Tomas

Auth/pricing systems by day. Building AI-agent tools and workflows by night.

[Tomas @tmsdnl]

@jxnlco LOOPD?

[Tomas @tmsdnl]

The gap between GPT-5.5 Medium and High in Codex seems larger than the names imply.

Medium constantly tests my patience. High just runs with it.

[Tomas @tmsdnl]

@alain Does something like this exist in desktop/mobile Safari?

[Tomas @tmsdnl]

@seanrobins_ The whole Bay Area tbh. And don’t even think about tethering to your laptop on Caltrain or in the car.

[Tomas @tmsdnl]

@ThePrimeagen And I wish the conversations started from top right X post corner would appear in conversation logs in main app so that I could continue instead of losing everything or having to copy paste

[Tomas @tmsdnl]

@zebassembly They just ask codex to do it

[Tomas @tmsdnl]

@Dimillian Also submitted a related feature request a week ago: https://t.co/WQYl1du1XZ

[Tomas @tmsdnl]

@Dimillian Works well with skills

/side $review or /side $commit

[Tomas @tmsdnl]

Rectangle v0.96 is out with my shortcut-cycling patch.

One shortcut can now be assigned to multiple actions, and repeated presses cycle through them.

Tiny workflow fix. I wouldn’t have made time for it without Codex.

https://t.co/zrFWJltj1p

[Tomas @tmsdnl]

@pbakaus @Nick_Davidov Got asked that last Friday, word for word

[Tomas @tmsdnl]

@Nick_Davidov “I live in San Francisco” - confidently declares the colleague from Tracy

[Tomas @tmsdnl]

@tenderizzation @knowclarified I remember the days when it was only 4

[Tomas @tmsdnl]

Another case where 32GB of RAM can feel tight: compiling Codex from source locally.

I caught a single rustc process at 28.76GB in Activity Monitor. System memory went from about 8GB before the compile to almost 28GB during it, so the compile likely added roughly 20GB of system memory pressure.

It finished, but 32GB leaves little headroom.

[Tomas @tmsdnl]

Note to future self and PSA for anyone buying a Mac Mini to sandbox coding agents inside a macOS VM:
- 32GB RAM is tight. A macOS VM for Xcode + simulators realistically needs ~16GB allocated to it, leaving little room for the host and additional Linux VMs/containers for agents.
- 256GB SSD disappears fast between Xcode, simulators, caches, and VM disks (1TB upgrade already on the way)
- Get 10GbE (if your network supports it). Once the Mini becomes a remote VM host, 1GbE screen sharing feels choppy. You cannot upgrade it later.

*Does not apply if you are yolo-running agents directly on the host.

[Tomas @tmsdnl]

@fcoury Yes, I see

[Tomas @tmsdnl]

Huge respect to @fcoury, Eric Traut, and the Codex team. My feature request went from GitHub issue to done and merged in 10h 41m. You rarely see that kind of speed on GitHub, even for a relatively easy fix. OpenAI is moving incredibly fast with Codex.

I wonder what the Codex team’s internal issue-to-merge leaderboard looks like.

PR: https://t.co/esqAfZFujM

Tested - works!

[Tomas @tmsdnl]

Filed this as a Codex feature request with a slightly broader framing: parallel side sessions with agent-like inspectability + clean result handoff.

Core idea: keep the main thread clean, let side work stay inspectable, and make it easy to bring back only the results that matter.

https://t.co/oZDWIudcLM

[Tomas @tmsdnl]

One Codex workflow gap I keep running into: keeping the main thread clean while still using noisy one-shot skills.

Since /side landed in Codex, I’ve been using /side $commit and /side $review for this.

It works, but the handoff is manual: run it in /side, wait, copy the review findings back into the main thread.

It’d be nice if skills could run like workers: off the main thread, inspectable while running, with a clean result handoff when done.

Conceptually: /side isolation + agent-style inspectability.

[Tomas @tmsdnl]

One Codex workflow gap I keep running into: keeping the main thread clean while still using noisy one-shot skills.

Since /side landed in Codex, I’ve been using /side $commit and /side $review for this.

It works, but the handoff is manual: run it in /side, wait, copy the review findings back into the main thread.

It’d be nice if skills could run like workers: off the main thread, inspectable while running, with a clean result handoff when done.

Conceptually: /side isolation + agent-style inspectability.

[Tomas @tmsdnl]

I always thought "a" vs. "an" was based on the written first letter: use "an" before words starting with vowels a, e, i, o, u.

But some words felt weird when I wrote them, and I never noticed which article I was using when speaking.

Turns out the actual rule is based on pronunciation, not spelling:
- "an hour" because hour sounds like "our"
- "a university" because it starts with a "yoo" sound
- "an MBA" because M is pronounced "em"
- "a European" because it starts with a "yuh" sound
- "an honest mistake" because the H is silent

The more you know.

[Tomas @tmsdnl]

Thanks for saying sorry, Codex, for shutting down my server. I probably shouldn’t have given you full SSH access to the homelab either.

“I ran a remote log search with an unescaped | pattern. The remote shell interpreted it as a pipeline, and one of the words in the pattern was shutdown, so it invoked the host shutdown command. I immediately tried shutdown -c, but the host was already entering shutdown and refused/couldn’t cancel cleanly.

That reboot was caused by my command, not by your broker issue.

I’m sorry. I should have wrapped the remote command safely instead of passing shell metacharacters through SSH. I’ll stop running commands unless you explicitly tell me to continue.”