weekly-audit-patterns

Documents the invariants of the nightly Claude audit GitHub Actions workflow.

1.6k|168|Updated Dec 16, 2024
One-click install
npx skills add https://github.com/amd/gaia --skill weekly-audit-patterns-amd
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: weekly-audit-patterns
Source: https://github.com/amd/gaia/tree/main/.claude/skills/weekly-audit-patterns
Command: npx skills add https://github.com/amd/gaia --skill weekly-audit-patterns-amd

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Editing the proactive nightly Claude audit workflow (claude-nightly-audit.yml) without knowing its non-obvious invariants causes duplicate issues, missed findings, and silently failing runs. This Skill captures the design decisions a future editor would otherwise break. ## Core Features & Use Cases - Dedup scheme documentation: Explains the two-key system (cluster_key per defect, dedup_key per location) that keeps one defect from becoming N issues, and the deterministic prepare_synthesis.py dedup pass. - Dimension boundaries: Defines the four mutually exclusive audit dimensions (correctness, docs, tests, features) and why security lives in its own workflow. - Operational invariants: Covers severity rules, the bug-label auto-fix promotion path, ISO day-of-week deep-mode selection, allowed_bots requirements, and dry_run validation. - Use Case: Before changing the audit workflow's cadence, filing logic, or adding an audit dimension, load this Skill to avoid regressions like the day-of-month deep-sweep bug or re-filing 130 open findings. ## Quick Start Ask the AI to explain the dedup and severity invariants before editing the nightly audit workflow file.

Frequently Asked Questions about weekly-audit-patterns

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How does the nightly audit workflow avoid filing duplicate issues?▼

Each finding carries a cluster_key identifying the root-cause defect and a dedup_key identifying its location. A deterministic script, prepare_synthesis.py, merges findings by cluster_key and searches the whole open backlog for existing issues before any filing happens.

What are the four audit dimensions in the nightly Claude audit?▼

The dimensions are correctness, docs, tests, and features, run as a matrix of one Claude job each. Security is deliberately excluded and handled by a separate workflow, claude-security-audit.yml, which reports to the private code-scanning tab.

How do I trigger auto-fix for an issue filed by the audit?▼

Apply the bug label to the filed issue. The existing auto-fix job in claude.yml, gated on the bug label, then creates a branch and PR. The audit workflow itself contains no PR-creation code.

Why did the audit workflow stop filing per-run report issues?▼

Per-run receipt issues accumulated as non-actionable bookkeeping, so the report now goes to the GitHub job summary via $GITHUB_STEP_SUMMARY. The workflow still never auto-closes issues; only humans close them.

When does the audit run in deep mode versus normal mode?▼

Normal mode audits the last N days of diff and skips nights with no commits. Deep mode sweeps the whole codebase and auto-selects on Sundays using ISO day-of-week (date -u +%u), not day-of-month.