Runtime Telemetry 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-runtime/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 gate matrix, bench_mode-first diff, severity and advisory rules, harness/baseline-integrity checks, and report format — this audit is live, not static: it drives the engine headless on on-disk game data via .claude/commands/audit-runtime/capture.sh and diffs against the checked-in baselines under .claude/audit-baselines/runtime/.

Adapt Claude Conventions to Codex

• Treat the user's text after the skill request as the command arguments described by the canonical skill's argument-hint.
• Resolve each /audit-<name> reference to that audit's SKILL.md under .claude/commands/, and load a sibling audit only when a finding's resolution chain routes there (e.g. /audit-nifal after a tex_missing_base_color bump).
• Prefer the repository's codebase-memory graph tools for symbol discovery and metric-source location; use text search for metric keys, log-line formats, capture scripts, docs, and graph gaps.
• Interpret Claude Task-agent instructions as Codex sub-agent delegation only when delegation is available and allowed. Otherwise run captures and gates serially, exactly as the canonical command prescribes.
• Advisory metrics are never findings, and a SKIP is never folded into a pass; 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.