What problem does it solve? AI-assisted AscendC operator-generation sessions repeatedly drift into the same failure modes — reward hacking, infrastructure bypass, premature platform-blame conclusions, unverified runtime-state claims, and premature stops — forcing the user to correct the same mistakes across sessions. This Skill audits the current working session against a catalog of those patterns before the user has to correct them again. ## Core Features & Use Cases - 43-item check catalog: Evaluates the session against categorized checks (priority drift, protocol bypass, premature conclusions, claim-without-verify, cheating-by-claim, KB pipeline integrity) with pass/warn/block verdicts. - Trigger-scoped subsets: Maps orchestrator triggers (pre-spawn, iter-cap warning, pre-finalize, pre-commit) to only the relevant catalog items, reducing audit cost per invocation. - Critic-bypass gating: Mandatorily runs before writing waiver or exception artifacts that would silence a specific check. - Use Case: Before spawning a multi-agent kernel-generation chain or finalizing an operator, invoke the audit to catch a planned direct agent spawn that bypasses the state machine, or a "hardware floor" claim lacking probe evidence. ## Quick Start Ask the assistant to run /aog-self-critic to audit the current session state and the last few actions before making a major decision.
