Failure captures write raw session content to disk, secrets included

[ebrett]

My agent Charlie uncovered a secrets issue and recommended the following, which I have implemented locally. It also suggested it could make a PR to share, if there is interest.

https://gist.github.com/ebrett/3ddeeef342729daef219aa2bc473cc3b

A small fix for a gap in PAI's learning/capture path: PAI/TOOLS/FailureCapture.ts persists raw session content to disk, so any secret that appeared in a tool's output during a session gets written verbatim into the failure record — and those records are exactly what gets synced to git or backed up.

[xg-gh-25]

This is a critical security anti-pattern. Crash dumps should never include raw session content — especially when that content may contain API keys, auth tokens, or PII.

Why this happens:
Most AI frameworks log full context on failure for debugging, without scrubbing. The assumption is "crash = developer sees it", but crash reports often end up in:

• GitHub issue attachments
• Shared debugging channels (Slack, Discord)
• Log aggregation services (CloudWatch, Datadog)

Fix approach:

1. Pre-crash sanitization:

def sanitize_for_crash_log(context):
    # Redact known secret patterns
    context = re.sub(r'(api[_-]?key|token|password)[:=]\s*[^\s]+', r'\1=REDACTED', context)
    # Truncate long values
    return context[:5000] + '...' if len(context) > 5000 else context

2. Use structured logging with field-level control:

logger.error("Session crashed", extra={
    "session_id": session.id,
    "error": str(e),
    "context_hash": hash(context),  # Not the content!
    "message_count": len(messages)
})

3. Never write raw context to disk without encryption:
   If you must persist for debugging, encrypt first:

encrypted = fernet.encrypt(json.dumps(context).encode())
Path("crash.enc").write_bytes(encrypted)

Immediate mitigation:

1. Audit your failure capture code for .write_text(context) patterns
2. Add secrets scanner pre-commit hook (truffleHog, detect-secrets)
3. Update docs to warn users: "Never share crash dump files publicly"

Would a PR for sanitized crash logging be helpful?

---

Shared from SwarmAI. Discussion: T-MEM: Memory Persistence

[ebrett]

The gist in the original post shows how I fixed it using 'redactSecrets', similiar to you suggested 'satinize_for_cash_log' function. I'm not sure PR's are being accepted right now. Not my call to make.