Agent 能力配置 · OpenProgram Docs

OpenProgram Docs

Get started Install Capabilities Interfaces Models Integrations Server & Ops Reference Design

Overview

Design Documents

UI · Foundations

Web UI OpenProgram · macOS App 图标规范 Cross-module UI invariants Surface system (dark mode) Web state layer: per-session stores OpenProgram 主题系统 Web style organization Interaction feedback — the 0ms rule Indicator dot system 布局密度设计 v4 — 左栏结构 / 纵向空间 / 真浏览器

UI · Chat and composer

OpenProgram · 附件处理设计 (viz) Attachment Handling Design (Web Chat) 聊天附件：双向流转与预览聊天轮次视觉规范 — 执行时间线 + 消息导航 Composer interaction modes — the input box as the unified catch point for "user decisions" Composer local paths OpenProgram · Composer 底部控制栏响应式设计 OpenProgram · Composer Tool Profile 菜单 fn-form compact — 设计示例 GUI agent — call structure & context flow OpenProgram · 聊天发送队列可靠性 WebSocket 命令与交互等待生命周期

UI · Browser and tabs

OpenProgram 内置浏览器扩展设计 OpenProgram 内置浏览器设计 Center Tabs 与复合分屏 Tab 设计 OpenProgram 远程 Web 访问设计 (viz) Remote Web Access — Owner-Only, Self-Hosted Control 内置浏览器主页按钮内置浏览器原生页面边界

UI · Settings and catalog

OpenProgram · Avatar randomization Programs source categories Settings 二级导航栏头

UI · Workspace and sidebar

OpenProgram 内置终端与 Claude Code OpenProgram 浏览器式界面原型 v3 Project workspace — files, tabs, and multi-session OpenProgram 右侧文件树层级设计左侧栏层级设计 mock

Overview

Community discoverability design Design Docs Site (Unified Documentation Website) Error Handling — Discipline and Gate IndexNow deployment discovery IndexNow implementation ledger Open Items Where Design Docs Are Out of Sync With Code OpenProgram implementation status and handoff Repository structure Repository structure implementation ledger Site discoverability and performance design TODO：待讨论的改进项 Usage Metering Subsystem Design Framework Overview — One Conversation from Input to Output 整框架对标：我们和十二家放在一起看功能清单对标：设计参考审计快照

Channels

Channels Channel Subsystem Design Channel Subsystem — Requirements and Invariants

CLI / TUI

CLI / TUI CLI / TUI Configuration Surface — Design CLI Naming Convention Config Write Safety — Design Single-Port Architecture Slash Commands — Reference Implementations Snapshot Slash Commands — Unified Design TUI Transcript Rendering and Interaction Unified `@agentic_function` Execution Path Web UI Port — Configuration and Conflict Handling

Context

Context Context — The Context Layer Context Compaction Context Composition — Registry-Based Three Layers Context Composition Comparison — Reference Projects vs. Us (organized by the three layers) 上下文压缩记忆的自我认知：agent怎么知道自己记得什么

Distribution

OpenProgram 安装、打包、发布与升级设计 OpenProgram 正式版本自动更新 Installation and distribution implementation plan

Extension Gating

Extension Gating Future work — explicitly NOT built Implementation map Reference-implementation comparison

Function Design

Function Design Function calling Agentic Program：工具、技能、agentic function是同一个概念

Function Design · Calling

@agentic_function Calling Sub-Functions in a Fixed Order

Integrations

Integrations Editor integration: LSP tools and ACP server Harness Standard — how an agentic program plugs into OpenProgram MCP Tool Integration Skills & Plugins Web Use 实施与验证记录 Web Use 技术规格 Web Use 设计概览 MCP 服务端：现状、九家对标与最小能力集

Memory — Memory System Design

Memory — Memory System Design Memory subsystem Knowing which turns memory has written 哪些轮次已经写进记忆 (viz) 记忆子系统：机制图解记忆写入：八个框架横向对比对标之后抄什么：四条做法的适配方案 Authority: Two Tiers and Pairing Entity Memory Entity Memory on Git: Session-Git + Project-Git OpenProgram Memory — UI proposal Virtual Memory 发言人身份：记忆里谁说的权限与归因：我们做了什么、别人做了什么、接下来做什么

Plans

Cache-Control Passthrough — Design Credential Connection Unification MCP Server Implementation Plan Wiring the Proactive Layer into the Code — Design

Proactive Layer — Design

Proactive Layer — Design Commitments migration Commitments 迁移说明 (viz) Event Layer Events and State (The Foundation) Framework Signal Routing: One Bus as the Hub How Rules Are Written and How They Run Invariants the Framework Must Enforce Itself OpenProgram · 事件层 API Reference OpenProgram · 事件层设计 (viz) OpenProgram · 框架演进 (viz) Overview: Walk Through One Scenario End to End Three Template Rules

Proactive Layer — Design · Research Archive

回放即 policy 测试威胁模型评估

Providers

Providers Model list and the Bailian provider OpenProgram · JSON Schema Structured Output 设计 OpenProgram · Provider metadata 加载诊断 OpenProgram · Provider 请求录制与严格离线回放 (viz) OpenProgram · 报错与超时机制（现状） Outbound network proxy Recording and replaying provider calls Request building: Context → per-provider parameters

Providers · Auth

Account management and pool rotation API-key / credential resolution claude-code direct subscription connection Credential connection unification (one Credential, one resolution point) Credential status — usable or stopped Credential validation OpenProgram · 已禁用的文件凭据加固设计 Self-contained auth storage

Providers · Models

Fast tier — detection, storage, and wires Model Catalog and Provider Configuration Thinking / Effort Subsystem Design

Providers · Reliability

Error Handling and Retry Decision Logic for Model Calls Error-taxonomy propagation — structured LLM errors up to the UI LLM-call fault tolerance & timeout management OpenProgram · 报错与超时机制（现状）

Research

Execution Trace Data Model — Choosing Span

Runtime

Runtime Additional working directories for a session Agent 基础配置 Agent 工具配置 — Tools & Programs Agent 能力配置 Agent 运行配额与任务生命周期治理 Agent 配置框架 Controllability of long-running agent tasks + three-surface consistency llm 与 agent：编排层的两个原语 P3 — Three-Surface Sync (filling gaps) + P1/P2 onto all three surfaces Permission System Design Runtime outbound URL and SSRF protection Runtime — Design Rationale Self-Programming Agentic Workflow — 设计总览 Self-Programming Workflow — DAG 与管理网页 Self-Programming Workflow — 创建、隔离与发布 Self-Programming Workflow — 编程模型与组合 Session goals — the /goal continuation loop Source checkout upgrades Tool toggles and toolset management Unified session context creation Workflow Control Primitives — 控制流原语设计 Sandbox 沙箱：现状、八家对标与完整修复顺序 Agent collaboration: one cross-branch communication primitive Agent协作：工具面 Agent协作：八家实现对照

Runtime · Dag

Branch Collaboration (Communication · Service · Merge) DAG Rendering Spec (Layout · Edges · Legend · Default Visibility) DAG 实时布局（真算法） DAG 布局视觉规范 DAG 持久化失败可观察性 Session DAG — Design 外部 Program 运行节点 · DAG 与聊天窗口设计

Runtime · Execution

`agent/dispatcher` — a responsibility-scoped package Agent Call Flow (Authoritative Design) Agent Worktree Tool Agentic 函数防自递归机制 — 设计文档 (viz) Async Task Lifecycle Next-step decision making(decision.make / exec(choices=)) Self-Recursion Guard for Agentic Functions Turn Cancellation — Design

Runtime · Operations

Automatic Branch Naming — Design DAG Edge Field Naming File Modification Management — Industry Analysis and OpenProgram Design Multi-agent branches and per-turn revert Persisting In-Flight State / Reconnect Recovery Rewind — Roll Back to Any Historical Message Unifying parent_id and called_by — Design User input requests: pausing a run to ask the user 分支自动命名设计 (viz)

Runtime · Session Subsystem

Session Subsystem Distill — sessions into reusable procedures LLM Title Generation OpenProgram · merged_heads 原子更新 OpenProgram · SessionStore 锁范围 Session Context Session Data Model Session Management Comparison Session Operations Session 注册表时间语义与并发一致性

Scheduler

OpenProgram Scheduler 与 Memory 接口

Testing

OpenProgram 测试系统

Design › Runtime › Agent 能力配置

Last updated · 2026-08-17

PreviousAgent 工具配置 — Tools & Programs NextAgent 运行配额与任务生命周期治理