Open-Source Library Uses Deterministic Folding to Manage LLM Agent Context Limits

·1 views

A developer has released Context Warp Drive, an open-source TypeScript library designed to help long-running LLM agents stay within context window limits without relying on summarization. The library uses a technique called deterministic folding, which compacts older transcript sections into stable artifacts while preserving precise references and keeping recent messages intact. Current features include rolling fold compaction, provider prefix caching support, episodic memory extraction, and helpers for Anthropic, OpenAI, and Gemini. The project has 459 deterministic tests and is available on GitHub, though its npm release is pending account authorization. The developer is seeking community feedback on whether deterministic folding is the right low-level abstraction for long-lived agents or whether such logic should sit at the application layer instead.

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)

Google's A2A Protocol Finds Niche Use Case a Year After Divisive Launch

Google introduced its Agent2Agent (A2A) protocol in April 2025, positioning it as an open standard for communication between independent AI agent systems built on different frameworks or vendor stacks. Unlike the Model Context Protocol (MCP), which connects agents to tools and data sources, A2A is designed to handle task delegation between agents that have their own capabilities and trust boundaries. The announcement drew mixed reactions from developers, many of whom questioned the need for a new standard when MCP already existed and most teams were still solving basic single-agent challenges. By 2026, A2A has neither faded away nor achieved universal adoption, but is gaining traction in specific scenarios involving genuinely independent agent systems. Its practical value hinges on understanding the distinction between tool integration and agent-to-agent delegation, which the protocol was specifically built to address.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Beginner's First Week With Power BI: Key Concepts and Takeaways

A self-taught learner has shared their initial experience exploring Microsoft Power BI, describing it as a more powerful alternative to Excel designed for larger datasets and interactive dashboards. During the first week, they studied how Power BI connects to multiple data sources, including Excel files, CSVs, web pages, and databases. They also learned the importance of data types and explored the Power Query Editor, which allows users to clean and transform raw data before analysis. An introduction to DAX (Data Analysis Expressions), Power BI's formula language, covered basic functions such as SUM() for calculating metrics like total revenue. The learner noted that the tool has already shifted their perspective on data, moving from viewing it as rows and columns to recognizing patterns and insights.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

LangChain Structured Output: Forcing LLMs to Return Reliable, Machine-Readable Data

Building production-grade LLM applications requires more than plain text responses — enterprise systems need consistent, machine-readable output for tasks like API integration, ticket classification, and workflow automation. LangChain's structured output feature addresses this by constraining LLMs to return data in predefined formats such as JSON, Pydantic objects, or typed dictionaries. Developers can use the with_structured_output() method with a Pydantic model to ensure the LLM's response is automatically validated and parsed into a usable Python object. Internally, LangChain converts the schema into model instructions, receives the response, validates it, and returns a structured object rather than raw text. This approach eliminates unpredictable formatting issues that cause backend failures in production environments.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Three Keyless Public APIs That Can Power a Stock Dashboard for Free

A developer has identified three public, API-key-free data sources that together cover the core needs of a stock dashboard: price history, company fundamentals, and an earnings calendar. Nasdaq's quote API returns daily OHLCV data for any ticker via a simple HTTP request, while the SEC's XBRL database provides structured financial filings including revenue, EPS, and net income for all public companies. The SEC endpoint requires only a descriptive User-Agent header with a contact address to comply with its usage guidelines. All three feeds return plain JSON over standard HTTP, keeping per-run costs minimal compared to solutions requiring headless browsers or proxies. The developer has packaged each feed as a pay-per-use scraper on Apify, with the first rows of every run available free to verify data shape before committing.

0 comments Read more at DEV Community