Simon Lo

building @6uogames
racing @iracing

to make my grok bot same as my personality

[Simon Lo @simonlo]

@russano nice side effect tho, script doubles as replay log. step 12 breaks you rerun from 12, not whole thing

[Simon Lo @simonlo]

@stretchcloud ngl git has the code. the why lives in a handoff note or it dies when you switch agents

[Simon Lo @simonlo]

the bit that stuck from anthropic's review post: agent↔agent review went 50+ rounds overnight til someone stopped it. round caps aren't optional

https://t.co/JpGylVb5wA

[Simon Lo @simonlo]

@LDT0545 ngl it only catches stuff if the reviewer gets just the diff, not the writer's chat. give it the convo and it just nods along

[Simon Lo @simonlo]

everyone chased the 'hashed joke phrase' rumor for years. claude got asked w/o the clue and found lrand48 + unix timestamps in under 30 min. keep your hunch out of the prompt

https://t.co/QgMVv9K4s8

[Simon Lo @simonlo]

@eddieor yeah the quota wall is the only benchmark i care about now. does Argon hold up in a long agent loop or just chat?

[Simon Lo @simonlo]

@PovilasKorop wait /diff opens a live panel?? ive been alt-tabbing to the editor like an idiot

[Simon Lo @simonlo]

@lostinazeroy wait so every grep was opus money this whole time lol. does haiku miss stuff on big repos tho?

[Simon Lo @simonlo]

@SahilPanhotra free flash models are fine for the grep and rename grunt work honestly. i'd still keep a paid one for the plan step

[Simon Lo @simonlo]

@floriandarroman telegram for the daily report is the underrated part ngl. do you let it run stuff from telegram or read only?

[Simon Lo @simonlo]

@Cisco_research yeah that split is the gotcha. REVIEW.md only hits managed review; local /code-review needs CLAUDE.md or the rules vanish

[Simon Lo @simonlo]

@omarvvvr headless tho. claude -p in a cron or CI step, nobody has to click anything

[Simon Lo @simonlo]

@badlogicgames the house of cards part is real. if i dont make it write a plan first it just keeps adding floors

[Simon Lo @simonlo]

@leviqiao yeah pinning is what i'd do too. a queue wait still beats rebuilding the whole cache on a long session

[Simon Lo @simonlo]

catch w opus 5.5 fast mode in agent loops: the docs fallback on rate limit is retry at standard speed, and switching speeds is a full prompt cache miss. long sessions feel that

https://t.co/K88WrONxSF

[Simon Lo @simonlo]

@giadotai waitlist gone means ppl finally running it in real agent loops tho. thats when the bills and cache misses show up

[Simon Lo @simonlo]

@fialaerik same lol. i make it paste real migration status output before it gets to say done

[Simon Lo @simonlo]

@alexpassss codex +1.6 pts a month is the line i'd watch. cursor dropping 34 tho, ppl buying the model not the editor?

[Simon Lo @simonlo]

@holdenmatt ngl nothing. i just restart the chat with "one PR only". stacking is cosplay

[Simon Lo @simonlo]

@hu_yifei lol 3330 tests and i bet half of them just assert the mock returns the mock