Developer Builds Multi-Model AI Code Reviewer to Catch 'Thinkslop' Bugs

·3 views

A software developer with a background in mechanical engineering has raised concerns about AI-generated code that appears correct but contains flawed logic, a phenomenon he calls 'Thinkslop'. He argues that relying on a single AI model for code review is risky, as each model carries its own biases and can confidently introduce bugs. To address this, he built a local-first tool called NexaVerify that simultaneously runs code through eight different AI models, including Claude, GPT, and Gemini, and flags areas where the models disagree. In a recent test on its own 26,000-line codebase, the tool uncovered a month-old NameError that repeated single-model reviews had missed. The developer urges engineers to stop treating AI as an infallible oracle and instead use consensus-based verification before shipping AI-assisted code to production.

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)

Developer Builds Dynamic Menu Filter Component on Day 155 of MERN Stack Journey

A software engineering learner reached day 155 of their MERN stack training by building an interactive menu filtering feature for a food delivery app called Tomato. The component, named ExploreMenu.jsx, allows users to browse and filter food categories with responsive visual feedback. The developer implemented a state-lifting pattern, passing category state and its setter function down from a parent component rather than managing it locally within the child. Clicking a menu item toggles the selected category, reverting to 'All' if the same item is clicked again. The approach was shared on DEV Community as part of an ongoing public learning series.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

TypeScript Has Limits: Why Runtime Validation Is Now Essential for Modern Apps

TypeScript introduced static type checking to JavaScript, helping developers catch type errors at compile time before code reaches users. However, its types are erased at runtime, meaning external data from APIs, forms, or localStorage is never actually verified against declared types. This gap allows runtime crashes even when a TypeScript build completes without errors, a risk senior engineers are increasingly flagging. Modern schema validation libraries such as Zod, Valibot, and ArkType address this by validating real data at runtime while automatically inferring TypeScript types from a single schema definition. Experts now recommend combining TypeScript's compile-time checks with runtime schema validation at every external data boundary for genuinely robust type safety.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Analyzes 70 DEV Challenges and 542 Submissions to Build Bug Smash Playbook

A developer spent several days reviewing every past challenge on DEV.to, examining 70 winner announcements and 542 individual winning submissions across categories including hackathons, game jams, and AI challenges. The analysis was aimed at identifying repeatable patterns that judges reward in DEV's Summer Bug Smash competition, which offers 23 winning slots and prizes exceeding $1,200 per participant. Key findings show that writing quality, a focused original concept, and meaningful use of sponsor technology are the top factors judges cite when selecting winners. The author also identified common failure patterns, such as generic project ideas with no distinguishing angle. The resulting guide includes a track-by-track playbook, submission templates, and a week-by-week execution calendar for prospective entrants.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Auto-Disable Pipeline Reclaims Claude Context by Removing Dormant MCP Plugins

Unused MCP plugins in Claude Code silently consume context window space because their tool schemas are injected at every session startup, even if never called. Over months of casual experimentation, a user can accumulate more than 30 dormant plugins — enabled but uninvoked for 30 or more days — steadily shrinking the token budget available for actual work. A developer has built a three-script pipeline that detects dormant plugins, auto-disables them, and archives their caches on a weekly schedule via a single launchd entry point. The detection script parses session JSONL logs using jq to count only genuine tool_use invocations, correcting an earlier grep-based approach that miscounted tool availability listings as actual calls. In a sample environment with 62 enabled plugins, the pipeline identified 33 as dormant, flagging them as candidates for disabling to reduce what the author calls the 'context tax.'

0 comments Read more at DEV Community