aisweep

AI attribution artifacts, swept out of your git history.

Find and block the session URLs, Generated with Claude Code lines and AI co-author trailers that AI coding agents leave in your commits — for developers who use Claude Code, Cursor, Copilot, Aider, Gemini CLI or Codex and publish clean repositories.

English | 简体中文

Why this exists

AI coding agents sign your commits by default, whether you asked for it or not. Claude Code appends a session URL (https://claude.ai/code/session_…) plus a Generated with Claude Code line and a Co-Authored-By: Claude trailer to every commit and PR description it writes — silently, opt-out, discoverable only after it is already in your history (anthropics/claude-code#66504, 89 👍 as of August 2026, and a lively HN thread).

Claude Code now has a setting to turn this off per tool. That still leaves:

1. history already written — the artifacts stay in every past commit;
2. every other tool — Cursor, Copilot, Aider & friends each have their own behaviour and their own switch;
3. teammates — whose ~/.claude/settings.json you do not control.

aisweep is the one tool-agnostic gate: audit what is already in your history, and block polluted commits before they land.

Features

🔍 Audit	scan full history (or any range) for AI artifacts, per-commit report
🛡️ Prevent	a commit-msg hook stops polluted commits before they land
🧰 Tool-agnostic	pattern-based — works with any agent, including future ones
🧩 Configurable	.aisweep.json: disable rules, tune severity, allowlist, add your own
📤 Reports	Markdown + JSON output for CI, PR review or pre-open-sourcing checks
🪶 Zero dependencies	Python ≥ 3.10 standard library only; pip install and go
🤝 Non-destructive	never rewrites, amends or touches a single commit

Quick start

Prerequisites: Python 3.10+ and git.

# install
pip install git+https://github.com/w1977-0/ai-commit-sweeper.git
# or isolated:  pipx install git+https://github.com/w1977-0/ai-commit-sweeper.git
# or without installing:
git clone https://github.com/w1977-0/ai-commit-sweeper.git
cd ai-commit-sweeper
python -m aisweep --help

# 1. audit a repository
aisweep scan ./my-project

# 2. prevent future artifacts
cd ./my-project
aisweep hook install

Usage

Audit existing history

$ aisweep scan ./my-project
aisweep 0.1.0 — scanning main in ./my-project
  a4e8601 2026-08-29 ✖ block claude-session-url     https://claude.ai/code/session_a1b2c3d4e5f6
  a4e8601 2026-08-29 ✖ block claude-generated-with  🤖 Generated with [Claude Code](https://claude.com/claude-code)
  a4e8601 2026-08-29 ✖ block claude-coauthored      Co-Authored-By: Claude <[email protected]>
  24a220d 2026-08-26 ⚠ warn  ai-coauthored          Co-Authored-By: Cursor <[email protected]>
2/12 commit(s) contain AI attribution artifacts (4 finding(s)).
Prevent future artifacts in this repo: aisweep hook install

Exit code is 1 when artifacts are found, so it drops into CI and scripts directly. Useful flags: --all (scan all refs), --rev, --since, --max N, --json, -o REPORT.md.

Block polluted commits before they land

$ aisweep hook install
commit-msg hook installed at .git/hooks/commit-msg (default (block-level rules only))

$ git commit --allow-empty -m "fix: quick patch" -m "https://claude.ai/code/session_zz9x8y7w6v"
aisweep: commit message contains AI attribution artifacts:
  ✖ block claude-session-url     https://claude.ai/code/session_zz9x8y7w6v
commit blocked. Edit the message (drop the artifacts or configure .aisweep.json), then commit again.

$ git commit --allow-empty -m "fix: quick patch"   # clean messages pass

The hook enforces your policy: --strict also blocks warn-level rules, and .aisweep.json can flip any rule either way.

Shareable reports

$ aisweep report ./my-project -o REPORT.md --json-output REPORT.json
report: 4 finding(s) across 12 commit(s) -> REPORT.md

Markdown for humans, JSON for machines — handy before open-sourcing a repo that was built with AI assistance.

Default rules

Rule	Severity	Catches
claude-session-url	block	session URLs like https://claude.ai/code/session_… (#66504)
claude-generated-with	block	the Generated with Claude Code attribution line
claude-coauthored	block	Co-Authored-By: Claude <[email protected]> trailers
ai-coauthored	warn	Co-Authored-By trailers naming other agents (Cursor, Copilot, Aider, Gemini, Codex, …)
ai-generated-with	warn	Generated with <tool> lines for other agents

Matching is case-insensitive and line-anchored, so ordinary prose (a link to the docs, a colleague named "Claude" in a normal co-author line) is not swept up by accident. See configuration for the escape hatches.

Configuration

aisweep init writes a starter .aisweep.json at the repository root:

{
  "disable": ["ai-coauthored"],
  "severities": {"ai-coauthored": "block"},
  "allow_patterns": ["^co-authored-by:.*\\bmy-team-bot\\b"],
  "extra_rules": [
    {
      "id": "acme-deploy-bot",
      "description": "our deploy bot trailer",
      "pattern": "^deployed-by: acme-bot",
      "severity": "warn"
    }
  ]
}

• disable — switch built-in rules off (e.g. your team wants attribution).
• severities — per-rule block / warn override.
• allow_patterns — lines matching these regexes never produce findings.
• extra_rules — your own patterns, compiled like the built-ins (case-insensitive, multiline).

How it works

flowchart LR
    A[git commit] --> B[commit-msg hook]
    B --> C{aisweep check}
    C -- clean --> D[commit lands]
    C -- artifacts --> E[commit blocked]
    F[aisweep scan] --> G[findings / MD+JSON report]

Loading

scan reads commit messages via git log and matches every line against the effective ruleset; check does the same for a single message file inside the commit-msg hook. aisweep is read-only: it never rewrites, amends or stages anything.

FAQ

Does it rewrite my history? No — that is out of scope by design. Use git-filter-repo for rewrites; aisweep gives you the precise map of what to rewrite and keeps new pollution out afterwards.

Is this anti-AI? No. The request behind #66504 is that attribution should be opt-in. aisweep enforces exactly that: your repository, your policy. If your team wants attribution, configure it in.

What about false positives? Patterns are narrow and line-anchored; anything that slips through can be allowed via allow_patterns or downgraded via severities.

Does it work on Windows? scan, check and report run anywhere Python 3.10+ runs. The hook is a POSIX shell script and works in Git Bash environments. CI tests on Linux.

Which tools does it cover? Any tool whose artifacts end up in commit messages — the rules match shapes, not vendors, and extra_rules covers anything we missed.

Roadmap

• Publish to PyPI
• GitHub Action wrapper (scan a repo in CI without installing)
• HTML report with filtering
• PR-description audit via GitHub API
• pre-commit framework integration

Contributing

Issues, bug reports and PRs are welcome — see CONTRIBUTING.md for the development setup, commit style and test requirements. Please note the Code of Conduct and the security policy.

License

MIT © 2026 w1977-0

Acknowledgments

• anthropics/claude-code#66504 — the feature request that documented the default-on session URL behaviour
• the Hacker News thread that surfaced how many people care about this
• git-filter-repo — the right tool for the rewriting part aisweep deliberately leaves out
• Contributor Covenant — our Code of Conduct template