FlightBox — AI skill for Claude Code

**Black-box flight recorder for AI agents** — record every LLM call your agent makes, replay sessions deterministically, and export a redacted evidence report when something breaks.

FlightBox is local-first. Recordings live in SQLite. No hosted dashboard is required.

Why

An agent failed and nobody can reproduce it. The final answer is in a log, but the interesting evidence is scattered across LLM requests, tool calls, model responses, timing, tokens, and local notes.

FlightBox gives you a deterministic debugging trail:

• record OpenAI / Anthropic / LiteLLM calls
• replay the same responses later
• diff two runs
• export JSONL or pytest replay tests
• generate a redacted Markdown / HTML report for PRs, CI notes, and teammates

Quick Start

pip install flightbox

Record

import flightbox
from openai import OpenAI

client = OpenAI()

with flightbox.record("debug-session") as rec:
    response = client.chat.completions.create(
        model="gpt-4o",
        messages=[{"role": "user", "content": "What is 2+2?"}],
    )
    print(response.choices[0].message.content)

print(f"Recorded as run: {rec.run_id}")

Replay

import flightbox

with flightbox.replay("abc123def4"):
    response = client.chat.completions.create(
        model="gpt-4o",
        messages=[{"role": "user", "content": "What i