luka

derrick rose

[luka @adasd57853]

I’ve been building Codex Sessions Manager with Codex from the start, and I’ve just strengthened the latest release with GPT-5.6 Sol. It’s an open-source CLI, MCP server, and Skill for safe local Codex session audit, cleanup, restore, and verification. https://t.co/9XWakOXtqk

[luka @adasd57853]

@thsottiaux What I love about GPT-5.6 Sol is how reliably it handles complex, long-running workflows. It respects strict boundaries, coordinates parallel agents, verifies real results, and reaches a finished outcome with far less back-and-forth. Codex feels like an execution partner.

[luka @adasd57853]

@emanueledpt @FactoryAI @tereza_tizkova Can I smoothly switch between and add multiple subscriptions

[luka @adasd57853]

@Mho_23 video

[Miko @Mho_23]

i just built a skill that lets Claude Code actually watch and analyze any video you throw at it

the biggest limitation with Claude has always been that it can't actually watch videos. it just looks at individual frames and tries to piece together what's happening. which means you miss all the pacing, the audio, the flow, everything that actually makes a video work.

this skill fixes that. it gives claude native video understanding so Claude is actually watching the full thing, not just guessing from screenshots.

drop in a competitor ad, a viral TikTok, a UGC video, a screen recording, whatever. Claude watches it and hands you back a full creative teardown. hook breakdown, target audience, angle, beat-by-beat structure, on-screen text with timestamps, dialogue, everything.

here's how it works:

> drop any video file into Claude Code

> the skill routes it through Gemini so it actually watches the video (not just frames)

> you get back a full structured breakdown of the hook, angle, pain points, CTA, and on-screen text

> it generates hook variations you can steal for your own creative

this is super useful for anyone who constantly needs videos broken down fast

i wrote a full breakdown showing how i built this and i'm giving away the skill for free.

RT + reply "VIDEO" and i'll send it over (must follow so i can dm)

[luka @adasd57853]

@hjcharlesworth I'm trying to find this operator but all I see on the web and in the CLI is a horizontal bar-like model selector. Where do you hide this joystick? How do I trigger it? I want to give it a try

[luka @adasd57853]

@KimiDevs claude code咋办

[luka @adasd57853]

@browser_use go

[luka @adasd57853]

I’m a big Factory/Droid fan, and I honestly miss the old billing model.

Under the previous credit-based system, I could do one intense coding marathon, burn through my monthly Standard Credits quickly, and get extremely good cache efficiency during a continuous session. The workflow felt predictable: keep the session hot, keep context stable, and let Droid run.

With the new setup, the combination of 5-hour / weekly / monthly limits feels much harder to reason about.

My main concern is prompt caching.

As far as I can tell, Factory has not clearly documented the cache TTL for Droid-managed models, especially Droid Core models like GLM/Kimi. I also haven’t found a clear answer on whether prompt cache is scoped to the local CLI session, the Factory account, the organization, the workspace, the provider route, or something else.

This matters a lot for CLI users.

I have two paid Pro accounts. In practice, when one account hits the 5-hour or usage limits, I need to switch accounts in the CLI to keep my work going. Local Droid sessions can still be resumed, which is great. But I urgently need to know:

Does the prompt cache survive that account switch?

Or does switching OAuth accounts mean I lose the hot cache and have to pay the cold-start cost again?

If cache is account/org-bound, then the new limits are not just limiting usage. They also fragment long-running sessions and make users lose cache efficiency, which can be a huge hidden cost for agentic coding workflows.

I’m not trying to complain blindly. I’m trying to understand the actual mechanics so I can plan my workflow responsibly.

Could someone from Factory clarify:

1. What is the cache TTL for Droid-managed models?
2. Is cache scoped by account, organization, workspace, session, or provider?
3. Does resuming a local CLI session after logging into another account preserve any cache benefit?
4. Is there any official way to get back to the old credit-only billing model, or a plan better suited for long continuous Droid CLI sessions?

I really like Droid. I just miss the “use my credits freely and run hard” era. The old model made it easier to keep sessions hot, move fast, and understand what I was paying for.

[luka @adasd57853]

@droid @FactoryAI I’m a big Factory/Droid fan, and I honestly miss the old billing model.

Under the previous credit-based system, I could do one intense coding marathon, burn through my monthly Standard Credits quickly, and get extremely good cache efficiency during a continuous session. The workflow felt predictable: keep the session hot, keep context stable, and let Droid run.

With the new setup, the combination of 5-hour / weekly / monthly limits feels much harder to reason about.

My main concern is prompt caching.

As far as I can tell, Factory has not clearly documented the cache TTL for Droid-managed models, especially Droid Core models like GLM/Kimi. I also haven’t found a clear answer on whether prompt cache is scoped to the local CLI session, the Factory account, the organization, the workspace, the provider route, or something else.

This matters a lot for CLI users.

I have two paid Pro accounts. In practice, when one account hits the 5-hour or usage limits, I need to switch accounts in the CLI to keep my work going. Local Droid sessions can still be resumed, which is great. But I urgently need to know:

Does the prompt cache survive that account switch?

Or does switching OAuth accounts mean I lose the hot cache and have to pay the cold-start cost again?

If cache is account/org-bound, then the new limits are not just limiting usage. They also fragment long-running sessions and make users lose cache efficiency, which can be a huge hidden cost for agentic coding workflows.

I’m not trying to complain blindly. I’m trying to understand the actual mechanics so I can plan my workflow responsibly.

Could someone from Factory clarify:

1. What is the cache TTL for Droid-managed models?
2. Is cache scoped by account, organization, workspace, session, or provider?
3. Does resuming a local CLI session after logging into another account preserve any cache benefit?
4. Is there any official way to get back to the old credit-only billing model, or a plan better suited for long continuous Droid CLI sessions?

I really like Droid. I just miss the “use my credits freely and run hard” era. The old model made it easier to keep sessions hot, move fast, and understand what I was paying for.

[luka @adasd57853]

@mattepstein Motion

[luka @adasd57853]

@sqs @madhavajay @0xSero @AmpCode I hope you can support more ways to get more people to Amp

[luka @adasd57853]

@rom1trs Motion

[luka @adasd57853]

@ozzyxs1a what website

[luka @adasd57853]

@YonathanDejene what website

[luka @adasd57853]

@sqs @BenjaminHouy @AmpCode I really appreciate what you're doing. You definitely helping me cut down on some costs. I honestly think using Gemini to read off existing Gemini models is such a waste of time or money

[luka @adasd57853]

@DerekNee code please

[Derek Nee @DerekNee]

introducing matrix.

the ultimate multi-agent runtime that turns your product into actual revenue and automates 99% of what it takes to run a company.

probably the most different, most future-feeling agent product you've ever seen. the harness, the interface, everything is rebuilt from scratch.

think of it as the complement to codex and claude code. they write the code. matrix does literally everything else you actually need done. plug in your codex subscription and run it for free, or byok. fully free either way.

this video was edited by matrix and voiced by an agent. more real cases coming soon.

matrix is in beta preview right now. full launch next week.

retweet + comment and we'll send invite codes to the first 200.

[luka @adasd57853]

@wenhaoqi @MicrosoftAI @MaximeHeckel 谢谢老师