ProofTicket: Turning Agent Work Into Reviewable Agent Tickets

This is a submission for the GitHub Finish-Up-A-Thon Challenge.

What I Built

ProofTicket is a finish-up layer for unfinished projects: ticket the improvement, let someone else spend the tokens, and approve the proof before it lands.

The idea is simple: stop hiding the half-finished project because you do not have the energy to reopen the whole repo. A friend can suggest the improvement they keep telling you to add. An AI coworker can spend the tokens to make a scoped attempt. Each ticket carries its own instructions, approval path, artifacts, and proof. A human still gets the final approval step before the work becomes shared project state. Next time your friend has an idea, send a ticket, not the Git repo.

Opening a ticket with a company already has the right shape: here is the problem, here is the context, here is who owns the next move. ProofTicket applies that same shape to AI coworking on GitHub.

The finished slice includes agent-submitted tickets, human approval, receipt inspection, evidence export, proof-style profile cards, signed GitHub and machine-webhook ingestion, a local MCP adapter, a five-minute demo path, and a deployed login path for reviewers.

Demo

Live demo: https://prooft.vercel.app

Repo: https://github.com/dodge1218/pt

npm install
npm run setup:local
npm run dev

In another terminal:

bash examples/five-minute-demo/print-demo-commands.sh

The demo path:

1. Register a local demo agent.
2. Submit an agent-created ticket with evidence.
3. Inspect the pending action receipt.
4. Approve the action.
5. Open the resulting ticket in the app.
6. Export the evidence bundle.

The Comeback Story

ProofTicket started as the core concept: proof attached to a ticket. The broader idea is still coordination, but the useful primitive is sharper: a ticket that lets someone else propose the finish-up work while you keep the approval step.

The old project grew into a broader collaboration product with too many futures attached: matching, context manifests, payouts, ZK proofs, GitHub sync, hosted SaaS controls.

The useful problem was smaller:

If humans and agents are both doing real work, where does that work live once it needs review, approval, evidence, or follow-up?

The finish-up work was cutting the project down to that workflow. Before, it was a broad coordination app. After, it is a focused prototype where a person or agent proposes a concrete improvement, spends the tokens or effort to make it reviewable, evidence stays attached, and the project owner approves or rejects the result.

Also, genuinely, thank you to the people who made this challenge. Of course I waited until the last minute to submit. People like me needed this exact kind of finish-up-a-thon: not another blank-page hackathon, but an extra nudge to pull an old thing out of the half-finished pile and make it coherent enough to show someone.

My Experience With GitHub Copilot

In April, I bought 14 Copilot Pro+ subscriptions for about $40 each. I have thought about it a lot, and I am probably the exact customer that makes a loss-leading pricing model stop working. I am not the best customer. Thanks again, though.

I wrote about that whole rabbit hole in The Best $40 Addendum. Then I was locked out of GitHub Copilot for about a month after writing multiple articles about how useful it was. It felt like the Vitalik warlock story in miniature: one platform switch changes, and you suddenly remember how much of your workflow depends on a centralized service.

GitHub Copilot was material here. The first half of ProofTicket came together in a Copilot Opus 4.6 assisted pass. The finish-up half came together in a Copilot GPT 5.5 assisted pass with a tighter target: stop expanding the idea and make the actual demo path work.

I am very happy Copilot is back up again, even if it is a bit more pricey.

The lesson was not "AI wrote the app." The useful part was turning vague ambition into testable steps: local setup, typed agent actions, human approval, receipt inspection, evidence export, and public copy that does not overclaim what the prototype is.

"Build an AI coworking app" is vague. "Make a five-minute flow where an agent submits work, a human approves it, and an evidence bundle exports cleanly" is concrete enough to finish.