BeforeDone runs configured checks, creates receipts bound to declared file scopes, rejects stale evidence, and builds incident reports from observable events and artifacts. It is a local guardrail for normal agent mistakes, not a sandbox or trust system.

What could you build with GPT-5.6 that wasn’t practical before?

Using GPT-5.6 Sol Ultra, I built BeforeDone, an open-source completion gate for coding agents. GPT-5.6 helped carry the project from a first-principles correctness model through cross-platform Go implementation, adversarial review, regression tests, packaging, documentation, and release. BeforeDone binds check receipts to the declared files that were actually verified, rejects stale evidence after those files change, requests a corrective continuation when Codex tries to stop without fresh evidence, and reconstructs failed runs from observable events. Sustaining that end-to-end engineering and verification loop—while keeping the tool small and auditable—wasn’t practical for me with earlier models.

Hi Product Hunt - I built BeforeDone after watching coding agents treat "I ran the tests earlier" as equivalent to "the current code is verified." It is not. BeforeDone turns configured checks into local receipts bound to their declared file scopes. Change a relevant file and the receipt becomes stale. The Codex Plugin uses those receipts as a Stop gate; the CLI can also build incident reports and conservative replay plans from observable evidence. I kept the boundary narrow on purpose. BeforeDone does not choose the right tests for you, recover hidden reasoning, sandbox a verifier, or defend against a malicious process with the same OS permissions. It is a guardrail for capable agents that sometimes stop too early. v1.0.1 is Apache-2.0, runs locally, and has no hosted account or BeforeDone telemetry. I would value feedback on two things: whether configuring a credible check set is clear, and whether a stale receipt explains itself well enough when it blocks completion. Repo: https://github.com/rrrrrredy/bef... Guide: https://rrrrrredy.github.io/befo...

What could you build with GPT-5.6 that wasn’t practical before?

BeforeDone

Fresh evidence before coding agents say done

Fresh evidence before coding agents say done