I always bounce between Windows, Mac, and Linux, and I use coding agents from different providers.
Couldn’t find a tool I liked that worked the same on all three. So I built Yardsort and open-sourced it.
https://t.co/CSVVvf7ULo
Building Flotta's monthly client report: a page, an email and a PDF. One model, three renderers.
The PDF is made on the server, no browser. A PNG or JPEG logo is drawn. A WebP one becomes the agency's initial. PDFs don't carry WebP.
Sending it is next. Not on production yet.
My Claude usage would be thousands a month at API prices. 3 things that keep token waste down: compact your context, keep sessions short and scoped to one task, and let agents suggest memories you approve so they don't solve the same problem twice. Built into Yardsort, my multi-agent IDE.
#claudecode #ai #programming #coding #developer #techtok
@Iboumbaye_real J'ai créé rustunnel, une alternative open source à ngrok écrite en Rust. Gratuit avec 2 tunnels et 1 Go/mois (HTTP, TCP, UDP). Pour un sous-domaine fixe, c'est 3 $/mois minimum en pay-as-you-go, et tu peux aussi l'héberger toi-même. https://t.co/BUZNc3muCO
To prove Flotta's billing end to end I signed up from the pricing page and bought my own Studio plan. Single-use 99%-off code, refunded after.
Webhook set the limits and approved the agency. I added a client and an agent. No operator step.
Cancel suspended it. No invoice.
https://t.co/qbSTkQmnvh
Giving each Flotta agent its own screen. First spike: the browser tooling installed onto the agent's 1 GB volume. 749 MB, disk 84% full, next install died with "No space left on device".
A restart wipes the rootfs, so it's all baked in. Image: 857 MB.
Lately I falling into this pattern of using Goal more with my agents, especially when I am getting close to a reset and I dont have to worry about token limits.
Flotta's slow Settings tab, part 2. Swapped the flyctl subprocess for Fly's Machines API: 545ms to 131ms. Still about 1s.
Timed both reads. Machine: 20ms. The rest was `flyctl releases`, cached for 60s. My clicks were 67s apart.
Now it serves the cache and refreshes behind it.
@mo_tawfik89 Idle side yes, the free list keeps at most 16. A burst never waits on it though. No free connection, it opens a fresh one, and anything past 16 closes when it comes back instead of piling up. Each reused one gets a SELECT 1 first.
I removed a per-request Postgres lock from Flotta's control plane. Prod didn't get faster. Settings tab: still 1.4s.
py-spy: 82% of CPU went to opening connections. Render gives it half a CPU.
Now it reuses idle ones, no pool to queue behind. Locally 11 requests: 322ms to 18ms.
Yardsort ate a third of a CPU core on Linux while agents worked.
Culprit: a 6px pulsing status dot. In WebKitGTK every animation frame repaints the whole window. UI main thread at 57 to 60%.
Now the dot sits still and a clock counts the burst. About 1%.
I ran Yardsort on a Hetzner CX33 with no screen, then typed at the agent from my iPhone.
yardsort-host: no WebKit, no GTK. Sign-in by paste, serve under tmux, pair at the box.
A shell without XDG_RUNTIME_DIR started a second daemon. After restart, the real agents looked ended.
Flotta agent's chat died on "oi". Every turn: 403 Blocked.
Once it had run `sleep 120; curl ...` to wait on a deploy. Render's Cloudflare WAF reads bodies, saw shell in the model call, flagged injection. History resends every turn. Dead forever.
Gateway's on Fly now. No edge.