Skip to content

LocalCodeForge v1.0.0 - local-first 30B coding agent kernel

Latest

Choose a tag to compare

@BMDPLAinsong BMDPLAinsong released this 15 Oct 10:03
· 17 commits to main since this release

LocalCodeForge v1.0.0 - local-first 30B coding agent kernel

A fully-local AI coding agent kernel for C#/.NET: it plans, writes, builds, fixes compile errors, and verifies your project - powered by a 30B Qwen3-Coder model on your own GPU. No cloud, no code leaves your machine.

Design bet: engineering facts as data, not prompts

  • Type inventory with member signatures + access modifiers injected before generation - the model can't hallucinate members it can't see
  • Compile errors located to the exact file before the fix round (LocateErrorFile)
  • Pre-write type-conflict guard + task-id isolated backups with automatic rollback on repeated failure
  • Advisory expected-output semantics: strict equality was misfiring on real math (1.03 vs a planned 100.00) - real computation is verified by compile+run, not guessed strings

Proven on real incremental tasks (E2E 7/7)

  • "Add ShippingService reusing TaxService/DiscountService, wire into Program.cs" - 3 consecutive green runs: 68s / 69s / 121s
  • New-class-calling-service-chain and static-utility-class variants complete in ~1 round
  • "Modify existing method" variant safely blocked by the conflict guard - honest failure, no corruption

Stack

.NET 10 · Semantic Kernel · Ollama (single tag law-unified, request-level temperature) · SQLite · web dashboard (live task stream, backup diff, rollback)

864 deterministic tests · docs in repo · MIT