What problem does it solve? Long-running AI agent sessions repeatedly fall into the same failure patterns: polling without waiting, treating gate rejections as deliverables, stampeding external research channels, reading large files directly into context, and silently stalling on dispatched subagents. This Skill turns those recurring anti-patterns into hard, auditable execution rules backed by read-only gate scripts. ## Core Features & Use Cases - Ten Iron Rules: Zero-polling waits, gate results as diagnostic signals, probe-before-dispatch external research, subagent-based exploration with summary-only returns, progress-reporting protocols, zero-dangling round cleanup, context budgeting, takeover consistency gates, watchdog binding, and structured handoff for long tasks. - Executable Gates: flow_check.py validates nine-stage problem-solving flows, NASA AAR retrospectives, and repeated tool-call detection from session JSONL; takeover_check.py runs six read-only consistency checks (baseline lock, registry consistency, stale cache, privacy-tracked paths, oversized files, workspace hygiene) when taking over a project. - Use Case: When resuming a multi-hour task with background subagents and CLI requests, load this Skill to run the 10-second pre-round self-check, avoid re-reading the same files, and close out dangling supervisors before ending each round. ## Quick Start Use the execution-discipline skill to audit my current long-running session for polling, repeated tool calls, and dangling subagents before I continue.
