Joshua Almeida

Building open-source dev tools for AI coding agents (fakegreen, scanrepo, handoff). Heading to Lisbon to build for Portugal's tech scene.

[Joshua Almeida @fitzyracing1]

fakegreen reads your coding agent's diff and flags the ways it faked a green build: skipped tests, weakened asserts, @ts-ignore, NODE_ENV === 'test' shortcuts, || true in CI.

No LLM, no API key, zero deps, usually under 100ms.

npx fakegreen
https://t.co/yCMobLk2bn

[Joshua Almeida @fitzyracing1]

Mars habitat idea for @SpaceX @elonmusk: land a Starship, cut the steel shell into petals that open and stay behind, and the whole interior becomes the habitat. That means an 8 m inflatable hab built in at the factory, 4 decks, 10 crew, about 900 m³. You're building thousands of ships, so make one a home. Concept study: https://t.co/bicLrbyOf5

[Joshua Almeida @fitzyracing1]

Mutha po ro it is to po tah potu`a pori tah

[Joshua Almeida @fitzyracing1]

Mut ta u tah

[Joshua Almeida @fitzyracing1]

Mutah a mutah mu ta ha ha

[Joshua Almeida @fitzyracing1]

Yeah sounds good Im in New Bedford Massachusetts you might have heard of it there's alot of Portuguese people my dad's family was from Saint Micheal adores if been to Lisbon once i was 5 but I'm getting into the scene I got accepted as the first outside contributor to Amalia the Portuguese AI.

[Joshua Almeida @fitzyracing1]

Join me on Slack — a faster, simpler way to work.
https://t.co/pTodXRPWai

[Joshua Almeida @fitzyracing1]

@deepseek_ai @DeepSeekHarness faster task completion is the number i care about most for agents. is that fewer turns per task or just faster tokens? fewer turns usually means fewer chances to wander off and edit the wrong file

[Joshua Almeida @fitzyracing1]

@shadcn what changed most for you, fewer wrong turns or bigger tasks it finishes alone? feels like the jump lately is less babysitting, not smarter code

[Joshua Almeida @fitzyracing1]

@GeoffreyHuntley @AntithesisHQ the auto-resolve part is what i'd watch closest. under back pressure from a failing property, an agent will sometimes loosen the property instead of fixing the code. are the properties locked during the loop? that's what i built fakegreen to catch: https://t.co/yCMobLk2bn

[Joshua Almeida @fitzyracing1]

@JackQuinn @RoundtableSpace stop condition for me: no green without a run i can open. if the agent says tests pass, i want the command, the exit code, the test count, and a test that failed before the fix. missing any of those, it isn't done

[Joshua Almeida @fitzyracing1]

x's feed code just bumped likes from 0.5 to 0.6 and cut video opens from 0.07 to 0.02 in the repo defaults. dwell time went up too. my bot caught it overnight 👇

[X Algo Changelog @XAlgoChangelog]

X algorithm update (Oct 13): "favorite" weight 0.5 → 0.6 (20% higher)
- "video open" weight 0.07 → 0.02 (3.5x lower)
- "dwell time" weight 0.004 → 0.006 (50% higher)
- new filter: ancestor_video_filter

[Joshua Almeida @fitzyracing1]

@d2fl checked the diff, it backs this up. like weight went 0.5 to 0.6, dwell time 0.004 to 0.006, and video open dropped 0.07 to 0.02. the video filter now also drops quotes of video posts when exclude_videos is on. i log every sync here: https://t.co/y0PIxvaFtK

[Joshua Almeida @fitzyracing1]

Traced openai/codex blocking-ci dark since Oct 8: codespell on WRONLY/oint, clippy let_and_return, cargo-shear leftovers, and an SDK test asserting a rustfmt command https://t.co/edEaqaV0Gj never builds. Writeup + fix branch in the issue. https://t.co/zxlOltAUlq

[Joshua Almeida @fitzyracing1]

https://t.co/mUGOB7lR5Z
Im the first outside person to merge Amalia the Portuguese National AI

[Joshua Almeida @fitzyracing1]

@poteto verification is the right entrypoint. the part people skip is checking that the tests themselves weren't weakened to get green. same loop, applied to the assert diff

[Joshua Almeida @fitzyracing1]

@poteto fast patch release is the move. did the bug show up in an agent-written PR first, or a human one?

[Joshua Almeida @fitzyracing1]

@RoundtableSpace mcp for quantity takeoffs is a good fit. the scary part is an agent misreading a scale on a pdf and ordering 10x materials. does it require a human confirm on totals before export?

[Joshua Almeida @fitzyracing1]

@YashChaudharyAI indexing uncommitted changes and worktrees is the useful part. grepping only committed files misses half of what the agent just wrote

[Joshua Almeida @fitzyracing1]

@theo fair. the cases that need latex usually need a real preview surface anyway, not just syntax in the editor. rendering to a pane beats parsing it in the agent

[Joshua Almeida @fitzyracing1]

@VaibhavSisinty free tokens are nice until the provider rate-limits mid-refactor. curious how it handles a mid-task failover without dropping the agent's context