xiaodcs-copilot-api-edge 2.3.9-edge.0

package/README.md

@@ -0,0 +1,888 @@
+# Copilot API
+
+<p align="center">
+  <img src="./docs/hero/copilot-api-hero.svg" alt="Copilot API - Universal AI Gateway" width="1600" />
+</p>
+
+<p align="center">
+  <strong>Universal AI Gateway</strong><br />
+  One Gateway. Any Client. Multiple AI Providers.<br />
+  Chat Completions &middot; OpenAI Responses &middot; Anthropic Messages
+</p>
+
+> **XiaoDcs Edge build:** this Edge npm artifact is built from a private
+> downstream repository and carries the latest Responses image-body budgeting,
+> stale compaction recovery, and Codex integrations. The original
+> MIT-licensed project is [caozhiyuan/copilot-api](https://github.com/caozhiyuan/copilot-api).
+
+<p align="center">
+  <a href="https://www.npmjs.com/package/xiaodcs-copilot-api-edge"><img src="https://img.shields.io/npm/v/xiaodcs-copilot-api-edge.svg" alt="npm version"></a>
+  <a href="https://github.com/caozhiyuan/copilot-api/blob/main/LICENSE"><img src="https://img.shields.io/badge/license-MIT-blue.svg" alt="License"></a>
+  <a href="https://github.com/caozhiyuan/copilot-api/stargazers"><img src="https://img.shields.io/github/stars/caozhiyuan/copilot-api.svg" alt="GitHub stars"></a>
+  <a href="https://bun.sh"><img src="https://img.shields.io/badge/Bun-%3E%3D1.2.x-orange.svg" alt="Bun >= 1.2.x"></a>
+  <a href="https://nodejs.org"><img src="https://img.shields.io/badge/Node-%3E%3D22.13.0-green.svg" alt="Node >= 22.13.0"></a>
+</p>
+
+<p align="center">
+  English | <a href="./README.zh-CN.md">简体中文</a>
+</p>
+
+## Quick Start
+
+The fastest way to get a working gateway:
+
+```sh
+npx xiaodcs-copilot-api-edge@latest start
+```
+
+The server listens on `http://localhost:4141` by default. Optionally authenticate with GitHub Copilot or configure a third-party provider first:
+
+```sh
+npx xiaodcs-copilot-api-edge@latest auth login
+```
+
+Verify the gateway is up:
+
+```sh
+curl http://localhost:4141/v1/models
+```
+
+> [!NOTE]
+> Token usage storage requires Node.js >= 22.13.0 or Bun. See [Using with npx](#using-with-npx) for details.
+
+From here, jump to the guide for your client: [Claude Code](#using-with-claude-code), [OpenCode](#using-with-opencode), [Codex](#using-with-codex), or run it with [Docker](#using-with-docker).
+
+## Highlights
+
+- **Unified API Gateway**: Serve OpenAI-compatible Chat Completions (`/v1/chat/completions`), the OpenAI Responses API (`/v1/responses`), and Anthropic-compatible Messages (`/v1/messages`) from one local endpoint.
+- **Multi-Provider**: Route GitHub Copilot, the built-in `codex` provider, and third-party providers (Kimi, DeepSeek, DashScope, OpenRouter, OpenCode Go, or a custom provider) behind the same gateway. GitHub Copilot is optional — with at least one enabled provider, the server starts in provider-only mode without a GitHub token.
+- **Coding Agent Ready**: First-class setups for Claude Code, OpenCode, and Codex, including the interactive `--claude-code` launcher and a merged model catalog for Codex.
+- **Streaming & WebSocket**: SSE streaming on all three client-facing protocols. Upstream Copilot Responses traffic selects WebSocket or HTTP from each model's advertised endpoints; streamed Responses traffic for the built-in `codex` provider uses WebSocket by default and uses HTTP when `useResponsesApiWebSocket` is disabled.
+- **Desktop App**: Electron GUI with GitHub Copilot sign-in, Codex OAuth, provider configuration, token usage, logs, and one-click start/stop.
+
+## Compatibility
+
+Every client talks to the same local endpoint. The gateway routes each request to GitHub Copilot, the built-in `codex` provider, or a configured third-party provider, translating between protocols when the provider speaks a different one.
+
+**Client / Protocol Matrix**
+
+| Client | Chat Completions | Responses | Anthropic Messages | Recommended |
+|---|:---:|:---:|:---:|---|
+| Claude Code | — | — | ✅ Native / Adapter | Anthropic Messages |
+| OpenCode | ✅ Native | ✅ Native / Adapter | ✅ Native / Adapter via `@ai-sdk/anthropic` | Anthropic Messages |
+| Codex | — | ✅ Native / Adapter | — | Responses |
+| OpenAI-compatible clients | ✅ Native | ✅ Native / Adapter | — | Chat Completions |
+| Anthropic-compatible clients | — | — | ✅ Native / Adapter | Anthropic Messages |
+
+**Providers and protocols.** Protocol support is model-specific. Chat Completions requires a native endpoint, while Responses and Messages can use supported adapters. The built-in `codex` provider uses Responses natively; third-party providers can use `anthropic`, `openai-compatible`, or `openai-responses`, with per-model overrides.
+
+## Desktop App
+
+Prefer a GUI? The Electron desktop app in `desktop/` covers GitHub Copilot sign-in, OpenAI Codex OAuth, and API-key configuration for Kimi, DeepSeek, DashScope, OpenRouter, or a custom provider — with one-click start/stop of the local server, and the local endpoint, auth header, available models, usage, and logs in one window.
+
+<p align="center">
+  <img src="./docs/screenshots/desktop-dashboard.png" alt="Copilot API desktop app dashboard" width="49%" />
+  <img src="./docs/screenshots/desktop-token-usage.png" alt="Copilot API desktop app token usage view" width="49%" />
+</p>
+
+Windows x64 (`.exe`), macOS Apple Silicon (`.dmg`), and Linux x64 (`.AppImage`) packages are published in [GitHub Releases](https://github.com/caozhiyuan/copilot-api/releases). See [Electron Desktop App](#electron-desktop-app) for full setup and advanced configuration.
+
+## Using with Claude Code
+
+This AI gateway can be used to power [Claude Code](https://docs.anthropic.com/en/claude-code), an experimental conversational AI assistant for developers from Anthropic.
+
+There are two ways to configure Claude Code to use this AI gateway:
+
+### Interactive Setup with `--claude-code` flag
+
+To get started, run the `start` command with the `--claude-code` flag:
+
+```sh
+npx xiaodcs-copilot-api-edge@latest start --claude-code
+```
+
+You will no longer be prompted to pick models manually. The gateway automatically detects the latest available model for each Claude Code size tier — opus maps to the newest Opus model, sonnet to the newest Sonnet model, and haiku to the newest Haiku model — and generates a command that sets `ANTHROPIC_DEFAULT_OPUS_MODEL`, `ANTHROPIC_DEFAULT_SONNET_MODEL`, and `ANTHROPIC_DEFAULT_HAIKU_MODEL` accordingly. Any tier without a matching model available is omitted. The command is copied to your clipboard and sets the environment variables needed for Claude Code to use the gateway.
+
+Paste and run this command in a new terminal to launch Claude Code.
+
+### Manual Configuration with `settings.json`
+
+Alternatively, you can configure Claude Code by creating a `.claude/settings.json` file in your project's root directory. This file should contain the environment variables needed by Claude Code. This way you don't need to run the interactive setup every time.
+
+Here is an example `.claude/settings.json` file:
+
+```json
+{
+  "env": {
+    "ANTHROPIC_BASE_URL": "http://localhost:4141",
+    "ANTHROPIC_AUTH_TOKEN": "dummy",
+    "ANTHROPIC_MODEL": "gpt-5.6-sol[1m]",
+    "ANTHROPIC_DEFAULT_OPUS_MODEL": "gpt-5.6-sol[1m]",
+    "ANTHROPIC_DEFAULT_SONNET_MODEL": "gpt-5.6-sol[1m]",
+    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "gpt-5.6-luna[1m]",
+    "CLAUDE_CODE_AUTO_COMPACT_WINDOW": "272000",
+    "CLAUDE_CODE_USE_VERTEX": "0",
+    "CLAUDE_CODE_USE_BEDROCK": "0",
+    "DISABLE_NON_ESSENTIAL_MODEL_CALLS": "1",
+    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1",
+    "CLAUDE_CODE_ATTRIBUTION_HEADER": "0",
+    "CLAUDE_CODE_ENABLE_PROMPT_SUGGESTION": "false",
+    "CLAUDE_CODE_DISABLE_TERMINAL_TITLE": "true",
+    "CLAUDE_CODE_ENABLE_AWAY_SUMMARY": "0",
+    "CLAUDE_CODE_TOTAL_TOKENS_REMINDER": "off",
+    "CLAUDE_CODE_EFFORT_LEVEL": "max",
+    "MCP_CONNECT_TIMEOUT_MS": "20000"
+  },
+  "alwaysThinkingEnabled": true,
+  "showThinkingSummaries": true
+}
+```
+
+- Replace `ANTHROPIC_MODEL`, `ANTHROPIC_DEFAULT_OPUS_MODEL`, `ANTHROPIC_DEFAULT_SONNET_MODEL`, and `ANTHROPIC_DEFAULT_HAIKU_MODEL` according to your needs. After configuration, please install the claude code plugin [Plugin Integrations](#plugin-integrations).  
+- `CLAUDE_CODE_TOTAL_TOKENS_REMINDER: "off"` disables Claude Code's total-tokens reminder, which injects a `<total_tokens>N tokens left</total_tokens>` block into the conversation to pace the model against a remaining token budget. The default budget is 15,000,000 (15M) tokens, which is not very meaningful, so it is turned off here.
+- If you are using the codex provider, it is recommended **not** to configure the model name in the `codex/xxx` format (e.g. `codex/gpt-5.6-sol`). Claude Code treats the `codex/` prefix as a special pattern and applies degraded behavior — for example, it strips all previously returned thinking blocks on every request. Use the plain model name (e.g. `gpt-5.6-sol`) instead, and add a `modelMappings` entry in `config.json` to route it back to the codex provider:
+  ```json
+  "modelMappings": {
+    "gpt-5.6-sol": "codex/gpt-5.6-sol",
+    "gpt-5.6-terra": "codex/gpt-5.6-terra",
+    "gpt-5.6-luna": "codex/gpt-5.6-luna"
+  },
+  ```
+- Setting CLAUDE_CODE_ATTRIBUTION_HEADER to 0 can prevent Claude code from adding billing and version information in system prompts, thereby avoiding prompt cache invalidation.
+- Turning off CLAUDE_CODE_ENABLE_PROMPT_SUGGESTION and CLAUDE_CODE_ENABLE_AWAY_SUMMARY can prevent quota from being consumed unnecessarily.
+- Claude Code WebSearch is supported for pure search requests. For Copilot, keep the global `messageApiWebSearchModel` set to a Responses-capable GPT model or a `provider/model` alias. For provider routes, use a native Anthropic provider or an `openai-responses` provider. Add `WebSearch` to `permissions.deny` only if you want to forbid this traffic.
+- If using a non-Claude model, do not enable ENABLE_TOOL_SEARCH. If using the Claude model, can enable ENABLE_TOOL_SEARCH. The current Claude Code uses the client tool search mode. In this mode, loading defer tools requires an additional request each time.
+- `CLAUDE_CODE_AUTO_COMPACT_WINDOW`: Set the context capacity in tokens used for auto-compaction calculations. Defaults to the model's context window: 200K for standard models or 1M for extended context models. Use a lower value like `500000` on a 1M model (e.g., `claude-opus-4-6[1m]`) to treat the window as 500K for compaction purposes. The value is capped at the model's actual context window. `CLAUDE_AUTOCOMPACT_PCT_OVERRIDE` is applied as a percentage of this value. Setting this variable decouples the compaction threshold from the status line's `used_percentage`, which always uses the model's full context window.
+
+You can find more options here: [Claude Code settings](https://docs.anthropic.com/en/docs/claude-code/settings#environment-variables)
+
+You can also read more about IDE integration here: [Add Claude Code to your IDE](https://docs.anthropic.com/en/docs/claude-code/ide-integrations)
+
+## Using with OpenCode
+
+OpenCode already has a direct GitHub Copilot provider. Use this section when you want OpenCode to point at this AI gateway through `@ai-sdk/anthropic` and reuse the agent behaviors described earlier in this README.
+
+### Minimal setup
+
+Start the AI gateway with the OpenCode OAuth app:
+
+```sh
+npx xiaodcs-copilot-api-edge@latest auth --oauth-app=opencode
+npx xiaodcs-copilot-api-edge@latest start
+```
+
+Then point OpenCode at the gateway with `@ai-sdk/anthropic`.
+
+Example `~/.config/opencode/opencode.json`:
+
+```json
+{
+  "$schema": "https://opencode.ai/config.json",
+  "provider": {
+    "local": {
+      "npm": "@ai-sdk/anthropic",
+      "name": "My Local",
+      "options": {
+        "baseURL": "http://localhost:4141/v1",
+        "apiKey": "dummy"
+      },
+      "models": {
+        "gpt-5.4": {
+          "name": "gpt-5.4",
+          "modalities": {
+            "input": ["text", "image"],
+            "output": ["text"]
+          },
+          "limit": {
+            "context": 400000,
+            "input": 272000,
+            "output": 128000
+          }
+        },
+        "claude-sonnet-4.6": {
+          "id": "claude-sonnet-4.6",
+          "name": "claude-sonnet-4.6",
+          "modalities": {
+            "input": ["text", "image"],
+            "output": ["text"]
+          },          
+          "limit": {
+            "context": 200000,
+            "output": 32000
+          },
+          "options": {
+            "thinking": {
+              "type": "adaptive"
+            },
+            "effort": "max"
+          }
+        }
+      }
+    }
+  }
+}
+```
+
+Why these fields matter:
+
+- `npm: "@ai-sdk/anthropic"` is the important part. OpenCode will speak Anthropic Messages semantics to this AI gateway instead of flattening everything into OpenAI Chat Completions.
+- `options.baseURL` should be `http://localhost:4141/v1`; the Anthropic SDK will append `/messages`, `/models`, and `/messages/count_tokens` automatically.
+- If you enable `auth.apiKeys` in this AI gateway, replace `dummy` with a real key. Otherwise any placeholder value is fine.
+
+## Using with Codex
+
+This AI gateway can also power Codex.
+
+### Codex `config.toml` Reference
+
+Add the following `[model_providers.copilot_api]` section to your Codex `~/.codex/config.toml`:
+
+```toml
+model_provider = "copilot_api"
+model_reasoning_summary = "auto"
+model_context_window = 272000
+model_auto_compact_token_limit = 244800
+web_search = "live"
+
+[model_providers.copilot_api]
+name = "OpenAI"
+base_url = "http://localhost:4141"
+env_key = "GITHUB_COPILOT_API_KEY"
+requires_openai_auth = true
+supports_websockets = false
+supports_standalone_web_search = true
+wire_api = "responses"
+request_max_retries = 3
+stream_max_retries = 3
+stream_idle_timeout_ms = 300000
+
+[features]
+remote_compaction_v2 = true
+# optional: set false only when the model does not support tool_search
+apps = false
+standalone_web_search = true
+
+[analytics]
+enabled = false
+```
+
+> [!NOTE]
+> `name` must be set to `"OpenAI"`.
+>
+> For third-party models that do not support `tool_search`, we recommend disabling features.apps. Otherwise, each prompt may consume an additional 20,000 or more tokens.
+>
+> `supports_standalone_web_search` and `[features] standalone_web_search` must both be enabled to expose the standalone `web.run` search tool.
+
+When Copilot exposes both a base Responses model and an exact `-fast` sibling
+(for example, `gpt-5.6-sol` and `gpt-5.6-sol-fast`), the gateway presents them
+to Codex as one model with a native **Fast** service tier. Selecting Fast makes
+Codex send `service_tier: "priority"`; the gateway routes that request to the
+paired Fast model and removes the unsupported field before forwarding it to
+GitHub Copilot. The raw `/v1/models` endpoint continues to list both model IDs.
+To make Fast the default for new Codex turns, add this top-level setting:
+
+```toml
+service_tier = "fast"
+```
+
+### If Codex Is Not Signed In to a GPT Account
+
+```toml
+[model_providers.copilot_api]
+name = "OpenAI"
+base_url = "http://localhost:4141"
+requires_openai_auth = false
+supports_websockets = false
+supports_standalone_web_search = true
+wire_api = "responses"
+request_max_retries = 3
+stream_max_retries = 3
+stream_idle_timeout_ms = 300000
+
+[features]
+standalone_web_search = true
+
+[model_providers.copilot_api.auth]
+command = "powershell.exe"
+args = [
+    "-NoProfile",
+    "-NonInteractive",
+    "-Command",
+    "[Console]::Out.Write($env:GITHUB_COPILOT_API_KEY)"
+]
+```
+
+macOS, replace the `auth` block with:
+
+```toml
+[model_providers.copilot_api.auth]
+command = "/bin/zsh"
+args = [
+    "-c",
+    "printf '%s' \"$GITHUB_COPILOT_API_KEY\""
+]
+```
+
+Without this configuration, Codex cannot fetch `/v1/models` while not signed in to a GPT account, so custom models are unavailable in the model picker.
+
+When a Codex client (`User-Agent` starts with `codex`) requests the top-level `GET /v1/models`, the gateway merges native Codex models with models available through the Messages adapter. The latter advertise `use_responses_lite: true`, except DeepSeek models, which use `use_responses_lite: false` and `tool_mode: null`. For other models, `/v1/responses` uses **Responses → Messages** for Anthropic providers, while OpenAI-compatible providers and Chat-only Copilot models reuse the existing Messages route for **Responses → Messages → Chat Completions**, then translate streaming or JSON results back to Responses.
+
+The merged catalog is what Codex shows in its model picker, including the models exposed by your configured providers:
+
+<img src="./docs/screenshots/codex-models.png" alt="Codex model picker showing models provided by the gateway" width="900" />
+
+For Codex clients, only `gpt-*` Copilot models use the native Responses API; non-GPT Copilot models always go through the adapter, even when they advertise native `/responses` support. The same Codex rule applies on provider `/v1/responses` routes (top-level `provider/model` aliases and `/:provider/v1/responses`): for `openai-responses` providers, non-`gpt-*` models fall back to the Messages adapter, while `gpt-*` models keep native Responses forwarding.
+
+Responses Lite tool definitions are read from `input.additional_tools`, without relying on top-level `tools`. Function, `namespace`, and custom tools are supported; clients must declare `apply_patch` as `type: "custom"`, and it is not handled as a standalone special tool type. Returned calls recover their original `name` and `namespace`. Tools are collected before old history is trimmed, so compaction requests retain them. The Messages fallback does not support Responses `tool_search` mode. Anthropic `output_config.effort` keeps the project's existing valid levels; Responses `minimal` maps to `low`, while `none` omits Anthropic effort.
+
+When Codex uses the top-level GitHub Copilot route with `approvals_reviewer = "auto_review"`, map its internal review model to a Responses-capable Copilot model in the gateway's `config.json`:
+
+```json
+{
+  "modelMappings": {
+    "codex-auto-review": "gpt-5.6-luna"
+  }
+}
+```
+
+This mapping only applies to the top-level GitHub Copilot route. Provider-scoped routes do not use `modelMappings`, so the built-in `/codex` provider continues to handle `codex-auto-review` natively.
+
+---
+
+## Project Overview
+
+A small AI gateway that can use GitHub Copilot, the built-in `codex` provider, or configured third-party providers such as DashScope. GitHub Copilot is optional: if no GitHub token is available, the server can still start in provider-only mode as long as at least one enabled provider is configured.
+
+The gateway exposes OpenAI- and Anthropic-compatible APIs from one local endpoint, so tools like [Claude Code](https://docs.anthropic.com/en/docs/claude-code/overview), OpenCode, Codex, and OpenAI-compatible clients can share the same local server.
+
+On the GitHub Copilot path, the gateway prefers Copilot's native Anthropic-style Messages API when available, preserving more Claude-native behavior for tool-heavy workflows.
+
+## Important Notes
+
+> [!IMPORTANT]
+> **Before using, please be aware of the following:**
+>
+> 1. **Codex configuration:** When using with Codex, add the gateway provider to `~/.codex/config.toml`. See [Codex `config.toml` Reference](#codex-configtoml-reference).
+>
+> 2. **Claude Code configuration:** When using with Claude Code, please configure the model ID as `claude-opus-4-8[1m]`. Example claude `settings.json` see [Manual Configuration with `settings.json`](#manual-configuration-with-settingsjson).
+>
+> 3. **OpenCode configuration:** When using with OpenCode, configure `~/.config/opencode/opencode.json` with `@ai-sdk/anthropic`. See [Using with OpenCode](#using-with-opencode).
+>
+> 4. **Built-in `copilot`, `codex` and third-party providers:** Run `npx xiaodcs-copilot-api-edge@latest auth` and choose `copilot`, `codex`, `deepseek`, `custom`, or other providers.
+>
+> 5. **Note:** See [GitHub Copilot Security Notice](./NOTICE.md#github-copilot-security-notice) for the warning removed from the README header.
+
+## Prerequisites
+
+- Bun (>= 1.2.x)
+- Node.js if you plan to run the published CLI with `npx`
+- GitHub account with Copilot subscription only if you want to use the GitHub Copilot provider
+- An API key or OAuth login for at least one configured provider if you want to run without GitHub Copilot
+
+## Installation
+
+To install dependencies, run:
+
+```sh
+bun install
+```
+
+## Running from Source
+
+The project can be run from source in several ways:
+
+### Development Mode
+
+```sh
+bun run dev start
+```
+
+### Production Mode
+
+```sh
+bun run start start
+```
+
+> The trailing `start` is the CLI subcommand passed to `src/main.ts`, not a typo: `bun run dev start` runs watch mode, `bun run start start` runs production.
+
+## Using with npx
+
+You can run the project directly using npx:
+
+> [!IMPORTANT]
+> Token usage storage uses Node's built-in `node:sqlite` module when running with `npx`. It is enabled on Node.js >= 22.13.0. On Node.js < 22.13.0, the CLI still starts, but token usage storage is disabled.
+>
+> If you want token usage storage without upgrading Node.js, run the published CLI with Bun instead: `bunx --bun xiaodcs-copilot-api-edge@latest start`.
+
+```sh
+npx xiaodcs-copilot-api-edge@latest start
+```
+
+With options:
+
+```sh
+npx xiaodcs-copilot-api-edge@latest start --port 8080
+```
+
+For authentication or provider configuration only:
+
+```sh
+npx xiaodcs-copilot-api-edge@latest auth
+```
+
+To run without GitHub Copilot, configure at least one provider first, then start the server normally:
+
+```sh
+npx xiaodcs-copilot-api-edge@latest auth login --provider dashscope
+npx xiaodcs-copilot-api-edge@latest start
+```
+
+## Using with Docker
+
+Build the image:
+
+```sh
+docker build -t copilot-api .
+```
+
+Run the container with a bind mount so auth data survives restarts:
+
+```sh
+mkdir -p ./copilot-data
+docker run -p 4141:4141 -v $(pwd)/copilot-data:/root/.local/share/copilot-api copilot-api
+```
+
+This stores GitHub auth data, provider config, and other gateway state in `./copilot-data` on the host, mapped to `/root/.local/share/copilot-api` in the container.
+
+Or pass a GitHub token directly:
+
+```sh
+docker run -p 4141:4141 -e GH_TOKEN=your_github_token_here copilot-api
+```
+
+## Electron Desktop App
+
+If you prefer a GUI, this repository also includes an Electron desktop app in `desktop/`. It supports GitHub Copilot sign-in, OpenAI Codex OAuth, and API-key configuration for Kimi, DeepSeek, DashScope, OpenRouter, or a custom provider. After authorization or provider configuration, it can start and stop the local proxy with one click and shows the local endpoint, auth header, available models, usage, and logs in the app.
+
+The settings screen also exposes `OAuth App`, `API Home`, `Enterprise URL`, verbose logging, and minimize-to-tray. Windows x64 (`.exe`), macOS Apple Silicon (`.dmg`), and Linux x64 (`.AppImage`) packages are published in GitHub Releases:
+
+https://github.com/caozhiyuan/copilot-api/releases
+
+On Linux, make the downloaded AppImage executable before launching it:
+
+```sh
+chmod +x Copilot-API-*-linux-x86_64.AppImage
+./Copilot-API-*-linux-x86_64.AppImage
+```
+
+Download the installer for your platform, authorize or configure a provider inside the app, choose a port, start the server, then point your client at the local endpoint shown in the app. Packaged desktop builds use the bundled Electron runtime, so normal desktop usage does not require installing Node.js separately. Token usage history is enabled when that bundled runtime supports SQLite.
+
+The desktop app's Advanced Config page reads and writes the shared model mappings through `GET/POST /admin/config/model-mappings`. The same mappings apply across `POST /v1/messages`, `POST /v1/messages/count_tokens`, `POST /v1/responses`, and `POST /v1/chat/completions` instead of being split per interface. It uses `auth.adminApiKey` instead of the regular `auth.apiKeys`, and the app reads that key directly from `config.json` after the server has generated it on startup.
+
+## GPT Tool Search
+
+For GPT Responses models such as `gpt-5.4+`, this AI gateway can expose Responses `tool_search` through a small MCP bridge. The same bridge can be used by Claude Code and opencode, as long as the client loads MCP servers and sends Anthropic Messages traffic through this gateway.
+
+Do not set Claude Code's native `ENABLE_TOOL_SEARCH` for GPT models. That flag enables Claude Code's own client-side tool search mode, and it may stop forwarding deferred tool definitions. This gateway needs the full tool definitions so it can keep the small always-loaded tool set eager and translate every other tool into Responses deferred namespaces.
+
+If you install `tool-search@copilot-api-marketplace`, Claude Code receives this MCP bridge automatically and you can skip the manual Claude Code MCP setup below.
+
+Add the tool search bridge to the MCP config used by Claude Code:
+
+```json
+{
+  "mcpServers": {
+    "tool_search": {
+      "type": "stdio",
+      "command": "npx",
+      "args": ["-y", "xiaodcs-copilot-api-edge@latest", "mcp"]
+    }
+  }
+}
+```
+
+Add the tool search bridge to the MCP config used by opencode:
+
+```json
+{
+  "mcp": {
+    "tool_search": {
+      "type": "local",
+      "command": ["npx", "-y", "xiaodcs-copilot-api-edge@latest", "mcp"]
+    }
+  }
+}
+```
+
+For local development, use `bun` as the command and `["run", "./src/main.ts", "mcp"]` as the args.
+
+Internally, the gateway now configures OpenAI Responses `tool_search` in client-executed mode. Deferred tools are still exposed as searchable namespaces, but the model is explicitly asked to return the exact deferred tool names it wants to load next.
+
+The bridge uses direct tool selection, not query search. Its tool input is `names`, a comma-separated list of exact deferred tool names, for example `TaskList,TaskGet,mcp__fetch__fetch`.
+
+## Plugin Integrations
+
+Plugin integrations are available for Claude Code and opencode.
+
+### Claude Code plugin integration (marketplace-based)
+
+The Claude Code integration is packaged as two plugins:
+
+- `agent-inject` injects `__SUBAGENT_MARKER__...` on `SubagentStart`, so the gateway can infer `x-initiator: agent`.
+- `tool-search` registers the `tool_search` MCP bridge used for GPT Responses deferred tool loading.
+
+- Marketplace catalog in this repository: `.claude-plugin/marketplace.json`
+- Plugin sources in this repository: `plugin/claude/agent-inject`, `plugin/claude/tool-search`
+
+Add the marketplace remotely:
+
+```sh
+/plugin marketplace add https://github.com/caozhiyuan/copilot-api.git
+```
+
+Install the plugins from the marketplace:
+
+```sh
+/plugin install agent-inject@copilot-api-marketplace
+/plugin install tool-search@copilot-api-marketplace
+```
+
+After installation, `agent-inject` injects `__SUBAGENT_MARKER__...` on `SubagentStart`, and the gateway uses it to infer `x-initiator: agent`.
+
+The `agent-inject` plugin also registers a `UserPromptSubmit` hook that returns `{"continue": true}`, and it can inject `SessionStart` reminder rules through environment variables:
+
+- `CLAUDE_PLUGIN_ENABLE_QUESTION_RULES=1` enables the two reminders about using the `question` tool automatically for Claude Code. Alternatively, you can add the same reminders manually in `CLAUDE.md`; see [CLAUDE.md or AGENTS.md Recommended Content](#claudemd-or-agentsmd-recommended-content).
+- `CLAUDE_PLUGIN_ENABLE_NO_BACKGROUND_AGENTS_RULE=1` enables the `run_in_background: true` avoidance reminder for agent hooks.
+
+The `tool-search` plugin bundles the same MCP bridge described in [GPT Tool Search](#gpt-tool-search), so Claude Code users do not need to add the `tool_search` server manually when they install that plugin.
+
+The plugin also auto-approves bridge calls through a `PermissionRequest` hook scoped exactly to `mcp__plugin_tool-search_tool_search__search`. The hook does not approve other MCP tools and does not override explicit `ask` or `deny` permission rules.
+
+### Opencode plugin
+
+The subagent marker producer is packaged as an opencode plugin located at `plugin/opencode/subagent-marker.js`.
+
+**Installation:**
+
+Copy the plugin file to your opencode plugins directory:
+
+```sh
+# Clone or download this repository, then copy the plugin
+cp plugin/opencode/subagent-marker.js ~/.config/opencode/plugins/
+```
+
+Or manually create the file at `~/.config/opencode/plugins/subagent-marker.js` with the plugin content.
+
+**Features:**
+
+- Tracks sub-sessions created by subagents
+- Automatically prepends a marker system reminder (`__SUBAGENT_MARKER__...`) to subagent chat messages
+- Sets `x-session-id` header for session tracking
+- Enables the gateway to infer `x-initiator: agent` for subagent-originated requests
+
+The plugin hooks into `session.created`, `session.deleted`, `chat.message`, and `chat.headers` events to provide seamless subagent marker functionality.
+
+## Using the Usage Viewer
+
+After starting the server, a URL to the Copilot Usage Dashboard will be displayed in your console. This dashboard is a web interface for monitoring your API usage.
+
+1.  Start the server. For example, using npx:
+    ```sh
+    npx xiaodcs-copilot-api-edge@latest start
+    ```
+2.  The server will output a URL to the usage viewer. Copy and paste this URL into your browser. It will look something like this:
+    `http://localhost:4141/usage-viewer?endpoint=http://localhost:4141/usage`
+    - If you use the `start.bat` script on Windows, this page will open automatically.
+
+The dashboard provides a user-friendly interface to view your Copilot usage data:
+
+> Token usage history requires Bun or Node.js >= 22.13.0. On Node.js < 22.13.0, the server runs normally but token usage storage is disabled.
+
+- **API Endpoint URL**: The dashboard is pre-configured to fetch data from your local server endpoint via a URL query parameter. You can manually switch this to any other compatible API endpoint.
+- **API Key Authentication**: If API Key authentication is enabled, enter a raw API key (sent as the `x-api-key` header) or `Authorization: Bearer <key>`. Credentials are remembered in the browser's local storage per endpoint origin, and switching to a different endpoint origin does not automatically send the previous credential.
+- **Period Selector**: Choose from Day, Week, or Month time ranges. The URL query parameter updates automatically when you switch, making it easy to bookmark and share.
+- **Fetch Data**: Click the "Refresh" button to load or refresh the usage data. The dashboard also fetches data automatically on page load.
+- **Copilot Quotas**: View quota usage for services such as Chat and Completions via progress bars. Hover over a card to see used/remaining details.
+- **Token Usage Metric Cards**: See a summary of Total, Input, Output, Cache Read, Cache Write, Requests, and estimated cost for the current period.
+- **Trend Chart (Week / Month)**: An interactive line chart with model and metric filters. Click a data point to inspect the usage breakdown for a specific day.
+- **Model Breakdown Table**: A per-model summary of requests, input/output/cache tokens, and estimated cost for the selected period.
+- **Request Events (Paginated)**: A time-sorted list of request event records with pagination support, showing timestamps, models, request IDs, and token counts.
+- **Detailed Information**: See the full JSON response from the API for a detailed breakdown of all available usage statistics.
+- **URL-based Configuration**: You can also specify the API endpoint and period directly via `endpoint` and `period` query parameters. For example:
+  `http://localhost:4141/usage-viewer?endpoint=http://your-api-server/usage&period=week`
+
+### Usage Viewer Screenshot
+
+<p align="center">
+  <img src="./docs/screenshots/usage-viewer.png" alt="Copilot API usage viewer" width="900" />
+</p>
+
+## Command Structure
+
+Copilot API now uses a subcommand structure with these main commands:
+
+- `start`: Start the gateway server. If a GitHub token is available, the server starts with Copilot enabled. If no GitHub token is available, it starts in provider-only mode when at least one enabled provider exists; otherwise it guides you through provider setup.
+- `auth`: Run provider login or configuration without starting the server. Use it for GitHub Copilot login, Codex OAuth, or third-party provider API key setup.
+- `debug`: Display diagnostic information including version, runtime details, file paths, and authentication status. Useful for troubleshooting and support.
+
+## Command Line Options
+
+### Global Options
+
+The following options can be used with any subcommand. When passing them before the subcommand, use the `--key=value` form:
+
+| Option            | Description                                            | Default | Alias |
+| ----------------- | ------------------------------------------------------ | ------- | ----- |
+| --api-home        | Path to the API home directory (sets COPILOT_API_HOME) | none    | none  |
+| --oauth-app       | OAuth app identifier (sets COPILOT_API_OAUTH_APP)      | none    | none  |
+| --enterprise-url  | Enterprise URL for GitHub (sets COPILOT_API_ENTERPRISE_URL) | none | none |
+
+### Start Command Options
+
+The following command line options are available for the `start` command:
+
+| Option         | Description                                                                   | Default    | Alias |
+| -------------- | ----------------------------------------------------------------------------- | ---------- | ----- |
+| --port         | Port to listen on                                                             | 4141       | -p    |
+| --verbose      | Enable verbose logging                                                        | false      | -v    |
+| --github-token | Provide GitHub token directly (must be generated using the `auth` subcommand) | none       | -g    |
+| --claude-code  | Generate a command to launch Claude Code with Copilot API config              | false      | -c    |
+| --show-token   | Show GitHub and Copilot tokens on fetch and refresh                           | false      | none  |
+| --proxy-env    | Initialize proxy from environment variables                                   | false      | none  |
+
+### Auth Command Options
+
+| Option       | Description               | Default | Alias |
+| ------------ | ------------------------- | ------- | ----- |
+| --provider   | Provider to log in with or configure (`copilot`, `codex`, `opencode-go`, `kimi`, `deepseek`, `dashscope`, `openrouter`, or `custom`) | prompt | none |
+| --verbose    | Enable verbose logging    | false   | -v    |
+| --show-token | Show GitHub token on auth | false   | none  |
+
+Use `copilot-api auth login --provider copilot` only when you want to enable the GitHub Copilot provider. Copilot is not required for `codex` or third-party provider-only usage.
+
+Use `copilot-api auth login --provider deepseek`, `--provider dashscope`, `--provider openrouter`, `--provider opencode-go`, or `--provider kimi` to add or update those common third-party providers from the CLI. DeepSeek prompts for masked `apiKey`, provider `type` (default `anthropic`), and `baseUrl` defaulting to `https://api.deepseek.com/anthropic`. DashScope prompts for masked `apiKey`, provider `type` (default `openai-compatible`), and prefilled `baseUrl`. OpenRouter prompts for masked `apiKey` and prefilled `baseUrl` only, and writes `type: "anthropic"`. OpenCode Go prompts for masked `apiKey` and prefilled `baseUrl` only, and writes `type: "openai-compatible"` (baseUrl `https://opencode.ai/zen/go`). Kimi prompts for masked `apiKey`, provider `type` (default `openai-compatible`), and `baseUrl` defaulting to `https://api.kimi.com/coding` (the same base URL serves both the Anthropic and OpenAI-compatible endpoints). OpenCode Go additionally routes built-in `qwen*` and `minimax*` models through Anthropic Messages and `gpt*`/`grok*`/`muse-spark*` models through OpenAI Responses; other models keep the OpenAI-compatible default. After a provider is configured and enabled, `copilot-api start` can run without any GitHub token.
+
+Use `copilot-api auth login --provider custom` to add or update another third-party provider from the CLI. The command prompts for the provider name, supported type (`anthropic`, `openai-compatible`, or `openai-responses`), `baseUrl`, masked `apiKey`, and `authType`; `authType` may be left as the type default or set to `x-api-key` / `authorization`.
+
+Gateway API keys live under `auth.apiKeys` in `config.json`. Manage them with `copilot-api auth keys` (one operation per invocation): add a key with `--add <key>`, remove one with `--remove <key>`, list all with `--list`, or clear them all with `--clear`. Clients authenticate with any configured key via `x-api-key` or `Authorization: Bearer`. When no keys are configured, `copilot-api start` starts with authentication bypassed and prints a startup info message.
+
+### Debug Command Options
+
+| Option | Description               | Default | Alias |
+| ------ | ------------------------- | ------- | ----- |
+| --json | Output debug info as JSON | false   | none  |
+
+## Configuration (config.json)
+
+- **Location:** `~/.local/share/copilot-api/config.json` (Linux/macOS) or `%USERPROFILE%\.local\share\copilot-api\config.json` (Windows).
+- **Default shape:**
+  ```json
+  {
+    "auth": {
+      "apiKeys": [],
+      "adminApiKey": "<auto-generated-on-startup>"
+    },
+    "providers": {},
+    "modelMappings": {},
+    "extraPrompts": {
+      "gpt-5-mini": "<built-in exploration prompt>"
+    },
+    "smallModel": "gpt-5-mini",
+    "contextManagement": {
+      "messages": true,
+      "responses": false
+    },
+    "modelResponsesApiCompactThresholds": {
+      "gpt-5.4": 217600,
+      "gpt-5.5": 217600
+    },
+    "modelReasoningEfforts": {
+      "gpt-5-mini": "low"
+    },
+    "useMessagesApi": true,
+    "useResponsesApiCompactionRecovery": false,
+    "useResponsesApiWebSocket": true,
+    "responsesTransport": {
+      "headersTimeoutMsV2": 300000,
+      "streamInactivityTimeoutMs": 300000,
+      "websocketOpenTimeoutMs": 30000,
+      "websocketPoolIdleTimeoutMs": 60000,
+      "websocketMaxBufferedBytes": 8388608,
+      "websocketMaxBufferedMessages": 1024
+    },
+    "useResponsesApiWebSearch": true,
+    "alphaSearchCodexPriority": true,
+    "alphaSearchModel": "gpt-5-mini",
+    "messageApiWebSearchModel": "gpt-5-mini"
+  }
+  ```
+- **auth.apiKeys:** API keys used for request authentication on non-admin routes. Supports multiple keys for rotation. Requests can authenticate with either `x-api-key: <key>` or `Authorization: Bearer <key>`. If empty or omitted, authentication for non-admin routes is disabled.
+- **auth.adminApiKey:** Single admin key used only for `/admin/*` routes. If missing, the server generates a random key at startup and writes it back to `config.json`. Requests use the same `x-api-key` or `Authorization: Bearer` headers, but regular `auth.apiKeys` never grant access to `/admin/*`.
+- **modelMappings:** Exact `sourceModel -> targetModel` rewrites shared by top-level `POST /v1/messages`, `POST /v1/messages/count_tokens`, `POST /v1/responses`, and `POST /v1/chat/completions` requests. Omit it or leave it as `{}` to disable rewrites. Both the source and target must be non-empty strings. Targets can be regular model IDs or `provider/model` aliases such as `dashscope/qwen3.6-plus`, and the rewrite happens before provider alias parsing. These mappings are not split per interface. The admin endpoints `GET/POST /admin/config/model-mappings` read and update only this field.
+- **extraPrompts:** Map of `model -> prompt` appended to the first system prompt when translating Anthropic-style requests to Responses API. Use this to inject guardrails or guidance per model. Missing default entries are auto-added without overwriting your custom prompts. For GPT-5.3+ models (e.g. `gpt-5.3-codex`, `gpt-5.4`, `gpt-5.5`), a built-in commentary prompt is used as fallback when not explicitly configured. The built-in prompts enable phase-aware commentary, which lets the model emit a short user-facing progress update before tools or deeper reasoning.
+- **providers:** Global upstream provider map. Each provider key (for example `dashscope`) becomes a route prefix (`/dashscope/v1/messages`). Supports `type: "anthropic"`, `type: "openai-compatible"`, and `type: "openai-responses"`. Top-level clients can also use `model: "dashscope/model-id"` with `/v1/messages`, `/v1/messages/count_tokens`, `/v1/responses`, and `/v1/chat/completions`; the gateway strips the `dashscope/` prefix before forwarding upstream. The `/v1/responses` route for `anthropic` and `openai-compatible` providers uses the Responses Lite → Messages adapter; `openai-compatible` providers then reuse the Messages → Chat translation. Codex clients (`User-Agent` starting with `codex`) also use the adapter for non-`gpt-*` models on `openai-responses` providers. `GET /v1/models` aggregates enabled provider models with `provider/model-id` IDs, while the top-level Codex-UA catalog also merges these adaptable models as `use_responses_lite` entries (except DeepSeek models, which use `use_responses_lite: false` and `tool_mode: null`). Use `GET /dashscope/v1/models` for a single provider's raw model list.
+  - `enabled` defaults to `true` if omitted.
+  - `baseUrl` should be provider API base URL without the final endpoint. For Anthropic providers, omit `/v1/messages`; for OpenAI-compatible providers, omit `/v1/chat/completions`; for OpenAI Responses providers, omit `/v1/responses`.
+  - `apiKey` is used as the upstream credential value and is required for regular providers.
+  - `authType` (optional): Controls how `apiKey` is sent upstream. Supports `x-api-key` and `authorization` for regular providers. Anthropic providers default to `x-api-key`; OpenAI-compatible and OpenAI Responses providers default to `authorization`. When set to `authorization`, the proxy sends `Authorization: Bearer <apiKey>`. `oauth2` is reserved for the built-in `codex` provider and is written automatically by `auth login --provider codex`.
+  - `pricingCurrency` (optional): Provider-level currency used for token cost calculation, for example `USD` or `CNY`. Quick providers default to `CNY` for DashScope and DeepSeek, and `USD` for Codex, Kimi, OpenCode Go, and OpenRouter. Costs are grouped by currency and are not exchange-rate converted.
+  - `models` (optional): Per-model configuration map. Each key is a model ID (matching the model name in requests), and the value is:
+    - `temperature` (optional): Default temperature value used when the request does not specify one.
+    - `topP` (optional): Default top_p value used when the request does not specify one.
+    - `topK` (optional): Default top_k value used when the request does not specify one.
+    - `extraBody` (optional): Dynamic fields merged into the upstream request body for that model. Request body fields with the same name take precedence. OpenAI-compatible providers can use this for fields such as `enable_thinking`, `preserve_thinking`, `reasoning_effort`. `thinking_budget` is a special OpenAI-compatible provider override: when configured in `extraBody`, it is forced after Anthropic `thinking.budget_tokens` translation and overrides the request-derived budget. For providers whose name is `dashscope` or whose `baseUrl` contains `aliyuncs.com`, the request-derived `thinking_budget` (from Anthropic `thinking.budget_tokens`) is forwarded upstream; for other OpenAI-compatible providers the request-derived `thinking_budget` is stripped, while an `extraBody` `thinking_budget` is still honored. For DashScope providers, `preserve_thinking` defaults to `true` when not explicitly set in `extraBody` or the request body.
+    - `pricing` (optional): Per-model token prices, in the provider `pricingCurrency`, per 1M tokens. Supported fields are `input`, `output`, `cachedInput` (implicit cache read), `explicitCachedInput` (explicit cache read), and `cacheCreationInput`. Use `tiers` with `maxInputTokens` for input-size tiered pricing.
+    - `contextCache` (optional): Defaults to `true` for providers whose name is `dashscope` or whose `baseUrl` contains `aliyuncs.com`; defaults to `false` for other OpenAI-compatible providers. This enables Alibaba Cloud Model Studio/DashScope explicit context cache by injecting `cache_control: { "type": "ephemeral" }` on up to 4 content blocks using the Context Cache format. The cache breakpoint strategy matches opencode's main provider flow: the first 2 system messages plus the last 2 non-system messages. Marked string content is converted to text content part arrays for `system` / `user` / `assistant` / `tool` messages; existing array content is marked on the last part. Set this to `false` when the model already supports implicit caching, or when the upstream does not accept this explicit-cache extension field. Set this to `true` for non-DashScope providers that support the same explicit-cache extension. Applied on both `/v1/messages` and `/v1/chat/completions` routes.
+    - `supportPdf` (optional): Controls whether the model supports PDF/document content. Defaults to `false`; unsupported PDFs are converted to a text notice. Set it to `true` to send PDF/document blocks as OpenAI Chat Completions file parts.
+    - `toolContentSupportType` (optional): Tool result content capabilities for that model, as an array of `array`, `image`, and `pdf`. Provider routes default to string-only tool content when omitted. If `supportPdf` is `true` but this list does not include `pdf`, file parts in tool results are moved to user role messages. The Copilot main flow uses the same string-only default, because some Copilot models do not support array or image tool content either.
+    - `type` (optional): Per-model override of the provider protocol type. Supports `anthropic`, `openai-compatible`, and `openai-responses`. When set, the provider's `/v1/messages` route uses this model's type instead of the provider-level type for request routing, auth header resolution, and upstream endpoint selection. This is useful for providers like OpenCode Go whose upstream supports both OpenAI-compatible and Anthropic Messages APIs for different models. When the type is overridden, the auth header is resolved from the overridden type's default (Anthropic defaults to `x-api-key`; OpenAI-compatible/Responses default to `authorization`).
+    - `contextWindow` (optional): Context window token limit advertised when this model is merged into the Codex-UA model catalog; for example, `1000000` declares a 1M-token context window. Missing configured values use upstream metadata first, then the built-in non-GPT model catalog, then `256000`.
+    - `maxOutputTokens` (optional): Maximum output token limit advertised in the Codex-UA model catalog. Missing configured values use upstream metadata first, then the built-in non-GPT model catalog, where defaults are capped at `64000`, then `32000`.
+    - `inputModalities` (optional): Supported Codex input types. Use `["text", "image"]` for a model that accepts both text and images. Missing configured values use upstream metadata before the built-in non-GPT model catalog. GPT models do not receive these built-in capability defaults and continue to use the native Codex catalog or upstream metadata.
+    - `reasoningEfforts` (optional): Reasoning levels advertised for Codex. Missing configured and upstream values use the built-in non-GPT model catalog before falling back to `["high", "xhigh", "max", "ultra"]`. Provider Responses requests with an unsupported effort are normalized to a supported level when these capabilities are known.
+    - `defaultReasoningEffort` (optional): Default Codex reasoning level. Built-in model metadata may provide a known default; otherwise it defaults to `max` when available, then the first configured level. Synthetic Codex models always enable parallel tool calls.
+- **smallModel:** Fallback model used for tool-less warmup messages (e.g., Claude Code probe requests); defaults to gpt-5-mini. The gateway forces this small model on no-tool warmup or probe requests to avoid consuming premium requests. This behavior only applies to non-token-based-billing GitHub Copilot accounts (`token_based_billing` is false); for token-based-billing accounts the warmup small-model fallback is skipped since there is no premium-request quota to preserve.
+- **contextManagement:** Controls whether the proxy adds Responses API `context_management` compaction instructions. `messages` applies when Anthropic-style `/v1/messages` requests are translated to Responses API, including `openai-responses` provider message routes, and defaults to `true`. `responses` applies to native `/v1/responses` traffic, including `provider/model` aliases and the built-in `codex` provider, and defaults to `false`. Enable `responses` only after checking that your client supports context management compaction. When enabled, the request includes `context_management` in the body and keeps only the latest compaction carrier on follow-up turns. The proxy only adds context management and compacts history for `gpt-*` models; both configuration switches have no effect on non-GPT models such as Grok. **Note:** Context management is also forcibly disabled for GPT-5.6 and above models (e.g. `gpt-5.6-sol`, `gpt-5.6-terra`, `gpt-5.6-luna`) because enabling it breaks prompt cache hits on those models. These overrides take precedence over the `contextManagement` and `modelResponsesApiCompactThresholds` settings.
+ - **modelResponsesApiCompactThresholds:** Per-model Responses API `compact_threshold` overrides used when the proxy adds `context_management`. These values take precedence over the fallback threshold from `resolveResponsesCompactThreshold` (`max_prompt_tokens * ratio`, or the default fallback). Defaults set `gpt-5.4` and `gpt-5.5` to `217600` (`272000 * 0.8`). Models not listed continue to use the normal fallback logic.
+- **modelReasoningEfforts:** Per-model fallback reasoning effort for `/v1/messages` requests. It is used only when the request does not provide `output_config.effort`.
+  - **Priority:** request `output_config.effort` > `modelReasoningEfforts[model]` > built-in default (`xhigh` for GPT-5.3+ models, otherwise `high`).
+  - **Forwarding:** the resolved value remains `output_config.effort` for the Copilot native Messages API and becomes `reasoning.effort` when translated to the Responses API.
+  - **Configuration values:** `none`, `minimal`, `low`, `medium`, `high`, `xhigh`, and `max`.
+- **useMessagesApi:** When `true`, models that advertise Copilot's native `/v1/messages` endpoint use the Messages API. If Messages is disabled or unavailable for the selected model, the gateway uses Responses when that model advertises a Responses endpoint, then falls back to Chat Completions when supported. Set this to `false` to skip native Messages routing. Defaults to `true`.
+- **useResponsesApiCompactionRecovery (experimental):** When `true`, successful remote Responses compactions enqueue a non-blocking, low-reasoning shadow-summary request and store only that summary under a hash of the opaque compaction item in the private `~/.local/share/copilot-api/compaction-recovery/cache.sqlite` directory. If Copilot later rejects that compaction or connection-bound history, HTTP requests progressively rebuild the request from cached summaries and visible messages. This option defaults to `false` because shadow summaries add one background model request and recovery is lossy. WebSocket errors are recovered on the client's next retry; automatic same-request recovery is HTTP-only.
+- **useResponsesApiWebSocket:** When `true`, Copilot Responses requests use WebSocket for models that advertise `ws:/responses`; models that advertise only `/responses` use HTTP. Streamed Responses requests for the built-in `codex` provider use WebSocket whenever this setting is enabled, while non-streaming Codex requests always use HTTP. Set this to `false` to make Copilot use HTTP `/responses` where the selected model advertises it and to send streamed Codex Responses requests over HTTP. WebSocket failures are not retried automatically over HTTP. Defaults to `true`. If a proxy, VPN, or network blocks or destabilizes WebSocket traffic, disable this setting or switch networks.
+- **responsesTransport:** Positive integer lifecycle and buffering limits for every upstream Responses transport. Invalid, zero, or negative values fall back to the defaults shown above. `headersTimeoutMsV2` covers connection setup through receipt of HTTP response headers; it is not a total generation deadline. `streamInactivityTimeoutMs` is reset by every HTTP body chunk or WebSocket message, allowing long generations to continue while they remain active. `websocketOpenTimeoutMs` limits the WebSocket handshake, while `websocketPoolIdleTimeoutMs` controls only completed, reusable pooled sockets. The byte and message limits bound queued WebSocket events; exceeding either limit fails that stream and invalidates its socket rather than dropping or reordering events.
+- **useResponsesApiWebSearch:** When `true`, the server keeps Responses API tools with `type: "web_search"` and forwards them upstream. Set to `false` to strip those tools from `/responses` payloads. Defaults to `true`.
+- **alphaSearchCodexPriority:** Defaults to `true`. Top-level alpha-search requests prefer the Codex alpha-search endpoint because it does not consume provider quota. If Codex is unavailable, or this setting is `false`, requests with a `provider/model` alias other than `codex/model` use that provider's `/v1/responses` endpoint, and requests without a provider prefix use GitHub Copilot Responses web search. The adapter recognizes every current Codex search command; unsupported `image_query` and `screenshot` operations return successful no-retry tool output.
+- **alphaSearchModel:** Native Responses search model used when a Messages-backed Responses Lite model cannot run Responses web search directly. Defaults to `gpt-5-mini`; it may be a regular Copilot model or an `openai-responses` `provider/model` alias. Set it to an empty string to disable this redirect, in which case alpha-search requests for those models return an invalid-request error.
+- **messageApiWebSearchModel:** Global fallback model used when a top-level Copilot `/v1/messages` request contains only the server-side `web_search` tool. Defaults to `gpt-5-mini`. If the value is a `provider/model` alias, the request is routed into that provider's Messages API path with the provider prefix stripped. For Copilot GPT models, web search runs through `/responses`. Mixed `web_search` plus custom tools are not supported and the server-side `web_search` tool is stripped.
+- **claudeAutoModel:** Model used for Claude Code background security-monitor requests on `/v1/messages` and provider message routes. A request is treated as a security-monitor request when it carries no tools, sets `stop_sequences` to `["</block>"]`, and contains a system text block starting with `You are a security monitor for autonomous AI coding agents.`; its model is then replaced with this value. For top-level requests, a `provider/model` alias is forwarded into that provider's Messages API; provider routes keep their current provider and use this configured value directly. Defaults to empty (disabled).
+- **claudeTokenMultiplier:** Multiplier applied to the fallback GPT-tokenizer estimate for Claude `/v1/messages/count_tokens` requests. Defaults to `1.15`. Increase it if your client is still compacting too late. This setting is only used when the proxy is estimating Claude tokens locally; if `anthropicApiKey` is configured and Anthropic token counting succeeds, the exact Anthropic count is returned instead.
+- **anthropicApiKey:** Anthropic API key used to forward Claude `/v1/messages/count_tokens` requests to Anthropic's real token counting endpoint, which returns exact counts instead of GPT tokenizer estimates. Can also be set via the `ANTHROPIC_API_KEY` environment variable. If not set, or if the upstream call fails, token counting falls back to local GPT tokenizer estimation controlled by `claudeTokenMultiplier`.
+
+Edit this file to customize prompts or swap in your own fast model. Restart the server (or rerun the command) after changes so the cached config is refreshed.
+
+## API Authentication
+
+- **Protected non-admin routes:** All routes except `/`, `/usage-viewer`, and `/usage-viewer/` require authentication when `auth.apiKeys` is configured and non-empty.
+- **Admin routes:** All `/admin/*` routes require `auth.adminApiKey`. If it is missing, the server generates one at startup and persists it to `config.json` before serving requests.
+- **Allowed auth headers:**
+  - `x-api-key: <your_key>`
+  - `Authorization: Bearer <your_key>`
+- **CORS preflight:** `OPTIONS` requests are always allowed.
+- **When no regular keys are configured:** Non-admin routes continue to allow requests. This does not apply to `/admin/*`, which only accepts `auth.adminApiKey`.
+
+Example request for a regular protected route:
+
+```sh
+curl http://localhost:4141/v1/models \
+  -H "x-api-key: your_api_key"
+```
+
+Example request for an admin route:
+
+```sh
+curl http://localhost:4141/admin/config/model-mappings \
+  -H "x-api-key: your_admin_api_key"
+```
+
+## API Endpoints
+
+The server exposes several OpenAI- and Anthropic-compatible endpoints. Requests can target GitHub Copilot, the built-in `codex` provider, or configured providers depending on the selected model and `provider/model` alias. Every `/v1/...` endpoint below also supports a provider-scoped path in the form `/:provider/v1/...`; those variants are omitted from the tables.
+
+### OpenAI Compatible Endpoints
+
+These endpoints mimic the OpenAI API structure.
+
+| Endpoint                    | Method | Description                                                      |
+| --------------------------- | ------ | ---------------------------------------------------------------- |
+| `POST /v1/responses`        | `POST` | OpenAI Most advanced interface for generating model responses. Supports `provider/model` aliases for `openai-responses` providers. |
+| `POST /v1/chat/completions` | `POST` | Creates a model response for the given chat conversation. Supports `provider/model` aliases for `openai-compatible` providers and can be used without Copilot when the target provider is configured. |
+| `GET /v1/models`            | `GET`  | Lists Copilot models plus enabled provider models using `provider/model-id` IDs. Requests from Codex clients (`User-Agent` beginning with `codex`) are forwarded to the Codex Models upstream. |
+| `POST /v1/embeddings`       | `POST` | Creates an embedding vector representing the input text.         |
+
+### Codex Backend Endpoints
+
+These endpoints implement Codex backend APIs. Top-level image requests require an active Codex login; alpha search can use either the Codex backend or a Responses web-search adapter.
+
+| Endpoint                                                       | Method | Description                                                     |
+| -------------------------------------------------------------- | ------ | --------------------------------------------------------------- |
+| `POST /v1/alpha/search`                | `POST` | Routes Codex alpha-search requests to the Codex backend, or handles supported commands locally and through Responses web search. |
+| `POST /v1/images/generations` | `POST` | Forwards a JSON image generation request to the Codex Images upstream. When the request omits `Content-Type`, the gateway defaults it to `application/json`. |
+| `POST /v1/images/edits` | `POST` | Forwards an image edit request to the Codex Images upstream. Send this request as `multipart/form-data` and let the HTTP client generate the `boundary`; the gateway preserves the incoming content type and streams the upload body. |
+
+For requests routed to the Codex backend, the gateway replaces client authorization and account headers with the active Codex login and preserves compatible request metadata. Responses-backed alpha search instead follows the selected Copilot or provider route.
+
+### Anthropic Compatible Endpoints
+
+These endpoints are designed to be compatible with the Anthropic Messages API.
+
+| Endpoint                         | Method | Description                                                  |
+| -------------------------------- | ------ | ------------------------------------------------------------ |
+| `POST /v1/messages`              | `POST` | Creates a model response for a given conversation. Supports `provider/model` aliases for configured providers, including translation through `openai-compatible` providers. |
+| `POST /v1/messages/count_tokens` | `POST` | Calculates the number of tokens for a given set of messages. Supports `provider/model` aliases for configured providers. |
+
+### Usage Monitoring Endpoints
+
+New endpoints for monitoring your Copilot usage and quotas.
+
+| Endpoint     | Method | Description                                                  |
+| ------------ | ------ | ------------------------------------------------------------ |
+| `GET /usage` | `GET`  | Get detailed Copilot usage statistics and quota information. |
+
+### Admin / Configuration Endpoints
+
+These endpoints are reserved for local administrative actions and only accept `auth.adminApiKey`.
+
+| Endpoint                              | Method | Description                                                                 |
+| ------------------------------------- | ------ | --------------------------------------------------------------------------- |
+| `GET /admin/config/model-mappings`    | `GET`  | Returns the current `config.json` path and the active `modelMappings` map.  |
+| `POST /admin/config/model-mappings`   | `POST` | Updates only the `modelMappings` field in `config.json` and returns it back. |
+
+## Example Usage
+
+Common `npx` commands:
+
+```sh
+# Start the gateway
+npx xiaodcs-copilot-api-edge@latest start
+
+# Start on a custom port with verbose logging
+npx xiaodcs-copilot-api-edge@latest start --port 8080 --verbose
+
+# Run the auth flow
+npx xiaodcs-copilot-api-edge@latest auth login
+
+# Configure a third-party provider, then run without GitHub Copilot
+npx xiaodcs-copilot-api-edge@latest auth login --provider dashscope
+npx xiaodcs-copilot-api-edge@latest start
+
+# Print debug information as JSON
+npx xiaodcs-copilot-api-edge@latest debug --json
+
+# Run the published CLI with Bun instead of Node.js
+bunx --bun xiaodcs-copilot-api-edge@latest start
+```
+
+OpenAI-compatible provider examples after configuring `dashscope`:
+
+```sh
+curl http://localhost:4141/v1/chat/completions \
+  -H "content-type: application/json" \
+  -d '{"model":"dashscope/qwen3.6-plus","messages":[{"role":"user","content":"hello"}]}'
+
+curl http://localhost:4141/dashscope/v1/messages \
+  -H "content-type: application/json" \
+  -d '{"model":"qwen3.6-plus","max_tokens":1024,"messages":[{"role":"user","content":"hello"}]}'
+```
+
+## Usage Tips
+
+### CLAUDE.md or AGENTS.md Recommended Content
+
+Same reminders as `CLAUDE_PLUGIN_ENABLE_QUESTION_RULES=1` in the `agent-inject` plugin, for when you don't use that plugin. Add to `CLAUDE.md` (Claude Code) or `AGENTS.md` (opencode/codex):
+
+```
+- Prohibited from directly asking questions to users, MUST use question tool.
+- Once you can confirm that the task is complete, MUST use question tool to make user confirm. The user may respond with feedback if they are not satisfied with the result, which you can use to make improvements and try again, after try again, MUST use question tool to make user confirm again.
+```