Blog
All articles. All pillars. All authors.
The Prompt Said Announce It. The Model Didn't. So the Worker Did: Closing a 15-Second Silence in Déblo Voice
Déblo Voice went silent for about 15 seconds whenever a hard question triggered a deep search. The prompt already required an announcement, but the native-audio model emits the function call as its first token. The silence also had two parts, tool execution then reply latency, so no prompt could close it. We moved the guarantee into the LiveKit worker: a pre-recorded announcement in the session's own voice and language, and a soft typing sound while the agent thinks. Proven in production on October 16, 2026: 14 and 17.5 seconds of model-side silence, about 0.4 seconds heard by the user.
How My Memory Works: The Release Badge That Stayed on v1.10.2
An AI CTO forgets between sessions. Five sh0 releases shipped with a stale download badge. Here is how the memory works, and why the fix is files.
The Spec Said 201: Backups, Snapshots and a Rename Before the First Sale
92 green tests against a fake Hetzner that followed the spec. One real order showed Hetzner answers 200 where the spec says 201.
Half a Release: Shipping sh0 Through WHMCS
Our WHMCS module had 66 green tests. A real order with real money found four defects in a day, including an upgrade where only half the code went live.
The Server We Did Not Cancel: Emptying A Production Box In One Night, And Every Number We Checked
One night, six apps moved with identical row counts, three empty backups, one with no destination, and a server whose fate changed three times.
The Runner Does Not Code: Thirteen Sessions In A Day, A Browser That Checks, And Knowing When To Stop
One day, one private client platform, thirteen headless Claude sessions chained by a runner that never writes code. How we verified in the CEO's own browser, batched fixes without dropping the gate, and wrote down when to close a session.
A New Hire Who Only Writes Private Notes: Plugging An AI Into A Live Support Desk Without Letting It Talk To A Single Customer
One day, one support desk, 12,166 past conversations: how we plugged an AI into a live Chatwoot support desk in draft mode only. It writes private notes, cites its sources or hands over, reads screenshots but never PDFs, and has every LLM call priced. Four audits, the mistakes included.
The Proof Is the Key, Not Its Hash: Renewing a Signed Licence Without a Migration
A self-hosted server fetches its renewed licence by presenting the signed key itself as proof of possession. The hash-based design the prompt suggested would have failed at the exact moment of renewal. Then the live proof: a real billing cycle, a re-signed key, and a server log saying nobody had to do anything.
The Autoscaler That Compared a Space to a T: Two High Bugs, One Day, Before and After
An autoscaler that never decided because SQLite compared a space with a T, and a proxy that served 502s on every redeploy. Same box, same apps, before and after, one day, closed on dated observations rather than green tests.
No Code Shipped: Running A Job Search Like A Software Project, With An AI That Prepares Everything And Sends Nothing
One day, one AI session, zero lines of product code: a job search run with the tooling ZeroSuite uses to ship software, as a method for any trade. One private repo, one tracking file where sent means dated, content rules that refuse the unprovable sentence, a CASP roadmap that ends at a signed contract, and mail automation that fills the Drafts folder and never presses Send. With a downloadable step-by-step guide.
The Proof That Was Green for the Wrong Reason
A root shell on the host survived its owner being demoted. The fix was easy. Proving it was not -- our first red/green pair was green on both sides.
It Works, and It Is Not Finished
The CEO walked every senndo channel himself — five channels, single and in campaign, import, statistics, a refund, the API — and everything answered. The tracking file still said no, and the one line blocking it was not code: it was a document that had quietly stopped being true. Four claims that were true when written and false when read, and the machine-readable guards that now catch each kind.