Lightbox

Flight recorder for AI agents. Append-only, tamper-evident tool execution logs.

What it is / isn't

• Records tool calls (send_email, read_file, api_call) - not LLM prompts/responses
• Append-only JSONL with SHA-256 hash chaining
• Offline verification - no network needed
• Local-first - sessions are portable directories you can zip and share
• Not a debugger, profiler, or execution controller - recording only

Quickstart

pip install lightbox-rec

SDK Usage

from lightbox import Session

session = Session()
session.emit("send_email", {"to": "user@example.com"}, {"success": True})
session.emit("read_file", {"path": "/data.txt"}, {"content": "Hello"})

LangChain Integration

from lightbox.integrations.langchain import wrap

agent = wrap(your_agent)  # Automatically records tool calls
agent.invoke({"input": "Do the thing"})

CLI

lightbox list                    # Show all sessions
lightbox show <session>          # Display events
lightbox replay <session>        # Animated playback
lightbox verify <session>        # Check integrity

Where logs live

~/.lightbox/sessions/<session_id>/
    events.jsonl     # Append-only event log
    meta.json        # Session metadata

Override with LIGHTBOX_DIR environment variable.

Exit codes (verify)

Redaction

from lightbox import Session, RedactionConfig

config = RedactionConfig(
    redact_keys=["api_key", "password"],  # Always redact
    max_inline_bytes=64 * 1024,           # Hash content over 64KB
)
session = Session(redaction_config=config)

Documentation

• Event Format - Event schema and canonicalization rules
• Integrations - LangChain and framework usage
• Security - Threat model and verification
• Changelog - Version history

License

MIT - see LICENSE