atlas：為想先讀文件再看程式碼的開發者提供中文翻譯

The original atlas README is in English, and it is long. If your working language is Chinese, that means you either read it slowly or you skim it and miss the parts that matter — the platform limits, where session data actually lives, what happens when you switch agents mid-task.

A Chinese translation now exists: yangshun2005/atlas-cn. It is a translation of pacifio/atlas, which currently has 0 stars. The Chinese version was published today and also has 0 stars. Neither number says anything about quality; I mention them because they are the only numbers I have.

Below is what the project is, what the translation adds, and where the translated docs are still thin.

What atlas actually is

atlas describes itself as source control for coding agents. The mechanism is a "checkpoint": when an agent produces a commit, atlas links that commit back to the session that produced it — the prompt, the tool calls, the reasoning trace, the file changes. All of it stays queryable months later.

That is the core claim. The problem it points at is real enough: agents write a growing share of code, and what survives is a commit message written by a model summarizing a diff. The prompt that caused the change, the tool calls, the approach the agent tried and abandoned — those live in a scrollback buffer until it scrolls away.

The README lists three consequences of that, all of which predate agents in some form:

Agents start from zero every session. Context you built yesterday is gone, and rebuilding it is your job.
Switching agents loses the thread. Claude Code cannot read Codex's history, and Codex cannot read Claude Code's.
Context is scattered across a knowledge base, CLAUDE.md, AGENTS.md, and each agent's own memory file. Nothing reads all of them.

atlas's answer is to run multiple agents against the same codebase in one window — Claude Code, Codex, atlas's own agent, or anything in the ACP registry — and have them share memory. A decision made in Claude Code shows up in Codex's next prompt. Plans, file changes, failures, and architecture notes are shared and matched on-device against whatever you are asking about.

Two design commitments are stated explicitly:

Nothing is locked in. Notes are markdown, the canvas is JSON, sessions are JSONL, the editor works on files on disk.
Local by default. Code, notes, and sessions stay on your machine. Cross-team sync requires signing in and creating an organization.

What the Chinese version adds beyond a translation

The repository is structured so the translated README body sits below a separate section labeled as supplementary — explicitly marked as not part of the original text. That distinction matters, because it lets the translator add things the upstream README does not spell out.

A quick-start path that the original does not give

The original README tells you to download a build and points at the docs site. The Chinese version walks through five steps:

Download from tryatlas.cc or the releases page — .dmg for macOS (Apple Silicon or Intel), .msi for Windows.
Install and launch.
Open your repository and run Claude Code, Codex, or atlas's built-in agent.
Commit normally; the commit is linked back to the agent session that produced it.
To trace something later, select a checkpoint and talk to it — it answers based on what actually happened in that session.

Step 3 is honest about a gap: the original does not provide the exact commands for launching each agent inside atlas. The translation says so rather than inventing them. I would rather see that than a guessed command that fails.

Environment requirements collected in one place

The upstream README states macOS 13+ and Windows 10+ (x64) and notes that Linux can be built from the same Tauri codebase but is untested. The Chinese version pulls this into a requirements block and adds the hardware detail — Apple Silicon or Intel on macOS, x64 on Windows — plus a note that Claude Code and Codex run as external subprocesses over ACP.

It also flags something the original does not: downloads may need a mirror or accelerator on mainland Chinese networks. Small thing, but it is the kind of friction that makes someone give up on step one.

One requirement is left open in both versions: whether Claude Code and Codex need to be installed separately beforehand. The translation says the original does not specify. That is a real gap in the upstream docs, not a translation failure.

Where your data actually lives

This is the section I would read first, and it is not in the upstream README in this form:

Session records: .atlas/sessions.db (SQLite), inside the project, gitignored
Checkpoint records: also under .atlas/
Notes: Markdown files in .atlas/knowledge/

The important detail is that these directories are git-ignored by default. If you assume your checkpoints are being committed alongside your code, you are wrong. They are local artifacts.

The translation also repeats the upstream privacy claim — sessions are scrubbed of secrets before being written to disk — and then adds the caveat that code and notes stay on the machine by default, so sensitive projects still warrant your own assessment. That is the right framing. A scrubbing step is not an audit.

Reading the translation critically

A few things to keep in mind before you rely on it.

It is a translation of a README, not of the product docs. The full documentation lives at docs.tryatlas.cc and is not translated here. If you need API-level detail, the Chinese repo will not have it.

Some sections are thin because the source is thin. Build-from-source is the clearest example. The upstream README says Linux builds from the same Tauri codebase but is untested, and does not give a concrete command sequence. The translation says exactly that and points you back to the README body. No invented cargo tauri build line, which I appreciate — a wrong build command costs more time than a missing one.

The translation date is 2026-10-23. If the upstream README changes after that, the Chinese version drifts. The repo asks for issues or PRs on inaccuracies, which is the only maintenance model that works for a project like this.

Zero stars on both sides. The original is early. The translation is one day old. Neither has been through the kind of use that surfaces the awkward edge cases. Treat the quick-start as a map, not a guarantee.

Who this is for

If you are evaluating atlas and your primary reading language is Chinese, start with the translated repo. The quick-start section gets you to a running install without reading the full README, and the data-location section tells you what you are agreeing to by using it — local SQLite, gitignored, no account needed until you want cross-team sync.

If you are already using atlas and hit something the Chinese docs get wrong, the repo takes PRs. That is the whole maintenance story for a translation project: it stays useful exactly as long as people correct it.

The original project is at pacifio/atlas. The Chinese version is at yangshun2005/atlas-cn.

Chinese documentation: https://github.com/yangshun2005/atlas-cn

I'm an independent enterprise AI delivery consultant — private LLM deployment, RAG, agent workflows, AIOps. If your team is stuck somewhere in that stack, my inbox is open.