Howaboua

Vibecoding End-boss ™. Chronically unemployed. Developing - the only Pi desktop app you want to use.

[Howaboua @Howaboua]

@maxjendrall Precisely my feeling I, It's just the bringback4o crowd that's against this, because they only see AI useful for one thing only and don't quite understand how tiny the difference between chat and work really is

[Howaboua @Howaboua]

BIG Pi Codex voice update coming today. This should actually feel like a true voice conversation.

Codex voice oauth doesn't allow custom tools. When the voice agent decides to delegate to the main agent with tools, to the model gets passed what you said last verbatim and loses the context of the conversation you just had.

This is how gpt-live and codex-live speech models work. There is no way to have it use tools. Gpt-realtime is different. You can add tools to it (this implementation is also coming to PCC given we will now get API credits), so it's a true voice agent not just a dumb chat. You can give it a delegate_task tool and it will rewrite what the main agent has to do into a prompt. Grok voice does that too.

Now enter @Jordy_vD_ who randomly suggested that there could be a separate session rephrasing the conversation into a user prompt.

So I'll be doing just that. Voice turns before and after

Before
User1=>Voice1=>User2=>Voice2=>User3=>
Voice decides to yield, it figured your last message is good enough to get passed as a task and passes to Coding. Coding only receives User3 and no other context for the task, apart from what it already has in its session, which is basically only the coding runs, and nothing you ever agreed with the voice session.

After
User1=>Voice1=>User2=>Voice2=>User3=>decides to pass to Coding=>Task delegator model receives everything between User1-User3 and rewrites it into a prompt=>Coding receives the rewritten prompt, with all the context of what you agreed with the voice session.

Works AMAZINGLY well. Uses Pi's streamsimple. I am using luna high for rewriting, it basically costs nothing and is fast enough to the point where you don't notice it. And your main Pi coding session receives a more detailed instruction.

Bosh. Dropping today. Alongside the option for API-based gpt-realtime with custom user defined tools, and an example how you can enable your extension tools to be available for the voice agent (I have enabled a switch for my websearch extension so the voice agent can "just use it").

[Howaboua @Howaboua]

@theo It's a decent orchestrator because of its speed, just never let it code anything. And I've been running my discord bot on it, it's the smoothest talker out of all the ones I've tested (muse spark 1.3, grok 4.5-7, luna 5.6-6)

[Howaboua @Howaboua]

@mjtechguy easy peasy :P

[Howaboua @Howaboua]

I got the mac dmg working on linux btw xD

[️ @blankspeaker]

SpaceXSI : Grok Bot has been updated to version 1.0.0

Direct download links are in the reply below. The public download page has not switched to this version yet.

Here are some of the new things they added:

💬 Messages and email
📧 Mail setup now asks separately to open emails and to press Send.

👥 Sharing and teams
🤝 A Shared Bots section lists bots others have shared with you.
🤖 Team bots can be given their own account or share yours.

🔌 Connectors
⚠️ Plugins can tell you they are not on your account yet.

💳 Payments and orders
✈️ Flight checkout uses a Book on [airline] call to action and title-case labels for Multi-city and Round trip.

Here are some things they have been working on but have not rolled out yet:
• Find, read, and save notes in a local Obsidian vault after you pick the vault folder on this Mac.

Some other things that were worked on since the last update:
• Search tells you to try a different query when a lookup does not work.

Note: These changes may or may not be available in this release. Based on unofficial 3rd-party SI analysis of prerelease builds.

The official changelog does not list 1.0.0.

For the official changelog see here: https://t.co/WJMjsooHBy

[Howaboua @Howaboua]

@rudavko @trq212 well Mario had a sad day yesterday

I'm not sure about the sad engineer part either, we do see plenty of really freaking angry ones tho, there's still a lot of anti ai warriors :> NGMI much :P

[Howaboua @Howaboua]

@DennisonBertram but the headscratch was definitely the correct approach!

esp these days

[Howaboua @Howaboua]

Interesting...

[Dennison @DennisonBertram]

Is there a security vulnerability in @bot @nikitabier?

I put in a API key that was malformed, and my Bot told me it didn't include "SK".

How could it know that if it's securely stored?

[Howaboua @Howaboua]

I had a pretty long post written, and then it dawned on me.

Just use Pi as the long running thread, forehead. Then you don't even have to change anything in t3code, you just clank up some extensions around your idea of recurring tasks, proactivity, optimal memory and context management...

None of the harnesses are really wired up for this out of the box, these are pretty low level things. While you can definitely make them organise their tasks in markdowns etc so they know what's up, and have them create their own cron jobs, the real issue is quota consumption burn.

It all revolves around compacting often, at small ctx windows, before cache expires, and using a harness that uses the least tokens to start up and is extremely controllable.

Anyway, it is extremely fun and rewarding when you get it right.

[Howaboua @Howaboua]

you do realise there is zero difference between chat/work modes apart from a few extra flags for tools, probably 3 lines in the sysprompt and a slightly different model?

the chat mode covers nearly everything that the work mode does. it never really made sense apart from a. copying anthropic (but anthropic has shared cowork/chat/code quota). b. given doing "work" related tasks normally costs more tools, their generous chat quota would've been abused, so they had to offload it to a mode that has less generous quota and is shared with codex.

and now with people creating connectors that wrap chat quota around coding agents, providing effectively unlmited sol for coding, the "hackers" have brought it upon themselves.

[Howaboua @Howaboua]

@TajemnikTV @thsottiaux well not until we know the quota mechanics behind it.

"but i want to chat and not work" is the outrage I am seeing, has nothing to do with quotas (and i do agree this would be a massive shift)

[Howaboua @Howaboua]

So apparently @bot now is aware of what time it is and receives timestamps with every user message? This feels new.

[Howaboua @Howaboua]

The bringback4o crowds are at it again. They saw Chat merging with Work and they are now raging that they dont want to do work, they just want to chat.

it's ironic how this very change would be for the benefit of people who do not understand what's going on in an agent loop beyond "it do be spitting out text and it's got emotions and a soul", and yet they decide to rage against the machine

Chill out peeps, it's probably a few lines of change in the system prompt, tops.

[Howaboua @Howaboua]

my discord bot (pi-backed) had a broken reload tool, got sent to investigate, then messaged another one of my agents to check its theory from the other side

two clankers comparing notes across projects, found a startup race, fixed it, reloaded itself

both running terra

astra never spotted this race btw

[Howaboua @Howaboua]

@poteto @ericzakariasson @bot that 1.0 version I already have installed :P

[Howaboua @Howaboua]

@thsottiaux We know its coming, the quotas being shared. I do not envy the PR backlash from this :X

[Howaboua @Howaboua]

@OpenAIDevs @bot Here's the answer btw for those curious. Pi extension coming soon xD

[Howaboua @Howaboua]

@OpenAIDevs @bot gimmie a rundown of how many proper voice agent hour that gives me so I can stop using the "meh" implementation via codex sub that allows no tools