Safety Audit

Use the Claude Code command as the canonical workflow instead of maintaining a duplicated snapshot.

Load the Canonical Command

1. Read .claude/commands/audit-safety/SKILL.md completely.
2. Read .claude/commands/_audit-common.md and .claude/commands/_audit-severity.md completely before auditing.
3. Follow the canonical command's unsafe census, dimension order (skip dimensions whose Paths have no commits since the last report), regression guards, evidence rules, deduplication, and report format.

Adapt Claude Conventions to Codex

• Treat the user's text after the skill request as any requested audit focus or depth.
• Resolve each /audit-<name> reference to that audit's SKILL.md under .claude/commands/, and load a sibling audit only when cross-referencing its contract is necessary.
• Prefer codebase-memory graph tools for symbol discovery, callers, resource lifetimes, and impact analysis. Use text search for unsafe blocks, safety comments, literals, configs, shader text, docs, and graph gaps.
• Interpret Claude Task-agent instructions as Codex sub-agent delegation only when delegation is available and allowed. Otherwise audit one dimension at a time.
• Keep unverified Vulkan claims explicitly labeled as requiring validation-layer or RenderDoc evidence, as required by the canonical command.
• Preserve the canonical output path and do not create GitHub issues; publishing remains a separate audit-publish action.

Do not copy the canonical command body into this skill. The Claude command is the single source of truth.