Developer cuts LLM token usage 42x with deterministic code-edit pipeline D-Engine

·1 views

Software developer Sergi Corruchaga published findings in September 2026 showing his open-source tool D-Engine completed the same coding task using 2,552 tokens, compared to over 107,000 tokens consumed by DeepSeek's official agent under default settings. Corruchaga attributed the gap not to the underlying model or thinking mode, but to architectural differences in how each tool manages context. Conventional agentic coding tools re-send the full conversation history on every loop turn, causing token costs to grow quadratically with the number of steps taken. D-Engine separates responsibilities by having the LLM only generate structured patch blocks, while a local deterministic runtime applies changes to an isolated copy of the repository and runs compiler checks before merging. The tool is available under the MIT licence as D-Engine v0.2.2.

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)

One Developer Spent a Full Day Using Only Expo-Built Apps — Here's What He Found

A developer challenged himself to spend an entire day using only apps built with Expo's open-source tooling, covering everything from sleep tracking and news to cooking and car washes. The experiment revealed that Expo-powered apps handle a surprisingly wide range of daily tasks, including weather, air quality, recipes, and even Tesla car controls. However, a significant gap emerged around messaging, as no viable Expo-built messaging app exists that could replace everyday tools like iMessage. The developer resorted to paying a celebrity via Cameo — which runs on Expo — to deliver a message to his wife, highlighting the platform's creative but uneven coverage. Overall, the test showed Expo apps can sustain a full day's routine, while also pointing to clear opportunities for developers to fill underserved categories.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Claude Task Master CLI Automates Full Pull Request Lifecycle for Developers

Claude Task Master is an open-source command-line tool designed to automate the entire pull request workflow, from planning and code writing to merging. The tool reads a repository, generates a task list, executes changes in isolation, and opens pull requests automatically. If CI checks fail, it updates the code and pushes new commits without human intervention; it can also fetch and apply GitHub review comments on its own. Additional features include session state persistence, multi-profile support, and external integrations via a REST API, MCP server, and signed webhooks. Available under the MIT license, it can be installed through PyPI or Docker and works with Claude Code or any Anthropic-compatible API.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

VernLLM: Open-Source TypeScript Framework Built to Handle Large-Scale LLM Calls

VernLLM is an open-source TypeScript framework designed to manage high-volume calls to large language models. The tool offers features including adaptive AIMD rate-limiting, advanced circuit-breakers, and multi-provider fallbacks. It also supports configurable retries with budget controls and built-in observability for monitoring call performance. Released under the MIT license, VernLLM is publicly available on GitHub and as an npm package at vern-llm.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Connect Power BI to Local and Cloud SQL Databases for Dashboards

A tutorial outlines the process of connecting Microsoft Power BI to both a local PostgreSQL database and a cloud-based SQL database hosted on Aiven, a cloud database services platform. Setting up the Aiven connection requires creating a service, noting credentials such as host, port, and username, and installing an SSL certificate in Windows' Trusted Root Certification Authorities store. Once connected, raw data is imported via DBeaver and loaded into Power BI, where Power Query is used to clean issues including missing values, mixed capitalisation, and inconsistent date formats. The cleaned dataset is then used to build an interactive dashboard featuring charts, slicers, and key business measures such as total revenue, gross profit margin, and average delivery days. The resulting dashboard is intended to support management in making data-driven decisions.

0 comments Read more at DEV Community