fleet-debrief

The morning debrief for your AI agent fleet. One command. It reads the session transcripts Claude Code already writes to disk and answers three questions, with nothing to install into your workflow and nothing to configure:

1. What did my agents do while I was away? Every session in the window — cost, duration, files touched, subagent activity, and the runs that need a second look (hook errors, denials, unparseable records) sorted to the top.
2. What exactly happened in this run? A step-by-step timeline of any session, subagent turns included, with its blast radius: every file the run edited, including edits that were later reverted — and the ones you then fixed by hand.
3. Why did run A work when run B didn't? Pick two runs and get an aligned, side-by-side diff: the step where they diverge, the tool-call deltas, the files only one of them touched, and the cost difference.

Local-only. Zero config. Nothing to sign up for, nothing leaves your machine.

Above: the same task run twice. The debrief lists both, you tick two boxes, and the diff shows where they parted company — one edited the checkout client, the other went hunting and retried at the wrong layer for three times the money. The comparison aligns runs on tool-call structure rather than wording, so the divergence it reports is a difference in what the agent did. Note what that means: two runs with the same shape and different content read as fully aligned, so the step count is shown alongside the verdict rather than a bare percentage.

The recording uses a synthetic corpus, because real transcripts carry paths and project names that cannot be published.

Quickstart

npx fleet-debrief          # -> http://127.0.0.1:7317

That's it. If you've ever run Claude Code on this machine, the debrief is already populated — it reads ~/.claude/projects retroactively, so your last weeks of sessions are there before you configure anything.

It installs three packages and binds in well under a second. The server ships compiled, so nothing is built at startup, and the only runtime dependencies are the HTTP layer.

Without a browser

npx fleet-debrief --print              # the debrief as terminal output
npx fleet-debrief --print --since 7d   # a wider window

Output is plain text with no colour, so it survives a pipe, a redirect, and a paste into an issue.

Options

--print          write the debrief to stdout and exit, no browser
--since <window> window to report on, e.g. 12h, 48h, 7d  (default 24h)
--port <n>       port for the local UI and API            (default 7317)
--version        print the version and exit
-h, --help       print this message and exit

--since also decides the window the UI opens on, so npx fleet-debrief --since 7d lands on a week rather than a day.

From a checkout

git clone https://github.com/kpachhai/fleet-debrief  # pii-allow:own-repo-url
cd fleet-debrief
npm install && npm run build
npm start

Privacy posture

Session transcripts are among the most intimate records a developer keeps: every prompt you typed, every file the agent touched. So:

• The server binds 127.0.0.1 only, and /api/* rejects requests whose Host or Origin is not this machine (DNS-rebinding and CSRF defense).
• No outbound network calls, at all. Prices come from a table vendored into the repo with its as-of date shown on screen, never fetched.
• Read-only: every filesystem call is a stat, a readdir, or a read. Your transcripts are someone else's files; this tool never writes to them.
• No accounts, no telemetry, no CDN assets. The UI works with the network off.

Status

v0. Claude Code is the only supported source today; the transcript reader, session model, and pricing engine are extracted from agentic-os, where they run against real multi-hundred-megabyte corpora. The run-vs-run diff aligns on tool-call structure (deliberately, not on prose — see the comment in server/diff.ts).

Roadmap, in order: budgets and spend alerts on top of the debrief; more coding CLIs as sources; then hash-chained run logs with optional public-ledger anchoring — turning the debrief you read into evidence you can hand an auditor.

Development

npm run dev:server   # API with reload on :7317
npm run dev:ui       # Vite dev server, proxies /api
npm run typecheck && npm test

Contributions welcome. Two rules inherited from the parent project: never commit fixtures captured from real transcripts (synthetic only — see tests/), and nothing may make an outbound network call.

License

Apache-2.0.