trace-sanitizer 0.1.0

Sanitize AI agent trajectories for traced.run

Trace-Sanitizer

Sanitize your AI agent trajectories and publish misaligned or funny trajectories on traced.run.

Supports Claude Code, Codex, Gemini CLI, and OpenCode.

Install

pip install trace-sanitizer

1. Single trajectory

Export one trajectory interactively — no configuration needed.

trace-sanitizer export

This auto-detects all sources, lets you pick a project and session, asks about tool output inclusion, and exports a sanitized JSONL file. Review it for PII, then publish it on traced.run.

Agent prompt

Paste this into Claude Code or any coding agent:

Help me export a conversation using trace-sanitizer.

1. pip install trace-sanitizer
2. trace-sanitizer update-skill claude
3. trace-sanitizer export
   (Interactive menu — pick a project and session.)
4. Review the exported JSONL for PII.
5. Tell the user how to publish it at https://traced.run

2. BETA: Donate multiple trajectories

This is beta.

Real world coding trajectories are important for early detection of AI misalignment, failures, and scheming in the wild. Your trajectories might contain relevant examples that can make a real difference for research. Trace-sanitizer is maintained by Dr. Joachim Schaeffer. I am currently a MATS 9.0 Fellow. I am talking to AI safety research institutes about formal data contribution pipelines. A more structured process will follow.

But if you want to get active now already, bulk donation works like this:

# Export all trajectories (auto-detects sources, includes tool outputs)
trace-sanitizer export --all

# Review the export for PII
trace-sanitizer confirm \
  --full-name "YOUR NAME" \
  --attest-full-name "Scanned for full name." \
  --attest-sensitive "Checked for company names and private URLs." \
  --attest-manual-scan "Sampled 20 sessions manually."

This creates a traced_export/ folder containing trajectories.jsonl and a LICENSE-CDLA-Permissive-2.0.md. Zip the folder and email the link to [e-mail].

You can customize the export with additional flags:

trace-sanitizer export --all --source claude          # only Claude Code
trace-sanitizer export --all --no-tool-outputs        # exclude tool outputs
trace-sanitizer export --all -o my_export.jsonl       # custom output path
trace-sanitizer config --exclude "secret-project"     # exclude specific projects
trace-sanitizer config --redact "CompanyName"          # redact custom strings

All commands

Command	Description
trace-sanitizer export	Interactive: pick a project and session
trace-sanitizer export --all	Export all sessions to traced_export/ folder
trace-sanitizer export --source claude	Filter by source
trace-sanitizer export --include-tool-outputs	Include tool results
trace-sanitizer export --no-thinking	Exclude extended thinking
trace-sanitizer prep	Discover projects (JSON)
trace-sanitizer list	List projects with exclusion status
trace-sanitizer config --source all	Set source scope
trace-sanitizer config --exclude "a,b"	Exclude projects
trace-sanitizer config --redact "str"	Add custom redaction strings
trace-sanitizer config --redact-usernames "u"	Anonymize usernames
trace-sanitizer config --confirm-projects	Confirm project selection
trace-sanitizer confirm ...	PII scan + review attestations
trace-sanitizer status	Show current stage (JSON)
trace-sanitizer update-skill claude	Install skill for Claude Code

What gets redacted

1. Paths — stripped to project-relative
2. Usernames — macOS username + configured handles replaced with hashes
3. Secrets — JWT tokens, API keys (Anthropic, OpenAI, HF, GitHub, AWS, etc.), private keys, webhooks
4. Entropy — high-entropy strings flagged as potential secrets
5. Emails — personal addresses removed
6. Custom — your configured strings and usernames

Not foolproof. Always review before sharing.

Data schema

Each JSONL line is one session:

{
  "session_id": "abc-123",
  "project": "my-project",
  "model": "claude-opus-4-6",
  "start_time": "2025-06-15T10:00:00+00:00",
  "messages": [
    {"role": "user", "content": "Fix the login bug"},
    {"role": "assistant", "content": "I'll investigate...", "tool_uses": [{"tool": "Read", "input": "src/auth.py"}]}
  ],
  "stats": {"user_messages": 5, "assistant_messages": 8, "tool_uses": 20}
}

License

MIT — forked from dataclaw by Banodoco.