Developer fixes AI agent bug that triggered redundant context compression on every restart

·1 views

A software developer discovered that their long-running AI agent was repeatedly compressing conversation history on every process restart, even when compression provided no benefit. The root cause was an in-memory counter tracking ineffective compression attempts, which reset to zero each time the process restarted. To fix this, the developer introduced a persistent 'thrashing guard' that stores the ineffective compression count in a database, allowing it to survive restarts. A centralized method was added to update both the in-memory counter and the database atomically, ensuring consistent state. On session resumption, the agent now reloads the persisted counter, preventing redundant compression loops that waste tokens and processing time.

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)

Browser-Based Image Compression Needs No Server, Just Ten Lines of JavaScript

Modern browsers have supported full client-side image compression since around 2012, using APIs like createImageBitmap, OffscreenCanvas, and convertToBlob — eliminating the need to upload files to external servers. A binary search loop targeting a desired file size generally outperforms a quality slider, as JPEG quality is non-linear and low settings introduce visible artifacts. Developers must handle several pitfalls, including canvas stripping EXIF orientation data and color profiles, transparent PNGs rendering with black backgrounds when converted to JPEG, and large images consuming hundreds of megabytes of RAM. Processing should run inside a Web Worker with a queue to prevent large or multiple files from crashing the browser tab. The key privacy advantage of this approach is architectural: the file never leaves the user's device, which is a fundamentally stronger guarantee than a server-side deletion promise.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer builds 'facet' CLI to manage multi-repo workspaces for AI coding agents

A DevOps engineer created a Go command-line tool called 'facet' to solve the problem of multiple AI agents and developers conflicting over shared Git working trees across several repositories. The tool creates isolated, disposable workspaces tied to individual GitHub issues, automatically cloning relevant repos and generating a CLAUDE.md context file for the agent. Facet infers which repositories belong to a task by reading issue fields and cross-references rather than relying on labels, which the author found too broad. An early version that integrated with the terminal multiplexer Zellij failed badly, causing lost work, and was scrapped in favour of a simpler design after deleting around 1,500 lines of code. Notably, much of the tool's later hardening was itself carried out by Claude Code agents working through an issue backlog under the author's review.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

What Is RAG and Why It Matters for Building Reliable AI Chatbots

Large language models (LLMs) like ChatGPT rely on fixed training data and can confidently produce incorrect answers when asked about organisation-specific information. Retrieval-Augmented Generation, or RAG, addresses this by fetching relevant documents from a company's own data sources before the model generates a response. Instead of guessing from memory, the model reads and reasons over the actual content provided to it at query time. This approach significantly reduces hallucinations and makes AI outputs easier to audit, since incorrect answers can be traced back to specific retrieved content. RAG has become the dominant architecture for enterprise AI applications that require accurate, up-to-date, and domain-specific responses.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to model Germany's 16 varying property transfer tax rates without hardcoding errors

Germany's Grunderwerbsteuer (property transfer tax) varies by state, ranging from 3.5% in Bavaria to 6.5% in North Rhine-Westphalia, creating up to a €12,000 difference on a €400,000 property. A common developer mistake is storing these rates as static constants, which silently produces incorrect historical data when rates change over time. The correct approach treats rates as time-scoped records tied to specific date ranges, ensuring past analyses remain reproducible and accurate. Additional pitfalls include floating-point errors in monetary calculations, which can erode user trust, and the lack of a standard definition for property yield across listing platforms. Developers building real estate tools are advised to surface all assumptions — such as vacancy rates and maintenance reserves — so users can adjust inputs rather than rely on opaque default outputs.

0 comments Read more at DEV Community