AIFR — AI Flight Recorder

OpenTelemetry for AI Software Development

Record, replay, and analyze AI-assisted coding sessions. Not chat transcripts — structured execution graphs.

AIFR captures every prompt, terminal command, code diff, test result, and retry in an AI coding session, producing a replayable, auditable event stream.

Screenshots

Home	Timeline

Prompt-to-Diff	Execution Graph

Session Overview	Terminal Replay

Install

npm i -g https://github.com/GeziP/AI-Flight-Recorder/releases/download/v0.2.0/aifr-0.2.0.tgz

Quick Start

1. Import existing sessions

If you already use Claude Code, Codex CLI, or Cursor, import history in one command:

cd your-project
aifr init
aifr import claude          # Import all Claude Code sessions
aifr import codex           # Import all Codex CLI sessions
aifr import cursor          # Import all Cursor AI sessions
aifr import claude --limit 5

2. Launch Web UI

aifr ui
# Opens http://localhost:3000

The Web UI provides these views:

View	Description
Timeline	Chronological event stream with search and filters
Graph	Execution graph with Dagre layout and node detail panel
Prompt-to-Diff	Which AI prompt caused which code change
Overview	Session metrics: attribution rate, retries, test results
Diff	Side-by-side code diff for session-scoped changes
Replay	Terminal playback with speed control and event markers
Report	Generated markdown session summary

3. Search across sessions

aifr search "authentication"
aifr search "failed" --type test
aifr search "src/auth" --type diff

Full-text search powered by SQLite FTS5, across all events in all sessions.

4. Other commands

aifr status                # List all sessions
aifr diff [session]        # View code changes in a session
aifr replay [session]      # Replay terminal output
aifr report [session]      # Generate markdown report
aifr redact [session]      # Strip secrets from events
aifr export [session]      # Export as .aifz archive

Supported AI Tools

Tool	Import	Live Recording
Claude Code	Yes	Yes
Codex CLI	Yes	Yes
Cursor	Yes	Planned

Command Reference

aifr init                Initialize .aifr/ in current project
aifr start               Start terminal recording
aifr status              List recorded sessions
aifr import <agent>      Import sessions (claude, codex, cursor)
aifr search <query>      Full-text search across sessions
aifr graph [session]     Build execution graph
aifr analyze [session]   Run session analysis
aifr report [session]    Generate markdown report
aifr redact [session]    Redact secrets from events
aifr replay [session]    Replay terminal output
aifr diff [session]      View code changes
aifr export [session]    Export as portable archive
aifr ui                  Launch Web UI

Data Storage

All data stays local. Nothing is uploaded.

.aifr/sessions/
  20260526_120011/
    events.jsonl          # Structured event stream (JSONL)
    terminal.log          # Raw terminal output
    metadata.json         # Session metadata (timestamps, agent, git ref)
    graph.json            # Execution graph
    analysis.json         # Session analysis results
    report.md             # Generated report
    git/
      before.patch        # Git diff at session start
      after.patch         # Git diff at session end

Event Schema

All events are normalized to a unified JSONL stream:

• PromptEvent — AI request content
• CommandEvent — Terminal command execution
• DiffEvent — Code changes (file list + patch)
• ToolEvent — Tool/API calls
• TestEvent — Test results
• TerminalOutputEvent — Terminal stdout/stderr
• RetryEvent — Retry attempts
• SessionEvent — Session lifecycle (start/end)

Secret Redaction

aifr redact scans events for 12 categories of secrets:

AWS keys, GitHub tokens, JWTs, Bearer tokens, API keys, private keys, passwords in URLs, password assignments, email addresses, Slack tokens, Google API keys.

Development

pnpm install             # Install dependencies
pnpm build               # Build all packages
pnpm dev:web             # Start Web UI dev server
pnpm test                # Run all tests

Known Limitations

• Imported session git patches reflect project state at import time, not the original session time
• Codex sessions contain commands but no user prompts, and cannot locate the original project directory
• Prompt-to-Diff mapping is inferred (temporal proximity), not provenance-tracked
• aifr start requires a real terminal (no pipe stdin)

License

MIT