@mausbot_ is there a way to write in a different S2T model on the mic button? The default is not the greatest and impacts how easy it is to use. Loving everything else. Voice holding it back though.
Something is wrong if you’re burning 3% on a tweet. Only thing I’d suggest is try writing a tweet and tell it explicitly to not use any other connectors except X, just in case it’s grounding itself in every connector every time. If it’s not that, it sounds back end metering issue. Hope you get fixed soon.
@itsdanalex@umar_who_code Well, that’s the problem with strategy it’s contrived. Just share and if somebody finds it useful, then great. Stay authentic.
@nvandewetering I’m using voice in Grok Bot and Dots. I find Bot a bit more responsive. Biggest challenge with voice interface at the moment is in an open-plan work environment as it feels a bit unnatural talking to a computer in public.
@nvandewetering Totally agree with this. All models and harnesses are useful, but the main decision is where you allocate funds. ChatGPT represents an army of midtier workers that are great for the job.
When I first got Grok Bot driving Codex CLI and Claude Code CLI, I didn't know how it would find out a job had finished. Mostly it didn't. A background shell didn’t wake an idle agent, and a detached job gave no signal, so I kept asking "has it finished?".
The first fix was a curl at the end of each job's https://t.co/X024Ddwc0w to a Grok Bot Webhook Routine. That only works if the job reaches the curl. On the old helper, 11 of 31 runs never sent a wake because the webhook env vars weren't there. The jobs often finished fine; nobody was told. A killed job would just go quiet. And leaving Grok Bot sat in a turn polling instead burns usage on waiting.
So we built a small watcher that doesn't rely on the job behaving. Jobs launch through job-run, which keeps the webhook secrets out of the worker and writes an exit receipt even when the job is killed. A separate notifier delivers from a SQLite outbox to the Webhook Routine, and the lead gets woken with the result.
Within the first couple of hours, 12 real jobs (9 Codex CLI, 3 Claude Code CLI) plus a canary. 8 succeeded, 4 failed, and all 13 wakes were accepted first time, HTTP 200. The median finish-to-webhook was 3.7s and the case it was built for turned up too; a Codex job got SIGTERM after 58.5s and its failure wake was still accepted just over four seconds later. Three more pushed cleanly, failed post-push checks, and still sent failure wakes.
Caveats, because it's super early: a 200 means the routine accepted the wake, not that the agent acted on it. The same wake can occasionally arrive twice, so each one carries an ID and the agent should ignore repeats. After a full reboot, the watcher needs a nudge to restart. The webhook keys sit in a locked-down file that other programs on the box could still read, but at worst that means a fake or wasted wake. And it spots hung jobs but only flags them for now; it won't kill a job that might just be slow.
The setup prompt is in the first reply if you want to build your own.
@thejakebuzz You might have already done this, but create a usage bot to scan your activity each week looking for repeated things that could be done by a script on the computer. Or even a skill to take the labour (usage) out of redoing everything.
@lukedewolf@trueadm It is good if you like to drive everything yourself, and dots is going in the right direction. It doesn’t agentically scale though yet. Grok Bot is well in the lead on that front - albeit far from perfect.
@zaneccampbell Maintaining bi-directional triggers and hoping they wake, behave, and act accordingly. So maintaining routing, task state and delivery across 3 harnesses often fell apart. The simpler approach was let Grok Bot drive, but then we had nerfageddon and usage tightening.