Avibe v3.0.8

Highlights

• Agent work is now easier to track and harder to lose. The new interactive run graph shows live and finished sessions, parent/child lineage, and Task/Watch triggers, with search and in-graph details. Claude Code background work that would otherwise die with its CLI session is routed through the Avibe Harness, while orphaned runs and stale watch workers are recovered or settled instead of remaining active forever. (#956, #986, #1018, #1000, #1002, #1005)
• You can now annotate Show Pages and send that context back to agents. Authenticated annotation controls, screenshot attachments, reverse marks, replies, and read receipts support a complete review loop from the Workbench or Show Page. Annotation messages now use plain language and recover safely after interrupted dispatch. (#948, #949, #955, #960, #964, #1015, #1026, #1039, #1052)
• Project sessions are easier to organize. Pin important sessions to the top, hide work to the background from desktop or mobile, and use foreground/background visibility without changing a session’s scope or working directory. Updates sync across clients. (#957, #970, #984, #988, #1038)
• Model Hub is available as an explicit opt-in. When enabled with VIBE_MODEL_HUB_ENABLED, it can migrate native backend configuration, connect OAuth or API-key sources, resolve models per turn, and run its managed routing engine on macOS and Linux across x64 and arm64. It remains off by default, and its APIs and UI stay unavailable until enabled. (#962, #963, #966, #975, #976, #977, #987, #995, #997, #998, #999, #1019)

Changes

Added

• Added per-topic Telegram forum settings with group inheritance for access, mentions, agent routing, working directory, and message visibility. (#961)
• Added History API routing for newly scaffolded Show Pages, enabling direct links under private and public Show Page paths. (#980)
• Added Claude Opus 5 model metadata, including its 1M context label and full effort-level range. (#992)
• Added automatic startup into a delegated user cgroup scope so Avibe’s existing agent memory and process limits can take effect on supported Linux systems. (#968)
• Added --visible and --hidden CLI shortcuts for session visibility. (#969)

Changed

• Made Harness Run rows describe the work performed and link to their sessions; Task and Watch rows now lead with state and timing information. (#1023, #1033)
• Improved the Activity panel with clearer tool summaries, optional tool-call visibility, bounded expansion, raw JSON details, and distinct queued-run styling. (#946, #973, #985)
• Standardized agent-facing collection commands on pagination: 20 rows by default, --page and --limit, and a maximum of 100 rows per page. (#1013)
• Derived session titles from clean user content instead of decorated backend prompts, excluding injected identity, time, and attachment-error text. (#1032)
• Kept Model Hub’s mobile layout usable without changing its desktop layout. (#1003, #1008)
• Centralized message-type policy across Python, storage, and the Web UI to prevent transcript and activity behavior from drifting. (#1046, #1048, #1049, #1050)
• Updated internal release-note generation, CI coverage, workflow pinning, and message-policy documentation. (#1040, #1051, #1059)

Fixed

• Recovered ordinary Workbench messages stranded as pending after an interrupted UI dispatch without sending them twice. (#1034)
• Restored realtime Activity updates after terminal or hung event-stream connections, with bounded retries and immediate reconciliation on focus, visibility, or network recovery. (#947)
• Marked silent and reply-less completed turns as Done instead of Interrupted. (#951)
• Made Claude OAuth/API-key switching ignore inherited Anthropic credentials and refresh the running backend after authentication changes. (#1017)
• Made OpenCode provider probes use the same model precedence as real turns and return safe upstream diagnostics instead of a generic timeout. (#945)
• Removed stale Model Hub references when sources or custom models are deleted, and aligned displayed supply with the route the next turn will actually use. (#1016, #1029)
• Revalidated remote model catalogs against upstream cache timing while retaining conditional requests and last-known-good data. (#993, #996)
• Fixed direct Tailscale Web UI access with modern tailscale whois Node payloads. (#944)
• Kept iOS Home Screen launches inside the PWA during authentication and prevented remote phones from opening dead loopback browser sheets. (#972)
• Synchronized Telegram command menus for private chats and removed the obsolete discovery step from setup. (#989, #990)
• Refined session pin controls so desktop rows stay compact and mobile uses the overflow menu while retaining a passive pinned indicator. (#1004, #1010, #1012)
• Prevented stale source-session provenance from creating dead links in chat and tightened source-chip spacing and wording. (#967, #974, #1024)

Breaking

• Removed unpaginated --all collection output; use normal pagination with --include-finished or --include-disabled where applicable. (#1013)
• Removed duplicate task, tasks, watch, and watches JSON aliases from task/watch definition responses; use definition and definitions. (#1042)

Full Changelog: https://github.com/avibe-bot/avibe/compare/v3.0.7...v3.0.8

Avibe v3.0.8

亮点

• Agent 的执行过程更清楚，也更不容易丢。 新的运行图覆盖进行中和已结束的 Session，可查看父子关系、Task/Watch 触发来源，并支持搜索和图内详情。Claude Code 中原本会随 CLI 会话消失的后台工作现在会交给 Avibe Harness；失去执行者的 Run 和残留 Watch 进程也会被回收或正确结束，不再一直卡在运行中。 (#956, #986, #1018, #1000, #1002, #1005)
• 现在可以直接在 Show Page 上批注，并把上下文交给 Agent。 已加入登录态批注、截图附件、Agent 反向标记、回复和已读回执，Workbench 与 Show Page 之间可以完成一整轮审阅。批注消息改为面向用户的自然文本，派发中断后也能安全恢复。 (#948, #949, #955, #960, #964, #1015, #1026, #1039, #1052)
• 项目里的 Session 更好整理了。 常用 Session 可以置顶，也可以从桌面端或手机端移到后台；前台与后台状态不再影响 Scope 和工作目录，并会实时同步到其他客户端。 (#957, #970, #984, #988, #1038)
• Model Hub 现已提供显式启用入口。 设置 VIBE_MODEL_HUB_ENABLED 后，可迁移本机后端配置、接入 OAuth 或 API Key 来源、按 Turn 解析模型，并在 macOS、Linux 的 x64 和 arm64 主机上运行托管路由引擎。该功能默认关闭，启用前 API 和界面均不可用。 (#962, #963, #966, #975, #976, #977, #987, #995, #997, #998, #999, #1019)

变更

新增

• 新增 Telegram 论坛 Topic 独立设置，可继承群组配置，并覆盖访问权限、@ 规则、Agent 路由、工作目录和消息可见性。 (#961)
• 为新创建的 Show Page 加入 History API 路由，私有页面和公开分享页都支持直接打开深层链接。 (#980)
• 加入 Claude Opus 5 模型信息，包括 1M 上下文标识和完整的 effort 档位。 (#992)
• 在支持的 Linux 系统上自动进入已委派的用户 cgroup，让 Avibe 现有的 Agent 内存和进程数限制真正生效。 (#968)
• 为 Session 可见性新增 --visible 和 --hidden CLI 简写。 (#969)

调整

• Harness Run 列表现在直接说明执行内容并可跳转到对应 Session；Task 和 Watch 列表优先展示状态与时间。 (#1023, #1033)
• 改进 Activity 面板的工具调用摘要、显示开关、展开上限和原始 JSON 查看，并区分排队中与运行中的 Agent Run。 (#946, #973, #985)
• 统一面向 Agent 的集合命令分页规则：默认 20 条，支持 --page、--limit，单页最多 100 条。 (#1013)
• Session 标题改为从干净的用户消息生成，不再混入注入给后端的时间、身份信息和附件错误。 (#1032)
• 完善 Model Hub 手机端布局，同时保持桌面端布局不变。 (#1003, #1008)
• 统一 Python、存储层和 Web UI 的消息类型策略，避免会话记录与 Activity 展示规则各自漂移。 (#1046, #1048, #1049, #1050)
• 更新 Release Notes 生成模型、UI 测试 CI、工作流版本固定和消息策略文档。 (#1040, #1051, #1059)

修复

• 恢复因 Workbench 派发中断而卡在 pending 的普通消息，同时避免重复发送。 (#1034)
• 修复 Activity 实时连接终止或握手卡住后无法恢复的问题；页面重新可见、获得焦点或网络恢复时会立即重连并校准状态。 (#947)
• 将静默完成和无回复但正常完成的 Turn 显示为“完成”，不再误报为“已中断”。 (#951)
• 修复 Claude 在 OAuth 与 API Key 之间切换时继承旧 Anthropic 环境变量的问题，并在认证变更后刷新运行中的后端。 (#1017)
• OpenCode Provider 探测现在与真实 Turn 使用相同的模型优先级，并能返回经过脱敏的上游错误，不再只等到通用超时。 (#945)
• 删除来源或自定义模型时同步清理失效的 Model Hub 引用，并让界面显示的供给状态与下一个 Turn 的实际路由一致。 (#1016, #1029)
• 按上游缓存周期重新校验远程模型目录，同时保留条件请求和最后一次可用数据。 (#993, #996)
• 兼容新版 tailscale whois 的 Node 数据结构，恢复 Tailnet 内直连 Web UI。 (#944)
• 修复 iOS 主屏 PWA 在认证时意外弹出浏览器页面的问题，并阻止远程手机打开无效的本机回环地址。 (#972)
• 同步 Telegram 私聊范围内的命令菜单，并移除已经无用的发现引导步骤。 (#989, #990)
• 调整 Session 置顶控件，保持桌面列表紧凑；手机端通过更多菜单操作，同时保留只读的置顶标识。 (#1004, #1010, #1012)
• 避免已删除的来源 Session 在聊天中生成失效链接，并优化来源标签的文案与间距。 (#967, #974, #1024)

破坏性变更

• 集合命令不再支持不分页的 --all；请使用普通分页，并按需搭配 --include-finished 或 --include-disabled。 (#1013)
• Task/Watch 定义响应不再输出重复的 task、tasks、watch、watches JSON 别名，请改用 definition 和 definitions。 (#1042)

Full Changelog: https://github.com/avibe-bot/avibe/compare/v3.0.7...v3.0.8