who's talking?

Somasundaram Mahesh

• Founding AI Engineer at AGI.INC
• Formally Android engineer. Building mobile products for 8+ years.
• Ships ToonGen, The Roaming King and many more with agents in the loop.
• @msomuin on X · msomu.com · github.com/msomu/receipt

the new failure mode

It writes a paragraph about itself

this morning, on this plugin

Two lies. Same CLI. Two reviews.

• screenshot always install -r + am start, then ok if any PNG existed
• failed install could still exit 0
• logcat always returned "ok": true
• on a real app, verify test was just assemble

The CLI looked like infrastructure. It was a story.

Verification is infrastructure. Lauren Tan · The Complete Guide to pstack Pt. 1 · 31 Aug 2026

what it is

A Cursor plugin. A local CLI. A Feature Map.

github.com/msomu/receipt · MIT · Android and Kotlin Multiplatform

the loop

Install is not the product

• Install the plugin once
• /create-receipt writes .cursor/skills/verify-<app>/ — real CLI + 3–5 feature files
• /receipt on every bug, feature, or KMP test — no proof, not done
• /maintain-receipt daily so the map stays true to the app
• Next agent starts cold with an honest harness

done means files

.receipt/proof/<id>/

honest CLI

The commands that cannot flatter

cd sample
./verify doctor
./verify test
./verify assemble
./verify screenshot --install --launch   # cold start
./verify screenshot                      # current frame only
./verify logcat                          # pidof the package

JSON. --dry-run. --help. AdbHarbor lease if present — never --force. Proof survives cleanup.

this Mac

Shell to the phone. No MCP.

Cursor on this Mac
  --shell-->  ./verify  /  adbharbor
  --lease-->  phone

Harbor owns :5037. Never raw adb.

cloud agent

The phone stays on the desk

Cursor Cloud
  --https-->  tunnel
  -->  harbor /mcp on this Mac
  --lease-->  same phone

adbharbor expose --via cloudflare. Paste URL + token.

three repos

Harbor serves. Receipt asks.

demo · 15 minutes

Fifteen minutes. Files on disk.

• 0:00 — read the first sentence of the README out loud
• 1:00 — ./verify doctor — JSON, not a paragraph
• 3:00 — ./verify test — XML + log under .receipt/proof/
• 5:00 — assemble, screenshot --install --launch, increment, screenshot again
• 8:00 — /create-receipt on github.com/msomu/base-mvvm-android-kt (Compose MVVM template)
• 10:00 — stop. Clone it. Watch the agent prove the change.

what you should see

Counter 1. Not Counter 0.

If the PNG does not open as a PNG, ok is false. Size is not a check. Magic bytes are.

in production

Two real apps.

Android screenshots on ToonGen are still unproven on a device. The CLI will not fake one.

daily

A maintain pass that never fails is untested

• Cursor Automation — schedule /maintain-receipt
• or a cloud agent, same prompt, once a day
• re-drive the Feature Map, fix drift, one PR or “clean”
• the first real pass must be able to go red

homework

One repo. One harness.

• Clone github.com/msomu/receipt
• Cursor → Plugins → Add from GitHub → msomu/receipt
• In your Android or KMP repo, run /create-receipt
• Next task: /receipt. No .receipt/proof/, not done.
• Tomorrow: /maintain-receipt