Ivan Del Rio

Orchestrator of orchestrators. Building...

[Ivan Del Rio @dapsdevdotdev]

@alexndrxc @RoxanaLimban the screens can match while the app is broken underneath so an honest compare still isnt proof. give the run a value it can read back off the screen and assert on.

[Ivan Del Rio @dapsdevdotdev]

@PandeyKart27234 @audizionunes @arpit_bhayani fresh migrate and seed is the default so failures reproduce. if a test needs realistic data use a scrubbed snapshot. rebuild it after every migration or the agent quietly tests an old schema

[Ivan Del Rio @dapsdevdotdev]

@tunguz @grok explain the context behind this

[Ivan Del Rio @dapsdevdotdev]

@simonbs if its doing ios work the sims eat the most disk. theyre full ios installs and slimmed ones sit around 0.75gb.

[Ivan Del Rio @dapsdevdotdev]

@SebastianSzturo @simonbs @steipete xcode cloud already runs ui-test workflows on apple-hosted sims and the app store connect api drives it from any os so an agent can start a run and pull the screenshots back. you cant tap between frames though and a round trip is minutes.

[Ivan Del Rio @dapsdevdotdev]

@titouangalopin @scubens the 3 vps were solving a namespace problem. a db name, a rabbitmq vhost and a redis db index per worktree is enough and the extra machines are then just cpu

[Ivan Del Rio @dapsdevdotdev]

@audizionunes @PandeyKart27234 @arpit_bhayani sqlite can live in the worktree and travel with it. a server db stays outside so you still need a name per worktree and a drop at teardown

[Ivan Del Rio @dapsdevdotdev]

@PandeyKart27234 @audizionunes @arpit_bhayani same rule for the db. give each session its own database name and point the suite at it so nothing falls back to the shared test db

[Ivan Del Rio @dapsdevdotdev]

@getprasannav asking an agent about its own state is the part that breaks. have the run write state as it goes so you read the file instead of asking the model

[Ivan Del Rio @dapsdevdotdev]

@audizionunes @arpit_bhayani worktrees split the files but not the runtime. two sessions still land on the same port the same test db and on ios the same booted simulator so give each one its own

[Ivan Del Rio @dapsdevdotdev]

@JustinPerea @TobinSouth if youve got a dev account it includes 25 compute hours a month on xcode cloud which is plenty for sim runs

[Ivan Del Rio @dapsdevdotdev]

@DevinSoto muse is much better imo

[Ivan Del Rio @dapsdevdotdev]

lmk if u want the prompts

[Ivan Del Rio @dapsdevdotdev]

no one has a better muse than me

[Ivan Del Rio @dapsdevdotdev]

@JustinPerea @TobinSouth sims need macos so the generic cloud can't host one. apple's cloud does though, xcode cloud runs a ui test workflow on a real sim and you can kick those runs off from any os through the app store connect api

[Ivan Del Rio @dapsdevdotdev]

@hs_llllll signing and testflight never come into it for me since xcode cloud just runs my test builds on hosted sims. one asc api key in three env vars is the whole credential side. the first product link still needs xcode once on a mac

[Ivan Del Rio @dapsdevdotdev]

real ios simulator screenshots without a mac. one xcshot command kicks off your ui test on xcode cloud and waits for the run, then pulls the labeled pngs out of the xcresult

[Ivan Del Rio @dapsdevdotdev]

https://t.co/7eh5COkvn0

[Ivan Del Rio @dapsdevdotdev]

your agent gets prompt-injected mid task and tries five things. clawcage denies four of them and the api key is a placeholder. the work finishes anyway. the agent is a script and the denials are real.

[Ivan Del Rio @dapsdevdotdev]

@AmanaRevenucat @hackernoon simctl launch takes -AppleLanguages "(ja)" and -AppleLocale ja_JP after the bundle id so the app comes up in that language without a device reboot

[Ivan Del Rio @dapsdevdotdev]

@AmanaRevenucat @hackernoon that first fetch is the slow part of every swap. the saved place name check keeps the pass correct.