## intrikata-stack skills (session bootstrap + self-audit) Two executable skill procedures are installed under `praxis/skills/`. `HARNESS.md` alongside them maps their Claude-Code-specific primitives (ToolSearch, AskUserQuestion, subagent dispatch, MCP) to generic equivalents — follow it whenever the skill text names a tool you don't have. - **Orientation requests** ("orient me", "catch me up", "what's the status", session start): read `praxis/skills/megapraxis/SKILL.md` and execute it. Without a praxis backend (BACKEND.md) use its filesystem-fallback path — a designed mode, not an error. - **Bootstrap-audit requests** ("audit the bootstrap", "trap roll-call"): read `praxis/skills/metamegapraxis/SKILL.md` and execute audit mode. Contract for both: every emitted claim corresponds to a truth on disk; uncomputed counts render as `count-skipped`, never guesses; undecidable items are named, never absorbed.