错误示例：deploy_and_monitor(app) —— 将判断逻辑捆绑在代码中 错误示例：classify_and_organize_files(files) —— 智能体应决定分类方式 正确示例：eve build create, eve build run, eve env deploy, eve job follow —— 智能体决定执行顺序

在 Eve 上：清单定义的是 WHAT（服务、流水线），智能体决定 HOW 和 WHEN 来组合它们。

设计测试：要改变行为，你是编辑文本（提示词/技能）还是重构代码？如果是代码——你的工具还不够原子化。

3. 可组合性 —— 新功能 = 新提示词

当工具是原子化的且存在对等性时，你通过编写提示词而非代码来添加能力。

Eve 示例：eve-pipelines-workflows 技能增加了流水线组合能力。无需新的 CLI 命令——该技能教会智能体组合现有的 eve pipeline 和 eve workflow 命令。

你的应用：如果添加功能需要新的 API 端点，你可能在捆绑逻辑。考虑现有的原语是否可以通过不同的方式组合。

4. 涌现能力 —— 智能体带来惊喜

构建原子化工具。智能体组合出意想不到的解决方案。你观察模式。优化常见模式。重复此过程。

Eve 示例：智能体组合 eve job create --parent + eve job dep add + 深度传播来构建任意的工作层次结构。平台并未规定这种做法——智能体从原子化原语中发现了它。

面向智能体应用的平台能力

除了四大原则，Eve 还提供（或正在构建）特定的原语，使得智能体应用的构建显著简化：

任务附件 —— 结构化上下文传递

在智能体之间传递计划、报告和见解，无需复杂的文件操作。将文本文档（markdown、JSON、YAML）附加到任务上。下游智能体从父任务读取附件。这解决了 80% 的“智能体传递结构化上下文”问题。

服务账户认证 —— 后端到 API 通信

使用 eve auth mint（当前）或服务账户（即将推出）来认证应用后端。每个有后端的应用都需要一个非用户令牌来进行 API 调用。作用域权限强制执行最小权限原则。

组织文档存储 —— 持久的智能体知识

智能体积累的知识可以超越单个任务：架构报告、风险评估、约定。基于数据库，支持全文搜索，并通过 PATCH 操作提供智能体原生的搜索/替换编辑功能。

网页聊天 —— 两种机制

根据你的需求选择：

• 网关提供者（机制 A）：通过 WebSocket 连接到 Eve 网关。零后端。最适合简单的聊天小部件和管理控制台。
• 后端代理（机制 B）：应用后端调用 POST /internal/orgs/:id/chat/route。完全控制增强、存储、路由。最适合生产级 SaaS。

决策：如果你的应用需要拦截、增强或存储对话 → 使用 B。否则 → 使用 A。

跨项目查询 —— 组织级智能

组合视图、仪表盘以及任何跨越多个项目的工具。组织级端点消除了 N+1 API 调用问题。

完整目录（包含 API 模式和优先级排名）请参见 references/eve-horizon-primitives.md。

Eve 特定的设计模式

文件作为通用接口

• 智能体知道 cat、grep、mv、mkdir
• 使用 .eve/manifest.yaml 作为单一事实来源 —— 智能体读取并编辑它
• 智能体配置存放在仓库文件中（agents.yaml、teams.yaml）—— 而非隐藏在数据库中
• 目录结构 = 信息架构：{entity_type}/{entity_id}/content
• 按用途命名文件：{entity}.json、{type}.md、agent_log.md

上下文注入（三层）

系统提示词应包含：

1. 可用资源 —— 存在什么，以及数量："/notes 中有 12 条笔记，3 个项目"
2. 能力 —— 智能体能做什么："创建、编辑、标记、删除笔记"
3. 近期活动 —— 发生了什么："用户 2 小时前创建了 '项目启动'"

Eve 将 EVE_API_URL、EVE_PROJECT_ID、EVE_ORG_ID、EVE_ENV_NAME 注入到每个环境中。技能提供领域词汇。

显式完成信号

• 任务返回带有 eve.status（"success"、"failed"、"waiting"）的 json-result
• 没有启发式完成检测 —— 始终使用显式信号
• 在任务级别通过阶段转换跟踪进度
• 支持 shouldContinue 用于多步骤操作
• 为中断恢复检查点状态

动态能力发现

• eve job list 发现可用工作
• eve agents list 发现可用智能体
• 技能系统在安装时自动发现能力
• 网关通过 slug 将消息路由到智能体 —— 新智能体立即可寻址
• 优先使用运行时发现（list_available_types()），而非静态的端点-工具映射

智能体到 UI 的通信

• 事件类型：thinking、toolCall、toolResult、textResponse、statusChange
• 没有静默操作 —— 变更立即可见
• 执行期间显示进度，而不仅仅是结果
• 考虑为嘈杂的内部操作使用 ephemeralToolCalls

成功检查清单

架构：

• 每个 UI 操作都有对应的 CLI/API 操作（对等性）
• 每个实体都有完整的 CRUD（完整性）
• 工具只做一件事；智能体决定组合方式（粒度化）
• 添加能力 = 添加技能/提示词，而非代码（可组合性）
• 智能体能处理你未明确设计的请求（涌现性）

实现：

• 清单和配置文件是事实来源（文件作为接口）
• 系统提示词包含资源、能力和近期活动（上下文注入）
• 完成是显式的，而非启发式的（带有 eve.status 的 json-result）
• 智能体操作立即反映在 UI 中（可见性）
• 动态发现适用于新能力（可扩展性）

如果构建带后端的应用：

• 配置服务账户认证以进行 API 访问
• 选择网页聊天机制（网关 vs 代理）
• 使用任务附件进行智能体间的结构化上下文传递
• 使用组织文档存储持久化知识（当可用时）

反模式

参考

完整平台原语目录（包含 API 模式、数据库设计和实现优先级排名）请参见 references/eve-horizon-primitives.md。

设计哲学来源：../eve-horizon/docs/ideas/agent-native-design.md 平台原语分析：../eve-horizon/docs/ideas/platform-primitives-for-agentic-apps.md

每周安装数

139

仓库

incept5/eve-skillpacks

首次出现

2026 年 2 月 12 日

安全审计

Gen Agent Trust HubPassSocketPassSnykWarn

安装于

gemini-cli139

codex139

claude-code137

pi53

opencode31

github-copilot31

CRUD Completeness : For every entity in your app, verify agents have Create, Read, Update, and Delete paths. Missing any one breaks parity.

2. Granularity — Atomic Primitives, Not Bundled Logic

Features emerge from agent loops, not monolithic tools.

Wrong : deploy_and_monitor(app) — bundles judgment into code Wrong : classify_and_organize_files(files) — agent should decide classification Right : eve build create, eve build run, eve env deploy, eve job follow — agent decides the sequence

On Eve : The manifest defines WHAT (services, pipelines), the agent decides HOW and WHEN to compose them.

Design test : To change behavior, do you edit prose (prompts/skills) or refactor code? If code — your tools aren't atomic enough.

3. Composability — New Features = New Prompts

When tools are atomic and parity exists, you add capabilities by writing prompts, not code.

Eve example : The eve-pipelines-workflows skill adds pipeline composition capability. No new CLI commands needed — the skill teaches agents to compose existing eve pipeline and eve workflow commands.

Your app : If adding a feature requires new API endpoints, you may be bundling logic. Consider whether existing primitives can be composed differently.

4. Emergent Capability — Agents Surprise You

Build atomic tools. Agents compose unexpected solutions. You observe patterns. Optimize common patterns. Repeat.

Eve example : Agents compose eve job create --parent + eve job dep add + depth propagation to build arbitrary work hierarchies. The platform didn't prescribe this — agents discovered it from atomic primitives.

Platform Capabilities for Agentic Apps

Beyond the four principles, Eve provides (or is building) specific primitives that make agentic apps dramatically simpler:

Job Attachments — Structured Context Passing

Pass plans, reports, and insights between agents without file gymnastics. Attach text documents (markdown, JSON, YAML) to jobs. Downstream agents read attachments from parent jobs. This solves 80% of the "agents passing structured context" problem.

Service Account Auth — Backend-to-API Communication

Use eve auth mint (today) or service accounts (emerging) to authenticate app backends. Every app with a backend needs a non-user token for API calls. Scoped permissions enforce least privilege.

Org Document Store — Persistent Agent Knowledge

Agents accumulate knowledge that outlives individual jobs: architecture reports, risk assessments, conventions. DB-backed with full-text search and agent-native search/replace editing via PATCH operations.

Web Chat — Two Mechanisms

Choose based on your needs:

• Gateway Provider (Mechanism A): WebSocket to Eve gateway. Zero backend. Best for simple chat widgets and admin consoles.
• Backend-Proxied (Mechanism B): App backend calls POST /internal/orgs/:id/chat/route. Full control over enrichment, storage, routing. Best for production SaaS.

Decision : If your app intercepts, enriches, or stores conversations → use B. Otherwise → use A.

Cross-Project Queries — Org-Level Intelligence

Portfolio views, dashboards, and any tool spanning multiple projects. Org-level endpoints eliminate N+1 API calls.

See references/eve-horizon-primitives.md for the full catalog with API schemas and priority ranking.

Eve-Specific Design Patterns

Files as Universal Interface

• Agents know cat, grep, mv, mkdir
• Use .eve/manifest.yaml as single source of truth — agents read and edit it
• Agent configs live in repo files (agents.yaml, teams.yaml) — not hidden in databases
• Directory structure = information architecture: {entity_type}/{entity_id}/content
• Name files by purpose: {entity}.json, {type}.md, agent_log.md

Context Injection (Three Layers)

System prompts should include:

1. Available resources — what exists, with counts: "12 notes in /notes, 3 projects"
2. Capabilities — what agents can do: "Create, edit, tag, delete notes"
3. Recent activity — what happened: "User created 'Project kickoff' 2 hours ago"

Eve injects EVE_API_URL, EVE_PROJECT_ID, EVE_ORG_ID, EVE_ENV_NAME into every environment. Skills provide domain vocabulary.

Explicit Completion Signals

• Jobs return json-result with eve.status ("success", "failed", "waiting")
• No heuristic completion detection — explicit signals always
• Track progress at task level with phase transitions
• Support shouldContinue for multi-step operations
• Checkpoint state for resume on interruption

Dynamic Capability Discovery

• eve job list discovers available work
• eve agents list discovers available agents
• Skills system auto-discovers capabilities at install time
• Gateway routes messages to agents by slug — new agents are instantly addressable
• Prefer runtime discovery (list_available_types()) over static tool-per-endpoint mapping

Agent-to-UI Communication

• Event types: thinking, toolCall, toolResult, textResponse, statusChange
• No silent actions — changes visible immediately
• Show progress during execution, not just results
• Consider ephemeralToolCalls for noisy internal operations

The Success Checklist

Architecture:

• Every UI action has a CLI/API equivalent (parity)
• Every entity has full CRUD (completeness)
• Tools do one thing; agent decides composition (granularity)
• Adding capability = adding a skill/prompt, not code (composability)
• Agent can handle requests you didn't explicitly design for (emergent)

Implementation:

• Manifest and config files are the source of truth (files as interface)
• System prompt includes resources, capabilities, and recent activity (context injection)
• Completion is explicit, not heuristic (json-result with eve.status)
• Agent actions reflect immediately in UI (visibility)
• Dynamic discovery works for new capabilities (extensibility)

If building an app with a backend:

• Service account auth configured for API access
• Web chat mechanism chosen (gateway vs. proxied)
• Job attachments used for structured inter-agent context
• Org document store for persistent knowledge (when available)

Anti-Patterns

Reference

See references/eve-horizon-primitives.md for the full platform primitives catalog with API schemas, DB designs, and implementation priority ranking.

For the source philosophy: ../eve-horizon/docs/ideas/agent-native-design.md For platform primitives analysis: ../eve-horizon/docs/ideas/platform-primitives-for-agentic-apps.md

Weekly Installs

139

Repository

incept5/eve-skillpacks

First Seen

Feb 12, 2026

Security Audits

Gen Agent Trust HubPassSocketPassSnykWarn

Installed on

gemini-cli139

codex139

claude-code137

pi53

opencode31

github-copilot31