A real recorded run · nothing edited

Watch an AI agent’s change get checked.

An AI agent got a real ticket. A second AI, from a different company, reviewed its work. The tests actually ran. Then OpenThunder said what was proven, what was not, and why.

Every number on this page is read from the run’s own files, and you can download all of them below. You can also run it yourself.

Run it yourselfSee the raw evidence
Ran a command actually executed Checked deterministic code analysis AI opinion a model’s judgment, not a test Not checked no evidence either way
1

Understand the code before anything changes

Before an agent touches the repo, OpenThunder reads it: what it does, which files are sensitive, and whether the tests pass today. No AI and no account are involved; this is the free local check.

2

The ticket the agent got

This is the exact request, written the way a teammate would write it. It does not hint at edge cases.

3

What the agent changed

4

What the checks found

Three different kinds of checks, each labeled for what it is: the tests that actually ran, an independent AI review, and OpenThunder’s own deterministic analysis.

5

What happened next

6

The verdict: what is proven, and what is not

→

Why this matters when you build with AI

A green check is not proof

An agent saying “done, tests pass” tells you nothing about the cases nobody tested. OpenThunder separates what ran from what was only claimed.

A second opinion that is not the author

The reviewer here is a different model from a different company, reading the actual diff, so it is not grading its own homework.

Evidence you can hand to a reviewer

Every run leaves files a teammate can check: the diff, the test logs with exit codes, the review, and the verdict.

Download OpenThunder freeCheck your own change, no install
⇓

The raw evidence

Everything above comes from these files. Nothing was rewritten for the page. The one change: local file paths from the recording machine are replaced with placeholders such as <repo> and <server-data>.

▶

Run it yourself

The demo repo is small: one service, one test file. Try the keyless checks first, with no account and no AI key.

curl -LO https://openthunder.dev/proof/agent-run/session-tokens-demo.zip
unzip session-tokens-demo.zip && cd agent-run-demo
git init -q && git add -A && git commit -qm baseline
npm test                                   # the tests, before any change
npx @skillstech/openthunder brief          # what this repo is
npx @skillstech/openthunder risks          # which files are sensitive

To run the agents, open the folder in OpenThunder Desktop and start a mission with the ticket from step 2, with npm test as the check. That needs your own AI provider keys. A model writes different code each time, so your run will not match this one line for line; the checks and the evidence work the same way.

Recording this caught bugs in OpenThunder itself. Our first recordings exposed problems in our own checks, so we fixed them before recording the run on this page. Every earlier attempt is still listed below.
  • The OpenAI reviewer crashed on current models, and it was shown only the agent’s summary instead of the code.
  • Reviewers marked every requirement as met, whatever they had found.
  • Every project with a test command was reported as “missing tests”, even when the new code was fully covered.
  • A filename guess (“no test file named after this file”) outranked measured coverage, and was mislabeled as a security finding.
  • Secret redaction hid ordinary code like token = newToken() from the reviewer.
  • After a fix round, the reviewer saw only that round’s files, and problems it had already seen fixed still blocked the change.