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

conversation-atlas

一个实时侧边面板,描绘漫长 Claude Code 会话中的目标、当前路径、可能的绕路、决定、未决问题、检查点、工作文件和下一步恢复位置;它会自行观察,并在改变意图前询问。

NeelAPatel@NeelAPatel

NeelAPatel/Claude-Mod-ConversationAtlas

已翻译

关于这个 mod

Conversation Atlas 是一个 Claude Code mod,增加侧边面板,在漫长会话中检测并帮助管理目标、决定和绕路。它显示目标和主题路径、带快照的绕路、决定与未决问题、检查点(提交、通过的测试、用户标记)、正在使用的文件,以及用于恢复工作的建议下一步。没有明确的用户确认,观察结果永远不会变成意图。可以通过 claude --plugin-dir . 或 Windows PowerShell 安装程式从源码安装;需要 Claude Code 2.1.287+。提供 /atlas 命令,包括 goal、next、mark、decision、detour、outcome、return、promote、scan、recover、observer、hide/show、setup 和 reset,并提供可配置的 observer 与 scanOnLaunch 设置。它完全在本机运行,没有遥测,资料储存在 Claude Code 的插件储存区和 <project>/.claude/atlas/。包含有少量用量成本的 Claude observer 模式,也包含免费的 engine-only 模式。

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 NeelAPatel/Claude-Mod-ConversationAtlas
claude plugin install conversation-atlas
原文 / README
<div align="center">

🗺️ Conversation Atlas

A live map of your Claude Code session: where you started, where you are, and how to get back.

License: MIT Version Claude Code Telemetry Status

</div> <div align="center"> <video src="https://github.com/user-attachments/assets/112dd1f4-972a-4970-9733-741a9e0cf5d2" width="720" controls muted></video> </div>

Tame your Claude's conversation history. Long Claude Code sessions drift: a quick tangent turns into an hour, a decision made early gets lost in scrollback, and after a break nobody remembers what "next" was. Conversation Atlas is a mod that hooks into your chats and adds a side pane that detects and helps you manage your goals, decisions and even detours while you work:

  • 🎯 your goal, and the path of topics that led here
  • ↳ detours you took, with a snapshot of where you left off
  • ◆ decisions and open questions you would otherwise lose
  • 📍 checkpoints (commits, passing tests, your own marks) and the files in play
  • ▸ what to resume next

Observation never becomes intent without you. Atlas can suggest a goal, a detour or a return. Only you turn a suggestion into intent, with a button press or an /atlas command.

Installation

Marketplace: coming soon. For now, install from source. Please Note that as it stands, this project is hardly v0.1.0 yet

Requires Claude Code 2.1.287+ (terminal or desktop Code tab). To dock the pane beside the transcript, use /tui fullscreen in a terminal at least 144 columns wide; otherwise open it from the footer button or with /atlas.

Try it from source (the repository root is the plugin):

git clone https://github.com/NeelAPatel/Claude-Mod-ConversationAtlas.git
cd Claude-Mod-ConversationAtlas
claude --plugin-dir .

Install for your user (Windows): .\scripts\install-atlas.ps1 validates the plugin and copies it to ~/.claude/skills/conversation-atlas. Run /reload-plugins in a session that is already open. Don't combine this with --plugin-dir in the same session, or two copies load.

Feature tour

1. Set up

On first launch the pane asks how Atlas should observe. Nothing costs Claude usage until you choose.

  • Claude observer (minimal usage cost, read below): Claude reports topic shifts, decisions and questions through a tiny tool.
  • Engine only (free, no usage cost): Atlas uses just files, activity, tests, commits and your own wording.

Change it any time with /atlas observer [claude|engine].

2. Goal, detour, return

  • /atlas goal ship the login refactor sets the goal. Atlas also shows the goal it detected; Use this as my goal makes it yours.
  • /atlas detour check the flaky test snapshots where you left off and starts a side trip. Decisions made during it are kept as its outcomes.
  • /atlas return ends the detour and sends Claude one short packet: the goal, topic, next step and settled decisions from the moment you left. /atlas promote makes the detour the new goal.

3. The pane

| Tab | Key | What it shows | |---|---|---| | Map | m | Goal, current topic path, detours, live activity, working set, resume next | | Trail | t | Topic tree and a timeline of prompts, shifts, decisions, detours, checkpoints | | Open | o | Everything waiting on you: suggestions, decisions to settle, open questions | | Evidence | e | Checkpoints, settled-decision ledger, past detours, files, earlier sessions |

The bottom bar holds Legend (l), decisions (d), open questions (q) and + Mark (k). Add to message on any item puts an [Atlas #n: …] chip in your draft; only chips still there when you send reach Claude. Details: docs/REFERENCE.md.

4. Recovering and scanning earlier conversations

Atlas can pick up a chat it wasn't watching, or one you left yesterday.

| Situation | What happens | Cost | |---|---|---| | Atlas installed or reloaded mid-chat, or you resume an old chat | Atlas replays the earlier messages: your prompts, files read and edited, test and commit checkpoints, a goal suggestion, and a question left open | Free | | You want topics and decisions for that history | /atlas scan (or Map earlier conversation in Trail) asks Claude once over the session's own transcript. Results are observations; confirm what is true in the Open tab | One mostly cached request | | A new session in the same project | Atlas offers to Resume the last goal | Free | | Older sessions | /atlas recover lists them; /atlas recover 2 resumes number 2. Evidence → Earlier sessions does the same with buttons | Free |

Two ways to resume an earlier session:

  • Resume this brings back its goal, next step and settled decisions and keeps your current map.
  • Resume full (Atlas saves only) replaces the current map with the saved one, after a confirm: /atlas recover <n> full confirm.

Nothing is adopted automatically; every resume is an explicit press or command. Set scanOnLaunch to claude to also run the one-request scan whenever Atlas joins a chat that already has history.

Commands and configuration

Everything is under /atlas (/atlas help lists them).

| Command | Description | |---|---| | /atlas | Open the pane | | /atlas goal <text> | Set your goal | | /atlas next <step> | Pin the next step | | /atlas mark [name] | Add a marked checkpoint | | /atlas decision <text> [--reason <why>] | Record a settled decision | | /atlas detour <reason> | Take a detour | | /atlas outcome <finding> / /atlas exclude <material> | Keep or set aside a finding made during a detour | | /atlas return | End the detour and send the return packet | | /atlas promote | Promote the detour to the goal | | /atlas scan | Map the conversation so far (one cached Claude request) | | /atlas recover [n] / /atlas recover <n> full [confirm] | List earlier sessions, resume one, or replace the map with one | | /atlas observer [claude\|engine] | Show or change the observer mode | | /atlas hide <b\|h> / show <b\|h> / filters | Hide, show or list report-backs (b) and hand-offs (h) in Trail | | /atlas setup | Show the setup screen again | | /atlas reset | Clear this session's map (saved files are kept) |

Settings live in /config → conversation-atlas, or pluginConfigs in your settings:

| Setting | Values | Default | Effect | |---|---|---|---| | observer | claude and engine, engine only | claude and engine | Preselected choice on the setup screen. After setup, your stored choice wins | | scanOnLaunch | engine, claude, off | engine | What Atlas does when it joins a chat that already has history |

Privacy and cost

  • Local only. Atlas makes no network calls and has no telemetry. It saves to Claude Code's plugin store and to <project>/.claude/atlas/ (add that to .gitignore if you don't want session maps committed).
  • Engine-only mode costs nothing, and neither does anything before you pick a mode.
  • Claude observer mode has a small usage cost. It adds one cached rules section to the system prompt and a short observe tool call on turns where something changed. Each call re-reads the cached context, so the cost grows with chat length: in one long-session measurement it came to roughly 10% of that session's tokens. That is a known rough edge, and some hooks and updates will be optimized in future releases to bring it down. Until then, engine-only mode is the free option, and the observer is worth it when topic and decision tracking matters to you.
  • /atlas scan is always an explicit action: one forked, mostly cached request.
  • What reaches Claude: the rules section (Claude mode), a one-line note of your confirmed goal and detour (Claude mode), a pending return packet, and chips you left in your message. Observed items are never presented as settled facts.

Limitations

  • Engine-only mode has no topic path; topic quality depends on Claude calling observe.
  • Wording cues like "btw" or "quick tangent" are English-only and only create suggestions.
  • One detour at a time; nested tangents show as topics inside it.
  • Matching a return to an earlier topic is fuzzy. A wrong match moves the observed path, never your intent.
  • Docking needs the fullscreen layout and a wide terminal; scroll limits are estimates.
  • Recovery reads at most 80 files per folder and lists 12 sessions.

Roadmap

Ideas that build directly on what Atlas already captures:

  • [ ] End-of-session recap: an opt-in summary of goal, decisions and next step, shown before it is saved as a resume note
  • [ ] Smarter resume: open a new session with the recap and return packet ready to review
  • [ ] Prompt and output summaries when you open an event in Trail (desktop first)
  • [ ] Decision ledger export to a project file Claude can read at session start
  • [ ] Checkpoint detail: attach test output or a diff stat to a checkpoint
  • [ ] Topic to transcript links: jump from a pane topic to the messages it covers

Contributing

Contributions are welcome: bug reports, ideas and pull requests. Read CONTRIBUTING.md first; architecture and invariants live in docs/ARCHITECTURE.md and AGENTS.md, and security reports go through SECURITY.md. Everyone participating follows the Code of Conduct.

This is my first open-source project, so ideas, feedback and discussions are very welcome.

License

MIT © 2026 NeelAPatel.

更多类似作品