AMD TokenFactory Provider for Copilot

English

A VS Code extension that integrates the AMD TokenFactory (Radeon AI Developer, free endpoint) models into GitHub Copilot Chat.

Architecture based on tokenrhythm-copilot (AGPL-3.0), which itself is based on opencode-go-copilot (MIT) and oai-compatible-copilot (MIT).

Two features only — nothing else:

OpenAI-compatible endpoint (https://developer.amd.com.cn/radeon/api/v1)
- Standard Chat Completions protocol with SSE streaming
- Reasoning/thinking content (reasoning_content) shown in Copilot's thinking UI
- Vision input (image_url) for vision-capable models
- Tool calling passthrough
- Automatic model discovery from the /models endpoint (with a built-in fallback list)
- Token usage reported to Copilot's native token indicator
Multi-key rotation
- Configure any number of rc-... API keys; every request picks the next available key (rotation) or sticks to one key until it fails (sticky)
- Keys failing with rotation errors (401 / 429 / 5xx, configurable status codes and text patterns) are skipped and the request retries with the next key
- Transient failures (429 / 5xx) cool the key down for a configurable period instead of disabling it permanently, plus a whole-round exponential-backoff retry
- Key management UI: add / batch import / delete / reset / availability check (masked display)

Install the extension and run the command AMD TokenFactory: Manage API Keys (multi-key rotation) (or click the provider in the Copilot Chat model picker — it opens the key manager).
Add at least one key.
Open the Copilot Chat model picker and choose a model under the AMD TokenFactory vendor.

Setting	Default	Description
`amdTokenFactory.baseUrl`	`https://developer.amd.com.cn/radeon/api/v1`	API base URL
`amdTokenFactory.apiKeyMode`	`rotation`	`rotation` (next key per request) or `sticky` (keep current key until it fails)
`amdTokenFactory.apiKeyRotationStatusCodes`	`[401,429,500,502,503,504]`	HTTP status codes that trigger switching to the next key
`amdTokenFactory.apiKeyRotationErrorPatterns`	rate limit / concurrency / invalid bearer token …	Error text patterns that trigger switching to the next key
`amdTokenFactory.transientRetryStatusCodes`	`[429,500,502,503,504]`	Status codes treated as transient (cooldown + whole-round retry)
`amdTokenFactory.exhaustedCooldownMin`	`5`	Cooldown minutes for transiently failed keys
`amdTokenFactory.transientRetryTimes`	`3`	Whole-round automatic retries when all keys are cooling down
`amdTokenFactory.requestTimeout`	`300000`	Request timeout (ms)
`amdTokenFactory.enableAutoModelDiscovery`	`true`	Discover the model list from `/models`
`amdTokenFactory.maxInputTokensRatio`	`1.0`	Declared input window ratio (controls Copilot auto-compaction)
`amdTokenFactory.temperature` / `amdTokenFactory.top_p`	`null`	Sampling params (null = not sent)
`amdTokenFactory.retry.*`	enabled / 3 / 1000	Per-request retry (network + retryable status codes)

npm install
npm run compile
npm run build      # package as amd-tokenfactory-copilot-<version>.vsix

MIT. Architecture based on tokenrhythm-copilot (AGPL-3.0), based on opencode-go-copilot (MIT) and oai-compatible-copilot (MIT).

把 AMD TokenFactory（Radeon AI 开发者平台，免费端点）的模型接入 GitHub Copilot Chat 的 VS Code 扩展。

只有两个功能，没有别的：

OpenAI 兼容端点（https://developer.amd.com.cn/radeon/api/v1）
- 标准 Chat Completions 协议 + SSE 流式输出
- 思考内容（reasoning_content）显示在 Copilot 的思考 UI 里
- 视觉模型支持图片输入（image_url）
- 工具调用透传
- 自动从 /models 端点发现模型列表（带内置兜底清单）
- token 用量上报到 Copilot 原生 token 指示器
多 Key 轮询
- 配置任意数量的 rc-... API Key；每次请求取下一个可用 Key（轮询模式），或固定用一个 Key 直到失效再切换（固定模式）
- Key 遇到轮换错误（401 / 429 / 5xx，状态码与文本 patterns 可配置）时跳过该 Key 换下一个重试
- 瞬态故障（429 / 5xx）只做冷却不永久禁用，外加整轮指数退避自动重试
- Key 管理界面：添加 / 批量导入 / 删除 / 重置失效 / 可用性检测（全部脱敏显示）

在 https://developer.amd.com.cn/radeon/tokenfactory 登录并创建 API Key（格式 rc-...）。

安装扩展，运行命令 AMD TokenFactory: 管理 API Keys（多 Key 轮询）（或在 Copilot Chat 模型选择器里点本提供商，会直接打开 Key 管理界面）。
添加至少一个 Key。
打开 Copilot Chat 的模型选择器，选择 AMD TokenFactory 下的模型。

npm install
npm run compile
npm run build      # 打包为 amd-tokenfactory-copilot-<version>.vsix

MIT。架构基于 tokenrhythm-copilot（AGPL-3.0），其基于 opencode-go-copilot（MIT）与 oai-compatible-copilot（MIT）。