ai-safe-edit

ai-safe-edit

The only safe entry point for an AI agent to edit files. Back up, validate, roll back on failure — and never leave a half-broken file behind.

中文说明 · MIT · Python 3.8+ · zero dependencies

---

The problem

When an agent edits a file, it writes text. When that text is assembled from shell heredocs, regex substitutions, or model-generated code, three things go wrong often enough to actually hurt:

The usual response is to write the lesson down. That is not a constraint. The next session, or the next model, repeats the mistake.

Real numbers from the project this came out of: 10 nested-quote errors in a single day, all of them landing in the actual product files, because the edits bypassed the tool that existed to prevent exactly that.

And a small confession, because it is the clearest illustration there is: while writing the commit message for this repository, the same class of failure hit again. The message contained `--selftest` inside a double-quoted shell string. The shell executed the backticks and deleted the word. The commit went in with a sentence missing its subject — and would have shipped that way, because nothing checks commit messages.

Knowing about a failure mode is not protection from it. Only a gate is.

What this does

It turns the lesson into a gate:

1. Back up before touching anything → .safe_edit_backup/<timestamp>/
2. Validate immediately after writing: .py → compile() · .json → json.loads · .yaml → yaml.safe_load · .bat → ASCII-only check
3. Roll back automatically if validation fails, and say so
4. Non-zero exit code on failure — a model or a script cannot miss it

A half-broken file is never left behind.

Quick start

# edit with the safety net
python safe_edit.py --file app.py --replace "old text" "new text"

# same replacement across several files
python safe_edit.py --files a.py b.py --replace "old" "new"

# large blocks with quotes or non-ASCII text: read both sides from files
# (this is the mode to use when shell argument escaping is unreliable)
python safe_edit.py --file app.py --replace-with-file old.txt new.txt

# validate without writing
python safe_edit.py --check-only a.py b.py

# roll back to the most recent backup
python safe_edit.py --rollback app.py

# self-test: positive and negative controls
python safe_edit.py --selftest

On failure you get an unambiguous report and a clean file:

x  app.py   改 1 处后**校验失败** ⇒ **已自动回滚**：语法错 line 207: invalid syntax
            （备份 .safe_edit_backup/20261015-213510/app.py）
==> 1 个失败（已回滚）—— 调用方必须知道这次没改成     (exit code 1)

Use it from an AI agent (MCP)

mcp_server.py exposes it over MCP — stdio, JSON-RPC 2.0, no dependencies to install. Tools: safe_edit, edit_check, edit_rollback.

{
  "mcpServers": {
    "ai-safe-edit": {
      "command": "python",
      "args": ["/absolute/path/to/ai-safe-edit/mcp_server.py"]
    }
  }
}

Works with any MCP client — including DeepSeek Harness, Claude Code, Cursor and others.

Why an agent should be forced through this

An agent that can write files freely will, eventually, write a broken one. The question is whether that failure surfaces immediately, with a rollback, or later, as a mysterious error in a file nobody remembers touching.

This project's answer, learned the hard way:

• A lesson written down is not a constraint. A lesson turned into a gate is.
• The caller must be told when it failed. Silent success is worse than loud failure — an agent that thinks its edit landed will build on top of it.

Honest limits

• It cannot judge whether an edit is correct — only that the result still parses. Semantic correctness still needs tests, gates, and a human.
• It does not prevent a bypass. If you want bypasses to at least leave a trace, see the companion idea below.
• .py / .json / .yaml / .bat are the validated formats today. Other extensions are backed up and written, but not checked.

Related work

This tool was built independently inside a hardware-design agent project. After it was working, we read DeepSeek Harness's source and found it solves the same class of problem from the other end: hardening an experiment requires explicit user approval (install_bundle + pendingBuilds), so nothing the model writes reaches persistent configuration without a human.

The two are complementary:

A companion idea in the same spirit: after every successful edit, record a hash of the file. A checker can then compare the current hash against the last record and flag "this file was changed without going through the tool" — turning a silent bypass into a visible one.

License

MIT