面向开源贡献的 AI Agent 桌面工作台 AI Agent desktop workbench for open-source contribution

CodeInsights

把开源贡献从“让模型试试”变成可计划、可审查、可验证的桌面工作流。Pipeline 安排阶段和证据，Agent 在本地工作区执行，JSONL 留下可回放记录。 Turn “ask a model to try” into planned, reviewable, verifiable contribution work.Pipeline structures the stages, Agent executes in your workspace, and JSONL keeps the replay trail.

GitHub 查看真实运行 Watch Real Run

01 / Plan 先形成方案 Shape the plan

02 / Run 再让 Agent 执行 Let agents work

03 / Prove 最后留下证据 Keep evidence

Shift

从“模型给答案”转向“过程可批准” From model answers to approvable process

CodeInsights 把任务拆成计划、执行和证据三层，让每一步都能被人接管、暂停、复查或恢复。 CodeInsights splits work into planning, execution, and evidence so each step can be paused, reviewed, resumed, or taken over by a human.

Problem

复杂修改容易失控 Complex edits drift easily

直接让 Agent 改仓库，很容易把风险、假设和验证缺口混进一段长输出里。 Letting an agent edit directly can bury risk, assumptions, and missing verification inside one long answer.

Control

把风险变成检查点 Turn risk into checkpoints

Pipeline 在计划、权限、AskUser 和人工 gate 处暂停，Agent 只在明确边界内执行。 Pipeline pauses at planning, permissions, AskUser, and human gates while Agent executes inside explicit boundaries.

Evidence

保留可回放证据 Keep replayable evidence

会话、事件、checkpoint、artifacts 和工作区文件优先落到本机，方便恢复、对比和排障。 Sessions, events, checkpoints, artifacts, and workspace files stay local first for recovery, comparison, and debugging.

plan.md review.md test evidence JSONL events checkpoint

真实产品界面 Real product surface

先看实际工作台，再看它背后的流程 See the workbench first, then the workflow behind it

以下素材来自真实 Electron 开发窗口，采集时使用隔离配置目录，避免读取本机已有渠道、会话和凭证。 These captures come from a real Electron development window using an isolated config directory to avoid local channels, sessions, and credentials.

6 秒真实录屏 6-second real run

Pipeline、Agent、设置与回到 Pipeline 的实际桌面流 A real desktop pass through Pipeline, Agent, settings, and back to Pipeline

首页首屏使用真实 Pipeline 截图建立产品信号，这段录屏提供更完整的界面节奏。 The hero uses a real Pipeline screenshot for product signal; this video shows the broader interface rhythm.

Pipeline 阶段轨道 Pipeline stage rail
Agent 工作区 Agent workspace
模型与 MCP 配置 Model and MCP settings

CodeInsights Pipeline desktop workbench — **Pipeline 工作台 Pipeline workbench** 六阶段路线、人工 gate、产物和运行日志集中展示。 Six-stage route, human gates, artifacts, and execution logs in one surface.

CodeInsights Agent workspace — **Agent 工作区 Agent workspace** 会话、命令入口、资源面板和工作区文件保持在同一流程中。 Sessions, command entry, resource panel, and workspace files stay in the same flow.

CodeInsights model and Codex settings — **模型配置 Model configuration** 多 Provider 渠道、Agent 供应商和 Pipeline Codex 认证来源集中配置。 Provider channels, Agent vendors, and Pipeline Codex auth sources are configured together.

CodeInsights Agent MCP and Skills settings — **MCP / Skills MCP / Skills** 工作区隔离的 MCP Server、Skills 和内置工具配置。 Workspace-scoped MCP servers, Skills, and built-in tool settings.

Pipeline v2 overview: Explorer and Planner lead into Codex-backed Developer, Reviewer, Tester, and Committer stages with human review gates. — Pipeline v2 starts with Explorer and Planner, pauses for human review, then routes implementation, review, testing, and submission drafting through Codex-backed stages. Rejection and blocked-test paths return to earlier stages instead of silently continuing.

发现与计划 Discover and plan

Explorer 阅读目标与仓库上下文，Planner 输出计划、风险、验证路径和可审查文档。 Explorer reads the goal and repository context; Planner produces the plan, risks, verification path, and reviewable documents.

实现与审查 Build and review

Developer 通过 Codex SDK / CLI 修改工作区，Reviewer 只读审查 diff、回归、安全和测试缺口。 Developer changes the workspace through Codex SDK / CLI; Reviewer checks diffs, regressions, security, and test gaps read-only.

验证与提交材料 Verify and draft

Tester 运行验证命令并收集证据，Committer 生成提交信息、PR 标题和正文草稿。 Tester runs verification commands and collects evidence; Committer drafts commit messages, PR titles, and PR bodies.

全局监听，不丢后台事件 Global listeners preserve background events

Agent IPC 监听器挂在 Renderer 顶层，页面切换时流式输出、权限请求和 AskUser 仍会进入 Jotai 状态。 Agent IPC listeners mount at the renderer root, so streams, permission requests, and AskUser prompts still enter Jotai state across view changes.

工作区隔离的 MCP / Skills Workspace-scoped MCP / Skills

每个工作区有独立 MCP 配置、Skills、workspace files 和 session cwd，适合长期任务与可审计自动化。 Each workspace has isolated MCP config, Skills, workspace files, and session cwd for long-running auditable automation.

Agent 兼容渠道边界清晰 Clear Agent-compatible provider boundary

Agent 模式当前面向 Anthropic Messages 兼容渠道：Anthropic、DeepSeek、Kimi API 和 Kimi Coding。 Agent mode targets Anthropic Messages compatible providers: Anthropic, DeepSeek, Kimi API, and Kimi Coding.

Agent Runtime overview: Renderer, preload, main IPC services, Agent runner, Claude Agent SDK compatible stream, JSONL persistence, and Jotai state. — Agent Runtime requests move from the Renderer through the preload bridge into main-process IPC handlers, Agent service, and runtime runner. Streamed SDK events are persisted as JSONL and reflected back into renderer state through global listeners.

System architecture overview: Electron main process, preload bridge, renderer UI, shared packages, Agent, Pipeline, Bridge, and local storage services. — **系统架构 System architecture** Electron 主进程负责服务和运行时，Preload 暴露安全 API，Renderer 通过 Jotai 展示 Agent、Pipeline、Bridge 和本地存储状态。 Electron main owns services and runtimes, preload exposes safe APIs, and renderer uses Jotai to display Agent, Pipeline, Bridge, and local-storage state.

Local storage overview: JSON configuration, JSONL conversations and events, attachments, workspaces, checkpoints, artifacts, and SDK config. — **本地存储 Local storage** JSON 保存配置和索引，JSONL 保存会话与事件，工作区、checkpoint、artifacts 和 SDK 配置都有明确目录。 JSON stores config and indexes; JSONL stores sessions and events; workspaces, checkpoints, artifacts, and SDK config have explicit directories.

IPC and state flow overview: shared contracts, main-process handlers, preload bridge, renderer APIs, global listeners, and Jotai atoms. — **IPC 与状态流 IPC and state flow** 共享类型定义通道契约，主进程处理 IPC，Preload 桥接白名单 API，Renderer 全局监听器写入 Jotai atoms。 Shared types define channel contracts, main handles IPC, preload bridges allowlisted APIs, and renderer global listeners write into Jotai atoms.

Model Layer

Provider Adapters

Anthropic、OpenAI、DeepSeek、Google、Kimi、智谱、MiniMax、豆包、通义千问和自定义 OpenAI 兼容端点。 Anthropic, OpenAI, DeepSeek, Google, Kimi, Zhipu, MiniMax, Doubao, Qwen, and custom OpenAI-compatible endpoints.

Review Layer

Human Gates

工具权限、AskUser、ExitPlan、Pipeline gate 都按 session 隔离，后台会话不丢请求。 Tool permissions, AskUser, ExitPlan, and Pipeline gates are isolated per session, preserving background requests.

Bridge Layer

IM Bridge

飞书、钉钉、微信 Bridge 可以把远程消息转成本地桌面 Agent / Chat 会话请求。 Feishu, DingTalk, and WeChat bridges can turn remote messages into local desktop Agent / Chat session requests.

Context Layer

Documents

附件、PDF、Office、文本解析和文件预览贯穿 Chat 回退、Agent 工作区和任务资料。 Attachments, PDF, Office, text parsing, and file preview support Chat fallback, Agent workspaces, and task context.

Runtime Layer

Runtime Checks

Bun、Git、Node、WSL、Shell 环境、代理和自动更新由主进程统一检测与管理。 Bun, Git, Node, WSL, shell environment, proxy settings, and updater state are checked and managed from the main process.

Safety Layer

Security Boundary

Electron contextIsolation、preload 白名单、safeStorage 加密和 Codex Git guard 共同约束高风险操作。 Electron contextIsolation, preload allowlists, safeStorage encryption, and Codex Git guards constrain high-risk operations.

准备环境 Prepare runtime

Bun 1.2.5+ · Git · macOS / Windows / Linux

配置渠道 Configure channels

在设置中创建 Agent 兼容渠道，并按需选择 Pipeline Codex 渠道或本机 Codex auth。 Create an Agent-compatible channel, then choose a Pipeline Codex channel or local Codex auth as needed.

创建工作区 Create a workspace

按工作区隔离 MCP Server、Skills、workspace files 和会话工作目录。 Scope MCP servers, Skills, workspace files, and session cwd per workspace.

bash GitHub

本地开发入口 Local development entry

git clone https://github.com/zcxGGmu/CodeInsights.git
cd CodeInsights

bun install
bun run dev

# build and launch after compilation
bun run electron:build
bun run electron:start