Dev builds open-source local LLM coding harness with Qwen3 27B, shares hard-won lessons
xornullvoid · reddit · 2026-08-27
A Reddit user shared warpdrv, an open-source, telemetry-free harness for running local LLMs, built with local Qwen 3.x 27B (>90% human-supervised, not vibe-coded). Works on Windows/Linux.
Distinctive features: just-in-time code review before tool calls, three-way sub-agent threads where both user and agents can chat (nested conversations supported), voice dictation annotations, built-in MCP tools (filesystem, web fetch, code graph, To-Dos), review of all AI messages by a second adversarial AI, work-modes for planning/building/researching/reviewing, and custom llama.cpp compilation from any GitHub branch (CUDA/ROCm/Vulkan).
Lessons learned:
- One chat per feature/bug keeps context focused and improves quality
- In large codebases, a short project description plus a script-generated full file tree in CLAUDE.md beats listing all features
- Serious apps need senior-dev oversight; unattended vibe-coding only suits small cloud-model projects
- Reject file-read requests you know are unrelated to avoid context pollution
Related event: Developer Builds Open-Source Coding Harness warpdrv on Local Qwen 27B(2 posts)→
More from coding & agent
- Dev pushes back on "AI coding means insecure apps": humans always shipped bugs — dreamwieber · 2026-10-15
- Eight non-negotiable principles for building AI that actually learns from experience — tokenbender · 2026-10-15
- Animated teardown of XBOW's Chrome full-chain exploit that won Google's first $250K VRP bounty — moyix · 2026-10-15
- pstack's durable core: 10-20 line skills that distill agent best practices into principles — HankYeomans · 2026-10-15
- 'Children of Elon' forum lets AI agents post and teach each other while humans watch — AiJohnAllen · 2026-10-15
- VoidZero open-sources Void: SDK, CLI and platform to deploy Vite apps on Cloudflare — cnakazawa · 2026-10-15