Owner-run pilot / October 14, 2026

From a finding
to a checked fix.

Your agent fixes the code. arnx runs the tests that check it.

On a real Python pipeline backed by Postgres, arnx generated tests for suspected bugs, the owner judged the findings and answered questions, and their coding agent made the fixes. The final sandbox check recorded 54 passed, zero failing and one skipped. The accepted test files stayed unchanged after acceptance.

The change

A token-callout prediction pipeline spanned 12 commits and 29 files, with 3,311 lines added. It involved hourly snapshots, polling, model training, live scoring and advisory messages. arnx set up five projects; Postgres was available to the aggregator tests, and the owner approved the environment.

The review used Opus, additional hunters and a second-model assessment with a fallback. A dropped connection was handled by resuming the review.

See the failure.

Tests reproduced an inverted Gini calculation, callouts lost beyond a 2,000-item page cap, stale prices used in evaluation and training folds that split the same hour. Where behavior needed a database, tests ran against Postgres in the approved sandbox.

The owner marked 18 findings real, including four the second model had disputed. Those are finding marks, not a count of distinct bugs.

Answers about intent mattered too. The owner wanted the price at the time a signal fired. A generated test showed that the implementation didn't match that answer, so arnx retained it as a failing test. A response about callout ID types didn't settle intent; that question stayed open and produced no test.

Check the fix.

arnx accept retained tests and wrote a fix prompt containing their locations, the findings, owner answers and the unresolved question. The owner's coding agent received that prompt, with instructions to change product code and leave the accepted tests intact.

The agent changed nine product files. arnx then ran the accepted tests in the approved environment, including Postgres.

The skipped test corresponded to a finding not confirmed. Accepted test files were unchanged after acceptance. A temporary schema-path workaround was removed after an arnx acceptance-path fix; the check still passed afterward.

Keep the test.

The product fix was accompanied by regression tests in the repository. The tests used to expose the accepted failures helped check the changes. These remain ordinary, editable test files, ready to run as the project evolves.

Time and cost in this run

Setup, review and acceptance totaled $5.36 in reported model spend. Owner review time and the fixing agent's time and cost were not recorded. Local execution still uses compute resources.

What this establishes

The complete workflow worked on one real, owner-judged development change. The branch had been used to tune arnx, and the run crossed builds as acceptance defects were repaired.

Two known issues were missed. The owner's marks do not establish a general precision rate. A separate held-out comparison tied the earlier build at 8 of 11 confirmed catches.

The next step is a controlled pilot with selected developers: repeat the workflow on their changes, measure the effort, and learn whether they keep the tests and come back for another review.