Skip to content

v2.1.3 — Layer 3 7B upgrade, static analysis, eval suite

Latest

Choose a tag to compare

@stalzkie stalzkie released this 21 Jun 13:07

What's new in v2.1.3

Layer 3 — Qwen2.5-Coder-7B

Upgraded from the 1.5B model to the 7B model. Measured on a 145-diff labeled corpus:

Metric 1.5B 7B
Precision 0.649 0.982
F1 0.726 0.745
FPR 0.727 0.018
False positives 40 1

localforge --install now auto-downloads the 7B model and falls back to 1.5B if already present.

Clean-diff fast path

The advisory engine now skips the model entirely on commits with no risky additions (refactors, deletions, whitespace changes) — keeping latency near zero for clean commits.

Post-inference false positive filter

A new suppressor (filter_false_positives) checks findings against known-safe patterns — parameterized queries, list-form subprocess calls, type-annotated variables — and removes them before writing to the report.

Layer 3.5 — Static Analysis

Deterministic tools now run alongside the LLM for categories that require AST-level reasoning:

  • Python: bandit (security), pylint (dead code, unused imports)
  • JavaScript / TypeScript: eslint (no-unused-vars, no-eval)
  • Go: go vet, staticcheck
  • Rust: cargo clippy (suspicious, correctness, perf, dead_code)

localforge --install auto-installs all available tools.

Layer 3 Eval Suite

  • 145 labeled diffs across 18 categories and 9 languages (90 TP, 55 TN)
  • Precision / recall / F1 / FPR metrics, per-category and per-language breakdowns
  • 4 matplotlib charts + JSON artifact
  • Run with: python3 tests/layer3_eval.py

Download

Asset Description
LocalForge-v2.1.3-arm64.dmg macOS app — drag to Applications
localforge-v2.1.3-macos-arm64 CLI binary — for terminal/script installs

Quick install:

curl -L https://github.com/stalzkie/local-forge/releases/latest/download/localforge-v2.1.3-macos-arm64 \
  -o /tmp/localforge && chmod +x /tmp/localforge && /tmp/localforge --install