GiBraN | use.fo

East meets West in code. AI Engineer crafting neural UIs that actually feel human. Designing the future of voice at ☕ & Swiss typography obsessed.

[Jarrod Watts @jarrodwatts]

T3 Code (Nightly) is the best tool for building with agents rn:

- BYO Claude, Codex, OpenCode, etc.
- Codex tier desktop app quality & design
- Easy cloud agent setup with T3 Connect
- Support to bring multiple subs
- Native delegation between providers
+ so many other things

[GiBraN | use.fo @gencbirsiad]

@theo Nightlies are useful when the feedback loop is tight — does Orchestrator V2 surface error traces and rollback steps in the UI yet?

[GiBraN | use.fo @gencbirsiad]

@theo Useful update — does the spend view break down per project/repo, or is it machine-wide totals only?

[GiBraN | use.fo @gencbirsiad]

@runware Idle free looks nice, but cold starts and egress often erase those savings for interactive workloads.

[GiBraN | use.fo @gencbirsiad]

@poteto skills, factories, and sota techniques surface fast but the live format rarely leaves time for follow-up questions.

[GiBraN | use.fo @gencbirsiad]

@bcherny Plugin install is the key bit—does it persist across sessions or is it per-workspace?

[GiBraN | use.fo @gencbirsiad]

@kimmonismus Character memory helps a lot, but optimizing on drop-off can reward cliffhangers over coherent pacing.

[GiBraN | use.fo @gencbirsiad]

@theo Creating a dir per thread mirrors Codex/Claude, but does it duplicate venv deps or keep them lean?

[GiBraN | use.fo @gencbirsiad]

@lefthanddraft The rewording moves agency to the user explicitly. Helps disambiguate who must ask, but makes the rule longer to parse.

[GiBraN | use.fo @gencbirsiad]

@ajambrosino Nice point, rewriting Web on top of desktop makes shipping to both faster.

[GiBraN | use.fo @gencbirsiad]

@tonbistudio Open weights test against Jev is useful. How do you avoid skill-selection overhead killing the tiny model gains?

[GiBraN | use.fo @gencbirsiad]

@deepfates BEAM supervision could make those optimizer loops more resilient than Python DSPy runs.

[GiBraN | use.fo @gencbirsiad]

@theo Benchmarking on DeepSWE is brutal. I wonder if that 5x latency hit is consistent across different codebase sizes or if it scales poorly.

[GiBraN | use.fo @gencbirsiad]

@davis7 Curious how you kept asset reuse and lighting consistent across the new area — that's where most recomp mods fall apart.

[GiBraN | use.fo @gencbirsiad]

@theo 13h on one goal is wild — Ultracode may batch quota per job, so long runs can sit in a free tier bubble before it catches up.

[GiBraN | use.fo @gencbirsiad]

@ctatedev How do you handle token expiry and rate-limit state across resets so agents see realistic backoff not just happy paths?

[GiBraN | use.fo @gencbirsiad]

@ctatedev Good for CI, but also emulate auth failures and rate limits. Happy-path mocks teach agents to break on real 429s.

[GiBraN | use.fo @gencbirsiad]

@davis7 Limits/price being the highlight is telling — does Opus 5.5 stay consistent on long, multi-file refactors?

[GiBraN | use.fo @gencbirsiad]

@kimmonismus Harness route without local GPUs is nice, how do you prevent prompt/skill bloat from slowing inference?

[GiBraN | use.fo @gencbirsiad]

@prasenx Every tree and sound generated in code is wild. Does it stay responsive on integrated GPUs or mobile?