ai-decision-tracer was built for production autonomous AI systems where understanding agent behavior is critical for security and compliance.

All steps are auto-saved to JSONL as they complete. If your agent crashes mid-execution, you still have a full trace up to the failure point. Uses atomic writes (temp file + rename) to prevent partial/corrupt output.

All trace data stays local. No network calls. No telemetry. No cloud sync. Your agent decisions, inputs, and outputs never leave your machine.

Stack traces from exceptions are captured in the trace file for postmortems, but the library never logs raw prompts or full model responses unless you explicitly pass them to step.log(). You control what gets recorded.

Zero third-party dependencies — nothing to supply-chain attack. Pure Python stdlib only.

Please report security vulnerabilities via GitHub Issues (mark as "security"). Do NOT include sensitive data (API keys, prompts) in public issues.

Expected response time: 48 hours.