v3.2.1

Avibe 3.2.1

Avibe 3.2.1 gives Vibey built-in web search, X search and image generation, keeps the service steadier under load, and fits reasoning effort to each Source. Upgrading from 3.2.0 runs two small database migrations; see Upgrading from 3.2.0 below.

Highlights

• Vibey can search the web, search X and make images. These built-in tools are on by default. Vibey uses its model provider's own search when there is one, otherwise another of your Sources, and generates images with a model you choose.
• Steadier under load. Several multi-second service stalls are gone, and Model Hub now cuts a stream that has gone silent and retries it before any output.
• Reasoning effort that matches each Source. Levels a Source does not accept are no longer offered, so "ultra" can no longer fail a turn.

Vibey's built-in tools

• On by default. Web search, X search (with an xAI Source) and image generation are available as soon as a Source can run them. Turn each one off in Models › Tools.
• Search.
  • When the provider of Vibey's model offers its own search, Vibey uses it.
  • Otherwise Vibey searches through another of your Sources. The order is automatic, or set it under Advanced.
  • A proxied search runs on a light model the Source lists, such as gpt-6-luna or claude-haiku-5-5, with an older one as fallback.
  • If a Source fails or doesn't have the model, the next one is tried. An overloaded provider is reported as unavailable, not as refusing.
  • If the provider refuses its own search, Vibey switches to another Source and says so: <Source> doesn't support it, now using <provider>. You can clear that in Models › Tools.
• Image generation. Vibey calls an image model, such as gpt-image-2, when it needs a picture. Choose the default model on the Tools tab. Images are saved as files in ~/.avibe/state/agent_core/media/ and linked from the reply; Vibey itself only sees the path.
• Long answers. An Anthropic answer that pauses mid-turn while its server tools run continues as the same answer (#2587).
• What you see. Each call shows as a card in the Web chat and as a line in chat apps. It says how the call ran, which sources it found, and an approximate cost. Citations become links, and cited answers read as one paragraph.
• Costs. Pay-as-you-go Sources are billed per call, estimated from the provider's list price; a relay's actual bill may differ. Subscriptions are not billed per call. Claude subscriptions cannot run these tools.

Reliability

• Silent streams. A streamed Anthropic or OpenAI Responses request that receives nothing from the engine for 120 s is ended and, if no output reached you yet, retried (#2538).
• Fewer stalls. The following no longer block the service, and Runs look up callbacks and heartbeats by key (one new database index, #2653):
  • scheduled turns, Harness callbacks and a run's storage work (#2650, #2654);
  • saving and reading the Model Hub config (#2648, #2612);
  • starting work while messages are being delivered (#2662).
• Source identity. A Source keeps its identity after you replace its key, and Vibey's provider-specific content stays bound to the Source that produced it (#2628, #2626). Replacing a key works even when the old credential file is damaged (#2672).
• Failure kinds are kept as they happened, instead of guessed (#2544).
• Lower long-context cost on OpenAI. On GPT-5.4 and later models, Vibey compacts a long conversation before it reaches 272K input tokens, where OpenAI bills the whole request at long-context rates, instead of near the 1M window (#2659).
• Saved API keys migrate without a connection test. Moving saved keys into Model Hub no longer checks each provider, so an expired key or an offline provider no longer blocks the others. New Sources show as Saved until they are verified (#2602).

Reasoning effort

• Efforts a client maps to something else, such as Codex's "ultra", are no longer offered as levels Vibey sends (#2551).
• Each Source model's levels have one owner. They are derived again at startup, so dated model ids such as gpt-5.6-sol-2026-09-01 get their model's levels, and Vibey's own model list follows (#2583, #2632).
• Anthropic's xhigh is sent natively where the Source supports it (#2575).
• The picker leaves out levels every Source refuses, names the Sources that refuse the others, and keeps a refused saved level visible (#2590, #2594).

Other changes

• Desktop: Computer Use runs JavaScript scripts with vibe computer run (#2633), and its driver exits after its parent is lost (#2614). Avibe Mini talks to a per-person Vibey conversation, with voice input (#2546). Show pages work again in the macOS app, whose install path contains a space (#2597).
• Vaults: multi-field secrets (#2563); search, tags and compact rows (#2553); a compact field editor (#2621). An agent deleting a secret needs your approval (#2554).
• Discord: messages only notify the people they name (#2675).
• Claude Code: checking its sign-in status in Settings no longer launches Claude or refreshes its credentials (#2617).
• Show pages: the public vendor-asset path rejects .. segments, so it cannot reach other Show Runtime routes (#2600).
• Chat apps: a message delivered only in part counts as undelivered (#2601).
• Pull request watches: an agent watching a pull request can wait for both a passing review and green CI and be told once (#2663).

Upgrading from 3.2.0

• The first start runs two additive database migrations, 20261013_0065 (multi-field Vault secrets) and 20261015_0066 (a callback index), after a verified backup of your state database. On an SSD, a 1.9 GB database took about 21 s in our test.
• Reasoning levels are derived again once at startup. Vibey's built-in tools start on.

Known issues

• xAI reasoning context is not yet carried between requests (#2620).
• Grok Imagine is not supported for image generation.
• Images on a Codex subscription are off.
• Retry-After from an upstream is not honored yet (#2666).
• An Anthropic relay on a custom address that accepts only x-api-key verifies when you add it, but its calls fail (#2667).

Full Changelog: v3.2.0...v3.2.1

---

Avibe 3.2.1

Avibe 3.2.1 让 Vibey 内置了网页搜索、X 搜索和图片生成，服务在负载下更稳，推理档位也与每个模型来源匹配。从 3.2.0 升级会执行两次小的数据库迁移，详见下方 从 3.2.0 升级。

亮点

• Vibey 能搜网页、搜 X、生成图片了。 这些内置工具默认开启。模型厂商自带搜索时 Vibey 直接用，否则改用你的其他模型来源；图片由你选定的模型生成。
• 负载下更稳。 好几处数秒级的服务卡顿已经消除；上游流长时间没有输出时，Model Hub 会中断它，并在输出开始前重试。
• 推理档位与模型来源匹配。 不再提供来源不接受的档位，「ultra」不会再让一轮对话失败。

Vibey 的内置工具

• 默认开启。 只要有模型来源能运行，就能用网页搜索、X 搜索（需要 xAI 来源）和图片生成。可以在 模型 › 工具 里逐个关闭。
• 搜索。
  • Vibey 所用模型的厂商自带搜索时，直接用它。
  • 否则通过你的另一个模型来源搜索。执行顺序默认自动，也可以在 高级 里指定。
  • 代理搜索使用来源列出的轻量模型，例如 gpt-6-luna 或 claude-haiku-5-5，并以更早的型号作为备选。
  • 某个来源失败或没有这个模型时会换下一个。厂商过载显示为暂时不可用，而不是拒绝。
  • 如果厂商拒绝自带搜索，Vibey 会改用其他来源，并注明「<来源> 不支持，已改用 <厂商>」，可以在「模型 › 工具」里清除。
• 图片生成。 Vibey 需要图片时会调用图片模型（例如 gpt-image-2），默认模型在「工具」页选择。图片存为 ~/.avibe/state/agent_core/media/ 下的文件，并在回复里给出链接；Vibey 自己只看到路径。
• 长回答。 Anthropic 的回答在服务端工具运行时中途暂停，会作为同一个回答继续下去（#2587）。
• 能看到什么。 每次调用在 Web 聊天里显示为一张卡片，在聊天应用里显示为一行，写明调用方式、找到的来源和大致费用。引用会变成链接，带引用的回答连成一段。
• 费用。 按量付费的来源每次调用都计费，按厂商 API 标价估算，中转的实际账单可能不同。订阅不按次计费。Claude 订阅不能运行这些工具。

稳定性

• 静默的流。 发往 Anthropic 或 OpenAI Responses 的流式请求，如果 120 秒内收不到引擎的任何数据就会被中断；如果还没有输出给你，会自动重试（#2538）。
• 卡顿更少。 以下操作不再阻塞服务；Run 按键查找回调和心跳（新增一个数据库索引，#2653）：
  • 定时对话、Harness 回调、运行的存储操作（#2650、#2654）；
  • 保存和读取 Model Hub 配置（#2648、#2612）；
  • 消息投递时启动新工作（#2662）。
• 模型来源身份。 更换 Key 后模型来源的身份不变，Vibey 的厂商专有内容绑定在产生它的来源上（#2628、#2626）。旧凭据文件损坏时也能替换 Key（#2672）。
• 失败类型 按实际情况保留，不再猜测（#2544）。
• OpenAI 长上下文更省钱。 在 GPT-5.4 及更新的模型上，Vibey 会在对话达到 272K 输入 token 之前压缩上下文（超过这个量 OpenAI 会按长上下文价格计费整次请求），而不是等到接近 1M 窗口（#2659）。
• 已保存的 API Key 迁移时不再做连接测试。 把已保存的 Key 迁入 Model Hub 时不再逐个检查厂商，过期的 Key 或离线的厂商不会再挡住其他 Key。新的来源在验证前显示为「已保存」（#2602）。

推理档位

• 客户端会映射成别的档位（例如 Codex 的「ultra」），不再作为 Vibey 发送的档位提供（#2551）。
• 每个模型来源的档位只有一个归属，启动时重新推导。带日期的模型 ID（如 gpt-5.6-sol-2026-09-01）会拿到所属模型的档位，Vibey 自己的模型列表也会跟着更新（#2583、#2632）。
• 在支持的来源上，Anthropic 的 xhigh 按原生方式发送（#2575）。
• 档位选择器会去掉所有来源都拒绝的档位，标出拒绝其他档位的来源，已保存但被拒绝的档位仍然可见（#2590、#2594）。

其他变更

• 桌面： Computer Use 可以用 vibe computer run 运行 JavaScript 脚本（#2633），父进程退出后它的驱动也会退出（#2614）。Avibe Mini 对接每个人自己的 Vibey 会话，支持语音输入（#2546）。macOS 应用的安装路径带空格，Show 页面现在可以正常使用了（#2597）。
• Vault： 支持多字段密钥（#2563）；新增搜索、标签和紧凑列表（#2553）；字段编辑器改为紧凑列表（#2621）。Agent 删除密钥需要你批准（#2554）。
• Discord： 消息只提醒其中点名的人（#2675）。
• Claude Code： 在设置里查看登录状态不会再启动 Claude 或刷新它的凭据（#2617）。
• Show 页面： 公开的 vendor 资源路径会拒绝 .. 路径段，不能再借此访问 Show Runtime 的其他路由（#2600）。
• 聊天应用： 只投递了一部分的消息按未投递处理（#2601）。
• PR 监视： Agent 监视 PR 时，可以等评审通过和 CI 全绿都到齐后只收到一次通知（#2663）。

从 3.2.0 升级

• 首次启动会在完整备份并校验状态数据库后，执行两次只增不改的数据库迁移：20261013_0065（Vault 多字段密钥）和 20261015_0066（回调索引）。我们的测试中，SSD 上 1.9 GB 的数据库约需 21 秒。
• 推理档位会在启动时重新推导一次；Vibey 的内置工具默认开启。

已知问题

• xAI 的推理上下文还不能在请求之间延续（#2620）。
• 图片生成暂不支持 Grok Imagine。
• Codex 订阅的图片生成已关闭。
• 暂不遵循上游返回的 Retry-After（#2666）。
• 只接受 x-api-key 的自定义地址 Anthropic 中转：添加时能通过验证，但实际调用会失败（#2667）。

Full Changelog: v3.2.0...v3.2.1