Beyond Code (AJ ONeal)

2026 needs a new CSV format that supports multiple sheets and that doesn't require escaping. ASCII US (Unit Separator) for comma ASCII RS (Record Separator) for newline ASCII GS (Group Separator) between worksheets *maybe* ASCII FS (File Separator) for embedded images Three header rows: 1. name of worksheet 2. Names (like usual) 3. Types (string, int, date, etc) Editors should treat Unit Separator like tab, except that it should be calculated to the data-width and data can wrap within the column based on maximum display width or column width. RS should obviously represent rows, and GS for tabs. Binary data must be hex encoded (because it compresses better than base64). Any data containing those control characters or other non-printing, non-UTF characters is considered binary. Metadata can be stored in a group before the worksheet begins. Lines starting with `#` and followed by `.` and no other separator until GS can be used to pad groups to a rounded standard size to avoid full-file rewrites on simple metadata or single-value update. No formulas or user-oriented spreadsheet stuff. This is just a data format for APIs and AIs.

In short: any task complex enough to require a SoTA / frontier model should be immediately captured on first use and parameterized for use with (and later training of) on-device models.

What does that look like? Well, I'm new to this but I think that it's: - the default login shell is an agent-enabled shell - the agent does expansion and gates what it runs not as a security feature, but to help the model produce better results - the agent edits failed commands, and file-content that didn't pass test, out of the context automatically - the agent captures generated scripts as part of skills - thousands of them - and sidesteps the model when vector search indicates a particular script fits We're in an era of automating tasks slowly, manually, and inconsistently when we could be automating _automated_ tasks.

I think that the next generation of Agent tools is still 5 years behind. They're delivering CD mp3 players in an era that we've already leapfrogged the mini hard drive straight through to the 3D nand. It's time to move past the imitation game and rethink from first principles with the current technology.

Also, we need git-style configuration for Ai agents.

We need an Ai-enabled shell. Meaning that when you type a command that's in the PATH, it executes the command, but when you start with a capital letter and it's not a PATH in the command, it invokes the Ai. Same rules apply that a leading space hides the command and the data from the Ai.

@ollama @SuperGrok GLM-5 does seem really great for code review.

GLM-5 is a little too eager to just do stuff (at least in OpenCode). It'll say "I need to ask the user if X or Y, and pause for 0.5s, and then move on. It always wants to remove my comments. It overgeneralizes - I tell it to pass the salt and it invents an arbitrary condiment passing algo that breaks stuff by passing anything that starts with "s" and ends with "t"... and then edits the test fixtures to pass the tests rather than the other way around. I'll have exactly the correct CLI command in my skill file and it will decide to "improve" upon it. It's done some things better, but I wish it would... chill - take an Addie or something.

This month @SuperGrok lost its stickiness for me because there's no coding agent tools, and no easy way to configure it with OpenCode and such. It's been almost a year together. I'll truly miss it. I'm now trying out Claude and looking forward to testing Ollama Cloud, and OpenCode Zen.

@souly9999 @SuperGrok I still have a few days on my subscription, but I can't figure out how to connect it to a coding agent or get an API key. Even though I'm subscribed the only way I can find to get an API key is to buy credits.

@_beyondcode @SuperGrok At least since a few days, there is API for Grok 4.20 its very straight forward used agentically.

@ollama @SuperGrok I'll have to figure out which are the right jobs for which models and update again once I have more experience.

Well, yes, but also Claude Code / Claude CLI kinda sucks. The tooling pieces are there... but they're so buggy and clanky. Have you tried OpenCode? That feels like it's written by senior-level engineers. It's snappy, it displays the correct information... it's polished. I'm so confused as to how these companies spend millions to billions on models and training, but they can't fork out a few hundred thousand to hire a 2- or 3-man team of experienced software developers to write tooling. I guess it fits in with the whole "Python: geared towards childen and PhDs" vibe: youtube.com/watch?v=YnL9vA…

@_beyondcode @SuperGrok lost supergrok stickiness because no coding agent tools - this is the make or break for most ai products. without agent integration, its just a fancy chat interface. claude code handles that part well

Reasoning Theatre. Interesting.

OpenCode is dope. Before I was fine with all the buggy, clunky interactions in Claude Code / Claude CLI because, well, it got the code done... but this is professional. I'd pay for this (as a tool).

I am suddenly okay with RAM costing $900. This made it worth it.

Is it just me, or did upgrading from Claude Pro to Claude Max make it much slower? Like, now they got me, let the value extraction begin! This is a serious question.

@goddek Why were you following and recording me? Thanks for doing the Ai face swap though.

The web was amazing prior to the reactification of everything 😄. Hotwire's Turbo framework is based on Chris Wanstrath's original Pjax concept (that powered GitHub in those early years). This kind of speed is still possible!

This is why you give a little more time than you think is reasonable for the login redirect flow. re: Claude login code expiration is 15s, but it's taking 20-40s to complete. Also why you should implement your own auth, in Go.

C 👏 S 👏 V 👏 Code agents love it. Customers love it. Your database loves it. Your customers' spreadsheets are it. I'm making ALL of my new APIs with it.