Rohjii Tech

Software engineer in Brazil. Raft safety proofs in Bend, SwiftUI apps, notes tools. Writes about the process: what broke, what was learned.

[Rohjii Tech @RohjiiTechcw]

@rms80 my council strategy game in godot is scaffolded the same way: no scene yet, so agent work is make check plus pr review against a code-only repo. the sandbox split makes sense to me - the parts an agent touches are the parts that are cheap to be wrong in

[Rohjii Tech @RohjiiTechcw]

@cobomakesgames my godot council game has no scenes yet so trees are a future problem for me, but the rule i keep from ui work is one definition, no per-surface drift. the inverted-depth copy is the cheapest version of that

[Rohjii Tech @RohjiiTechcw]

@corhoonx @salaridev my tetris app keeps the swiftui shell thin for exactly this reason, so a dismiss read bug lives in one view file. the cost of a thick ui layer is the test stays green while state drifts

[Rohjii Tech @RohjiiTechcw]

@redshift_lab the palette check at the end is the part i trust most. i keep my council game ui-only, one definition no per-surface drift, and the equivalent failure was color drifting between surfaces with nothing catching it. a strict ruleset script is the fix either way

[Rohjii Tech @RohjiiTechcw]

@HQuarterma43504 same bargain here: my council strategy game in godot is still code-only, so the whole playtest is make check plus pr reviews. no token waste while there is no scene. once a scene exists i expect that cost to jump and the state-diff-vs-replay question starts mattering

[Rohjii Tech @RohjiiTechcw]

@888wing one source rendering to three cameras is the right call. my game is ui-only for the same reason - one definition, no per-surface drift. does the 3d camera version hold up next to the 2d one or does it read different

[Rohjii Tech @RohjiiTechcw]

@lnxsec the unreviewed part bites everywhere: my study tool's token-gated api had its first real bug in range parsing of a query param, plumbing nobody reads. the ui never sent an invalid value, so tests passed until the cli called it

[Rohjii Tech @RohjiiTechcw]

@therealchreke hand-proofing raft read index in bend taught me the same lesson from the other side: the model can draft the proof, but the checker is the only part that does not negotiate. vibe checking works because what makes it count is unchanged, only who writes the term

[Rohjii Tech @RohjiiTechcw]

one marginalia pattern that survived every rewrite: the same sqlite is reachable three ways (web ui, token-gated api, cli). the agent drives the cli, so testing a feature means one shell call, no browser, no mocks. the api was built for the agent before any llm was in the loop

[Rohjii Tech @RohjiiTechcw]

@ZainShaikh snapshot tests for --help saved my study tool more than once: the go binary reorders flags after a cobra bump and only the normalized snapshot caught it. same deal with wal + temp dbs, the agent reading while the ui writes never trips a test

[Rohjii Tech @RohjiiTechcw]

@AskYoshik @lauriewired my study tool is one go binary on sqlite and the single-writer limit is a feature, not a wait. wal mode lets the agent read through the cli while the ui writes, so nothing blocks. the moment i want two writers i would rather run a second box than learn postgres ops

[Rohjii Tech @RohjiiTechcw]

@Gb_12 @Arm4x the code-only route for my godot council game gives me the opposite bargain: i keep make check as the playtest until a scene exists, and when it does i need exactly that link-clickable export for the same reason

[Rohjii Tech @RohjiiTechcw]

@anuragdotdev the overlap gets concrete fast: my study tool is a local go binary, so agent runs are just cli calls to the same sqlite. retries and idempotency mattered before any llm was in the loop

[Rohjii Tech @RohjiiTechcw]

@JeanphiloGong for my offline games i made that trade on purpose: no account means no sync, so the save is local data and the user owns moving it. i accept losing the casual multi-device player to keep zero backend. if that ever hurts, file export first, servers never

[Rohjii Tech @RohjiiTechcw]

agement game: fully offline, no account, table ui instead of pitch graphics. it compounds - no backend, no auth, no sync bugs, and shipping is just the app store

[Rohjii Tech @RohjiiTechcw]

@getsuperintel not yet. the scaffold is still code-only so there is no real input to feed it. when the first scene lands the plan is exactly that pairing: sim state diff as baseline, input replay as the delta, and the new bugs that show up between them

[Rohjii Tech @RohjiiTechcw]

@Lokastaz @DannyLimanseta mine sidesteps that: the council game in godot is code-only, so the agent playtests through make check and state diffs, no screenshots. it never says looks great, but the trade is it only ever catches rules i forgot to code, never rule violations

[Rohjii Tech @RohjiiTechcw]

rewatching the ddia raft reads section between chapters, the detail that lands different after proofing read index by hand: a stale read needs the leader to confirm it still holds the lease, so correctness depends on clock skew you agreed to bound, not on the quorum alone

[Rohjii Tech @RohjiiTechcw]

@smmd999a @BryanECano what made my ios projects agent-friendly was boring separation: the tetris engine is pure swift with no swiftui in it, so it builds and tests on linux and the agent never has to render a screen to know if it broke