agent-recorder

agent-flight-recorder

Every agent session you run disappears the moment it ends. This one doesn't.

A CLI that records every Claude Code and Codex session - prompts, tool calls, shell commands, file changes, errors and token costs into a searchable SQLite database. Find what you worked on, see what failed and extract reusable workflow patterns.

---

The problem this solves

AI coding sessions are opaque. When a session ends, all you have is changed files and a vague memory of what the agent tried. There's no way to answer:

• What tools did the agent call, and in what order?
• Which shell commands failed, and what was the error?
• How much did that session actually cost in tokens?
• Why does this same class of problem keep taking 3 sessions to fix?

afr records all of that and keeps it queryable - locally, permanently, without sending anything to a server.

---

Install

git clone https://github.com/AravindKurapati/agent-flight-recorder
cd agent-flight-recorder
pip install -e .

Requirements: Python 3.11+ - no API keys, no accounts, everything stays on your machine.

---

Usage

Ingest your sessions

afr ingest claude     # reads ~/.claude/projects/**/*.jsonl
afr ingest codex      # reads ~/.codex/sessions/**/*.jsonl

Or wire a Claude Code hook to ingest automatically after every session (add to ~/.claude/settings.json):

{
  "hooks": {
    "Stop": [{ "matcher": "", "hooks": [{ "type": "command", "command": "afr ingest claude" }] }]
  }
}

Browse recent runs

afr list --days 7

Inspect a session

afr show 50c3f2a1

Search across all sessions

afr search "authentication"
afr search "modal" --days 30

Resume a session

afr list/afr search show an 8-char id prefix, but claude --resume needs the full session id and must run from the directory the session started in. afr resume resolves the prefix to the full id plus the recorded working directory and hands you the exact command:

afr resume 6c84c429
#   Session: 6c84c429-b7cb-4685-af6d-a318a2def2fc  (claude)
#   branch: main
#   Resume with:
#     cd "D:\Aru\NYU\prosper-podcast" && claude --resume 6c84c429-b7cb-4685-af6d-a318a2def2fc

afr resume 6c84c429 --run    # launch it now instead of printing

An ambiguous prefix lists the matches so you can add characters. Codex sessions get a codex resume <id> command instead. Sessions recorded before v0.2.0 backfill their directory on the next afr ingest; if the original session file is gone, resume still gives you the full id to run from the project directory yourself.

See patterns across sessions

afr stats --days 30

See your 5-hour usage windows

afr config set weekly-reset "Wed 00:00"      # when your weekly cap resets
afr config set timezone "America/New_York"   # your display timezone (IANA)
afr windows

afr windows reconstructs the Anthropic 5-hour usage windows you actually opened (from recorded run timestamps) and shows how many fresh windows fit before your weekly reset. This is the time/window view; for the token-budget forecast, use claude-burnrate with /usage.

Export a session as markdown

afr export 50c3f2a1              # prints to stdout
afr export 50c3f2a1 --out report.md   # writes to file

Valid outcomes: shipped, blocked, abandoned, exploratory

Extract reusable workflow skills

After a few sessions solving the same class of problem, run:

afr extract-skills --min-runs 3

Candidate: deployment-modal-debug
  4 sessions | tools: Bash, Read | errors: 6
  Generate SKILL.md? [y/n]: y
  Written → generated_skills/deployment-modal-debug/SKILL.md

It clusters sessions by keyword similarity, finds ones that ended in shipped, and drafts a SKILL.md from the successful tool sequence. You approve before anything is written.

---

All commands

<id> accepts the first 8 characters from afr list output.

---

What gets recorded

For each session:

• Goal - first user message
• Location - the working directory and git branch the session ran in (enables afr resume)
• Tool calls - every tool fired, input summary, success or error
• Shell commands - command, exit code, stdout/stderr excerpt
• File events - every read, write, patch, or delete
• Errors - failed tool calls and non-zero shell exits
• Token counts - input, output, cache read, cache write
• Outcome - you tag this: shipped, blocked, abandoned, exploratory

Secrets are redacted before anything is written to the database (API keys, bearer tokens, private keys, .env contents).

---

How data is stored

Everything lives at ~/.afr/afr.db - a single SQLite file on your machine. No data leaves your machine.

You can query it directly with any SQLite client:

sqlite3 ~/.afr/afr.db "SELECT user_goal, outcome, tokens_in FROM runs ORDER BY started_at DESC LIMIT 10"

---

Supported agents

---

License

MIT