srivatsa-source/DockDesk Neural Auditor
Local-first semantic audit of code vs documentation using Ollama. Auto-tunes model selection for your codebase.
View on GitHubTrust Signals
- Scorecard Score
- not yet scored
- Maintenance Recency
- Stale
- License
- None
Inputs
| name | description | required | default |
|---|---|---|---|
| model | Default LLM model (used for both detection and fix if dedicated models are not set). Recommended audit-suitable models: - qwen2.5-coder:3b (fast, <5k LOC) - qwen2.5-coder:7b (default, balanced) - codellama:7b (alternative, code-focused) - qwen2.5-coder:14b (thorough, >50k LOC) Use 'auto' for automatic selection based on codebase size. | no | qwen2.5-coder:7b |
| detect_model | [DEPRECATED] Use reasoning_model instead. Model for Phase 1 (detection / risk assessment). | no | "" |
| fix_model | [DEPRECATED] Use reasoning_model instead. Model for Phase 2 (fix generation). | no | "" |
| reasoning_model | DeepSeek-R1 reasoning model for logical risk assessment and safety judgement. Default: deepseek-r1-distill-qwen:1.5b This model acts as the "brain" — it judges risk, validates fixes, and decides push safety. | no | deepseek-r1:1.5b |
| discord_webhook | Discord webhook URL for posting audit summaries and push-guard notifications. Leave empty to disable Discord notifications. | no | "" |
| auto_tune | Automatically select the best model based on codebase LOC (overrides model input) | no | false |
| ollama_host | Ollama API endpoint (for self-hosted instances) | no | http://localhost:11434 |
| fail_on_risk | Risk level that triggers workflow failure (HIGH, MEDIUM, LOW) | no | HIGH |
| output_format | Output format: md (markdown), json, sarif (for VS Code/IDE integration) | no | md |
| auto_fix | Automatically apply documentation fixes | no | false |
| workspace_path | Path to the workspace to audit (relative to repository root) | no | . |
| skip_rag | Skip RAG retrieval for faster audits and reduced disk usage (recommended for CI) | no | false |
| custom_rules | Comma-separated custom audit rules injected into LLM prompts. Example: "Flag hardcoded secrets,Ensure all public functions have docstrings" | no | "" |
| python_version | Python version to use | no | 3.11 |
Outputs
| name | description |
|---|---|
| report_path | Path to the generated audit report |
| status | Audit status (pass/fail) |
| files_audited | Number of files audited |
| high_risk_count | Number of HIGH risk findings |
| fixes_applied | Number of fixes automatically applied |
| sarif_path | Path to SARIF report (for GitHub Code Scanning integration) |