Repository navigation
Releases: CCDawn/Vibelution
Release list
v2.0.0
Release notes (English)
Vibelution is a local-first workbench where AI agents work as a team: an AI research team that plans, builds and reviews together, a finance research agent, and virtual humans with a life of their own — all in an Electron desktop app. This release is version 2.0.0 (the desktop shell package itself remains 1.0.16) and covers the user-visible changes since 1.1.2.
Upgrade note (breaking): persisted LLM operator configuration must be schema v2. Recognizable schema v1 files — including role_bindings — are upgraded in place by a one-shot upgrader when the app first reads them. Files that cannot be upgraded are left unchanged and are not read at runtime, so check your model provider settings after updating. Rolling back to the 1.1.2 label does not rewrite a config that has already been upgraded.
Finance agent: a workbench of its own (aaa0908)
The finance agent (炒股智能体) now opens from the specialist menu or /finance into its own desktop workspace, with dedicated navigation, session tabs and a native chat layout. The first visit creates a general agent, an empty report library and a session — without calling a model. Public news shown in that chat is reference only and is never filed into your library. The assistant does not collect accounts, holdings or amounts, and it cannot place orders.
Virtual humans: a life, a mood, a memory (1ec76d4, d9a15e0)
Any agent can bind the virtual-human life plugin. Only an explicitly enabled binding gets a daily schedule, moods, memory and proactive messages — the companion initiates conversation instead of only answering. Agents without the binding are completely unchanged.
Research teams that keep moving (6810399, 50ba2c1)
Long-running team research got two robustness upgrades. A stage failure planner now reviews self-blocked source-collection tasks with the LLM and proposes how to unblock them, instead of letting a stage stall silently. The meeting fence was fixed end to end: opening-round dispatch evidence is persisted, running bound rounds are no longer closed as orphans, and dev cold-start challenges are recognized correctly.
LLM provider configuration: schema v2 (06223a8, 8f7cf2e)
Model service configuration was fixed and simplified across all providers, and it is the reason for the schema v2 breaking change above: one consistent configuration shape, with v1 files upgraded in place where possible. Provider usage records are now normalized through a single path, piloted in Rust with a Python implementation kept at parity (8f7cf2e).
Chat and assistant polish (fb19182, 32f91b8, 2ecd924)
- The assistant now makes it clear that a model pinned for one session coexists with the default model, so per-session pins no longer look like a configuration mistake.
- Agent log diagnosis leads with the scene summary, how many times it repeated and the primary reason, and names the evidence files to open. Nested provider errors are classified properly and stale protocol markers are downgraded instead of crying wolf.
- The screenshot tool respects restricted provider access.
Launcher and desktop stability (f2292e0, 043872b, c65f5ab)
The Linux deployment bootstrap is hardened, the sandbox auto-selects the native Codex backend when available, and the launcher no longer waits on non-essential work during shutdown — closing the app is snappier and more reliable.
Faster frontend (1b4fa20)
The canvas engine and the research process pack now load on demand, trimming what the workbench downloads and parses up front.
Removed compatibility routes (8022371)
The workbench no longer redirects the legacy paths /chat-rooms, /agents/teams, /agents/memory/* and /evolution. Use /chat, /teams, /memory/*, /supervised-evolution or /self-evolution instead. Chat panel widths are now stored only in paneLayouts.chat; a leftover shell.chatPanelWidths value is copied once and then dropped.
Release notes（中文）
Vibelution 是一个本地优先的多智能体工作台：AI 科研团队负责规划、开发与评审，炒股智能体负责金融研究，虚拟人陪你日常相处——全部运行在 Electron 桌面端。本版本为 2.0.0（桌面外壳包版本仍为 1.0.16），收录自 1.1.2 以来所有用户可感知的变化。
升级提示（破坏性变更）： 持久化的 LLM 运营配置必须是 schema v2。可识别的 schema v1 文件（含 role_bindings）会在首次读取时由一次性升级器就地升级；无法升级的文件保持原样、运行时不再读取，因此升级后请检查一次模型服务商设置。已升级到 schema v2 的配置，回滚到 1.1.2 也不会被改写。
炒股智能体：有了自己的工作台（aaa0908f9）
炒股智能体现在从专项菜单或 /finance 进入自己的桌面工作区，带独立的导航、会话标签页和原生聊天布局。首次进入会创建通用 Agent、空的报告库和一个会话，全程不调用模型。聊天中展示的公开新闻仅作参考，不会归档进报告库；助理不收集账号、持仓和金额，也不能下单。
虚拟人：有生活、有情绪、有记忆（1ec76d487, d9a15e0）
任何 Agent 都可以绑定虚拟人生插件。只有显式启用的绑定才会拥有作息、情绪、记忆和主动消息——虚拟人会主动找你聊天，而不是只等你开口。未绑定的 Agent 完全不受影响。
研究团队：持续推进不卡壳（681039994, 50ba2c1）
长时间运行的团队研究加了两道保险。阶段失败规划器会用 LLM 复核自行受阻的资料收集任务并提出解法，而不是让阶段默默卡死；meeting fence 的三处缺陷也一次修齐：开场轮的派发证据会持久化、运行中的绑定轮次不再被当作孤儿关闭、开发冷启动的挑战身份能被正确识别。
LLM 服务商配置：schema v2（06223a8c9, 8f7cf2e）
模型服务配置在各服务商之间做了修复与简化，也是上文 schema v2 破坏性变更的来源：统一为一致的配置结构，v1 文件能升则就地升。各服务商的用量记录现在走同一条归一化链路，并以 Rust 实现试点、与 Python 实现保持结果一致（8f7cf2e9e）。
聊天与助理体验（fb1918232, 32f91b8, 2ecd924）
- 助理现在会明确提示：会话内固定的模型与默认模型是并存的，按会话固定不再像配置出错。
- Agent 日志诊断改为先给场景摘要、重复次数和首要原因，并点名应打开的证据文件。嵌套的服务商错误被正确分类，过期的协议标记会被降级，不再误报。
- 截图工具遵守受限的服务商访问边界。
启动器与桌面稳定性（f2292e0df, 043872b, c65f5ab）
Linux 部署引导得到加固；沙箱在可用时自动选择原生 Codex 后端；启动器关闭不再等待非关键路径——退出更快也更可靠。
前端更快（1b4fa20f6）
canvas 引擎与研究流程包改为按需加载，减少了工作台启动时需要下载和解析的内容。
移除兼容路由（80223713a）
工作台不再重定向旧路径 /chat-rooms、/agents/teams、/agents/memory/* 和 /evolution，请改用 /chat、/teams、/memory/*、/supervised-evolution 或 /self-evolution。聊天面板宽度只存储在 paneLayouts.chat；遗留的 shell.chatPanelWidths 值会被复制一次后弃用。
v1.1.2
v1.1.1
v1.1.0
1.1.0
Public baseline after the post-1.0.16 workbench stretch.
Highlights
- Teams research home: flow strip + organization canvas
- Chat projection / paint hardening; per-session scroll memory when switching threads
- Shell + Launcher: stable beforeunload (no status-poll rebind flash); intentional refresh one-shot pass
- Challenge Cup ops workspace lazy-loaded out of the eager Teams graph
- Open-source packaging (MIT, CONTRIBUTING, SECURITY) and plainer README
- Version narrative aligned: VERSION / CHANGELOG / web package / AppShell pill
See CHANGELOG.md for the full note.