hg552 · Nov 8, 2026 · Nov 9, 2026 · Nov 9, 2026
diff --git a/README.md b/README.md
@@ -35,6 +35,7 @@ subscription and makes its requests. The packages work in OpenCode and in
 | [minimax](packages/minimax) | MiniMax Code (China): account credits and M Plan, with MiniMax Code's device sign-in | `minimax-code` |
 | [minimax](packages/minimax) | MiniMax Code (international): account credits and M Plan, with MiniMax Code's device sign-in | `minimax-code-global` |
 | [mimo](packages/mimo) | Xiaomi MiMo, with a Xiaomi account | `mimo-app` |
+| [mistral-vibe](packages/mistral-vibe) | Mistral Vibe (Vibe Code, unofficial): the CLI's model catalog plus DeepSeek, each model on both wallets — the plan's Vibe budget, and the API budget as `<id>--api` — with both budgets as quota windows | `mistral-vibe` |
 | [nous](packages/nous) | Nous Portal models, with Nous Portal's device sign-in | `nous-portal` |
 | [opendesign](packages/opendesign) | Open Design Cloud (open-design.ai), with the Open Design CLI's device sign-in | `opendesign` |
 | [qoder](packages/qoder) | Qoder subscription (qoder.com), with Qoder's device sign-in | `qoder` |

diff --git a/packages/mistral-vibe/LICENSE b/packages/mistral-vibe/LICENSE
@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 magpie-community contributors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
diff --git a/packages/mistral-vibe/README.md b/packages/mistral-vibe/README.md
@@ -0,0 +1,175 @@
+# Mistral Vibe
+
+[简体中文](README.zh-CN.md) | English
+
+Mistral Vibe's coding plan in [magpie](https://usemagpie.ai) and
+[OpenCode](https://opencode.ai). **Unofficial:** not affiliated with,
+endorsed by, or supported by Mistral AI.
+
+One provider, one model group, two wallets. The Vibe key signs in and the
+model list is the Vibe plan's own catalog plus DeepSeek; every model is
+offered on the Vibe budget and, if a Studio key is given too, on the API
+budget as `<id>--api`. The plan's monthly Vibe Code and API budgets are read
+into magpie's usage card, `magpie quota`, the quota alert and
+`GET /v1/magpie/quotas`.
+
+**Provider id:** `mistral-vibe`
+
+| | |
+|---|---|
+| API | `https://api.mistral.ai/v1` (`@ai-sdk/openai-compatible`) |
+| Sign-in | one credential holding both keys: `vibe=…;api=…` (the Vibe key alone works) |
+| Models | the Vibe Code catalog (4 models on 2026-11-08) plus DeepSeek |
+| Quota | `GET /v1/billing/budget` → `vibe_budget`, `api_budget`, `reset_at` |
+
+## Sign-in
+
+```sh
+magpie plugin login mistral-vibe
+```
+
+Paste one line:
+
+```
+vibe=<Vibe key>;api=<Studio key>
+```
+
+- **Vibe key** — the key the official CLI writes to `~/.vibe/.env` after
+  `vibe --setup`, or chat.mistral.ai/code/extensions → **Vibe CLI** card →
+  **Advanced** → **Vibe API Key**. Its requests spend the plan's Vibe budget.
+- **Studio key** (optional) — console → API › API keys. It adds the same
+  models on the API budget.
+
+The Vibe key alone is enough for the Vibe wallet. Neither key expires. JSON
+(`{"vibe":"…","api":"…"}`) and two bare keys separated by `;` are accepted
+too.
+
+The line holds both keys: treat it like two secrets, and don't paste it into
+issues, screenshots or logs.
+
+## Models
+
+On 2026-11-08 the Vibe Code catalog held four models:
+
+| id | label | context | effort levels | vision |
+|---|---|---|---|---|
+| `mistral-large-4-0` | Mistral Large 4 | 1M | low / medium / high | yes |
+| `mistral-medium-latest` | Mistral Medium 3.5 | 262k | none / high | yes |
+| `zai-glm-5-3` | GLM 5.3 (Mistral Hosted) | 1M | low / high / max | no |
+| `zai-glm-5-3-flash` | GLM 5.3 Flash (Mistral Hosted) | 1M | low / high / max | yes |
+
+DeepSeek is added because the Vibe key serves it (measured), though the CLI's
+picker does not list it. With both keys every model is offered twice; the
+alias picks the wallet:
+
+| Model id | Name | Spends |
+|---|---|---|
+| `deepseek-v4-1-flash-1017` | `deepseek-v4-1-flash-1017 (Vibe budget)` | the plan's Vibe budget first, then PAYG |
+| `deepseek-v4-1-flash-1017--api` | `deepseek-v4-1-flash-1017 (API budget)` | the API budget / PAYG |
+
+The `--api` suffix is stripped before the request goes upstream; the fetch
+picks the key from it. With one key only, only that wallet's ids are listed.
+
+Measured on 2026-11-08 on an Education-plan account (150 USD Vibe Code
+budget, 15 USD API budget), same model through magpie's gateway:
+
+| Alias | Request | `vibe_budget` | Charged to the Vibe budget |
+|---|---|---|---|
+| `…--api` | 248 tokens, 200 | 0.000784 % → 0.000784 % | **$0.000000** |
+| plain | 266 tokens, 200 | 0.000784 % → 0.000986 % | **$0.000303** |
+
+If the catalog can't be read, the account's `/v1/models` is used instead,
+deduped and filtered: chats only (`capabilities.completion_chat`) and one id
+per family (64 entries → 14 for the account above). Families come from
+`billing_model_name` and the `aliases` links; the id kept is, in order, the
+one billing names, the one whose display name is the id, else `-latest`. A
+request naming an aliased id from an older config still works.
+
+### Show the models in magpie
+
+magpie shows only the models ticked for a provider, and a plugin's new ids
+start unticked. Tick them once — the app's provider page, or:
+
+```sh
+magpie provider models mistral-vibe \
+  +mistral-large-4-0 +mistral-large-4-0--api \
+  +mistral-medium-latest +mistral-medium-latest--api \
+  +zai-glm-5-3 +zai-glm-5-3--api \
+  +zai-glm-5-3-flash +zai-glm-5-3-flash--api \
+  +deepseek-v4-1-flash-1017 +deepseek-v4-1-flash-1017--api
+```
+
+## Quota
+
+`auth.usage` reads the three public endpoints the official CLI reads:
+`/v1/billing/budget` (both budgets, their reset), `/v1/billing/subscription`
+(plan name) and `console.mistral.ai/api/vibe/whoami` (the key's scope). Both
+windows are shown; the key's own budget comes first. A refused key (401)
+marks the account expired; a 403 on the budget or a plan without one says so
+instead of showing an empty card.
+
+```sh
+magpie quota mistral-vibe
+```
+
+## Request and reply fixes
+
+Written from the behaviour measured against Mistral's API — the fields it
+answers with 422, GLM's 4096-token cap without `max_tokens`, and its typed
+content parts, in `src/request.mjs`, `src/reply.mjs`, `src/effort.mjs` and
+`src/transport.mjs`. They are not a translation of another implementation. The
+alias table in `src/constants.mjs` is plain facts about Mistral's ids (the same
+ids mistral-gateway lists).
+
+- aliases are rewritten to the id Mistral serves;
+- `max_completion_tokens` → `max_tokens`, `seed` → `random_seed`, `user`
+  dropped (Mistral answers those with 422);
+- `max_tokens` is filled in for GLM only (131072): without it Mistral caps a
+  GLM completion at 4096 tokens, which its thinking alone can spend;
+- `reasoning_content` and `index` are stripped from messages a client sends
+  back as they came;
+- a GLM tool-result turn without an effort gets `reasoning_effort: "high"`,
+  which avoids GLM ending such a turn with no visible text;
+- `stream_options: {include_usage: true}` is added to a stream that names none;
+- a rejected `reasoning_effort` is retried once at the nearest level the model
+  lists (nested `reasoning.effort` on `/v1/responses`);
+- a 429 is returned as it came: magpie rests the account and falls back on it,
+  and `Retry-After` says when to try again;
+- GLM's typed content parts (`[{type:"thinking"…},{type:"text"…}]`) become
+  `content` (a string) and `reasoning_content`, streamed or buffered, and a
+  stream that delta'd tool calls but ends `stop` ends `tool_calls`;
+- a rebuilt reply drops the upstream's `content-encoding` and `content-length`,
+  which no longer describe its bytes.
+
+One account holds both wallets, so a refusal that rests the account rests the
+`--api` ids with it.
+
+## Privacy
+
+The keys stay in magpie's own store (`plugin-auth.json`, or OpenCode's
+`auth.json`). The plugin sends no telemetry and reads nothing else; its only
+network calls are the three quota reads above and the requests magpie makes
+for a model.
+
+## Disclaimer
+
+This is an unofficial community project, not affiliated with, endorsed by or
+supported by Mistral AI. It uses the credentials and endpoints the official
+Mistral Vibe CLI uses; Mistral's terms still apply, and the plan's limits stay
+in force. Mistral's consumer terms forbid sharing an account or its keys, and
+creating accounts to get around limits — keep this to your own account. Terms
+and endpoints can change at any time; Mistral decides how it enforces its own
+terms. The software is provided "as is", without warranty of any kind, and the
+authors accept no liability for any loss, including account suspension,
+restricted keys or lost subscription fees. Check the current terms before
+relying on it.
+
+## License
+
+MIT. See [LICENSE](LICENSE).
+
+## Test
+
+```sh
+bun test          # local stand-in for Mistral's endpoints; no network, no key
+```
diff --git a/packages/mistral-vibe/README.zh-CN.md b/packages/mistral-vibe/README.zh-CN.md
@@ -0,0 +1,158 @@
+# Mistral Vibe
+
+简体中文 | [English](README.md)
+
+把 **Mistral Vibe** 的编程套餐接入 [magpie](https://usemagpie.ai) 与
+[OpenCode](https://opencode.ai) 的供应商插件。**非官方**：与 Mistral AI 无隶属、
+无背书、无支持关系。
+
+一个供应商、一个模型组、两个钱包。Vibe key 登录后，模型来自 Vibe 套餐自己的
+目录并追加 DeepSeek；每个模型都出现在 Vibe 钱包里，若同时提供 Studio key，还会
+以 `<id>--api` 的形式出现在 API 钱包里。套餐每月的 Vibe Code 与 API 额度会显示
+在 magpie 的用量卡片、`magpie quota`、额度提醒和 `GET /v1/magpie/quotas` 中。
+
+**供应商 id：** `mistral-vibe`
+
+| | |
+|---|---|
+| API | `https://api.mistral.ai/v1`（`@ai-sdk/openai-compatible`） |
+| 登录 | 一条凭据装两把 key：`vibe=…;api=…`（只填 Vibe key 也可以） |
+| 模型 | Vibe Code 目录（2026-11-08 为 4 个）加上 DeepSeek |
+| 额度 | `GET /v1/billing/budget` → `vibe_budget`、`api_budget`、`reset_at` |
+
+## 登录
+
+```sh
+magpie plugin login mistral-vibe
+```
+
+粘贴一行：
+
+```
+vibe=<Vibe key>;api=<Studio key>
+```
+
+- **Vibe key**：官方 CLI 执行 `vibe --setup` 后写入 `~/.vibe/.env` 的那把，
+  或 chat.mistral.ai/code/extensions → **Vibe CLI** 卡片 → **Advanced** →
+  **Vibe API Key**。它的请求先花套餐的 Vibe 额度。
+- **Studio key**（可选）：console 的 **API › API keys**。它让同一批模型也出现在
+  API 额度上。
+
+只填 Vibe key 就能用 Vibe 钱包。两把 key 都不过期。也接受 JSON
+（`{"vibe":"…","api":"…"}`）或两把裸 key 以 `;` 分隔。
+
+这一行同时包含两把 key，等同于两个密码：不要贴到 issue、截图或日志里。
+
+## 模型
+
+2026-11-08 的 Vibe Code 目录有 4 个模型：
+
+| id | 名称 | 上下文 | 思考档位 | 视觉 |
+|---|---|---|---|---|
+| `mistral-large-4-0` | Mistral Large 4 | 1M | low / medium / high | 是 |
+| `mistral-medium-latest` | Mistral Medium 3.5 | 262k | none / high | 是 |
+| `zai-glm-5-3` | GLM 5.3 (Mistral Hosted) | 1M | low / high / max | 否 |
+| `zai-glm-5-3-flash` | GLM 5.3 Flash (Mistral Hosted) | 1M | low / high / max | 是 |
+
+额外加入 DeepSeek：CLI 选择器里没有它，但实测 Vibe key 可以调用。两把 key 都在
+时，每个模型出现两次，靠别名区分钱包：
+
+| 模型 id | 显示名 | 花哪个 |
+|---|---|---|
+| `deepseek-v4-1-flash-1017` | `deepseek-v4-1-flash-1017 (Vibe budget)` | 先花套餐的 Vibe 额度，再按量付费 |
+| `deepseek-v4-1-flash-1017--api` | `deepseek-v4-1-flash-1017 (API budget)` | API 额度 / 按量付费 |
+
+`--api` 后缀在请求发出前去掉，fetch 用它挑 key。只有一把 key 时，只列出该钱包的 id。
+
+2026-11-08 在 Education 套餐账号上实测（Vibe Code 额度 150 USD、API 额度
+15 USD），同一个模型经 magpie 网关：
+
+| 别名 | 请求 | `vibe_budget` | 扣在 Vibe 额度上 |
+|---|---|---|---|
+| `…--api` | 248 tokens，200 | 0.000784 % → 0.000784 % | **$0.000000** |
+| 裸 id | 266 tokens，200 | 0.000784 % → 0.000986 % | **$0.000303** |
+
+目录读不到时，退回账号的 `/v1/models` 并去重：只留对话模型
+（`capabilities.completion_chat`），同一底层模型只留一个 id（上面的账号是
+64 条 → 14 个）。归并依据 `billing_model_name` 与 `aliases` 互链；保留的 id
+依次优先：billing 名、显示名与 id 相同的那个、`-latest`。老配置里用别名的请求
+仍然可用。
+
+### 在 magpie 里显示模型
+
+magpie 只显示已勾选的模型，插件新增的 id 默认未勾选。勾一次即可——在应用里的
+供应商页面，或：
+
+```sh
+magpie provider models mistral-vibe \
+  +mistral-large-4-0 +mistral-large-4-0--api \
+  +mistral-medium-latest +mistral-medium-latest--api \
+  +zai-glm-5-3 +zai-glm-5-3--api \
+  +zai-glm-5-3-flash +zai-glm-5-3-flash--api \
+  +deepseek-v4-1-flash-1017 +deepseek-v4-1-flash-1017--api
+```
+
+## 额度
+
+`auth.usage` 读取官方 CLI 用的三个公开接口：`/v1/billing/budget`（两个额度和
+重置时间）、`/v1/billing/subscription`（套餐名）、
+`console.mistral.ai/api/vibe/whoami`（key 的作用域）。两个窗口都会显示，key
+自己的那个排在前面。key 被拒（401）会标记账号过期；额度接口 403 或套餐没有该
+额度时会说明原因，而不是显示空卡片。
+
+```sh
+magpie quota mistral-vibe
+```
+
+## 请求与回复修正
+
+按对 Mistral 接口的实测行为重写（它对哪些字段返回 422、GLM 不传
+`max_tokens` 时 4096 的输出上限、以及类型化 content），代码在
+`src/request.mjs`、`src/reply.mjs`、`src/effort.mjs`、`src/transport.mjs`，
+不是对其它实现的翻译。`src/constants.mjs` 的别名表是 Mistral id 的客观事实
+（mistral-gateway 也列出同一批 id）。
+
+- 别名改写为 Mistral 实际提供的 id；
+- `max_completion_tokens` → `max_tokens`、`seed` → `random_seed`、丢弃 `user`
+  （Mistral 对这些返回 422）；
+- 只给 GLM 补 `max_tokens`（131072）：不补时 Mistral 把 GLM 输出限制在
+  4096，思考就可能耗尽；
+- 客户端原样回传的消息里去掉 `reasoning_content` 和 `index`；
+- GLM 的 tool 结果轮次若未指定档位，补 `reasoning_effort: "high"`，避免它
+  无可见输出就结束；
+- 流式请求缺少 `stream_options: {include_usage: true}` 时补上；
+- 被拒的 `reasoning_effort` 会按模型支持的最接近档位重试一次（`/v1/responses`
+  下是嵌套的 `reasoning.effort`）；
+- 429 原样返回：由 magpie 停用账号并切换下一个，`Retry-After` 说明何时重试；
+- GLM 的类型化 content（`[{type:"thinking"…},{type:"text"…}]`）在流式与非流式
+  下都换算为字符串 `content` 和 `reasoning_content`；已产生 tool_calls 却以
+  `stop` 结束的流改为 `tool_calls`；
+- 重建的回复会去掉上游的 `content-encoding` 与 `content-length`，它们已不再
+  描述新的字节。
+
+一个账号同时装两个钱包，所以一次拒绝对账号的停用也会连带 `--api` 那些 id。
+
+## 隐私
+
+key 只保存在 magpie 自己的存储（`plugin-auth.json`，或 OpenCode 的
+`auth.json`）里。插件不上报遥测、不读取其他文件；网络请求只有上面三个额度
+接口，以及 magpie 为模型发出的请求。
+
+## 免责声明
+
+这是非官方社区项目，与 Mistral AI 无隶属、无背书、无支持关系。它使用官方
+Mistral Vibe CLI 相同的凭据与接口；Mistral 的条款仍然适用，套餐额度限制照常
+生效。Mistral 的消费者条款禁止共享账号或 key、禁止通过多账号绕过限制——请只在
+自己的账号上使用。条款与接口可能随时变化，如何执行由 Mistral 决定。本软件按
+"现状"提供，不附带任何担保；作者不对任何损失负责，包括账号被限制、key 被停用
+或订阅费用损失。依赖前请自行核对最新条款。
+
+## 许可证
+
+MIT，见 [LICENSE](LICENSE)。
+
+## 测试
+
+```sh
+bun test          # 本地模拟 Mistral 接口；不联网、不需要 key
+```