工作区边界
所有文件操作限制在指定工作区内;范围之外的路径一律拒绝。
能力
每项能力对应执行过程中的一个环节:边界、授权、回退、上下文与结果呈现。
边界、授权与回退。
所有文件操作限制在指定工作区内;范围之外的路径一律拒绝。
读、写、执行分别授权,默认每次确认;可按操作类型配置自动放行。
危险命令直接拒绝,部分规则可改为「需要确认」;与权限审批叠加生效。
每轮保留快照,可退回任意一步继续,无需重开会话。
文件读取、改动与命令执行实时呈现;改动以行级 diff 展示。
上下文开销与响应速度。
每轮提示词有硬上限,超出时收缩旧工具结果,避免上下文无限膨胀。
旧历史按轮次摘要,不调用 LLM;输出稳定,也不额外消耗额度。
解析函数、类与类型并精确到行区间,比全文搜索更准,也更省 token。
保持请求前缀字节稳定,命中服务商的 prompt 缓存。
多供应商可选,单个提供商故障时自动转移;支持 Ollama / LM Studio 等本地模型。
结果集中呈现。
对话、文件树、终端与改动同屏,不需要在窗口之间来回切换。
代码、图片与报告集中呈现,可随时预览,结束时不会散落于临时目录。
每处改动可查、可还原;关键节点留有检查点。
PDF、Office 与图片可直接预览;扫描件由本地 OCR 处理。
Windows、macOS 与 Linux 使用同一套工作方式。
Features
Bi does not chase more features. It makes every step more effective, more visible and more reversible.
Draw the boundary before adding capability.
Every file operation stays inside the workspace you chose; paths outside it are refused.
Reads, writes and commands are authorized separately and confirmed by default, with configurable auto-approval per action type.
Dangerous commands are rejected outright; some rules can require confirmation instead. Rules and approvals stack.
A snapshot is kept each turn, so you can return to any step and continue instead of restarting the conversation.
Reads, edits and command runs appear in real time; changes are shown as line-level diffs.
Less Tokens. More Work.
A hard per-turn prompt cap; stale tool results contract when it is exceeded, so context cannot grow without bound.
Older history is summarized by turn without calling an LLM: stable output, no extra quota.
Functions, classes and types are indexed to line ranges — more precise than full-text search, and cheaper.
The request prefix stays byte-stable so provider prompt caches keep hitting.
Choose among providers with automatic failover, or point at local models such as Ollama and LM Studio.
Results land on one screen.
Conversation, file tree, terminal and diffs share one view. No window hopping.
Code, images and reports gather in one place, ready to preview — not scattered across temp folders.
Every change is inspectable and reversible; key points keep checkpoints.
PDF, Office and images preview in place; scans are processed by local OCR.
Windows, macOS and Linux share the same way of working.
Bi 的核心是一个显式循环:模型读取上下文、决定调用哪个工具、执行,再将结果回灌。权限与检查点附着在循环的每一步,无需事后补救。
Bi is built around an explicit loop: the model reads context, chooses a tool, executes, and feeds the result back. Permissions and checkpoints attach to every step, not as an afterthought.
