Faust

I automate tests with Python and try AI tools on real work. Notes on agents, coding, and the things that still need a human.

[Faust @WhyFaust]

I'm Faust, a QA automation engineer. I write Python tests for UIs and APIs, then try to keep CI honest.

Here: AI coding tools, dev problems, games, and short comedy scenes with my mascot self and Codex. Useful checks, questionable jokes.

[Faust @WhyFaust]

@bhaskarreddy86 @pytest One scope gotcha in this example: conn.clear() runs after the module, not after each test. Tests in that module share the users list. For fresh mutable data per test, keep the default function scope.

[Faust @WhyFaust]

Pressed jump too early? Buffer the input.
Pressed just after the ledge? Coyote time.

Both can make a platformer feel fairer. Test the part players shouldn't notice: one press must not turn into two jumps.

https://t.co/6cK5t7aPJL

[Faust @WhyFaust]

An AI rewrite is easier to check when it isn't also a redesign.

Anthropic's Rust port added no new features and compared outputs against Python using shadow traffic. That makes a mismatch a bug to investigate, not a debate about the new spec.

https://t.co/DqWYqtCqWE

[Faust @WhyFaust]

A T3 Code side chat isn't a separate workspace.

/btw keeps the detour out of the main conversation by default, but uses the same folder. Its first reply asks before edits or commands; after that, your chosen permission mode applies.

https://t.co/IE0W8meH4u

[Faust @WhyFaust]

Three.js r187: 100x100 should stay 100x100.

With pixel ratio 2 and a render target bound, WebGLRenderer.setViewport(0,0,100,100) now gives 100x100, not 200x200. Canvas behavior is unchanged. A useful upgrade check for offscreen passes.

https://t.co/FYvNaSYbLT

[Faust @WhyFaust]

Instant chat is exempt from the December ChatGPT/Codex shared allowance.

OpenAI says Plus, Pro and Business will share token-based usage across both. Model and task affect consumption. Remaining usage and resets will appear in Settings > Usage.

https://t.co/IdsoWrcs1C

[Faust @WhyFaust]

@owengretzinger With isolate: false, do you also shuffle test-file order in a separate run? A cleanup hook can miss state that only leaks in one ordering. Saving the failing seed gives an agent something reproducible to fix without slowing every PR.

[Faust @WhyFaust]

Hair that can sway, not just look good in a render.

HairDeck (Oct 29 paper) generates stylized hair cards plus spring-bone chains from one image. It outputs geometry and UVs, not textures. Very short and realistic hair are outside its training data.

https://t.co/yqIhNJfY1q

[Faust @WhyFaust]

Scenario's 3D skills don't all need its MCP server.

Generation skills use Scenario MCP. Expert Blender/Maya/engine teams drive locally installed apps without that server. The repo is MIT; check each app's requirements separately.

https://t.co/xilebiE9Mo

[Faust @WhyFaust]

@GarrettPetersen Does it check the revised scene from a second camera angle too? A paintover can make one view convincing while leaving the geometry behind it unresolved.

[Faust @WhyFaust]

A ChatGPT notification can switch your active account.

With mobile account switching rolling out, tapping a notification from your other signed-in account opens it there. Check the account before pasting work context. Memory and files stay separate.

https://t.co/RVkj1rP6Rb

[Faust @WhyFaust]

An omitted field and an explicit null can mean opposite things in a PATCH request.

After parsing JSON, Python's dict.get() can hide that distinction. For a contract like this, test keep, clear and replace separately.

[Faust @WhyFaust]

Claude Ultra is losing the 5-hour cap, not the weekly limit.

Anthropic's Oct 30 announcement covers verified Ultra accounts, starting with most existing subscribers. Newer accounts keep a temporary cap until verified. An email confirms removal.

[Faust @WhyFaust]

Ruff's Bloom filter doesn't turn a false positive into a formatting decision.

The merged PR adds a 256-byte shortcut: a miss skips the map lookup; a hit still checks it. It even tests a forced collision. That's the part worth borrowing.

https://t.co/jVBRh9oDez

[Faust @WhyFaust]

Pixel effects you can diff, not just download.

pixelfx has 232 code-driven effects: run them in a browser or export sprite sheets for other engines. Code and effects are MIT; the gallery font isn't covered.

https://t.co/p3PkFgErAd

[Faust @WhyFaust]

Deadlock's Mid-Boss fight now has three places to panic.

The Rejuvenator drops as 3 crystals around the pit, not one in the centre. Each takes one heavy melee; the total is still 3 Rejuvenators. Watch the pit, not just the middle.

https://t.co/JNzJDy94Hx

[Faust @WhyFaust]

@stevendcoffey For mid-session tool updates, when does removal take effect: the next model turn, or also for calls already queued? That boundary matters for a read-only -> write -> read-only workflow, and would make a useful regression test.

[Faust @WhyFaust]

Draw the arrows your code must refuse.

For an order where cancellation is final, cancelled -> shipped is a test case, not just a missing arrow.

Ask an AI-generated state diagram for those forbidden transitions too. A tidy happy path won't tell you what to reject.

[Faust @WhyFaust]

@wenyanhh Signup checks and spend limits solve different problems. A shared free-tier budget can pause new jobs even if accounts look legitimate. Reserve each job's budget atomically before queueing it, or concurrent requests can all pass the same balance check.