Skip to content

About

See what your AI agent actually did. Execution replay built around App Composer contracts and Kujo evidence.

Topics

Resources

Security policy

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

Agent Black Box

See what your AI agent actually did.

An evidence-first execution replay: follow tool calls, inspect file changes, pause at a failure, and watch the recovery. Three independent Worker applications exchange versioned receipts using App Composer's real manifest and promotion contracts.

Replay Studio showing a failed verification

Narrated walkthrough

Watch the Siren / Eleven v4 walkthrough: follow a bug from failure to fix, then see how App Composer connects the three apps. Includes a mobile 480p download, transcript, captions and verification evidence.

Run locally

Node 22.13+ with node:sqlite (tested on Node 26.12.1), npm and Git:

npm ci
npm run dev

Open http://127.0.0.1:4173. The demo opens directly to the labeled failure/recovery example. Select Kujo verification · recorded for the genuine RunLedger receipt. Import a JSON receipt, inspect it, then choose Share report and explicitly approve the sanitized snapshot. A public report is read-only; “public” on localhost does not make it internet-accessible.

Private and public local stores live separately under ignored .private/data/. Restarting the server retains receipts, but rotates the local owner session. Reopen / for a fresh cookie.

Verify

npx playwright install chromium
npm test
npm run build
npm run test:e2e
npm run compose:check
node scripts/compose.mjs --promote-local

npm test includes real workerd service bindings, separate D1 databases, and an executable Studio update that retains existing receipts. Browser tests exercise import → failure → recovery → explicit publication → anonymous public replay, keyboard controls, reduced motion, safe text, mobile layout, accessibility and 1,000-event windowing.

Independent builds: npm run build --workspace @agent-black-box/replay-studio (likewise run-importer and receipt-publisher). Each app produces its own Worker artifact and manifest under dist/. UI assets belong to Studio. No cloud credentials are needed for these commands.

What is real

  • Composer's upstream MIT validators are vendored unchanged and actually executed. Worker service bindings exchange data in workerd. App-owned D1 persistence survives an independent component update.
  • The successful example comes from a real, isolated Kujo task recorded with RunLedger. Its command outputs are retained in evidence/recorded-execution.json; the receipt accurately discloses unavailable outputs and timing.
  • The failure/recovery example is synthetic, labeled throughout. Token usage and cost are unavailable. Nothing exposes hidden reasoning or executes imported code.
  • PackWrite compiled and validated the context pack through its documented offline manifest seam; no provider key was available. Loop ran the engineering gates and retained evidence.

What is not claimed

This repository includes the source, recorded walkthrough and a runnable local demo. No hosted application endpoint is deployed. Composer's current native stack editor restricts apps to a reviewed list; these three new apps require an operator-reviewed catalogue/mount extension. This project uses a small explicit local adapter rather than claiming native registration, provisioning, or production promotion.

Read more

MIT. Credit to Loftwah and App Composer for the independent-app architecture and reused contracts, and Kujo for execution evidence and engineering workflows. See THIRD_PARTY.md.

About

See what your AI agent actually did. Execution replay built around App Composer contracts and Kujo evidence.

Topics

Resources

Security policy

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages