CodeSmellAuditor

Local-first CLI that reviews C# source against markdown governance rules using a local Ollama model. It streams the critique live and saves a timestamped markdown report.

This is an AI-assisted semantic code reviewer, not a deterministic static analyzer. It does not use Roslyn or parse C# into a syntax tree; the LLM interprets plain source text against your rule packs.

Built while learning .NET architecture and agent-assisted workflows — the irony of using AI to audit AI-generated code is not lost on me.

Requirements

• .NET 10 SDK
• Ollama running locally with a pulled model (default: qwen3.5:4b)

Run

dotnet run --project CodeSmellAuditor.Cli

Drop .cs files into WorkstationStorage/Targets/, governance rules as .mdc into WorkstationStorage/Rules/. Reports land in WorkstationStorage/Reports/.

Configuration

Environment variables (all optional):

Variable	Default	Purpose
CODESMELL_STORAGE	WorkstationStorage under repo root	Root folder containing Rules/, Targets/, Reports/
CODESMELL_MODEL	qwen3.5:4b	Ollama model name
CODESMELL_NUM_CTX	8192	Model context window passed to Ollama
CODESMELL_NUM_PREDICT	1200	Max output tokens for the critique

How it works

1. Program.cs (composition root) resolves paths and wires MarkdownRuleRepository, OllamaAiOrchestrator, and AuditEngine.
2. Core loads compact .mdc rules, reads each target .cs file, builds a budgeted prompt, and streams Ollama's response.
3. Cli renders progress through SpectreAuditProgressReporter; Core stays presentation-neutral via IAuditProgressReporter.
4. Each audit saves a markdown report with YAML front matter. Pass/fail is parsed from a Status: line in the report body.

Sample report format: docs/sample-report-excerpt.md

More detail: architecture · known limitations · testing

Tests

dotnet test CodeSmellAuditor.slnx -c Release

15 unit tests covering rule loading, prompt building, pass/fail parsing, and AuditEngine orchestration with fakes. Ollama HTTP streaming is exercised manually, not in CI.

Current scope

• Single-file .cs audits against local markdown rule packs
• Local Ollama only (privacy-preserving; no cloud API)
• Interactive CLI with streaming terminal output
• Layered Core / Cli / Tests solution with constructor injection at the composition root

Future implementation

• CLI arguments (--model, --storage, --non-interactive) instead of env vars only
• Non-zero exit codes when any audit requires review
• Surface batch pass/fail summary in the CLI (engine already returns AuditRunResult)
• Roslyn-based deterministic pre-checks before the LLM pass
• Cloud model backend via a second IAiOrchestrator implementation

License

MIT