deepseek-r1", DisplayName: "DeepSeek-R1", Description: "DeepSeek reasoning model R1", Created: 1737331200, Thinking: &ThinkingSupport{Levels: []string{"low", "medium", "high"}}},
+		{ID: "deepseek-r1", DisplayName: "DeepSeek-R1", Description: "DeepSeek reasoning model R1", Created: 1737331200},
 		{ID: "deepseek-v3", DisplayName: "DeepSeek-V3-671B", Description: "DeepSeek V3 671B", Created: 1734307200},
 		{ID: "qwen3-32b", DisplayName: "Qwen3-32B", Description: "Qwen3 32B", Created: 1747094400},
 		{ID: "qwen3-235b-a22b-thinking-2507", DisplayName: "Qwen3-235B-A22B-Thinking", Description: "Qwen3 235B A22B Thinking (2507)", Created: 1753401600, Thinking: &ThinkingSupport{Levels: []string{"low", "medium", "high"}}},
--- a/internal/registry/model_registry.go
+++ b/internal/registry/model_registry.go
@@ -90,6 +90,9 @@ type ModelRegistry struct {
 	models map[string]*ModelRegistration
 	// clientModels maps client ID to the models it provides
 	clientModels map[string][]string
+	// clientModelInfos maps client ID to a map of model ID -> ModelInfo
+	// This preserves the original model info provided by each client
+	clientModelInfos map[string]map[string]*ModelInfo
 	// clientProviders maps client ID to its provider identifier
 	clientProviders map[string]string
 	// mutex ensures thread-safe access to the registry
@@ -104,10 +107,11 @@ var registryOnce sync.Once
 func GetGlobalRegistry() *ModelRegistry {
 	registryOnce.Do(func() {
 		globalRegistry = &ModelRegistry{
-			models:          make(map[string]*ModelRegistration),
-			clientModels:    make(map[string][]string),
-			clientProviders: make(map[string]string),
-			mutex:           &sync.RWMutex{},
+			models:           make(map[string]*ModelRegistration),
+			clientModels:     make(map[string][]string),
+			clientModelInfos: make(map[string]map[string]*ModelInfo),
+			clientProviders:  make(map[string]string),
+			mutex:            &sync.RWMutex{},
 		}
 	})
 	return globalRegistry
@@ -144,6 +148,7 @@ func (r *ModelRegistry) RegisterClient(clientID, clientProvider string, models [
 		// No models supplied; unregister existing client state if present.
 		r.unregisterClientInternal(clientID)
 		delete(r.clientModels, clientID)
+		delete(r.clientModelInfos, clientID)
 		delete(r.clientProviders, clientID)
 		misc.LogCredentialSeparator()
 		return
@@ -152,7 +157,7 @@ func (r *ModelRegistry) RegisterClient(clientID, clientProvider string, models [
 	now := time.Now()

 	oldModels, hadExisting := r.clientModels[clientID]
-	oldProvider, _ := r.clientProviders[clientID]
+	oldProvider := r.clientProviders[clientID]
 	providerChanged := oldProvider != provider
 	if !hadExisting {
 		// Pure addition path.
@@ -161,6 +166,12 @@ func (r *ModelRegistry) RegisterClient(clientID, clientProvider string, models [
 			r.addModelRegistration(modelID, provider, model, now)
 		}
 		r.clientModels[clientID] = append([]string(nil), rawModelIDs...)
+		// Store client's own model infos
+		clientInfos := make(map[string]*ModelInfo, len(newModels))
+		for id, m := range newModels {
+			clientInfos[id] = cloneModelInfo(m)
+		}
+		r.clientModelInfos[clientID] = clientInfos
 		if provider != "" {
 			r.clientProviders[clientID] = provider
 		} else {
@@ -287,6 +298,12 @@ func (r *ModelRegistry) RegisterClient(clientID, clientProvider string, models [
 	if len(rawModelIDs) > 0 {
 		r.clientModels[clientID] = append([]string(nil), rawModelIDs...)
 	}
+	// Update client's own model infos
+	clientInfos := make(map[string]*ModelInfo, len(newModels))
+	for id, m := range newModels {
+		clientInfos[id] = cloneModelInfo(m)
+	}
+	r.clientModelInfos[clientID] = clientInfos
 	if provider != "" {
 		r.clientProviders[clientID] = provider
 	} else {
@@ -436,6 +453,7 @@ func (r *ModelRegistry) unregisterClientInternal(clientID string) {
 	}

 	delete(r.clientModels, clientID)
+	delete(r.clientModelInfos, clientID)
 	if hasProvider {
 		delete(r.clientProviders, clientID)
 	}
@@ -871,3 +889,44 @@ func (r *ModelRegistry) GetFirstAvailableModel(handlerType string) (string, erro

 	return "", fmt.Errorf("no available clients for any model in handler type: %s", handlerType)
 }
+
+// GetModelsForClient returns the models registered for a specific client.
+// Parameters:
+//   - clientID: The client identifier (typically auth file name or auth ID)
+//
+// Returns:
+//   - []*ModelInfo: List of models registered for this client, nil if client not found
+func (r *ModelRegistry) GetModelsForClient(clientID string) []*ModelInfo {
+	r.mutex.RLock()
+	defer r.mutex.RUnlock()
+
+	modelIDs, exists := r.clientModels[clientID]
+	if !exists || len(modelIDs) == 0 {
+		return nil
+	}
+
+	// Try to use client-specific model infos first
+	clientInfos := r.clientModelInfos[clientID]
+
+	seen := make(map[string]struct{})
+	result := make([]*ModelInfo, 0, len(modelIDs))
+	for _, modelID := range modelIDs {
+		if _, dup := seen[modelID]; dup {
+			continue
+		}
+		seen[modelID] = struct{}{}
+
+		// Prefer client's own model info to preserve original type/owned_by
+		if clientInfos != nil {
+			if info, ok := clientInfos[modelID]; ok && info != nil {
+				result = append(result, info)
+				continue
+			}
+		}
+		// Fallback to global registry (for backwards compatibility)
+		if reg, ok := r.models[modelID]; ok && reg.Info != nil {
+			result = append(result, reg.Info)
+		}
+	}
+	return result
+}
--- a/internal/runtime/executor/aistudio_executor.go
+++ b/internal/runtime/executor/aistudio_executor.go
@@ -322,7 +322,7 @@ func (e *AIStudioExecutor) translateRequest(req cliproxyexecutor.Request, opts c
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
 	payload := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), stream)
-	payload = applyThinkingMetadata(payload, req.Metadata, req.Model)
+	payload = ApplyThinkingMetadata(payload, req.Metadata, req.Model)
 	payload = util.ApplyDefaultThinkingIfNeeded(req.Model, payload)
 	payload = util.ConvertThinkingLevelToBudget(payload)
 	payload = util.NormalizeGeminiThinkingBudget(req.Model, payload)
@@ -384,8 +384,16 @@ func ensureColonSpacedJSON(payload []byte) []byte {

 	for i := 0; i < len(indented); i++ {
 		ch := indented[i]
-		if ch == '"' && (i == 0 || indented[i-1] != '\\') {
-			inString = !inString
+		if ch == '"' {
+			// A quote is escaped only when preceded by an odd number of consecutive backslashes.
+			// For example: "\\\"" keeps the quote inside the string, but "\\\\" closes the string.
+			backslashes := 0
+			for j := i - 1; j >= 0 && indented[j] == '\\'; j-- {
+				backslashes++
+			}
+			if backslashes%2 == 0 {
+				inString = !inString
+			}
 		}

 		if !inString {
--- a/internal/runtime/executor/codex_executor.go
+++ b/internal/runtime/executor/codex_executor.go
@@ -54,9 +54,9 @@ func (e *CodexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("codex")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort", false)
-	body = normalizeThinkingConfig(body, upstreamModel, false)
-	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
+	body = ApplyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort", false)
+	body = NormalizeThinkingConfig(body, upstreamModel, false)
+	if errValidate := ValidateThinkingConfig(body, upstreamModel); errValidate != nil {
 		return resp, errValidate
 	}
 	body = applyPayloadConfig(e.cfg, req.Model, body)
@@ -152,9 +152,9 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 	to := sdktranslator.FromString("codex")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)

-	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort", false)
-	body = normalizeThinkingConfig(body, upstreamModel, false)
-	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
+	body = ApplyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort", false)
+	body = NormalizeThinkingConfig(body, upstreamModel, false)
+	if errValidate := ValidateThinkingConfig(body, upstreamModel); errValidate != nil {
 		return nil, errValidate
 	}
 	body = applyPayloadConfig(e.cfg, req.Model, body)
@@ -254,7 +254,7 @@ func (e *CodexExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth

 	modelForCounting := req.Model

-	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort", false)
+	body = ApplyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning.effort", false)
 	body, _ = sjson.SetBytes(body, "model", upstreamModel)
 	body, _ = sjson.DeleteBytes(body, "previous_response_id")
 	body, _ = sjson.SetBytes(body, "stream", false)
--- a/internal/runtime/executor/gemini_cli_executor.go
+++ b/internal/runtime/executor/gemini_cli_executor.go
@@ -11,6 +11,8 @@ import (
 	"fmt"
 	"io"
 	"net/http"
+	"regexp"
+	"strconv"
 	"strings"
 	"time"

@@ -784,20 +786,45 @@ func parseRetryDelay(errorBody []byte) (*time.Duration, error) {
 	// Try to parse the retryDelay from the error response
 	// Format: error.details[].retryDelay where @type == "type.googleapis.com/google.rpc.RetryInfo"
 	details := gjson.GetBytes(errorBody, "error.details")
-	if !details.Exists() || !details.IsArray() {
-		return nil, fmt.Errorf("no error.details found")
+	if details.Exists() && details.IsArray() {
+		for _, detail := range details.Array() {
+			typeVal := detail.Get("@type").String()
+			if typeVal == "type.googleapis.com/google.rpc.RetryInfo" {
+				retryDelay := detail.Get("retryDelay").String()
+				if retryDelay != "" {
+					// Parse duration string like "0.847655010s"
+					duration, err := time.ParseDuration(retryDelay)
+					if err != nil {
+						return nil, fmt.Errorf("failed to parse duration")
+					}
+					return &duration, nil
+				}
+			}
+		}
+
+		// Fallback: try ErrorInfo.metadata.quotaResetDelay (e.g., "373.801628ms")
+		for _, detail := range details.Array() {
+			typeVal := detail.Get("@type").String()
+			if typeVal == "type.googleapis.com/google.rpc.ErrorInfo" {
+				quotaResetDelay := detail.Get("metadata.quotaResetDelay").String()
+				if quotaResetDelay != "" {
+					duration, err := time.ParseDuration(quotaResetDelay)
+					if err == nil {
+						return &duration, nil
+					}
+				}
+			}
+		}
 	}

-	for _, detail := range details.Array() {
-		typeVal := detail.Get("@type").String()
-		if typeVal == "type.googleapis.com/google.rpc.RetryInfo" {
-			retryDelay := detail.Get("retryDelay").String()
-			if retryDelay != "" {
-				// Parse duration string like "0.847655010s"
-				duration, err := time.ParseDuration(retryDelay)
-				if err != nil {
-					return nil, fmt.Errorf("failed to parse duration")
-				}
+	// Fallback: parse from error.message "Your quota will reset after Xs."
+	message := gjson.GetBytes(errorBody, "error.message").String()
+	if message != "" {
+		re := regexp.MustCompile(`after\s+(\d+)s\.?`)
+		if matches := re.FindStringSubmatch(message); len(matches) > 1 {
+			seconds, err := strconv.Atoi(matches[1])
+			if err == nil {
+				duration := time.Duration(seconds) * time.Second
 				return &duration, nil
 			}
 		}
--- a/internal/runtime/executor/gemini_executor.go
+++ b/internal/runtime/executor/gemini_executor.go
@@ -83,7 +83,7 @@ func (e *GeminiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	body = applyThinkingMetadata(body, req.Metadata, req.Model)
+	body = ApplyThinkingMetadata(body, req.Metadata, req.Model)
 	body = util.ApplyDefaultThinkingIfNeeded(req.Model, body)
 	body = util.NormalizeGeminiThinkingBudget(req.Model, body)
 	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
@@ -178,7 +178,7 @@ func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)
-	body = applyThinkingMetadata(body, req.Metadata, req.Model)
+	body = ApplyThinkingMetadata(body, req.Metadata, req.Model)
 	body = util.ApplyDefaultThinkingIfNeeded(req.Model, body)
 	body = util.NormalizeGeminiThinkingBudget(req.Model, body)
 	body = util.StripThinkingConfigIfUnsupported(req.Model, body)
@@ -290,7 +290,7 @@ func (e *GeminiExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Aut
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("gemini")
 	translatedReq := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	translatedReq = applyThinkingMetadata(translatedReq, req.Metadata, req.Model)
+	translatedReq = ApplyThinkingMetadata(translatedReq, req.Metadata, req.Model)
 	translatedReq = util.StripThinkingConfigIfUnsupported(req.Model, translatedReq)
 	translatedReq = fixGeminiImageAspectRatio(req.Model, translatedReq)
 	respCtx := context.WithValue(ctx, "alt", opts.Alt)
--- a/internal/runtime/executor/iflow_executor.go
+++ b/internal/runtime/executor/iflow_executor.go
@@ -57,13 +57,13 @@ func (e *IFlowExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort", false)
+	body = ApplyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort", false)
 	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
 	if upstreamModel != "" {
 		body, _ = sjson.SetBytes(body, "model", upstreamModel)
 	}
-	body = normalizeThinkingConfig(body, upstreamModel, false)
-	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
+	body = NormalizeThinkingConfig(body, upstreamModel, false)
+	if errValidate := ValidateThinkingConfig(body, upstreamModel); errValidate != nil {
 		return resp, errValidate
 	}
 	body = applyPayloadConfig(e.cfg, req.Model, body)
@@ -148,13 +148,13 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 	to := sdktranslator.FromString("openai")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)

-	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort", false)
+	body = ApplyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort", false)
 	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
 	if upstreamModel != "" {
 		body, _ = sjson.SetBytes(body, "model", upstreamModel)
 	}
-	body = normalizeThinkingConfig(body, upstreamModel, false)
-	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
+	body = NormalizeThinkingConfig(body, upstreamModel, false)
+	if errValidate := ValidateThinkingConfig(body, upstreamModel); errValidate != nil {
 		return nil, errValidate
 	}
 	// Ensure tools array exists to avoid provider quirks similar to Qwen's behaviour.
--- a/internal/runtime/executor/openai_compat_executor.go
+++ b/internal/runtime/executor/openai_compat_executor.go
@@ -60,13 +60,13 @@ func (e *OpenAICompatExecutor) Execute(ctx context.Context, auth *cliproxyauth.A
 	}
 	translated = applyPayloadConfigWithRoot(e.cfg, req.Model, to.String(), "", translated)
 	allowCompat := e.allowCompatReasoningEffort(req.Model, auth)
-	translated = applyReasoningEffortMetadata(translated, req.Metadata, req.Model, "reasoning_effort", allowCompat)
+	translated = ApplyReasoningEffortMetadata(translated, req.Metadata, req.Model, "reasoning_effort", allowCompat)
 	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
 	if upstreamModel != "" && modelOverride == "" {
 		translated, _ = sjson.SetBytes(translated, "model", upstreamModel)
 	}
-	translated = normalizeThinkingConfig(translated, upstreamModel, allowCompat)
-	if errValidate := validateThinkingConfig(translated, upstreamModel); errValidate != nil {
+	translated = NormalizeThinkingConfig(translated, upstreamModel, allowCompat)
+	if errValidate := ValidateThinkingConfig(translated, upstreamModel); errValidate != nil {
 		return resp, errValidate
 	}

@@ -156,13 +156,13 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 	}
 	translated = applyPayloadConfigWithRoot(e.cfg, req.Model, to.String(), "", translated)
 	allowCompat := e.allowCompatReasoningEffort(req.Model, auth)
-	translated = applyReasoningEffortMetadata(translated, req.Metadata, req.Model, "reasoning_effort", allowCompat)
+	translated = ApplyReasoningEffortMetadata(translated, req.Metadata, req.Model, "reasoning_effort", allowCompat)
 	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
 	if upstreamModel != "" && modelOverride == "" {
 		translated, _ = sjson.SetBytes(translated, "model", upstreamModel)
 	}
-	translated = normalizeThinkingConfig(translated, upstreamModel, allowCompat)
-	if errValidate := validateThinkingConfig(translated, upstreamModel); errValidate != nil {
+	translated = NormalizeThinkingConfig(translated, upstreamModel, allowCompat)
+	if errValidate := ValidateThinkingConfig(translated, upstreamModel); errValidate != nil {
 		return nil, errValidate
 	}

--- a/internal/runtime/executor/payload_helpers.go
+++ b/internal/runtime/executor/payload_helpers.go
@@ -11,9 +11,9 @@ import (
 	"github.com/tidwall/sjson"
 )

-// applyThinkingMetadata applies thinking config from model suffix metadata (e.g., (high), (8192))
+// ApplyThinkingMetadata applies thinking config from model suffix metadata (e.g., (high), (8192))
 // for standard Gemini format payloads. It normalizes the budget when the model supports thinking.
-func applyThinkingMetadata(payload []byte, metadata map[string]any, model string) []byte {
+func ApplyThinkingMetadata(payload []byte, metadata map[string]any, model string) []byte {
 	budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(model, metadata)
 	if !ok || (budgetOverride == nil && includeOverride == nil) {
 		return payload
@@ -45,10 +45,10 @@ func applyThinkingMetadataCLI(payload []byte, metadata map[string]any, model str
 	return util.ApplyGeminiCLIThinkingConfig(payload, budgetOverride, includeOverride)
 }

-// applyReasoningEffortMetadata applies reasoning effort overrides from metadata to the given JSON path.
+// ApplyReasoningEffortMetadata applies reasoning effort overrides from metadata to the given JSON path.
 // Metadata values take precedence over any existing field when the model supports thinking, intentionally
 // overwriting caller-provided values to honor suffix/default metadata priority.
-func applyReasoningEffortMetadata(payload []byte, metadata map[string]any, model, field string, allowCompat bool) []byte {
+func ApplyReasoningEffortMetadata(payload []byte, metadata map[string]any, model, field string, allowCompat bool) []byte {
 	if len(metadata) == 0 {
 		return payload
 	}
@@ -73,6 +73,12 @@ func applyReasoningEffortMetadata(payload []byte, metadata map[string]any, model
 	if util.ModelUsesThinkingLevels(baseModel) || allowCompat {
 		if budget, _, _, matched := util.ThinkingFromMetadata(metadata); matched && budget != nil {
 			if effort, ok := util.OpenAIThinkingBudgetToEffort(baseModel, *budget); ok && effort != "" {
+				if *budget == 0 && effort == "none" && util.ModelUsesThinkingLevels(baseModel) {
+					if _, supported := util.NormalizeReasoningEffortLevel(baseModel, effort); !supported {
+						return StripThinkingFields(payload, false)
+					}
+				}
+
 				if updated, err := sjson.SetBytes(payload, field, effort); err == nil {
 					return updated
 				}
@@ -232,12 +238,12 @@ func matchModelPattern(pattern, model string) bool {
 	return pi == len(pattern)
 }

-// normalizeThinkingConfig normalizes thinking-related fields in the payload
+// NormalizeThinkingConfig normalizes thinking-related fields in the payload
 // based on model capabilities. For models without thinking support, it strips
 // reasoning fields. For models with level-based thinking, it validates and
 // normalizes the reasoning effort level. For models with numeric budget thinking,
 // it strips the effort string fields.
-func normalizeThinkingConfig(payload []byte, model string, allowCompat bool) []byte {
+func NormalizeThinkingConfig(payload []byte, model string, allowCompat bool) []byte {
 	if len(payload) == 0 || model == "" {
 		return payload
 	}
@@ -246,22 +252,22 @@ func normalizeThinkingConfig(payload []byte, model string, allowCompat bool) []b
 		if allowCompat {
 			return payload
 		}
-		return stripThinkingFields(payload, false)
+		return StripThinkingFields(payload, false)
 	}

 	if util.ModelUsesThinkingLevels(model) {
-		return normalizeReasoningEffortLevel(payload, model)
+		return NormalizeReasoningEffortLevel(payload, model)
 	}

 	// Model supports thinking but uses numeric budgets, not levels.
 	// Strip effort string fields since they are not applicable.
-	return stripThinkingFields(payload, true)
+	return StripThinkingFields(payload, true)
 }

-// stripThinkingFields removes thinking-related fields from the payload for
+// StripThinkingFields removes thinking-related fields from the payload for
 // models that do not support thinking. If effortOnly is true, only removes
 // effort string fields (for models using numeric budgets).
-func stripThinkingFields(payload []byte, effortOnly bool) []byte {
+func StripThinkingFields(payload []byte, effortOnly bool) []byte {
 	fieldsToRemove := []string{
 		"reasoning_effort",
 		"reasoning.effort",
@@ -278,9 +284,9 @@ func stripThinkingFields(payload []byte, effortOnly bool) []byte {
 	return out
 }

-// normalizeReasoningEffortLevel validates and normalizes the reasoning_effort
+// NormalizeReasoningEffortLevel validates and normalizes the reasoning_effort
 // or reasoning.effort field for level-based thinking models.
-func normalizeReasoningEffortLevel(payload []byte, model string) []byte {
+func NormalizeReasoningEffortLevel(payload []byte, model string) []byte {
 	out := payload

 	if effort := gjson.GetBytes(out, "reasoning_effort"); effort.Exists() {
@@ -298,10 +304,10 @@ func normalizeReasoningEffortLevel(payload []byte, model string) []byte {
 	return out
 }

-// validateThinkingConfig checks for unsupported reasoning levels on level-based models.
+// ValidateThinkingConfig checks for unsupported reasoning levels on level-based models.
 // Returns a statusErr with 400 when an unsupported level is supplied to avoid silently
 // downgrading requests.
-func validateThinkingConfig(payload []byte, model string) error {
+func ValidateThinkingConfig(payload []byte, model string) error {
 	if len(payload) == 0 || model == "" {
 		return nil
 	}
--- a/internal/runtime/executor/qwen_executor.go
+++ b/internal/runtime/executor/qwen_executor.go
@@ -51,13 +51,13 @@ func (e *QwenExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 	from := opts.SourceFormat
 	to := sdktranslator.FromString("openai")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
-	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort", false)
+	body = ApplyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort", false)
 	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
 	if upstreamModel != "" {
 		body, _ = sjson.SetBytes(body, "model", upstreamModel)
 	}
-	body = normalizeThinkingConfig(body, upstreamModel, false)
-	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
+	body = NormalizeThinkingConfig(body, upstreamModel, false)
+	if errValidate := ValidateThinkingConfig(body, upstreamModel); errValidate != nil {
 		return resp, errValidate
 	}
 	body = applyPayloadConfig(e.cfg, req.Model, body)
@@ -131,13 +131,13 @@ func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 	to := sdktranslator.FromString("openai")
 	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), true)

-	body = applyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort", false)
+	body = ApplyReasoningEffortMetadata(body, req.Metadata, req.Model, "reasoning_effort", false)
 	upstreamModel := util.ResolveOriginalModel(req.Model, req.Metadata)
 	if upstreamModel != "" {
 		body, _ = sjson.SetBytes(body, "model", upstreamModel)
 	}
-	body = normalizeThinkingConfig(body, upstreamModel, false)
-	if errValidate := validateThinkingConfig(body, upstreamModel); errValidate != nil {
+	body = NormalizeThinkingConfig(body, upstreamModel, false)
+	if errValidate := ValidateThinkingConfig(body, upstreamModel); errValidate != nil {
 		return nil, errValidate
 	}
 	toolsResult := gjson.GetBytes(body, "tools")
--- a/internal/translator/antigravity/gemini/antigravity_gemini_request.go
+++ b/internal/translator/antigravity/gemini/antigravity_gemini_request.go
@@ -122,6 +122,38 @@ type FunctionCallGroup struct {
 	ResponsesNeeded int
 }

+// parseFunctionResponse attempts to unmarshal a function response part.
+// Falls back to gjson extraction if standard json.Unmarshal fails.
+func parseFunctionResponse(response gjson.Result) map[string]interface{} {
+	var responseMap map[string]interface{}
+	err := json.Unmarshal([]byte(response.Raw), &responseMap)
+	if err == nil {
+		return responseMap
+	}
+
+	log.Debugf("unmarshal function response failed, using fallback: %v", err)
+	funcResp := response.Get("functionResponse")
+	if funcResp.Exists() {
+		fr := map[string]interface{}{
+			"name": funcResp.Get("name").String(),
+			"response": map[string]interface{}{
+				"result": funcResp.Get("response").String(),
+			},
+		}
+		if id := funcResp.Get("id").String(); id != "" {
+			fr["id"] = id
+		}
+		return map[string]interface{}{"functionResponse": fr}
+	}
+
+	return map[string]interface{}{
+		"functionResponse": map[string]interface{}{
+			"name":     "unknown",
+			"response": map[string]interface{}{"result": response.String()},
+		},
+	}
+}
+
 // fixCLIToolResponse performs sophisticated tool response format conversion and grouping.
 // This function transforms the CLI tool response format by intelligently grouping function calls
 // with their corresponding responses, ensuring proper conversation flow and API compatibility.
@@ -180,13 +212,7 @@ func fixCLIToolResponse(input string) (string, error) {
 					// Create merged function response content
 					var responseParts []interface{}
 					for _, response := range groupResponses {
-						var responseMap map[string]interface{}
-						errUnmarshal := json.Unmarshal([]byte(response.Raw), &responseMap)
-						if errUnmarshal != nil {
-							log.Warnf("failed to unmarshal function response: %v\n", errUnmarshal)
-							continue
-						}
-						responseParts = append(responseParts, responseMap)
+						responseParts = append(responseParts, parseFunctionResponse(response))
 					}

 					if len(responseParts) > 0 {
@@ -265,13 +291,7 @@ func fixCLIToolResponse(input string) (string, error) {

 			var responseParts []interface{}
 			for _, response := range groupResponses {
-				var responseMap map[string]interface{}
-				errUnmarshal := json.Unmarshal([]byte(response.Raw), &responseMap)
-				if errUnmarshal != nil {
-					log.Warnf("failed to unmarshal function response: %v\n", errUnmarshal)
-					continue
-				}
-				responseParts = append(responseParts, responseMap)
+				responseParts = append(responseParts, parseFunctionResponse(response))
 			}

 			if len(responseParts) > 0 {
--- a/internal/translator/antigravity/openai/chat-completions/antigravity_openai_request.go
+++ b/internal/translator/antigravity/openai/chat-completions/antigravity_openai_request.go
@@ -39,31 +39,13 @@ func ConvertOpenAIRequestToAntigravity(modelName string, inputRawJSON []byte, _
 	// Note: OpenAI official fields take precedence over extra_body.google.thinking_config
 	re := gjson.GetBytes(rawJSON, "reasoning_effort")
 	hasOfficialThinking := re.Exists()
-	if hasOfficialThinking && util.ModelSupportsThinking(modelName) {
-		switch re.String() {
-		case "none":
-			out, _ = sjson.DeleteBytes(out, "request.generationConfig.thinkingConfig.include_thoughts")
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.thinkingBudget", 0)
-		case "auto":
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.thinkingBudget", -1)
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.include_thoughts", true)
-		case "low":
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.thinkingBudget", 1024)
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.include_thoughts", true)
-		case "medium":
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.thinkingBudget", 8192)
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.include_thoughts", true)
-		case "high":
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.thinkingBudget", 32768)
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.include_thoughts", true)
-		default:
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.thinkingBudget", -1)
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.include_thoughts", true)
-		}
+	if hasOfficialThinking && util.ModelSupportsThinking(modelName) && !util.ModelUsesThinkingLevels(modelName) {
+		out = util.ApplyReasoningEffortToGeminiCLI(out, re.String())
 	}

 	// Cherry Studio extension extra_body.google.thinking_config (effective only when official fields are absent)
-	if !hasOfficialThinking && util.ModelSupportsThinking(modelName) {
+	// Only apply for models that use numeric budgets, not discrete levels.
+	if !hasOfficialThinking && util.ModelSupportsThinking(modelName) && !util.ModelUsesThinkingLevels(modelName) {
 		if tc := gjson.GetBytes(rawJSON, "extra_body.google.thinking_config"); tc.Exists() && tc.IsObject() {
 			var setBudget bool
 			var budget int
--- a/internal/translator/claude/gemini/claude_gemini_request.go
+++ b/internal/translator/claude/gemini/claude_gemini_request.go
@@ -114,14 +114,16 @@ func ConvertGeminiRequestToClaude(modelName string, inputRawJSON []byte, stream
 			}
 		}
 		// Include thoughts configuration for reasoning process visibility
-		if thinkingConfig := genConfig.Get("thinkingConfig"); thinkingConfig.Exists() && thinkingConfig.IsObject() {
-			if includeThoughts := thinkingConfig.Get("include_thoughts"); includeThoughts.Exists() {
-				if includeThoughts.Type == gjson.True {
-					out, _ = sjson.Set(out, "thinking.type", "enabled")
-					if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
-						out, _ = sjson.Set(out, "thinking.budget_tokens", thinkingBudget.Int())
-					}
-				}
+		// Only apply for models that support thinking and use numeric budgets, not discrete levels.
+		if thinkingConfig := genConfig.Get("thinkingConfig"); thinkingConfig.Exists() && thinkingConfig.IsObject() && util.ModelSupportsThinking(modelName) && !util.ModelUsesThinkingLevels(modelName) {
+			// Check for thinkingBudget first - if present, enable thinking with budget
+			if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() && thinkingBudget.Int() > 0 {
+				out, _ = sjson.Set(out, "thinking.type", "enabled")
+				normalizedBudget := util.NormalizeThinkingBudget(modelName, int(thinkingBudget.Int()))
+				out, _ = sjson.Set(out, "thinking.budget_tokens", normalizedBudget)
+			} else if includeThoughts := thinkingConfig.Get("include_thoughts"); includeThoughts.Exists() && includeThoughts.Type == gjson.True {
+				// Fallback to include_thoughts if no budget specified
+				out, _ = sjson.Set(out, "thinking.type", "enabled")
 			}
 		}
 	}
--- a/internal/translator/claude/openai/chat-completions/claude_openai_request.go
+++ b/internal/translator/claude/openai/chat-completions/claude_openai_request.go
@@ -16,6 +16,7 @@ import (
 	"strings"

 	"github.com/google/uuid"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
@@ -65,18 +66,23 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream

 	root := gjson.ParseBytes(rawJSON)

-	if v := root.Get("reasoning_effort"); v.Exists() {
-		out, _ = sjson.Set(out, "thinking.type", "enabled")
-
-		switch v.String() {
-		case "none":
-			out, _ = sjson.Set(out, "thinking.type", "disabled")
-		case "low":
-			out, _ = sjson.Set(out, "thinking.budget_tokens", 1024)
-		case "medium":
-			out, _ = sjson.Set(out, "thinking.budget_tokens", 8192)
-		case "high":
-			out, _ = sjson.Set(out, "thinking.budget_tokens", 24576)
+	if v := root.Get("reasoning_effort"); v.Exists() && util.ModelSupportsThinking(modelName) && !util.ModelUsesThinkingLevels(modelName) {
+		effort := strings.ToLower(strings.TrimSpace(v.String()))
+		if effort != "" {
+			budget, ok := util.ThinkingEffortToBudget(modelName, effort)
+			if ok {
+				switch budget {
+				case 0:
+					out, _ = sjson.Set(out, "thinking.type", "disabled")
+				case -1:
+					out, _ = sjson.Set(out, "thinking.type", "enabled")
+				default:
+					if budget > 0 {
+						out, _ = sjson.Set(out, "thinking.type", "enabled")
+						out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
+					}
+				}
+			}
 		}
 	}

--- a/internal/translator/claude/openai/responses/claude_openai-responses_request.go
+++ b/internal/translator/claude/openai/responses/claude_openai-responses_request.go
@@ -10,6 +10,7 @@ import (
 	"strings"

 	"github.com/google/uuid"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
@@ -52,20 +53,23 @@ func ConvertOpenAIResponsesRequestToClaude(modelName string, inputRawJSON []byte

 	root := gjson.ParseBytes(rawJSON)

-	if v := root.Get("reasoning.effort"); v.Exists() {
-		out, _ = sjson.Set(out, "thinking.type", "enabled")
-
-		switch v.String() {
-		case "none":
-			out, _ = sjson.Set(out, "thinking.type", "disabled")
-		case "minimal":
-			out, _ = sjson.Set(out, "thinking.budget_tokens", 1024)
-		case "low":
-			out, _ = sjson.Set(out, "thinking.budget_tokens", 4096)
-		case "medium":
-			out, _ = sjson.Set(out, "thinking.budget_tokens", 8192)
-		case "high":
-			out, _ = sjson.Set(out, "thinking.budget_tokens", 24576)
+	if v := root.Get("reasoning.effort"); v.Exists() && util.ModelSupportsThinking(modelName) && !util.ModelUsesThinkingLevels(modelName) {
+		effort := strings.ToLower(strings.TrimSpace(v.String()))
+		if effort != "" {
+			budget, ok := util.ThinkingEffortToBudget(modelName, effort)
+			if ok {
+				switch budget {
+				case 0:
+					out, _ = sjson.Set(out, "thinking.type", "disabled")
+				case -1:
+					out, _ = sjson.Set(out, "thinking.type", "enabled")
+				default:
+					if budget > 0 {
+						out, _ = sjson.Set(out, "thinking.type", "enabled")
+						out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
+					}
+				}
+			}
 		}
 	}

--- a/internal/translator/codex/claude/codex_claude_request.go
+++ b/internal/translator/codex/claude/codex_claude_request.go
@@ -12,6 +12,7 @@ import (
 	"strings"

 	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
@@ -214,7 +215,22 @@ func ConvertClaudeRequestToCodex(modelName string, inputRawJSON []byte, _ bool)

 	// Add additional configuration parameters for the Codex API.
 	template, _ = sjson.Set(template, "parallel_tool_calls", true)
-	template, _ = sjson.Set(template, "reasoning.effort", "medium")
+
+	// Convert thinking.budget_tokens to reasoning.effort for level-based models
+	reasoningEffort := "medium" // default
+	if thinking := rootResult.Get("thinking"); thinking.Exists() && thinking.IsObject() {
+		if thinking.Get("type").String() == "enabled" {
+			if util.ModelUsesThinkingLevels(modelName) {
+				if budgetTokens := thinking.Get("budget_tokens"); budgetTokens.Exists() {
+					budget := int(budgetTokens.Int())
+					if effort, ok := util.OpenAIThinkingBudgetToEffort(modelName, budget); ok && effort != "" {
+						reasoningEffort = effort
+					}
+				}
+			}
+		}
+	}
+	template, _ = sjson.Set(template, "reasoning.effort", reasoningEffort)
 	template, _ = sjson.Set(template, "reasoning.summary", "auto")
 	template, _ = sjson.Set(template, "stream", true)
 	template, _ = sjson.Set(template, "store", false)
--- a/internal/translator/codex/gemini/codex_gemini_request.go
+++ b/internal/translator/codex/gemini/codex_gemini_request.go
@@ -245,7 +245,22 @@ func ConvertGeminiRequestToCodex(modelName string, inputRawJSON []byte, _ bool)

 	// Fixed flags aligning with Codex expectations
 	out, _ = sjson.Set(out, "parallel_tool_calls", true)
-	out, _ = sjson.Set(out, "reasoning.effort", "medium")
+
+	// Convert thinkingBudget to reasoning.effort for level-based models
+	reasoningEffort := "medium" // default
+	if genConfig := root.Get("generationConfig"); genConfig.Exists() {
+		if thinkingConfig := genConfig.Get("thinkingConfig"); thinkingConfig.Exists() && thinkingConfig.IsObject() {
+			if util.ModelUsesThinkingLevels(modelName) {
+				if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
+					budget := int(thinkingBudget.Int())
+					if effort, ok := util.OpenAIThinkingBudgetToEffort(modelName, budget); ok && effort != "" {
+						reasoningEffort = effort
+					}
+				}
+			}
+		}
+	}
+	out, _ = sjson.Set(out, "reasoning.effort", reasoningEffort)
 	out, _ = sjson.Set(out, "reasoning.summary", "auto")
 	out, _ = sjson.Set(out, "stream", true)
 	out, _ = sjson.Set(out, "store", false)
--- a/internal/translator/gemini-cli/openai/chat-completions/gemini-cli_openai_request.go
+++ b/internal/translator/gemini-cli/openai/chat-completions/gemini-cli_openai_request.go
@@ -39,31 +39,13 @@ func ConvertOpenAIRequestToGeminiCLI(modelName string, inputRawJSON []byte, _ bo
 	// Note: OpenAI official fields take precedence over extra_body.google.thinking_config
 	re := gjson.GetBytes(rawJSON, "reasoning_effort")
 	hasOfficialThinking := re.Exists()
-	if hasOfficialThinking && util.ModelSupportsThinking(modelName) {
-		switch re.String() {
-		case "none":
-			out, _ = sjson.DeleteBytes(out, "request.generationConfig.thinkingConfig.include_thoughts")
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.thinkingBudget", 0)
-		case "auto":
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.thinkingBudget", -1)
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.include_thoughts", true)
-		case "low":
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.thinkingBudget", 1024)
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.include_thoughts", true)
-		case "medium":
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.thinkingBudget", 8192)
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.include_thoughts", true)
-		case "high":
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.thinkingBudget", 32768)
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.include_thoughts", true)
-		default:
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.thinkingBudget", -1)
-			out, _ = sjson.SetBytes(out, "request.generationConfig.thinkingConfig.include_thoughts", true)
-		}
+	if hasOfficialThinking && util.ModelSupportsThinking(modelName) && !util.ModelUsesThinkingLevels(modelName) {
+		out = util.ApplyReasoningEffortToGeminiCLI(out, re.String())
 	}

 	// Cherry Studio extension extra_body.google.thinking_config (effective only when official fields are absent)
-	if !hasOfficialThinking && util.ModelSupportsThinking(modelName) {
+	// Only apply for models that use numeric budgets, not discrete levels.
+	if !hasOfficialThinking && util.ModelSupportsThinking(modelName) && !util.ModelUsesThinkingLevels(modelName) {
 		if tc := gjson.GetBytes(rawJSON, "extra_body.google.thinking_config"); tc.Exists() && tc.IsObject() {
 			var setBudget bool
 			var budget int
--- a/internal/translator/gemini/claude/gemini_claude_request.go
+++ b/internal/translator/gemini/claude/gemini_claude_request.go
@@ -154,7 +154,8 @@ func ConvertClaudeRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 	}

 	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when enabled
-	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() && util.ModelSupportsThinking(modelName) {
+	// Only apply for models that use numeric budgets, not discrete levels.
+	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() && util.ModelSupportsThinking(modelName) && !util.ModelUsesThinkingLevels(modelName) {
 		if t.Get("type").String() == "enabled" {
 			if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
 				budget := int(b.Int())
--- a/internal/translator/gemini/openai/chat-completions/gemini_openai_request.go
+++ b/internal/translator/gemini/openai/chat-completions/gemini_openai_request.go
@@ -37,33 +37,17 @@ func ConvertOpenAIRequestToGemini(modelName string, inputRawJSON []byte, _ bool)

 	// Reasoning effort -> thinkingBudget/include_thoughts
 	// Note: OpenAI official fields take precedence over extra_body.google.thinking_config
+	// Only convert for models that use numeric budgets (not discrete levels) to avoid
+	// incorrectly applying thinkingBudget for level-based models like gpt-5.
 	re := gjson.GetBytes(rawJSON, "reasoning_effort")
 	hasOfficialThinking := re.Exists()
-	if hasOfficialThinking && util.ModelSupportsThinking(modelName) {
-		switch re.String() {
-		case "none":
-			out, _ = sjson.DeleteBytes(out, "generationConfig.thinkingConfig.include_thoughts")
-			out, _ = sjson.SetBytes(out, "generationConfig.thinkingConfig.thinkingBudget", 0)
-		case "auto":
-			out, _ = sjson.SetBytes(out, "generationConfig.thinkingConfig.thinkingBudget", -1)
-			out, _ = sjson.SetBytes(out, "generationConfig.thinkingConfig.include_thoughts", true)
-		case "low":
-			out, _ = sjson.SetBytes(out, "generationConfig.thinkingConfig.thinkingBudget", 1024)
-			out, _ = sjson.SetBytes(out, "generationConfig.thinkingConfig.include_thoughts", true)
-		case "medium":
-			out, _ = sjson.SetBytes(out, "generationConfig.thinkingConfig.thinkingBudget", 8192)
-			out, _ = sjson.SetBytes(out, "generationConfig.thinkingConfig.include_thoughts", true)
-		case "high":
-			out, _ = sjson.SetBytes(out, "generationConfig.thinkingConfig.thinkingBudget", 32768)
-			out, _ = sjson.SetBytes(out, "generationConfig.thinkingConfig.include_thoughts", true)
-		default:
-			out, _ = sjson.SetBytes(out, "generationConfig.thinkingConfig.thinkingBudget", -1)
-			out, _ = sjson.SetBytes(out, "generationConfig.thinkingConfig.include_thoughts", true)
-		}
+	if hasOfficialThinking && util.ModelSupportsThinking(modelName) && !util.ModelUsesThinkingLevels(modelName) {
+		out = util.ApplyReasoningEffortToGemini(out, re.String())
 	}

 	// Cherry Studio extension extra_body.google.thinking_config (effective only when official fields are absent)
-	if !hasOfficialThinking && util.ModelSupportsThinking(modelName) {
+	// Only apply for models that use numeric budgets, not discrete levels.
+	if !hasOfficialThinking && util.ModelSupportsThinking(modelName) && !util.ModelUsesThinkingLevels(modelName) {
 		if tc := gjson.GetBytes(rawJSON, "extra_body.google.thinking_config"); tc.Exists() && tc.IsObject() {
 			var setBudget bool
 			var budget int
--- a/internal/translator/gemini/openai/responses/gemini_openai-responses_request.go
+++ b/internal/translator/gemini/openai/responses/gemini_openai-responses_request.go
@@ -389,36 +389,16 @@ func ConvertOpenAIResponsesRequestToGemini(modelName string, inputRawJSON []byte
 	}

 	// OpenAI official reasoning fields take precedence
+	// Only convert for models that use numeric budgets (not discrete levels).
 	hasOfficialThinking := root.Get("reasoning.effort").Exists()
-	if hasOfficialThinking && util.ModelSupportsThinking(modelName) {
+	if hasOfficialThinking && util.ModelSupportsThinking(modelName) && !util.ModelUsesThinkingLevels(modelName) {
 		reasoningEffort := root.Get("reasoning.effort")
-		switch reasoningEffort.String() {
-		case "none":
-			out, _ = sjson.Set(out, "generationConfig.thinkingConfig.include_thoughts", false)
-			out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingBudget", 0)
-		case "auto":
-			out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingBudget", -1)
-			out, _ = sjson.Set(out, "generationConfig.thinkingConfig.include_thoughts", true)
-		case "minimal":
-			out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingBudget", 1024)
-			out, _ = sjson.Set(out, "generationConfig.thinkingConfig.include_thoughts", true)
-		case "low":
-			out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingBudget", 4096)
-			out, _ = sjson.Set(out, "generationConfig.thinkingConfig.include_thoughts", true)
-		case "medium":
-			out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingBudget", 8192)
-			out, _ = sjson.Set(out, "generationConfig.thinkingConfig.include_thoughts", true)
-		case "high":
-			out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingBudget", 32768)
-			out, _ = sjson.Set(out, "generationConfig.thinkingConfig.include_thoughts", true)
-		default:
-			out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingBudget", -1)
-			out, _ = sjson.Set(out, "generationConfig.thinkingConfig.include_thoughts", true)
-		}
+		out = string(util.ApplyReasoningEffortToGemini([]byte(out), reasoningEffort.String()))
 	}

 	// Cherry Studio extension (applies only when official fields are missing)
-	if !hasOfficialThinking && util.ModelSupportsThinking(modelName) {
+	// Only apply for models that use numeric budgets, not discrete levels.
+	if !hasOfficialThinking && util.ModelSupportsThinking(modelName) && !util.ModelUsesThinkingLevels(modelName) {
 		if tc := root.Get("extra_body.google.thinking_config"); tc.Exists() && tc.IsObject() {
 			var setBudget bool
 			var budget int
--- a/internal/translator/openai/claude/openai_claude_request.go
+++ b/internal/translator/openai/claude/openai_claude_request.go
@@ -10,6 +10,7 @@ import (
 	"encoding/json"
 	"strings"

+	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
@@ -60,6 +61,18 @@ func ConvertClaudeRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 	// Stream
 	out, _ = sjson.Set(out, "stream", stream)

+	// Thinking: Convert Claude thinking.budget_tokens to OpenAI reasoning_effort
+	if thinking := root.Get("thinking"); thinking.Exists() && thinking.IsObject() {
+		if thinkingType := thinking.Get("type"); thinkingType.Exists() && thinkingType.String() == "enabled" {
+			if budgetTokens := thinking.Get("budget_tokens"); budgetTokens.Exists() {
+				budget := int(budgetTokens.Int())
+				if effort, ok := util.OpenAIThinkingBudgetToEffort(modelName, budget); ok && effort != "" {
+					out, _ = sjson.Set(out, "reasoning_effort", effort)
+				}
+			}
+		}
+	}
+
 	// Process messages and system
 	var messagesJSON = "[]"

--- a/internal/translator/openai/gemini/openai_gemini_request.go
+++ b/internal/translator/openai/gemini/openai_gemini_request.go
@@ -13,6 +13,7 @@ import (
 	"math/big"
 	"strings"

+	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )
@@ -76,6 +77,17 @@ func ConvertGeminiRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 				out, _ = sjson.Set(out, "stop", stops)
 			}
 		}
+
+		// Convert thinkingBudget to reasoning_effort
+		// Always perform conversion to support allowCompat models that may not be in registry
+		if thinkingConfig := genConfig.Get("thinkingConfig"); thinkingConfig.Exists() && thinkingConfig.IsObject() {
+			if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
+				budget := int(thinkingBudget.Int())
+				if effort, ok := util.OpenAIThinkingBudgetToEffort(modelName, budget); ok && effort != "" {
+					out, _ = sjson.Set(out, "reasoning_effort", effort)
+				}
+			}
+		}
 	}

 	// Stream parameter
--- a/internal/translator/openai/openai/responses/openai_openai-responses_request.go
+++ b/internal/translator/openai/openai/responses/openai_openai-responses_request.go
@@ -2,6 +2,7 @@ package responses

 import (
 	"bytes"
+	"strings"

 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -64,7 +65,7 @@ func ConvertOpenAIResponsesRequestToOpenAIChatCompletions(modelName string, inpu
 			}

 			switch itemType {
-			case "message":
+			case "message", "":
 				// Handle regular message conversion
 				role := item.Get("role").String()
 				message := `{"role":"","content":""}`
@@ -106,6 +107,8 @@ func ConvertOpenAIResponsesRequestToOpenAIChatCompletions(modelName string, inpu
 					if len(toolCalls) > 0 {
 						message, _ = sjson.Set(message, "tool_calls", toolCalls)
 					}
+				} else if content.Type == gjson.String {
+					message, _ = sjson.Set(message, "content", content.String())
 				}

 				out, _ = sjson.SetRaw(out, "messages.-1", message)
@@ -189,23 +192,9 @@ func ConvertOpenAIResponsesRequestToOpenAIChatCompletions(modelName string, inpu
 	}

 	if reasoningEffort := root.Get("reasoning.effort"); reasoningEffort.Exists() {
-		switch reasoningEffort.String() {
-		case "none":
-			out, _ = sjson.Set(out, "reasoning_effort", "none")
-		case "auto":
-			out, _ = sjson.Set(out, "reasoning_effort", "auto")
-		case "minimal":
-			out, _ = sjson.Set(out, "reasoning_effort", "low")
-		case "low":
-			out, _ = sjson.Set(out, "reasoning_effort", "low")
-		case "medium":
-			out, _ = sjson.Set(out, "reasoning_effort", "medium")
-		case "high":
-			out, _ = sjson.Set(out, "reasoning_effort", "high")
-		case "xhigh":
-			out, _ = sjson.Set(out, "reasoning_effort", "xhigh")
-		default:
-			out, _ = sjson.Set(out, "reasoning_effort", "auto")
+		effort := strings.ToLower(strings.TrimSpace(reasoningEffort.String()))
+		if effort != "" {
+			out, _ = sjson.Set(out, "reasoning_effort", effort)
 		}
 	}

--- a/internal/util/gemini_thinking.go
+++ b/internal/util/gemini_thinking.go
@@ -152,6 +152,71 @@ func NormalizeGeminiCLIThinkingBudget(model string, body []byte) []byte {
 	return updated
 }

+// ReasoningEffortBudgetMapping defines the thinkingBudget values for each reasoning effort level.
+var ReasoningEffortBudgetMapping = map[string]int{
+	"none":    0,
+	"auto":    -1,
+	"minimal": 512,
+	"low":     1024,
+	"medium":  8192,
+	"high":    24576,
+	"xhigh":   32768,
+}
+
+// ApplyReasoningEffortToGemini applies OpenAI reasoning_effort to Gemini thinkingConfig
+// for standard Gemini API format (generationConfig.thinkingConfig path).
+// Returns the modified body with thinkingBudget and include_thoughts set.
+func ApplyReasoningEffortToGemini(body []byte, effort string) []byte {
+	normalized := strings.ToLower(strings.TrimSpace(effort))
+	if normalized == "" {
+		return body
+	}
+
+	budgetPath := "generationConfig.thinkingConfig.thinkingBudget"
+	includePath := "generationConfig.thinkingConfig.include_thoughts"
+
+	if normalized == "none" {
+		body, _ = sjson.DeleteBytes(body, "generationConfig.thinkingConfig")
+		return body
+	}
+
+	budget, ok := ReasoningEffortBudgetMapping[normalized]
+	if !ok {
+		return body
+	}
+
+	body, _ = sjson.SetBytes(body, budgetPath, budget)
+	body, _ = sjson.SetBytes(body, includePath, true)
+	return body
+}
+
+// ApplyReasoningEffortToGeminiCLI applies OpenAI reasoning_effort to Gemini CLI thinkingConfig
+// for Gemini CLI API format (request.generationConfig.thinkingConfig path).
+// Returns the modified body with thinkingBudget and include_thoughts set.
+func ApplyReasoningEffortToGeminiCLI(body []byte, effort string) []byte {
+	normalized := strings.ToLower(strings.TrimSpace(effort))
+	if normalized == "" {
+		return body
+	}
+
+	budgetPath := "request.generationConfig.thinkingConfig.thinkingBudget"
+	includePath := "request.generationConfig.thinkingConfig.include_thoughts"
+
+	if normalized == "none" {
+		body, _ = sjson.DeleteBytes(body, "request.generationConfig.thinkingConfig")
+		return body
+	}
+
+	budget, ok := ReasoningEffortBudgetMapping[normalized]
+	if !ok {
+		return body
+	}
+
+	body, _ = sjson.SetBytes(body, budgetPath, budget)
+	body, _ = sjson.SetBytes(body, includePath, true)
+	return body
+}
+
 // ConvertThinkingLevelToBudget checks for "generationConfig.thinkingConfig.thinkingLevel"
 // and converts it to "thinkingBudget".
 // "high" -> 32768
--- a/internal/util/openai_thinking.go
+++ b/internal/util/openai_thinking.go
@@ -5,15 +5,18 @@ package util
 //
 // Ranges:
 //   - 0            -> "none"
+//   - -1           -> "auto"
 //   - 1..1024      -> "low"
 //   - 1025..8192   -> "medium"
 //   - 8193..24576  -> "high"
 //   - 24577..      -> highest supported level for the model (defaults to "xhigh")
 //
-// Negative values (except the dynamic -1 handled elsewhere) are treated as unsupported.
+// Negative values other than -1 are treated as unsupported.
 func OpenAIThinkingBudgetToEffort(model string, budget int) (string, bool) {
 	switch {
-	case budget < 0:
+	case budget == -1:
+		return "auto", true
+	case budget < -1:
 		return "", false
 	case budget == 0:
 		return "none", true
--- a/sdk/api/handlers/handlers.go
+++ b/sdk/api/handlers/handlers.go
@@ -5,6 +5,7 @@ package handlers

 import (
 	"bytes"
+	"encoding/json"
 	"fmt"
 	"net/http"
 	"strings"
@@ -117,6 +118,16 @@ func (h *BaseAPIHandler) GetContextWithCancel(handler interfaces.APIHandler, c *
 	newCtx = context.WithValue(newCtx, "handler", handler)
 	return newCtx, func(params ...interface{}) {
 		if h.Cfg.RequestLog && len(params) == 1 {
+			if existing, exists := c.Get("API_RESPONSE"); exists {
+				if existingBytes, ok := existing.([]byte); ok && len(bytes.TrimSpace(existingBytes)) > 0 {
+					switch params[0].(type) {
+					case error, string:
+						cancel()
+						return
+					}
+				}
+			}
+
 			var payload []byte
 			switch data := params[0].(type) {
 			case []byte:
@@ -437,12 +448,53 @@ func (h *BaseAPIHandler) WriteErrorResponse(c *gin.Context, msg *interfaces.Erro
 			}
 		}
 	}
-	c.Status(status)
+
+	errText := http.StatusText(status)
 	if msg != nil && msg.Error != nil {
-		_, _ = c.Writer.Write([]byte(msg.Error.Error()))
-	} else {
-		_, _ = c.Writer.Write([]byte(http.StatusText(status)))
+		if v := strings.TrimSpace(msg.Error.Error()); v != "" {
+			errText = v
+		}
 	}
+
+	// Prefer preserving upstream JSON error bodies when possible.
+	buildJSONBody := func() []byte {
+		trimmed := strings.TrimSpace(errText)
+		if trimmed != "" && json.Valid([]byte(trimmed)) {
+			return []byte(trimmed)
+		}
+		errType := "invalid_request_error"
+		switch status {
+		case http.StatusUnauthorized:
+			errType = "authentication_error"
+		case http.StatusForbidden:
+			errType = "permission_error"
+		case http.StatusTooManyRequests:
+			errType = "rate_limit_error"
+		default:
+			if status >= http.StatusInternalServerError {
+				errType = "server_error"
+			}
+		}
+		payload, err := json.Marshal(ErrorResponse{
+			Error: ErrorDetail{
+				Message: errText,
+				Type:    errType,
+			},
+		})
+		if err != nil {
+			return []byte(fmt.Sprintf(`{"error":{"message":%q,"type":"server_error"}}`, errText))
+		}
+		return payload
+	}
+
+	body := buildJSONBody()
+	c.Set("API_RESPONSE", bytes.Clone(body))
+
+	if !c.Writer.Written() {
+		c.Writer.Header().Set("Content-Type", "application/json")
+	}
+	c.Status(status)
+	_, _ = c.Writer.Write(body)
 }

 func (h *BaseAPIHandler) LoggingAPIResponseError(ctx context.Context, err *interfaces.ErrorMessage) {
--- a/sdk/cliproxy/auth/manager.go
+++ b/sdk/cliproxy/auth/manager.go
@@ -375,10 +375,19 @@ func (m *Manager) executeWithProvider(ctx context.Context, provider string, req
 		}

 		accountType, accountInfo := auth.AccountInfo()
+		proxyInfo := auth.ProxyInfo()
 		if accountType == "api_key" {
-			log.Debugf("Use API key %s for model %s", util.HideAPIKey(accountInfo), req.Model)
+			if proxyInfo != "" {
+				log.Debugf("Use API key %s for model %s %s", util.HideAPIKey(accountInfo), req.Model, proxyInfo)
+			} else {
+				log.Debugf("Use API key %s for model %s", util.HideAPIKey(accountInfo), req.Model)
+			}
 		} else if accountType == "oauth" {
-			log.Debugf("Use OAuth %s for model %s", accountInfo, req.Model)
+			if proxyInfo != "" {
+				log.Debugf("Use OAuth %s for model %s %s", accountInfo, req.Model, proxyInfo)
+			} else {
+				log.Debugf("Use OAuth %s for model %s", accountInfo, req.Model)
+			}
 		}

 		tried[auth.ID] = struct{}{}
@@ -423,10 +432,19 @@ func (m *Manager) executeCountWithProvider(ctx context.Context, provider string,
 		}

 		accountType, accountInfo := auth.AccountInfo()
+		proxyInfo := auth.ProxyInfo()
 		if accountType == "api_key" {
-			log.Debugf("Use API key %s for model %s", util.HideAPIKey(accountInfo), req.Model)
+			if proxyInfo != "" {
+				log.Debugf("Use API key %s for model %s %s", util.HideAPIKey(accountInfo), req.Model, proxyInfo)
+			} else {
+				log.Debugf("Use API key %s for model %s", util.HideAPIKey(accountInfo), req.Model)
+			}
 		} else if accountType == "oauth" {
-			log.Debugf("Use OAuth %s for model %s", accountInfo, req.Model)
+			if proxyInfo != "" {
+				log.Debugf("Use OAuth %s for model %s %s", accountInfo, req.Model, proxyInfo)
+			} else {
+				log.Debugf("Use OAuth %s for model %s", accountInfo, req.Model)
+			}
 		}

 		tried[auth.ID] = struct{}{}
@@ -471,10 +489,19 @@ func (m *Manager) executeStreamWithProvider(ctx context.Context, provider string
 		}

 		accountType, accountInfo := auth.AccountInfo()
+		proxyInfo := auth.ProxyInfo()
 		if accountType == "api_key" {
-			log.Debugf("Use API key %s for model %s", util.HideAPIKey(accountInfo), req.Model)
+			if proxyInfo != "" {
+				log.Debugf("Use API key %s for model %s %s", util.HideAPIKey(accountInfo), req.Model, proxyInfo)
+			} else {
+				log.Debugf("Use API key %s for model %s", util.HideAPIKey(accountInfo), req.Model)
+			}
 		} else if accountType == "oauth" {
-			log.Debugf("Use OAuth %s for model %s", accountInfo, req.Model)
+			if proxyInfo != "" {
+				log.Debugf("Use OAuth %s for model %s %s", accountInfo, req.Model, proxyInfo)
+			} else {
+				log.Debugf("Use OAuth %s for model %s", accountInfo, req.Model)
+			}
 		}

 		tried[auth.ID] = struct{}{}
--- a/sdk/cliproxy/auth/types.go
+++ b/sdk/cliproxy/auth/types.go
@@ -157,6 +157,20 @@ func (m *ModelState) Clone() *ModelState {
 	return &copyState
 }

+func (a *Auth) ProxyInfo() string {
+	if a == nil {
+		return ""
+	}
+	proxyStr := strings.TrimSpace(a.ProxyURL)
+	if proxyStr == "" {
+		return ""
+	}
+	if idx := strings.Index(proxyStr, "://"); idx > 0 {
+		return "via " + proxyStr[:idx] + " proxy"
+	}
+	return "via proxy"
+}
+
 func (a *Auth) AccountInfo() (string, string) {
 	if a == nil {
 		return "", ""
--- a/test/thinking_conversion_test.go
+++ b/test/thinking_conversion_test.go
@@ -2,7 +2,6 @@ package test

 import (
 	"fmt"
-	"net/http"
 	"strings"
 	"testing"
 	"time"
@@ -10,20 +9,20 @@ import (
 	_ "github.com/router-for-me/CLIProxyAPI/v6/internal/translator"

 	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/runtime/executor"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	sdktranslator "github.com/router-for-me/CLIProxyAPI/v6/sdk/translator"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
 )

-// statusErr mirrors executor.statusErr to keep validation behavior aligned.
-type statusErr struct {
-	code int
-	msg  string
+// isOpenAICompatModel returns true if the model is configured as an OpenAI-compatible
+// model that should have reasoning effort passed through even if not in registry.
+// This simulates the allowCompat behavior from OpenAICompatExecutor.
+func isOpenAICompatModel(model string) bool {
+	return model == "openai-compat"
 }

-func (e statusErr) Error() string { return e.msg }
-
 // registerCoreModels loads representative models across providers into the registry
 // so NormalizeThinkingBudget and level validation use real ranges.
 func registerCoreModels(t *testing.T) func() {
@@ -34,14 +33,62 @@ func registerCoreModels(t *testing.T) func() {
 	reg.RegisterClient(uid+"-claude", "claude", registry.GetClaudeModels())
 	reg.RegisterClient(uid+"-openai", "codex", registry.GetOpenAIModels())
 	reg.RegisterClient(uid+"-qwen", "qwen", registry.GetQwenModels())
+	// Custom openai-compatible model with forced thinking suffix passthrough.
+	// No Thinking field - simulates an external model added via openai-compat
+	// where the registry has no knowledge of its thinking capabilities.
+	// The allowCompat flag should preserve reasoning effort for such models.
+	customOpenAIModels := []*registry.ModelInfo{
+		{
+			ID:          "openai-compat",
+			Object:      "model",
+			Created:     1700000000,
+			OwnedBy:     "custom-provider",
+			Type:        "openai",
+			DisplayName: "OpenAI Compatible Model",
+			Description: "OpenAI-compatible model with forced thinking suffix support",
+		},
+	}
+	reg.RegisterClient(uid+"-custom-openai", "codex", customOpenAIModels)
 	return func() {
 		reg.UnregisterClient(uid + "-gemini")
 		reg.UnregisterClient(uid + "-claude")
 		reg.UnregisterClient(uid + "-openai")
 		reg.UnregisterClient(uid + "-qwen")
+		reg.UnregisterClient(uid + "-custom-openai")
 	}
 }

+var (
+	thinkingTestModels = []string{
+		"gpt-5",           // level-based thinking model
+		"gemini-2.5-pro",  // numeric-budget thinking model
+		"qwen3-code-plus", // no thinking support
+		"openai-compat",   // allowCompat=true (OpenAI-compatible channel)
+	}
+	thinkingTestFromProtocols = []string{"openai", "claude", "gemini", "openai-response"}
+	thinkingTestToProtocols   = []string{"gemini", "claude", "openai", "codex"}
+
+	// Numeric budgets and their level equivalents:
+	// -1 -> auto
+	// 0 -> none
+	// 1..1024 -> low
+	// 1025..8192 -> medium
+	// 8193..24576 -> high
+	// >24576 -> model highest level (right-most in Levels)
+	thinkingNumericSamples = []int{-1, 0, 1023, 1025, 8193, 64000}
+
+	// Levels and their numeric equivalents:
+	// auto -> -1
+	// none -> 0
+	// minimal -> 512
+	// low -> 1024
+	// medium -> 8192
+	// high -> 24576
+	// xhigh -> 32768
+	// invalid -> invalid (no mapping)
+	thinkingLevelSamples = []string{"auto", "none", "minimal", "low", "medium", "high", "xhigh", "invalid"}
+)
+
 func buildRawPayload(fromProtocol, modelWithSuffix string) []byte {
 	switch fromProtocol {
 	case "gemini":
@@ -53,141 +100,10 @@ func buildRawPayload(fromProtocol, modelWithSuffix string) []byte {
 	}
 }

-// applyThinkingMetadataLocal mirrors executor.applyThinkingMetadata.
-func applyThinkingMetadataLocal(payload []byte, metadata map[string]any, model string) []byte {
-	budgetOverride, includeOverride, ok := util.ResolveThinkingConfigFromMetadata(model, metadata)
-	if !ok || (budgetOverride == nil && includeOverride == nil) {
-		return payload
-	}
-	if !util.ModelSupportsThinking(model) {
-		return payload
-	}
-	if budgetOverride != nil {
-		norm := util.NormalizeThinkingBudget(model, *budgetOverride)
-		budgetOverride = &norm
-	}
-	return util.ApplyGeminiThinkingConfig(payload, budgetOverride, includeOverride)
-}
-
-// applyReasoningEffortMetadataLocal mirrors executor.applyReasoningEffortMetadata.
-func applyReasoningEffortMetadataLocal(payload []byte, metadata map[string]any, model, field string) []byte {
-	if len(metadata) == 0 {
-		return payload
-	}
-	if !util.ModelSupportsThinking(model) {
-		return payload
-	}
-	if field == "" {
-		return payload
-	}
-	if effort, ok := util.ReasoningEffortFromMetadata(metadata); ok && effort != "" {
-		if updated, err := sjson.SetBytes(payload, field, effort); err == nil {
-			return updated
-		}
-	}
-	if util.ModelUsesThinkingLevels(model) {
-		if budget, _, _, matched := util.ThinkingFromMetadata(metadata); matched && budget != nil {
-			if effort, ok := util.OpenAIThinkingBudgetToEffort(model, *budget); ok && effort != "" {
-				if updated, err := sjson.SetBytes(payload, field, effort); err == nil {
-					return updated
-				}
-			}
-		}
-	}
-	return payload
-}
-
-// normalizeThinkingConfigLocal mirrors executor.normalizeThinkingConfig.
-func normalizeThinkingConfigLocal(payload []byte, model string) []byte {
-	if len(payload) == 0 || model == "" {
-		return payload
-	}
-
-	if !util.ModelSupportsThinking(model) {
-		return stripThinkingFieldsLocal(payload, false)
-	}
-
-	if util.ModelUsesThinkingLevels(model) {
-		return normalizeReasoningEffortLevelLocal(payload, model)
-	}
-
-	// Model supports thinking but uses numeric budgets, not levels.
-	// Strip effort string fields since they are not applicable.
-	return stripThinkingFieldsLocal(payload, true)
-}
-
-// stripThinkingFieldsLocal mirrors executor.stripThinkingFields.
-func stripThinkingFieldsLocal(payload []byte, effortOnly bool) []byte {
-	fieldsToRemove := []string{
-		"reasoning_effort",
-		"reasoning.effort",
-	}
-	if !effortOnly {
-		fieldsToRemove = append([]string{"reasoning"}, fieldsToRemove...)
-	}
-	out := payload
-	for _, field := range fieldsToRemove {
-		if gjson.GetBytes(out, field).Exists() {
-			out, _ = sjson.DeleteBytes(out, field)
-		}
-	}
-	return out
-}
-
-// normalizeReasoningEffortLevelLocal mirrors executor.normalizeReasoningEffortLevel.
-func normalizeReasoningEffortLevelLocal(payload []byte, model string) []byte {
-	out := payload
-
-	if effort := gjson.GetBytes(out, "reasoning_effort"); effort.Exists() {
-		if normalized, ok := util.NormalizeReasoningEffortLevel(model, effort.String()); ok {
-			out, _ = sjson.SetBytes(out, "reasoning_effort", normalized)
-		}
-	}
-
-	if effort := gjson.GetBytes(out, "reasoning.effort"); effort.Exists() {
-		if normalized, ok := util.NormalizeReasoningEffortLevel(model, effort.String()); ok {
-			out, _ = sjson.SetBytes(out, "reasoning.effort", normalized)
-		}
-	}
-
-	return out
-}
-
-// validateThinkingConfigLocal mirrors executor.validateThinkingConfig.
-func validateThinkingConfigLocal(payload []byte, model string) error {
-	if len(payload) == 0 || model == "" {
-		return nil
-	}
-	if !util.ModelSupportsThinking(model) || !util.ModelUsesThinkingLevels(model) {
-		return nil
-	}
-
-	levels := util.GetModelThinkingLevels(model)
-	checkField := func(path string) error {
-		if effort := gjson.GetBytes(payload, path); effort.Exists() {
-			if _, ok := util.NormalizeReasoningEffortLevel(model, effort.String()); !ok {
-				return statusErr{
-					code: http.StatusBadRequest,
-					msg:  fmt.Sprintf("unsupported reasoning effort level %q for model %s (supported: %s)", effort.String(), model, strings.Join(levels, ", ")),
-				}
-			}
-		}
-		return nil
-	}
-
-	if err := checkField("reasoning_effort"); err != nil {
-		return err
-	}
-	if err := checkField("reasoning.effort"); err != nil {
-		return err
-	}
-	return nil
-}
-
 // normalizeCodexPayload mirrors codex_executor's reasoning + streaming tweaks.
-func normalizeCodexPayload(body []byte, upstreamModel string) ([]byte, error) {
-	body = normalizeThinkingConfigLocal(body, upstreamModel)
-	if err := validateThinkingConfigLocal(body, upstreamModel); err != nil {
+func normalizeCodexPayload(body []byte, upstreamModel string, allowCompat bool) ([]byte, error) {
+	body = executor.NormalizeThinkingConfig(body, upstreamModel, allowCompat)
+	if err := executor.ValidateThinkingConfig(body, upstreamModel); err != nil {
 		return body, err
 	}
 	body, _ = sjson.SetBytes(body, "model", upstreamModel)
@@ -214,9 +130,10 @@ func buildBodyForProtocol(t *testing.T, fromProtocol, toProtocol, modelWithSuffi
 	)

 	var err error
+	allowCompat := isOpenAICompatModel(normalizedModel)
 	switch toProtocol {
 	case "gemini":
-		body = applyThinkingMetadataLocal(body, metadata, normalizedModel)
+		body = executor.ApplyThinkingMetadata(body, metadata, normalizedModel)
 		body = util.ApplyDefaultThinkingIfNeeded(normalizedModel, body)
 		body = util.NormalizeGeminiThinkingBudget(normalizedModel, body)
 		body = util.StripThinkingConfigIfUnsupported(normalizedModel, body)
@@ -225,13 +142,14 @@ func buildBodyForProtocol(t *testing.T, fromProtocol, toProtocol, modelWithSuffi
 			body = util.ApplyClaudeThinkingConfig(body, budget)
 		}
 	case "openai":
-		body = applyReasoningEffortMetadataLocal(body, metadata, normalizedModel, "reasoning_effort")
-		body = normalizeThinkingConfigLocal(body, upstreamModel)
-		err = validateThinkingConfigLocal(body, upstreamModel)
+		body = executor.ApplyReasoningEffortMetadata(body, metadata, normalizedModel, "reasoning_effort", allowCompat)
+		body = executor.NormalizeThinkingConfig(body, upstreamModel, allowCompat)
+		err = executor.ValidateThinkingConfig(body, upstreamModel)
 	case "codex": // OpenAI responses / codex
-		body = applyReasoningEffortMetadataLocal(body, metadata, normalizedModel, "reasoning.effort")
+		// Codex does not support allowCompat; always use false.
+		body = executor.ApplyReasoningEffortMetadata(body, metadata, normalizedModel, "reasoning.effort", false)
 		// Mirror CodexExecutor final normalization and model override so tests log the final body.
-		body, err = normalizeCodexPayload(body, upstreamModel)
+		body, err = normalizeCodexPayload(body, upstreamModel, false)
 	default:
 	}

@@ -287,83 +205,46 @@ func TestThinkingConversionsAcrossProtocolsAndModels(t *testing.T) {
 	cleanup := registerCoreModels(t)
 	defer cleanup()

-	models := []string{
-		"gpt-5",             // supports levels (low/medium/high)
-		"gemini-2.5-pro",    // supports numeric budget
-		"qwen3-coder-flash", // no thinking support
-	}
-	fromProtocols := []string{"openai", "claude", "gemini", "openai-response"}
-	toProtocols := []string{"gemini", "claude", "openai", "codex"}
-
 	type scenario struct {
 		name        string
 		modelSuffix string
-		expectFn    func(info *registry.ModelInfo) (present bool, budget int64)
 	}

-	buildBudgetFn := func(raw int) func(info *registry.ModelInfo) (bool, int64) {
-		return func(info *registry.ModelInfo) (bool, int64) {
-			if info == nil || info.Thinking == nil {
-				return false, 0
-			}
-			return true, int64(util.NormalizeThinkingBudget(info.ID, raw))
+	numericName := func(budget int) string {
+		if budget < 0 {
+			return "numeric-neg1"
 		}
+		return fmt.Sprintf("numeric-%d", budget)
 	}

-	levelBudgetFn := func(level string) func(info *registry.ModelInfo) (bool, int64) {
-		return func(info *registry.ModelInfo) (bool, int64) {
-			if info == nil || info.Thinking == nil {
-				return false, 0
-			}
-			if b, ok := util.ThinkingEffortToBudget(info.ID, level); ok {
-				return true, int64(b)
-			}
-			return false, 0
-		}
-	}
+	for _, model := range thinkingTestModels {
+		_ = registry.GetGlobalRegistry().GetModelInfo(model)

-	for _, model := range models {
-		info := registry.GetGlobalRegistry().GetModelInfo(model)
-		min, max := 0, 0
-		if info != nil && info.Thinking != nil {
-			min = info.Thinking.Min
-			max = info.Thinking.Max
-		}
-
-		for _, from := range fromProtocols {
+		for _, from := range thinkingTestFromProtocols {
 			// Scenario selection follows protocol semantics:
 			// - OpenAI-style protocols (openai/openai-response) express thinking as levels.
 			// - Claude/Gemini-style protocols express thinking as numeric budgets.
 			cases := []scenario{
-				{name: "no-suffix", modelSuffix: model, expectFn: func(_ *registry.ModelInfo) (bool, int64) { return false, 0 }},
+				{name: "no-suffix", modelSuffix: model},
 			}
 			if from == "openai" || from == "openai-response" {
-				cases = append(cases,
-					scenario{name: "level-low", modelSuffix: fmt.Sprintf("%s(low)", model), expectFn: levelBudgetFn("low")},
-					scenario{name: "level-high", modelSuffix: fmt.Sprintf("%s(high)", model), expectFn: levelBudgetFn("high")},
-					scenario{name: "level-auto", modelSuffix: fmt.Sprintf("%s(auto)", model), expectFn: levelBudgetFn("auto")},
-				)
+				for _, lvl := range thinkingLevelSamples {
+					cases = append(cases, scenario{
+						name:        "level-" + lvl,
+						modelSuffix: fmt.Sprintf("%s(%s)", model, lvl),
+					})
+				}
 			} else { // claude or gemini
-				if util.ModelUsesThinkingLevels(model) {
-					// Numeric budgets for level-based models are mapped into levels when needed.
-					cases = append(cases,
-						scenario{name: "numeric-0", modelSuffix: fmt.Sprintf("%s(0)", model), expectFn: buildBudgetFn(0)},
-						scenario{name: "numeric-1024", modelSuffix: fmt.Sprintf("%s(1024)", model), expectFn: buildBudgetFn(1024)},
-						scenario{name: "numeric-1025", modelSuffix: fmt.Sprintf("%s(1025)", model), expectFn: buildBudgetFn(1025)},
-						scenario{name: "numeric-8192", modelSuffix: fmt.Sprintf("%s(8192)", model), expectFn: buildBudgetFn(8192)},
-						scenario{name: "numeric-8193", modelSuffix: fmt.Sprintf("%s(8193)", model), expectFn: buildBudgetFn(8193)},
-						scenario{name: "numeric-24576", modelSuffix: fmt.Sprintf("%s(24576)", model), expectFn: buildBudgetFn(24576)},
-						scenario{name: "numeric-24577", modelSuffix: fmt.Sprintf("%s(24577)", model), expectFn: buildBudgetFn(24577)},
-					)
-				} else {
-					cases = append(cases,
-						scenario{name: "numeric-below-min", modelSuffix: fmt.Sprintf("%s(%d)", model, min-10), expectFn: buildBudgetFn(min - 10)},
-						scenario{name: "numeric-above-max", modelSuffix: fmt.Sprintf("%s(%d)", model, max+10), expectFn: buildBudgetFn(max + 10)},
-					)
+				for _, budget := range thinkingNumericSamples {
+					budget := budget
+					cases = append(cases, scenario{
+						name:        numericName(budget),
+						modelSuffix: fmt.Sprintf("%s(%d)", model, budget),
+					})
 				}
 			}

-			for _, to := range toProtocols {
+			for _, to := range thinkingTestToProtocols {
 				if from == to {
 					continue
 				}
@@ -402,7 +283,22 @@ func TestThinkingConversionsAcrossProtocolsAndModels(t *testing.T) {
 								}
 								return true, fmt.Sprintf("%d", *budget), false
 							case "openai":
-								if !util.ModelSupportsThinking(normalizedModel) {
+								allowCompat := isOpenAICompatModel(normalizedModel)
+								if !util.ModelSupportsThinking(normalizedModel) && !allowCompat {
+									return false, "", false
+								}
+								// For allowCompat models, pass through effort directly without validation
+								if allowCompat {
+									effort, ok := util.ReasoningEffortFromMetadata(metadata)
+									if ok && strings.TrimSpace(effort) != "" {
+										return true, strings.ToLower(strings.TrimSpace(effort)), false
+									}
+									// Check numeric budget fallback for allowCompat
+									if budget, _, _, matched := util.ThinkingFromMetadata(metadata); matched && budget != nil {
+										if mapped, okMap := util.OpenAIThinkingBudgetToEffort(normalizedModel, *budget); okMap && mapped != "" {
+											return true, mapped, false
+										}
+									}
 									return false, "", false
 								}
 								if !util.ModelUsesThinkingLevels(normalizedModel) {
@@ -427,14 +323,8 @@ func TestThinkingConversionsAcrossProtocolsAndModels(t *testing.T) {
 								}
 								return false, "", true // validation would fail
 							case "codex":
-								if !util.ModelSupportsThinking(normalizedModel) {
-									return false, "", false
-								}
-								if !util.ModelUsesThinkingLevels(normalizedModel) {
-									// Non-levels models don't support effort strings in codex
-									if from != "openai-response" {
-										return false, "", false
-									}
+								// Codex does not support allowCompat; require thinking-capable level models.
+								if !util.ModelSupportsThinking(normalizedModel) || !util.ModelUsesThinkingLevels(normalizedModel) {
 									return false, "", false
 								}
 								effort, ok := util.ReasoningEffortFromMetadata(metadata)
@@ -523,6 +413,314 @@ func TestThinkingConversionsAcrossProtocolsAndModels(t *testing.T) {
 	}
 }

+// buildRawPayloadWithThinking creates a payload with thinking parameters already in the body.
+// This tests the path where thinking comes from the raw payload, not model suffix.
+func buildRawPayloadWithThinking(fromProtocol, model string, thinkingParam any) []byte {
+	switch fromProtocol {
+	case "gemini":
+		base := fmt.Sprintf(`{"model":"%s","contents":[{"role":"user","parts":[{"text":"hi"}]}]}`, model)
+		if budget, ok := thinkingParam.(int); ok {
+			base, _ = sjson.Set(base, "generationConfig.thinkingConfig.thinkingBudget", budget)
+		}
+		return []byte(base)
+	case "openai-response":
+		base := fmt.Sprintf(`{"model":"%s","input":[{"role":"user","content":[{"type":"text","text":"hi"}]}]}`, model)
+		if effort, ok := thinkingParam.(string); ok && effort != "" {
+			base, _ = sjson.Set(base, "reasoning.effort", effort)
+		}
+		return []byte(base)
+	case "openai":
+		base := fmt.Sprintf(`{"model":"%s","messages":[{"role":"user","content":"hi"}]}`, model)
+		if effort, ok := thinkingParam.(string); ok && effort != "" {
+			base, _ = sjson.Set(base, "reasoning_effort", effort)
+		}
+		return []byte(base)
+	case "claude":
+		base := fmt.Sprintf(`{"model":"%s","messages":[{"role":"user","content":"hi"}]}`, model)
+		if budget, ok := thinkingParam.(int); ok {
+			base, _ = sjson.Set(base, "thinking.type", "enabled")
+			base, _ = sjson.Set(base, "thinking.budget_tokens", budget)
+		}
+		return []byte(base)
+	default:
+		return []byte(fmt.Sprintf(`{"model":"%s","messages":[{"role":"user","content":"hi"}]}`, model))
+	}
+}
+
+// buildBodyForProtocolWithRawThinking translates payload with raw thinking params.
+func buildBodyForProtocolWithRawThinking(t *testing.T, fromProtocol, toProtocol, model string, thinkingParam any) ([]byte, error) {
+	t.Helper()
+	raw := buildRawPayloadWithThinking(fromProtocol, model, thinkingParam)
+	stream := fromProtocol != toProtocol
+
+	body := sdktranslator.TranslateRequest(
+		sdktranslator.FromString(fromProtocol),
+		sdktranslator.FromString(toProtocol),
+		model,
+		raw,
+		stream,
+	)
+
+	var err error
+	allowCompat := isOpenAICompatModel(model)
+	switch toProtocol {
+	case "gemini":
+		body = util.ApplyDefaultThinkingIfNeeded(model, body)
+		body = util.NormalizeGeminiThinkingBudget(model, body)
+		body = util.StripThinkingConfigIfUnsupported(model, body)
+	case "claude":
+		// For raw payload, Claude thinking is passed through by translator
+		// No additional processing needed as thinking is already in body
+	case "openai":
+		body = executor.NormalizeThinkingConfig(body, model, allowCompat)
+		err = executor.ValidateThinkingConfig(body, model)
+	case "codex":
+		// Codex does not support allowCompat; always use false.
+		body, err = normalizeCodexPayload(body, model, false)
+	}
+
+	body, _ = sjson.SetBytes(body, "model", model)
+	body = filterThinkingBody(toProtocol, body, model, model)
+	return body, err
+}
+
+func TestRawPayloadThinkingConversions(t *testing.T) {
+	cleanup := registerCoreModels(t)
+	defer cleanup()
+
+	type scenario struct {
+		name          string
+		thinkingParam any // int for budget, string for effort level
+	}
+
+	numericName := func(budget int) string {
+		if budget < 0 {
+			return "budget-neg1"
+		}
+		return fmt.Sprintf("budget-%d", budget)
+	}
+
+	for _, model := range thinkingTestModels {
+		supportsThinking := util.ModelSupportsThinking(model)
+		usesLevels := util.ModelUsesThinkingLevels(model)
+		allowCompat := isOpenAICompatModel(model)
+
+		for _, from := range thinkingTestFromProtocols {
+			var cases []scenario
+			switch from {
+			case "openai", "openai-response":
+				cases = []scenario{
+					{name: "no-thinking", thinkingParam: nil},
+				}
+				for _, lvl := range thinkingLevelSamples {
+					cases = append(cases, scenario{
+						name:          "effort-" + lvl,
+						thinkingParam: lvl,
+					})
+				}
+			case "gemini", "claude":
+				cases = []scenario{
+					{name: "no-thinking", thinkingParam: nil},
+				}
+				for _, budget := range thinkingNumericSamples {
+					budget := budget
+					cases = append(cases, scenario{
+						name:          numericName(budget),
+						thinkingParam: budget,
+					})
+				}
+			}
+
+			for _, to := range thinkingTestToProtocols {
+				if from == to {
+					continue
+				}
+				t.Logf("═══════════════════════════════════════════════════════════════════════════════")
+				t.Logf("  RAW PAYLOAD: %s -> %s | model: %s", from, to, model)
+				t.Logf("═══════════════════════════════════════════════════════════════════════════════")
+
+				for _, cs := range cases {
+					from := from
+					to := to
+					cs := cs
+					testName := fmt.Sprintf("raw/%s->%s/%s/%s", from, to, model, cs.name)
+					t.Run(testName, func(t *testing.T) {
+						expectPresent, expectValue, expectErr := func() (bool, string, bool) {
+							if cs.thinkingParam == nil {
+								if to == "codex" && from != "openai-response" && supportsThinking && usesLevels {
+									// Codex translators default reasoning.effort to "medium" for thinking-capable level models
+									return true, "medium", false
+								}
+								return false, "", false
+							}
+
+							switch to {
+							case "gemini":
+								if !supportsThinking || usesLevels {
+									return false, "", false
+								}
+								// Gemini expects numeric budget (only for non-level models)
+								if budget, ok := cs.thinkingParam.(int); ok {
+									norm := util.NormalizeThinkingBudget(model, budget)
+									return true, fmt.Sprintf("%d", norm), false
+								}
+								// Convert effort level to budget for non-level models only
+								if effort, ok := cs.thinkingParam.(string); ok && effort != "" {
+									// "none" disables thinking - no thinkingBudget in output
+									if strings.ToLower(effort) == "none" {
+										return false, "", false
+									}
+									if budget, okB := util.ThinkingEffortToBudget(model, effort); okB {
+										// ThinkingEffortToBudget already returns normalized budget
+										return true, fmt.Sprintf("%d", budget), false
+									}
+									// Invalid effort does not map to a budget
+									return false, "", false
+								}
+								return false, "", false
+							case "claude":
+								if !supportsThinking || usesLevels {
+									return false, "", false
+								}
+								// Claude expects numeric budget (only for non-level models)
+								if budget, ok := cs.thinkingParam.(int); ok && budget > 0 {
+									norm := util.NormalizeThinkingBudget(model, budget)
+									return true, fmt.Sprintf("%d", norm), false
+								}
+								// Convert effort level to budget for non-level models only
+								if effort, ok := cs.thinkingParam.(string); ok && effort != "" {
+									// "none" and "auto" don't produce budget_tokens
+									lower := strings.ToLower(effort)
+									if lower == "none" || lower == "auto" {
+										return false, "", false
+									}
+									if budget, okB := util.ThinkingEffortToBudget(model, effort); okB {
+										// ThinkingEffortToBudget already returns normalized budget
+										return true, fmt.Sprintf("%d", budget), false
+									}
+									// Invalid effort - claude sets thinking.type:enabled but no budget_tokens
+									return false, "", false
+								}
+								return false, "", false
+							case "openai":
+								if allowCompat {
+									if effort, ok := cs.thinkingParam.(string); ok && strings.TrimSpace(effort) != "" {
+										normalized := strings.ToLower(strings.TrimSpace(effort))
+										return true, normalized, false
+									}
+									if budget, ok := cs.thinkingParam.(int); ok {
+										if mapped, okM := util.OpenAIThinkingBudgetToEffort(model, budget); okM && mapped != "" {
+											return true, mapped, false
+										}
+									}
+									return false, "", false
+								}
+								if !supportsThinking || !usesLevels {
+									return false, "", false
+								}
+								if effort, ok := cs.thinkingParam.(string); ok && effort != "" {
+									if normalized, okN := util.NormalizeReasoningEffortLevel(model, effort); okN {
+										return true, normalized, false
+									}
+									return false, "", true // invalid level
+								}
+								if budget, ok := cs.thinkingParam.(int); ok {
+									if mapped, okM := util.OpenAIThinkingBudgetToEffort(model, budget); okM && mapped != "" {
+										// Check if the mapped effort is valid for this model
+										if _, validLevel := util.NormalizeReasoningEffortLevel(model, mapped); !validLevel {
+											return true, mapped, true // expect validation error
+										}
+										return true, mapped, false
+									}
+								}
+								return false, "", false
+							case "codex":
+								// Codex does not support allowCompat; require thinking-capable level models.
+								if !supportsThinking || !usesLevels {
+									return false, "", false
+								}
+								if effort, ok := cs.thinkingParam.(string); ok && effort != "" {
+									if normalized, okN := util.NormalizeReasoningEffortLevel(model, effort); okN {
+										return true, normalized, false
+									}
+									return false, "", true
+								}
+								if budget, ok := cs.thinkingParam.(int); ok {
+									if mapped, okM := util.OpenAIThinkingBudgetToEffort(model, budget); okM && mapped != "" {
+										// Check if the mapped effort is valid for this model
+										if _, validLevel := util.NormalizeReasoningEffortLevel(model, mapped); !validLevel {
+											return true, mapped, true // expect validation error
+										}
+										return true, mapped, false
+									}
+								}
+								if from != "openai-response" {
+									// Codex translators default reasoning.effort to "medium" for thinking-capable models
+									return true, "medium", false
+								}
+								return false, "", false
+							}
+							return false, "", false
+						}()
+
+						body, err := buildBodyForProtocolWithRawThinking(t, from, to, model, cs.thinkingParam)
+						actualPresent, actualValue := func() (bool, string) {
+							path := ""
+							switch to {
+							case "gemini":
+								path = "generationConfig.thinkingConfig.thinkingBudget"
+							case "claude":
+								path = "thinking.budget_tokens"
+							case "openai":
+								path = "reasoning_effort"
+							case "codex":
+								path = "reasoning.effort"
+							}
+							if path == "" {
+								return false, ""
+							}
+							val := gjson.GetBytes(body, path)
+							if to == "codex" && !val.Exists() {
+								reasoning := gjson.GetBytes(body, "reasoning")
+								if reasoning.Exists() {
+									val = reasoning.Get("effort")
+								}
+							}
+							if !val.Exists() {
+								return false, ""
+							}
+							if val.Type == gjson.Number {
+								return true, fmt.Sprintf("%d", val.Int())
+							}
+							return true, val.String()
+						}()
+
+						t.Logf("from=%s to=%s model=%s param=%v present(expect=%v got=%v) value(expect=%s got=%s) err(expect=%v got=%v) body=%s",
+							from, to, model, cs.thinkingParam, expectPresent, actualPresent, expectValue, actualValue, expectErr, err != nil, string(body))
+
+						if expectErr {
+							if err == nil {
+								t.Fatalf("expected validation error but got none, body=%s", string(body))
+							}
+							return
+						}
+						if err != nil {
+							t.Fatalf("unexpected error: %v body=%s", err, string(body))
+						}
+
+						if expectPresent != actualPresent {
+							t.Fatalf("presence mismatch: expect %v got %v body=%s", expectPresent, actualPresent, string(body))
+						}
+						if expectPresent && expectValue != actualValue {
+							t.Fatalf("value mismatch: expect %s got %s body=%s", expectValue, actualValue, string(body))
+						}
+					})
+				}
+			}
+		}
+	}
+}
+
 func TestOpenAIThinkingBudgetToEffortRanges(t *testing.T) {
 	cleanup := registerCoreModels(t)
 	defer cleanup()
@@ -534,6 +732,7 @@ func TestOpenAIThinkingBudgetToEffortRanges(t *testing.T) {
 		want   string
 		ok     bool
 	}{
+		{name: "dynamic-auto", model: "gpt-5", budget: -1, want: "auto", ok: true},
 		{name: "zero-none", model: "gpt-5", budget: 0, want: "none", ok: true},
 		{name: "low-min", model: "gpt-5", budget: 1, want: "low", ok: true},
 		{name: "low-max", model: "gpt-5", budget: 1024, want: "low", ok: true},
@@ -541,7 +740,7 @@ func TestOpenAIThinkingBudgetToEffortRanges(t *testing.T) {
 		{name: "medium-max", model: "gpt-5", budget: 8192, want: "medium", ok: true},
 		{name: "high-min", model: "gpt-5", budget: 8193, want: "high", ok: true},
 		{name: "high-max", model: "gpt-5", budget: 24576, want: "high", ok: true},
-		{name: "over-max-clamps-to-highest", model: "gpt-5", budget: 24577, want: "high", ok: true},
+		{name: "over-max-clamps-to-highest", model: "gpt-5", budget: 64000, want: "high", ok: true},
 		{name: "over-max-xhigh-model", model: "gpt-5.2", budget: 50000, want: "xhigh", ok: true},
 		{name: "negative-unsupported", model: "gpt-5", budget: -5, want: "", ok: false},
 	}
Author	SHA1	Message	Date
Luis Pater	3af24597ee	docs: remove Amp CLI integration guides and update references	2025-12-15 23:50:56 +08:00
Luis Pater	88b101ebf5	Merge pull request #549 from router-for-me/log Improve Request Logging Efficiency and Standardize Error Responses	2025-12-15 20:43:12 +08:00
Luis Pater	d9a65745df	fix(translator): handle empty item type and string content in OpenAI response parser	2025-12-15 20:35:52 +08:00
hkfires	97ab623d42	fix(api): prevent double logging for streaming responses	2025-12-15 18:00:32 +08:00
hkfires	14aa6cc7e8	fix(api): ensure all response writes are captured for logging The response writer wrapper has been refactored to more reliably capture response bodies for logging, fixing several edge cases. - Implements `WriteString` to capture writes from `io.StringWriter`, which were previously missed by the `Write` method override. - A new `shouldBufferResponseBody` helper centralizes the logic to ensure the body is buffered only when logging is active or for errors when `logOnErrorOnly` is enabled. - Streaming detection is now more robust. It correctly handles non-streaming error responses (e.g., `application/json`) that are generated for a request that was intended to be streaming. BREAKING CHANGE: The public methods `Status()`, `Size()`, and `Written()` have been removed from the `ResponseWriterWrapper` as they are no longer required by the new implementation.	2025-12-15 17:45:16 +08:00
hkfires	3bc489254b	fix(api): prevent double logging for error responses The WriteErrorResponse function now caches the error response body in the gin context. The deferred request logger checks for this cached response. If an error response is found, it bypasses the standard response logging. This prevents scenarios where an error is logged twice or an empty payload log overwrites the original, more detailed error log.	2025-12-15 16:36:01 +08:00
hkfires	4c07ea41c3	feat(api): return structured JSON error responses The API error handling is updated to return a structured JSON payload instead of a plain text message. This provides more context and allows clients to programmatically handle different error types. The new error response has the following structure: { "error": { "message": "...", "type": "..." } } The `type` field is determined by the HTTP status code, such as `authentication_error`, `rate_limit_error`, or `server_error`. If the underlying error message from an upstream service is already a valid JSON string, it will be preserved and returned directly. BREAKING CHANGE: API error responses are now in a structured JSON format instead of plain text. Clients expecting plain text error messages will need to be updated to parse the new JSON body.	2025-12-15 16:19:52 +08:00
Luis Pater	f6720f8dfa	Merge pull request #547 from router-for-me/amp feat(amp): require API key authentication for management routes	2025-12-15 16:14:49 +08:00
Chén Mù	e19ab3a066	Merge pull request #543 from router-for-me/log feat(auth): add proxy information to debug logs	2025-12-15 15:59:16 +08:00
hkfires	8f1dd69e72	feat(amp): require API key authentication for management routes All Amp management endpoints (e.g., /api/user, /threads) are now protected by the standard API key authentication middleware. This ensures that all management operations require a valid API key, significantly improving security. As a result of this change: - The `restrict-management-to-localhost` setting now defaults to `false`. API key authentication provides a stronger and more flexible security control than IP-based restrictions, improving usability in containerized environments. - The reverse proxy logic now strips the client's `Authorization` header after authenticating the initial request. It then injects the configured `upstream-api-key` for the request to the upstream Amp service. BREAKING CHANGE: Amp management endpoints now require a valid API key for authentication. Requests without a valid API key in the `Authorization` header will be rejected with a 401 Unauthorized error.	2025-12-15 13:24:53 +08:00
hkfires	f26da24a2f	feat(auth): add proxy information to debug logs	2025-12-15 13:14:55 +08:00
Luis Pater	8e4fbcaa7d	Merge pull request #533 from router-for-me/think refactor(thinking): centralize reasoning effort mapping and normalize budget values	2025-12-15 10:34:41 +08:00
hkfires	09c339953d	fix(openai): forward reasoning.effort value Drop the hardcoded effort mapping in request conversion so unknown values are preserved instead of being coerced to `auto	2025-12-15 09:16:15 +08:00
hkfires	367a05bdf6	refactor(thinking): export thinking helpers Expose thinking/effort normalization helpers from the executor package so conversion tests use production code and stay aligned with runtime validation behavior.	2025-12-15 09:16:15 +08:00
hkfires	d20b71deb9	fix(thinking): normalize effort mapping Route OpenAI reasoning effort through ThinkingEffortToBudget for Claude translators, preserve "minimal" when translating OpenAI Responses, and treat blank/unknown efforts as no-ops for Gemini thinking configs. Also map budget -1 to "auto" and expand cross-protocol thinking tests.	2025-12-15 09:16:15 +08:00
hkfires	712ce9f781	fix(thinking): drop unsupported none effort When budget 0 maps to "none" for models that use thinking levels but don't support that effort level, strip thinking fields instead of setting an invalid reasoning_effort value. Tests now expect removal for this edge case.	2025-12-15 09:16:14 +08:00
hkfires	a4a3274a55	test(thinking): expand conversion edge case coverage	2025-12-15 09:16:14 +08:00
hkfires	716aa71f6e	fix(thinking): centralize reasoning_effort mapping Move OpenAI `reasoning_effort` -> Gemini `thinkingConfig` budget logic into shared helpers used by Gemini, Gemini CLI, and antigravity translators. Normalize Claude thinking handling by preferring positive budgets, applying budget token normalization, and gating by model support. Always convert Gemini `thinkingBudget` back to OpenAI `reasoning_effort` to support allowCompat models, and update tests for normalization behavior.	2025-12-15 09:16:14 +08:00
hkfires	e8976f9898	fix(thinking): map budgets to effort for level models	2025-12-15 09:16:14 +08:00
hkfires	8496cc2444	test(thinking): cover openai-compat reasoning passthrough	2025-12-15 09:16:14 +08:00
hkfires	5ef2d59e05	fix(thinking): gate reasoning effort by model support Only map OpenAI reasoning effort to Claude thinking for models that support thinking and use budget tokens (not level-based thinking). Also add "xhigh" effort mapping and adjust minimal/low budgets, with new raw-payload conversion tests across protocols and models.	2025-12-15 09:16:14 +08:00
Chén Mù	07bb89ae80	Merge pull request #542 from router-for-me/aistudio	2025-12-15 09:13:25 +08:00
hkfires	27a5ad8ec2	Fixed: #534 fix(aistudio): correct JSON string boundary detection for backslash sequences	2025-12-15 09:00:14 +08:00
Luis Pater	707b07c5f5	Merge pull request #537 from sukakcoding/fix/function-response-fallback fix: handle malformed json in function response parsing	2025-12-15 03:31:09 +08:00
sukakcoding	4a764afd76	refactor: extract parseFunctionResponse helper to reduce duplication	2025-12-15 01:05:36 +08:00
sukakcoding	ecf49d574b	fix: handle malformed json in function response parsing	2025-12-15 00:59:46 +08:00
Luis Pater	5a75ef8ffd	Merge pull request #536 from AoaoMH/feature/auth-model-check feat: using Client Model Infos;	2025-12-15 00:29:33 +08:00
Test	07279f8746	feat: using Client Model Infos;	2025-12-15 00:13:05 +08:00
Luis Pater	71f788b13a	fix(registry): remove unused `ThinkingSupport` from DeepSeek-R1 model	2025-12-14 21:30:17 +08:00
Luis Pater	59c62dc580	fix(registry): correct DeepSeek-V3.2 experimental model ID	2025-12-14 21:27:43 +08:00
Luis Pater	d5310a3300	Merge pull request #531 from AoaoMH/feature/auth-model-check feat: add API endpoint to query models for auth credentials	2025-12-14 16:46:43 +08:00
Luis Pater	f0a3eb574e	fix(registry): update DeepSeek model definitions with new IDs and descriptions	2025-12-14 16:17:11 +08:00
Test	bb15855443	feat: add API endpoint to query models for auth credentials	2025-12-14 15:16:26 +08:00
Luis Pater	14ce6aebd1	Merge pull request #449 from sususu98/fix/gemini-cli-429-retry-delay-parsing fix(gemini-cli): enhance 429 retry delay parsing	2025-12-14 14:04:14 +08:00
Luis Pater	2fe83723f2	Merge pull request #515 from teeverc/fix/response-rewriter-streaming-flush fix(amp): flush response buffer after each streaming chunk write	2025-12-14 13:26:05 +08:00
teeverc	cd8c86c6fb	refactor: only flush stream response on successful write	2025-12-13 13:32:54 -08:00
teeverc	52d5fd1a67	fix: streaming for amp cli	2025-12-13 13:17:53 -08:00
sususu	07d21463ca	fix(gemini-cli): enhance 429 retry delay parsing Add fallback parsing for quota reset delay when RetryInfo is not present: - Try ErrorInfo.metadata.quotaResetDelay (e.g., "373.801628ms") - Parse from error.message "Your quota will reset after Xs." This ensures proper cooldown timing for rate-limited requests. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>	2025-12-11 09:34:39 +08:00