LiuHe toolchain cuts LLM agent repo indexing from 30 seconds to 98ms

·1 views

Open-source project LiuHe v0.4.6 is a code-operation toolchain built specifically for LLM coding agents, addressing chronic slowdowns in repository symbol indexing. The team rebuilt their tooling around a Rust tree-sitter parse daemon, a SQLite symbol index, and incremental self-healing to replace a slow file-by-file parsing approach. Previously, indexing a 1,482-file Ansible repository took tens of seconds per request; after the redesign, a full index build takes 9.7 seconds and subsequent repo map queries return in 98ms. The toolkit also introduces transactional, journal-backed file writes that survive process kills, eliminating silently lost edits that plagued earlier agent workflows. Released under the MIT license with zero-build deployment, LiuHe offers 44 deterministic tools spanning code reading, analysis, editing, quality gating, and verification.

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)

Aaron Swartz faced prosecution for scraping data that Meta now does freely

A blog post circulating on Hacker News draws a sharp contrast between the legal treatment of Aaron Swartz and current practices by Meta. Swartz, a co-creator of RSS and internet activist, was federally prosecuted in the early 2010s for scraping academic articles from JSTOR. He died by suicide in 2013 before his case went to trial, facing decades in prison. Meta and other large tech companies now routinely scrape vast amounts of web data to train AI models, largely without legal consequence. The post has reignited debate about the unequal application of internet law and the selective enforcement of data access rules.

0 comments Read more at Hacker News

ProgrammingDEV Community ·

Proxmox GPU Passthrough Failure Traced to HBA Conflict, Not the GPU

A Proxmox 8 host configured for PCIe passthrough of an NVIDIA RTX A4000 into an Ubuntu 24.04 VM repeatedly failed to start the guest, with QEMU exiting silently and leaving no console output. Host-side dmesg logs revealed IRQ allocation failures and VFIO interrupt remapping errors at VM startup, pointing to a deeper hardware conflict rather than a misconfigured GPU setup. Despite following standard passthrough steps — including VT-d, IOMMU, OVMF BIOS, and q35 machine type — the issue persisted across multiple troubleshooting sessions. The root cause turned out to be a Host Bus Adapter (HBA) in the system, not the GPU itself, making the fault difficult to identify because error messages implicated the wrong device. The passthrough was being set up to accelerate Filecoin sealing workloads, specifically the TreeRC tree-building phase, which can drop from two to three hours down to fifteen to twenty minutes with GPU acceleration.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How PNG Screenshots Become a Defensible PDF Audit Trail for UI Regressions

A structured approach to archiving UI screenshots can prevent costly post-incident disputes when teams cannot trace which build produced a broken state. The method involves compiling PNG snapshots into a single PDF per change window, named with a build identifier, environment tag, and timestamp. Each PDF includes a cover page embedding key metadata such as the git commit SHA, CI job URL, viewport settings, and data seed used during the test run. PNG streams are embedded directly into the PDF without JPEG re-encoding to preserve pixel-exact reproducibility essential for reliable visual diffs. This discipline replaces unstructured screenshot folders with a lightweight, hashable, and ticketable audit log that can answer provenance questions months after a release.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

OpenAI Confirmed Using NVIDIA GB200 NVL72 for AI Training and Inference

NVIDIA has identified OpenAI as one of its Lighthouse model builders deploying the GB200 NVL72 rack-scale system for next-generation model training and production inference. The confirmation came from NVIDIA's Q2 2026 earnings call, which named OpenAI among key users of the integrated data-center-scale platform. Unlike traditional single-GPU setups, the GB200 NVL72 treats a large pool of compute resources as a coordinated system, designed to support both training and serving workloads. Separately, NVIDIA has unveiled its next-generation Vera Rubin platform, with Microsoft Azure announcing plans to integrate Rubin NVL72 racks across its AI superfactories, though no direct OpenAI-Rubin deployment has been confirmed. The developments reflect a broader industry shift from individual accelerators toward tightly integrated, rack-scale infrastructure for frontier AI development.

0 comments Read more at DEV Community