ClaudeMods
☰
ZH-CN
● 0 人在线 · 浏览 0 次
赞助提交作品
GitHub 仓库 · 发布者 qf-studio

navigator

完成你开始的事——持久的会话、会学习的 AI、能交付的功能。上下文工程 + Theory of Mind + Knowledge Graph + Loop Mode + Task Mode。包括项目知识图谱、统一工作流编排、会话开始时自动更新、ToM 验证检查点、双向建模、质量检测、增强上下文标记和 Loop Mode。凭借经过验证的策略减少 92% token。基于 Riedl 与 Weidmann 2025 Human-AI Synergy 研究。

已翻译

关于这个 mod

Navigator

完成你开始的事

持久的会话。会学习的 AI。能交付的功能。

MIT License Version


Navigator 是什么?

Navigator 是一个用于 AI 辅助开发、实现上下文工程的 Claude Code 插件。

不要一开始就载入所有文档(然后在第 7 次交互时崩溃),Navigator 会在需要时载入所需内容,让上下文窗口保持 94% 可用于实际工作。

面向使用 Claude Code、希望会话能完成整个功能而不是中途崩溃的开发者。


你被困在这样的循环里

你知道这个模式:

第 5 次交互:Claude 忘了你最近的改动
第 7 次交互:开始产生幻觉(“那个函数不存在”)
第 8 次交互:“已达到上下文上限”

重启。重新载入。重新解释一切。重复。

你“以防万一”载入了 150k 个文档 token。

你用了 8k。

剩下的呢?噪声淹没了信号。


打破循环

Navigator 实现了上下文工程——在需要时载入所需内容。

| 指标 | 没有 Navigator | 使用 Navigator | |--------|-------------------|----------------| | 载入的 token | 150,000 | 12,000 | | 会话长度 | 5-7 次交互 | 20+ 次交互 | | 结束时的上下文 | 95%(已崩溃) | 35%(从容) | | token 节省 | — | 92% |

结果:真正能完成开始之事的会话。


你的 AI 也会变得更聪明

Navigator v5.0 加入心智理论——Claude 会学习你。

nav-profile:记住你跨会话的偏好

“记住我偏好简洁的解释”
→ 应用于未来会话
→ 从修正中自动学习

nav-diagnose:捕捉协作偏移

同一个修正出现两次 → 触发质量检查
“你还是没理解” → 重新锚定提示

Loop Mode(v5.1):运行直到完成

“运行直到完成:添加用户身份验证”
→ 带进度跟踪的结构化完成
→ 双条件退出(启发式 + 明确信号)
→ 停滞检测防止无限循环

Knowledge Graph(v6.0):持久化的项目记忆

“关于身份验证,我们知道什么?”
→ 跨任务、SOP 和系统文档的统一搜索
→ 经验记忆:模式、陷阱、决策
→ 会话开始时自动呈现相关上下文

相同的工作流。更多能力。

Navigator 是一个超集。除了上下文工程,还包含你所期待的一切。

| 功能 | Navigator | 其他产品 | |---------|-----------|--------| | 结构化工作流 | ✅ 完整技能套件 | ✅ | | 组件生成 | ✅ | ✅ | | 测试生成 | ✅ | ✅ | | 会话持久度 | 20+ 次交互 | 5-7 次交互 | | token 节省 | 已验证 92% | 无 | | 代码简化 | ✅ | ❌ | | 任务模式 | ✅ | ❌ | | 心智理论 | ✅ | ❌ | | 知识图谱 | ✅ | ❌ | | TRIZ 发散式求解 | ✅ | ❌ | | Web 深度研究(带引用、通过门禁检查) | ✅ | ❌ | | 循环模式 | ✅ | ❌ | | OpenTelemetry 指标 | ✅ | ❌ | | Figma MCP 集成 | ✅ | ❌ |

相同的基础。更出色的上下文管理。


证据,而非承诺

不是估算。通过 OpenTelemetry 验证。

╔══════════════════════════════════════════════════════╗
║          NAVIGATOR EFFICIENCY REPORT                 ║
╚══════════════════════════════════════════════════════╝

📊 TOKEN USAGE
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Your project documentation:     150,000 tokens
Loaded this session:             12,000 tokens
Tokens saved:                   138,000 tokens (92% ↓)

📈 SESSION METRICS
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Context usage:                        35% (excellent)
Efficiency score:                  94/100 (excellent)

检查你的数据:安装后运行 /nav:stats。


快速开始

安装

# Claude Code plugin marketplace
/plugin marketplace add qf-studio/navigator
/plugin install navigator

# Restart Claude Code

代码智能(可选)。 为你的技术栈安装语言服务器插件(/plugin install pyright-lsp@claude-plugins-official、typescript-lsp、gopls-lsp、…;服务器二进制文件必须位于 PATH 中),研究代理就会用跳转到定义、引用和大纲来回答符号问题,而不是文本搜索。不安装时,它和以前一样使用 Grep。在这个仓库上测得:一个“谁调用它”的问题从 9 次请求、284k 个上下文 token 降至 5 次请求、142k 个,完整保留调用点列表。

初始化

“在这个项目中初始化 Navigator”

每次会话开始时

“开始我的 Navigator 会话”

就这样。Navigator 会处理其余事项。


可选:类型化提示评审器

Navigator 的循环触发器、复杂度和歧义检测器是关键词匹配器。使用 TypeSafe 密钥后,它们会获得类型化评审器(v7.7.0):每个提示发送一次请求,明确的回答会覆盖关键词,其余情况回退到关键词。根据 60 条已标注提示测量,层级准确率从 38 提升到 53。默认关闭,因为提示文本会发送到 API。

export TYPESAFE_API_KEY=...      # 或写入 ~/.config/typesafe/api_key (chmod 600)

然后说 enable judge,并在插件根目录运行 python3 hooks/nav_hook_lib/judge.py --check 进行验证。设置和调优见 .agent/sops/integrations/typesafe-judge-setup.md。


你会得到什么

会根据自然语言自动调用的技能:

“开始我的 Navigator 会话”              → 节省 92% 的会话
“为个人资料创建 React 组件”            → 组件 + 测试 + 样式
“为帖子添加 API 端点”                  → 端点 + 验证 + 测试
“创建上下文标记:checkpoint”            → 97% 上下文压缩
“运行直到完成:添加深色模式”            → 循环模式完成
“深入研究 WebGPU 浏览器支持”           → 带引用的报告 + 图谱记忆
“enable judge”                          → 受提示门控的类型化 Jev 判断

无需记忆命令。 技能会识别意图并执行。


停止重启。开始交付。

你的会话会崩溃,是因为默认的上下文管理有问题。

Navigator 用上下文工程解决这个问题——原则与 Anthropic 推荐的相同。

节省 92% token。20+ 次交互的会话。经过验证的指标。

/plugin marketplace add qf-studio/navigator
/plugin install navigator

完成你开始的事。


故障排除

技能已载入,却打印文档而不是执行

/plugin update navigator

然后重试。插件版本过旧。


链接

许可证

MIT License - 参见 LICENSE

安装

请先查看作者 README 确认 marketplace 和插件名称;命令可能随仓库结构改变。

claude plugin marketplace add qf-studio/navigator
claude plugin install navigator
原文 / README

Navigator

Finish What You Start

Sessions that last. AI that learns. Features that ship.

MIT License Version


What is Navigator?

Navigator is a Claude Code plugin that implements context engineering for AI-assisted development.

Instead of loading all your documentation upfront (and crashing at exchange 7), Navigator loads what you need, when you need it—keeping 94% of your context window available for actual work.

For developers using Claude Code who want sessions that finish features instead of crash mid-way.


The Loop You're Stuck In

You know the pattern:

Exchange 5:  Claude forgets your recent changes
Exchange 7:  Hallucinations start ("that function doesn't exist")
Exchange 8:  "Context limit reached"

Restart. Reload. Re-explain everything. Repeat.

You loaded 150k tokens of documentation "just in case."

You used 8k.

The rest? Noise drowning out signal.


Break The Loop

Navigator implements context engineering—load what you need, when you need it.

| Metric | Without Navigator | With Navigator | |--------|-------------------|----------------| | Tokens loaded | 150,000 | 12,000 | | Session length | 5-7 exchanges | 20+ exchanges | | Context at end | 95% (crashed) | 35% (comfortable) | | Token savings | — | 92% |

Result: Sessions that actually finish what they start.


And Your AI Gets Smarter

Navigator v5.0 adds Theory of Mind—Claude learns you.

nav-profile: Remembers your preferences across sessions

"Remember I prefer concise explanations"
→ Applied in future sessions
→ Auto-learns from corrections

nav-diagnose: Catches collaboration drift

Same correction twice → Quality check triggered
"You're not getting this" → Re-anchoring prompt

Loop Mode (v5.1): Run until done

"Run until done: add user authentication"
→ Structured completion with progress tracking
→ Dual-condition exit (heuristics + explicit signal)
→ Stagnation detection prevents infinite loops

Knowledge Graph (v6.0): Project memory that persists

"What do we know about auth?"
→ Unified search across tasks, SOPs, system docs
→ Experiential memory: patterns, pitfalls, decisions
→ Auto-surfaces relevant context on session start

Same Workflows. More Capabilities.

Navigator is a superset. Everything you'd expect, plus context engineering.

| Feature | Navigator | Others | |---------|-----------|--------| | Structured workflows | ✅ Full skill suite | ✅ | | Component generation | ✅ | ✅ | | Test generation | ✅ | ✅ | | Session longevity | 20+ exchanges | 5-7 exchanges | | Token savings | 92% verified | None | | Code simplification | ✅ | ❌ | | Task Mode | ✅ | ❌ | | Theory of Mind | ✅ | ❌ | | Knowledge Graph | ✅ | ❌ | | TRIZ divergent solving | ✅ | ❌ | | Web deep research (cited, gate-checked) | ✅ | ❌ | | Loop mode | ✅ | ❌ | | OpenTelemetry metrics | ✅ | ❌ | | Figma MCP integration | ✅ | ❌ |

Same foundation. Superior context management.


Proof, Not Promises

Not estimates. Verified via OpenTelemetry.

╔══════════════════════════════════════════════════════╗
║          NAVIGATOR EFFICIENCY REPORT                 ║
╚══════════════════════════════════════════════════════╝

📊 TOKEN USAGE
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Your project documentation:     150,000 tokens
Loaded this session:             12,000 tokens
Tokens saved:                   138,000 tokens (92% ↓)

📈 SESSION METRICS
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Context usage:                        35% (excellent)
Efficiency score:                  94/100 (excellent)

Check yours: Run /nav:stats after installing.


Quick Start

Install

# Claude Code plugin marketplace
/plugin marketplace add qf-studio/navigator
/plugin install navigator

# Restart Claude Code

Code intelligence (optional). Install a language-server plugin for your stack (/plugin install pyright-lsp@claude-plugins-official, typescript-lsp, gopls-lsp, …; the server binary must be on your PATH) and the research agent answers symbol questions with go-to-definition, references and outlines instead of text search. Without one it uses Grep exactly as before. Measured on this repo: a who-calls question went from 9 requests and 284k context tokens to 5 requests and 142k, with the complete call-site list.

Initialize

"Initialize Navigator in this project"

Start Every Session

"Start my Navigator session"

That's it. Navigator handles the rest.


Optional: Typed Prompt Judge

Navigator's loop-trigger, complexity and ambiguity detectors are keyword matchers. With a TypeSafe key they get a typed judge behind them (v7.7.0): one request per prompt, decisive answers override, everything else falls back to the keywords. Measured on 60 labeled prompts, tier accuracy went from 38 to 53. Off by default because the prompt text is sent to the API.

export TYPESAFE_API_KEY=...      # or: write it to ~/.config/typesafe/api_key (chmod 600)

Then say enable judge and verify with python3 hooks/nav_hook_lib/judge.py --check from the plugin root. Setup and tuning: .agent/sops/integrations/typesafe-judge-setup.md.


What You Get

Skills that auto-invoke on natural language:

"Start my Navigator session"              → Session with 92% savings
"Create a React component for profile"    → Component + tests + styles
"Add an API endpoint for posts"           → Endpoint + validation + tests
"Create context marker: checkpoint"       → 97% context compression
"Run until done: add dark mode"           → Loop mode completion
"Deep research on WebGPU browser support"  → Cited report + graph memories
"Enable judge"                             → Typed Jev judgments behind the prompt gates

No commands to memorize. Skills detect intent and execute.


Stop Restarting. Start Shipping.

Your sessions crash because context management is broken by default.

Navigator fixes this with context engineering—the same principles Anthropic recommends.

92% token savings. 20+ exchange sessions. Verified metrics.

/plugin marketplace add qf-studio/navigator
/plugin install navigator

Finish what you start.


Troubleshooting

Skill loads but prints docs instead of executing

/plugin update navigator

Then retry. Outdated plugin version.


Links

License

MIT License - See LICENSE

更多类似作品