Qarinah cuts AI coding-agent context tokens by 98.7% while retaining full retrieval

·1 views

Qarinah is an open-source tool that compresses project memory into a compact, cited pack rather than replaying an entire codebase history at the start of each AI coding-agent session. In a benchmark across six software-task fixtures, the full-history baseline required 442,113 estimated input-context tokens, while Qarinah's approach used just 5,682 — a reduction of 98.71%. Despite the compression, every required retrieval target was still found within the top five results, and the tool passed all 380 deterministic file-specific queries across projects of varying sizes. The estimated token savings translate to lower input costs across common provider pricing tiers, though the figures exclude caching, output tokens, and other real-world billing factors. Qarinah is Apache-2.0 licensed, local-first, and compatible with tools such as Codex, Claude Code, and Cursor, with all benchmark methodology and fixture data made publicly available.

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)

VectorWare Brings Rust Portable SIMD to GPUs, Unifying CPU and GPU Code

VectorWare has enabled Rust's portable SIMD library, core::simd, to run natively on GPUs, marking a significant shift in parallel programming. Until now, GPU development required separate codebases written in CUDA or OpenCL, which operate on fundamentally different paradigms from CPU programming. The breakthrough works by mapping Rust's SIMD vectors onto a GPU's 32-lane warp units, allowing a single warp instruction to process vector operations just as AVX does on x86 CPUs. This means developers can write one set of Rust SIMD code that compiles and runs across x86, ARM, and GPU targets without modification. The development lowers the barrier to GPU programming significantly, potentially allowing any Rust application using SIMD for numerical workloads to leverage GPU acceleration without a rewrite.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Cline VS Code AI Agent Is Free to Install but Costs Up to $90/Month in API Fees

Cline is an open-source VS Code extension with over 5 million installs that offers agentic AI coding capabilities using a bring-your-own-API-key model. After five months of daily use, reviewers found its Plan/Act workflow highly effective, enabling tasks like building a full notifications system across 11 files in roughly eight minutes. The extension itself is free under the Apache 2.0 license, but real-world API costs averaged $60–90 per month — exceeding the price of paid alternatives like Cursor Pro or GitHub Copilot. Cline also supports browser automation and MCP integrations with tools like Postgres, GitHub, and Slack, though these features require significant configuration time. Notable limitations include no inline autocomplete, occasional freezes during long sessions, and slower response times on large-context tasks.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developers port Python PEG parser to Rust in 72 hours, verify with 20,000 fuzz tests

A team built rs-parsimonious, a complete Rust port of the Python PEG packrat parser library 'parsimonious', during Port Mortem 2026, a hackathon focused on proving ports work rather than just writing them. The project was completed in 72 hours with zero unsafe code blocks, enforced through CI. Correctness was verified via differential fuzz testing across 20,041 cases with zero divergences, alongside the upstream Python test suite running unmodified. The Rust port delivered notable performance gains, cutting cold start time from 23ms to 2ms, p99 latency from 0.068ms to 0.016ms, and peak memory usage from roughly 13MB to 6.7MB. A key architectural decision was modeling grammar expressions as Arc-shared structs with a kind enum, enabling efficient packrat memoization through pointer-based cache keying.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Deterministic Context Batching Cuts LLM Token Usage by 81.7% for OpenAPI Testing

A developer built an agentic API testing prototype that reads OpenAPI specifications, generates test plans, executes them via a headless Dart runner, and reasons about failures. The core challenge was that large OpenAPI specs, often containing hundreds of endpoints and nested schemas, were too token-heavy to send repeatedly to a language model. To solve this, the developer created a deterministic Dart algorithm called OpenAPI Context Batching, which splits endpoints by root domain and recursively resolves only the schema references relevant to a given request. This approach reduced approximate token usage from around 50,000 to 9,200 tokens across a five-interaction workflow, an 81.7% reduction. The project was originally developed as part of a GSoC 2026 proposal for foss42, and though the developer was not selected, the working prototype was completed independently.

0 comments Read more at DEV Community