DEV Community: Kyalo460

Even I'm not using my APP

Kyalo460 — Wed, 07 Oct 2026 14:16:13 +0000

So I noticed something huge, even though I'm not using the app I created for my daily work. How and why do I expect other users to stick with my app if it doesn't even please me? I feel like I missed a foundational building block. I kept telling myself, I'll start using it soon; it's not ready yet; maybe other people will find it useful...

That changes today! I'm going to add a new feature that should attract more frequent users. I'm not sure if it is achievable, but I won't know until I try. The idea is: what if users were to give tasks to agents from my app??? This should transform it from a utility to a standalone app, which should equate to more users. But I need to do it right. Stay tuned>>>

Here is the link to the app tho:
https://aiwatchdog.vercel.app/

Most of my traffic is coming from USA!!!

Kyalo460 — Tue, 06 Oct 2026 02:47:01 +0000

It's day 4 after launching my website, and I noticed that most of my users are coming from the USA. This is surprising because I have done more marketing in Kenya, but since it's an AI tool, it makes sense that more USA residents would be interested. I will keep an eye on this and will probably focus my marketing efforts abroad as opposed to locally.

Here is the site I'm talking about:
https://aiwatchdog.vercel.app/

AI WatchDog - Give agents a task, and go on about your day.

Kyalo460 — Mon, 05 Oct 2026 15:57:56 +0000

I have made a website that monitors all your agents and alerts you when an agent is stuck, idle, or caught in a loop. This way, you can give prompts and go about your day without the need to check back every 5 mins to see if the AI is still generating code.

The app is open source and free to use; log in, run a command in the terminal, and go make dinner. You'll get notified when you are needed. The link to the site is:

https://aiwatchdog.vercel.app/

I appreciate any feedback. Thank you for your time.

I got some feedback that the app is too vague, and users don't really know what to do. So I added a 'How to use' tab that guides users on how to utilise it to be more productive. Here's the link to the latest build https://aiwatchdog.vercel.app/

Kyalo460 — Mon, 05 Oct 2026 06:52:25 +0000

AI Watchdog — Stop babysitting your AI agents

The open-source watchdog for autonomous AI workflows. Get told when your agents are done, stuck, idle, or waiting for you — and only then.

aiwatchdog.vercel.app

Is software overpriced nowadays?

Kyalo460 — Sun, 04 Oct 2026 02:58:46 +0000

I had an argument with a fellow software engineer recently.

He said: “Software isn’t overpriced.”
I said: “Bro… the game has changed.” 😂

Think about it. A few years ago, building a SaaS could mean weeks or months of development, expensive infrastructure, servers, databases, DevOps, and a small army of engineers arguing about whether to use tabs or spaces. Today? AI can write half your boilerplate before you finish your coffee. You have free AI tools. Free databases.Free hosting tiers. Open-source everything. Cloud platforms that let you deploy with three clicks. And with the right stack + AI agents, a solo developer can sometimes go from “I have an idea” → “here’s the MVP” in a few days — sometimes even hours.

So yes, good software still takes skill. Architecture still matters. Security still matters. Maintenance definitely matters. But the cost and time of getting software from idea to working product has fallen dramatically.

And I think that creates an uncomfortable question: If the cost of producing software keeps falling, should the price of software stay the same?

Maybe the future isn't about charging more because software is expensive to build. Maybe it's about charging for value, reliability, distribution, support, and the problems you solve. Because if two developers can build the same MVP over a weekend…“It took us 6 months to build” probably isn't the strongest sales pitch anymore. 😂

The software industry just got a massive productivity upgrade.

Now we're all trying to figure out what that does to the price tag.
hashtag#SoftwareEngineering hashtag#AI hashtag#SaaS hashtag#Startups hashtag#Programming hashtag#DevTools hashtag#ArtificialIntelligence

AiWatchDog, monitor your agents, increase productivity

Kyalo460 — Sat, 03 Oct 2026 01:50:17 +0000

I Built an Open Source Watchdog for AI Coding Agents
Live demo → https://aiwatchdog.vercel.app/ · Open source (MIT)

The problem
I run a lot of AI coding agents — Kilo, Claude Code, Codex, Aider, my own scripts. And I kept hitting the same wall:

I start a task, walk away, come back 20 minutes later, and stare at a terminal with no idea whether the agent is crushing it, waiting for me, or wedged in an infinite retry loop burning tokens.

So I built the thing I kept wishing existed.

AI Watchdog — real-time visibility and intervention for AI coding agents. It ingests everything your agent does, decides with a multi-signal confidence model whether it's working / waiting / genuinely stuck, and notifies you only when you actually matter to the outcome.

Why "stuck detection" is the hard part
The naive rule — no output for 5 minutes → STUCK — is wrong, and I refused to ship it. Agents legitimately go silent while compiling, installing dependencies, running Docker, or grinding through a long test suite. Alert on that and you'll mute the tool within a day.

So Watchdog runs a weighted ensemble of independent detectors, each producing an explainable signal:

Signal Weight Suppressed by
No heartbeat 20 explicit waiting_for_user
No tool activity 20 known long phases (build/test/docker)
Repeated similar errors 25 one-off failures
Repeated commands 15 parameterized commands
No filesystem changes 10 non-coding executions
Process looks idle 10 returns null if processes aren't visible

Confidence is the clamped sum of triggered weights — and every contributing weight is traceable, so the UI can always answer "why does it think you're stuck?"

Here's what I'm proudest of: the arithmetic works in your favor. Total silence scores 55 — five points below the 60 threshold. A coding execution with no file changes hits 65 and crosses. Silence is not evidence. That was a design rule, not a tuning accident.

What's in it
Live Command Center — every execution streaming over WebSocket
Trace view — the full event tree, so you see the exact moment it derailed
One command connects any agent: watchdog wrap -- kilo run "fix the failing test" Kilo, Claude Code, Codex, Aider, OpenCode, npm test, your own script. stdin passes through, exit code propagates, heartbeats every 15s.
TypeScript + Python SDKs — five lines to instrument an existing agent
VS Code extension — terminal commands and file changes
A stuck-agent simulator with 5 modes, so you can prove the detector before trusting it
Multi-tenant by construction — organizationId is required in every repo method, so a method that forgets it doesn't compile
Fail-closed redaction — if the redaction module can't load, ingest throws rather than persisting raw secrets
~1,500 tests, nothing mocked in e2e — real API process, real Postgres, real WebSocket, real child CLI
The honest hard parts
I shipped a cross-tenant read oracle and then found it. A global UNIQUE(dedupe_key) with an unscoped lookup meant anyone guessing stuck: could read another tenant's alert. Rescoped to (organization_id, dedupe_key).

My partition helper could delete the dev database. pg_class matched across all schemas, so a parallel test schema made a test drop production partitions. I observed this rather than theorized about it. Now pinned to current_schema().

State writes weren't compare-and-set — a read-then-write race. Now UPDATE ... WHERE id=$1 AND state=$2.

Deliberate deviations: no Next.js (SSR buys nothing for an auth-gated realtime SPA), no ORM (hand-written SQL, deterministic, full control of monthly RANGE partitioning + CAS), and the detection engine is pure — no DB, no network, no process.env — which is exactly what makes it exhaustively testable.

The load-bearing detail nobody sees: persisting hysteresis state is the write that matters. Skip it and STUCK detection silently never fires.

What's next
Extracting the sweeper to a real BullMQ worker with advisory locks, so multiple API replicas stop double-sweeping
Turning on real adapter capabilities — supportsPauseResume and supportsProcessMetrics are false for every adapter today. I'd rather ship an honest capability matrix than pretend
A real observability loop for editor-based agents — the VS Code extension can't see chat traffic inside another extension (Kilo, Cline, Roo); no API exposes it, so the localhost bridge plus terminal/file signals is the honest ceiling right now
SSO/MFA, audit logs, richer notification rules, deeper CI integrations
Why open source
The interesting part — the detection methodology, the weight table, the suppression rules, the hysteresis design — is more valuable shared than defended.

LLM tracing tools see model calls but not the shell, file, and git activity where agents actually get stuck. APM tools don't model agent execution at all. Agent platforms are closed and single-vendor. Nobody owns the neutral, execution-centric layer. I'd like this to be it.

→ Try it at https://aiwatchdog.vercel.app/

If you've ever lost 20 minutes wondering whether your agent was alive, I'd love to hear about it. And if your agent does something Watchdog doesn't understand yet, tell me — those reports are literally the roadmap.