Öge

01 The problem

Coding agents multiplied how much work gets executed. The bottleneck moved to the person supervising them: reading diffs, rerunning tests, and deciding whether "done" really means done.

An agent that checks its own work is grading its own exam. Its claims about passing tests are not evidence.

02 What I built

Öge drives the agents you already use (Claude Code, Codex, Antigravity) over their structured streams, on private copies of your repository, so nothing touches your working tree until you apply it.

In Standard mode a fresh QA agent sees only the task and the changed code, never the implementer’s reasoning, and can add held-out tests the implementer never sees. Öge runs the checks itself; failures feed a bounded repair loop.

A result is Accepted only on the evidence Öge collected. Every run produces a Receipt that states what was checked, what passed and what was not covered, exportable as Markdown for a PR or JSON for tooling.

03 Worth knowing

• Uses your existing agent login; never asks for, reads or stores provider credentials.
• Project checks and protected test patterns live in config the agents cannot change mid-run.
• Tests run against fakes and recorded sessions, so CI spends no model quota.
• Next: the Desk, one screen for several agent chats, designed from research on interruptions and resumption.