onfafanutifafa/getdebug-edge

The security-first, offline, code review Africa's laptops can actually run for free. # getdebug-edge **CodeRabbit-class code review — free, offline, and private — built for the laptops and realities African developers actually have.** An on-device, offline autonomous coding agent for the Africa Deep Tech Challenge 2026 — Laptop LLM Challenge. Point it at a project folder and a small AI model running entirely on your machine flags bugs, security vulnerabilities, and correctness issues with suggested fixes. No cloud, no account, no per-seat pricing, and your code never leaves your laptop — which for fintech and health-tech teams is not a convenience but a compliance requirement. getdebug-edge is a from-scratch, contest-eligible spinoff of getdebug's "analyze → flag → fix" workflow, rebuilt to run entirely on an 8 GB commodity laptop with integrated graphics and zero cloud dependency. Where getdebug is a hosted SaaS calling a cloud LLM, getdebug-edge runs a small quantized coding model locally via a single persistent `llama-server` process (model loaded once, reused for every chunk) and orchestrates its own observe → think → act loop over localhost — no cloud API, no network call once the model is downloaded. Hardware-tuning choices (context cap, KV-cache quantization, threading, RAM safety margin) are documented in `SCOPE.md` §7. See `SCOPE.md` for the full problem statement, constraints, model choice, architecture, and benchmark plan. See `REPORT.md` for the contest-required technical writeup (fill in once benchmarks are run). ## Status **Working end-to-end and measured by the canonical ADTC profiler.** Run through the official profiler in a container pinned to the contest spec (4 CPUs, 8 GB RAM, swap disabled): the shipping **Q4_K_M** model measures **2.21 GB peak RSS → S_eff = 68**, `throttled=false`, no OOM, no crash, and **acc_norm = 0.82** on lm-eval `arc_easy` (the automated accuracy stage). A full agent review over a real multi-file repo peaks around ~3.5 GB (the tool's end-to-end footprint — Python orchestrator + linters — vs the 2.21 G …