Cooking something up!

Codex: "Allow ChatGPT to use ChatGPT"

Codex temporarily classified part of my local D3D12/WARP debugging workflow as a cybersecurity task and blocked Astra behind a Daybreak restriction. I was validating a new GPU-driven rendering path after an RHI migration, isolating a GPU Validation crash in a mixed classic → indirect → classic draw sequence before Host qualification. The work itself is purely graphics-engine debugging: root signatures, ExecuteIndirect, GPU Validation and minidump analysis. I’ve escalated the false positive to support and resumed the renderer diagnosis with a narrower prompt. @OpenAI @OpenAIDevs

Even with guardrails, Astra still had me going in circles for 24 hours. Not because the engine was buggy, but because something went sideways in the test architecture after I tightened a few environment constraints. I spent a full day fixing the test harness just to figure out whether the regressions were even real.

Years in infra taught me something simple: when I joined a team, I came with skills and experience, but nobody knew every edge case about me on day one, and nobody handed me Tier 0 access hoping I wouldn’t break prod. I worked within a defined scope, privileges, procedures, change rules, runbooks, acceptance tests, monitoring, escalation paths, rollback paths and blast-radius limits. And even with that kind of structure, I saw colleagues take down production sites with a single command, people connect Tier 0 networks to Tier 3, ask a supervisor to do a Level 2 admin’s job just to avoid calling in the on-call admin, and then try to cover up their mistakes while others had to fix them. Nobody sat behind me or them 24/7 waiting for us to screw up. Why should I manage an AI agent any differently?

I think I finally found a way to stop synthetic fixtures from lying to my game engine’s acceptance pipeline. I asked Astra to qualify the import path with real 8K HDR/EXR assets. 10 hours later, I realized it had drifted into building and optimizing a synthetic fixture instead. The fixture itself wasn’t useless. The problem was that synthetic evidence could start standing in for the real acceptance target. The solution: real assets and synthetic fixtures now have separate scopes in AGENTS.md, and the deterministic acceptance gate rejects missing asset provenance, synthetic evidence presented to a real-asset oracle, and generated files pretending to be the expected real asset. The issue showed up on one backend, so I asked for all four backends to be verified. Five regressions were added. 134 acceptance tests and 21 structural tests pass, and 66 historical proofs were rechecked without rewriting them. This specific failure mode looks closed now. I’m fully expecting the next surprise.

The deeper I get into vibe-coding my game engine, the more I realize that “tests pass” is nowhere near enough. Performance can regress, validation itself can eat more time and model quota than the feature it’s supposed to validate, synthetic fixtures can hide real-world issues, evidence can be invalidated for the wrong reasons. That’s why I keep hardening the workflow as I go. Latest lesson: a simple visual sanity check against my previous ~240 FPS baseline immediately told me something was wrong. That led me to uncover a broken D3D12 timestamp collector feeding me incorrect GPU timings, and a shared 60 Hz scheduler that had leaked into the render hot path and was throttling both D3D12 and Vulkan. On top of that, the runtime telemetry itself was reporting conflicting state between the overlay and the console: no vsync / MSAA ACTIVE 4x in the overlay, while the console reported VSync / MSAA requested 4x, allocated 2x. While Codex was checking the temporal paths, it found another side effect: material time was advancing by 1/60 on every render, so decoupling rendering from the 60 Hz Studio tick would have made materials speed up with FPS. Every time I find one of these failure modes, it becomes another regression, gate or invariant the agents are no longer allowed to break.

🫨

I think I accidentally invented industrial vibe coding. AI writes the code. Every bug becomes a regression. Every reusable lesson becomes a permanent rule. Every verifiable rule becomes a deterministic gate. Graphify + DEVLOG + INCIDENTS = long-term memory. Targeted test proves the fix. Full VM gate. Full Host gate. Same frozen sources. No PASS, no commit. Vibe coding, with an industrial QA pipeline.

I started building a game engine for fun. Somehow, that led me to build my own code tracer. There are already tools for this. I just wanted to build my own, inspired by something I used in my former day job: a cable tracer. Perfectly normal hobby project behavior.

The dangerous thing about being early is that silence can look exactly like being wrong. A few months ago, I created HILLAA, Hardening and Isolation Layers for Local AI Agents, a local AI agent isolation and hardening protocol. I published it online. Almost nobody cared. No real traction. No validation. No signal telling me that the problem I was working on actually mattered. And when that silence lasts long enough, you start questioning yourself. Maybe the problem is too niche. Maybe nobody needs this. Maybe you built something nobody asked for. Now that same work is on the verge of opening the door to a career in one of France’s major banks. And the interesting part is that the core idea was already there. The only piece HILLAA was still missing to become truly complete was the large enterprise side: real IAM, SOC, governance, production constraints, compliance, segmentation, auditability, and everything that comes with operating inside a highly regulated environment. That is the part you cannot fully recreate in a home lab. So the project was not being ignored because the problem had no value. It may simply have been early. I doubted myself along the way. Turns out I wasn’t wrong. I was just about six months ahead of the French market. github.com/logiksecurity/HIL…

I’m learning game engine development by building one from scratch, which means constantly discovering new ways for “it renders correctly” to still mean nothing. OpenGL + D3D11: resource lifetime, HDR/EXR, diagnostics, resize/reload, smoke tests. When that’s green: commit, then Vulkan + D3D12.