Self-Taught Dev Uses Dual Implementations to Catch AI Coding Agent Bugs

·2 views

A self-taught engineer transitioning from restaurant work built an evaluation framework to verify the correctness of agentic AI code. The system runs identical match-3 game logic across two independently written implementations — a React build and a Java engine — and flags any disagreement as evidence of a bug. Logic-invariant tests validated across roughly 9,000 random boards complement the differential oracle, catching errors that conventional single-implementation test suites missed. The approach is embedded in an autonomous development harness featuring a self-review loop, an independent critic, and a human-approval gate for irreversible actions. The developer has open-sourced both the dual implementations and a deliberately buggy version to demonstrate the oracle's detection capability.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Engineer Documents ESP32 Smart Key, MikroTik QoS and Self-Hosted Server Project

A technical report has been published detailing a combined network infrastructure and embedded systems engineering project. The project includes a wireless SmartKey module built on an ESP32 microcontroller and a 433MHz RF receiver for motorcycle remote control. On the networking side, a MikroTik router running the CAKE QoS algorithm was deployed to reduce bufferbloat and improve latency stability on local networks. The infrastructure component uses Nginx and Cloudflared Tunnel inside local containers to securely expose services without a static public IP. Hardware schematics, network configurations, and test results are available through a scientific repository.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer upgrades two Laravel apps in layers, adds security headers and job reconciler

A developer documented a week of backend work that included migrating two Laravel applications off an outdated internal skeleton by applying changes in six discrete, independently revertable layers rather than a single large commit. Security response headers — including X-Frame-Options, Referrer-Policy, and HSTS in production — were added as a global middleware default rather than per-route decisions. A data retention purge command was built with explicit guard rails to prevent accidental deletion from high-risk tables. A queued-job reconciler was shipped to catch jobs that were queued but never executed. The week also included a custom 404 page and a QR-code redirect with a safe fallback URL, reinforcing an overarching theme of sensible defaults that reduce the risk of late-night incidents.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

The Reconciler Pattern: Fixing Silent Queue Job Failures That Never Show Up in Logs

When a queued job is dispatched only once at row-creation time, it can silently disappear due to worker downtime, redeployments, or queue clears — leaving database rows permanently stuck in a pending state with no error or failed-job entry. Unlike standard job retries, which only apply to jobs that actually ran, these lost dispatches are invisible to queue-level monitoring. The reconciler pattern addresses this by running a scheduled database sweep — typically every 15 minutes — that identifies stale, unclaimed pending rows and re-dispatches their associated jobs. Key implementation safeguards include touching the row before dispatch to prevent duplicate re-queuing, applying a batch limit to avoid flooding the queue after prolonged outages, and ensuring the job handler is idempotent so accidental re-runs are harmless no-ops. The core principle is that the database is the durable source of truth, while the queue is merely a delivery mechanism — and domain-level reconciliation must complement, not replace, queue-level retries.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Repo-rter v0.4.1 Ditches LocalStorage for Native File System in Tauri App

Repo-rter, a desktop app built with Next.js and Tauri v2 to extend GitHub's 14-day traffic history limit, released version 0.4.1 with a major overhaul of its data storage layer. The previous approach relied on browser-level localStorage inside a WebView, which proved unreliable due to OS cache purging, a 5MB storage cap, and difficult backup restoration. The team rewrote the data persistence layer in Rust, using native Tauri commands to read and write JSON files directly to the OS Application Data directory, which is protected from system cache sweepers. The updated TypeScript storage logic detects whether the app is running in Tauri and routes data accordingly, falling back to localStorage only in web browser development environments. The release also addressed GitHub API rate limit concerns for power users tracking 50 or more repositories by introducing a sliding-window sync strategy.

0 comments Read more at DEV Community