Cross-provider LLM judge using Codex (GPT-5) to evaluate implementations written by Claude. Internal skill invoked by the insistir lead during the review loop. Not user-facing.
Document a solved problem to compound team knowledge. Triggered by phrases like "that worked", "it's fixed", "problem solved", or explicitly via the /insistir:compound command. Orchestrates parallel sub-agents to analyze the problem, extract the solution, and suggest prevention — then assembles a single searchable solution file in docs/solutions/.
File-based TODO lifecycle for tracking review findings, managing technical debt, and organizing work items. Each TODO is a standalone markdown file with YAML frontmatter, named by convention: {id}-{status}-{priority}-{description}.md. Status transitions drive the lifecycle: pending (created from review) -> ready (triaged and approved) -> complete (resolved). Use when: (1) review findings need persistent tracking beyond a single session, (2) technical debt should be cataloged with solutions, (3) work items need triage before implementation. Do NOT use for: (1) simple one-off fixes that can be done immediately, (2) tasks already tracked in an external issue tracker.
Loop-engineering goal loop — implement → verify → judge → iterate until the goal is provably met. Separates worker from judge (different model instances) to prevent self-comforting optimism. Fresh implementer each iteration preserves mistakes as learning signal without context pollution. Use when: (1) user says /insistir:goal, (2) a task needs verifiable completion evidence, (3) "goal loop", "loop until done", "iterate until passing", "keep going until tests pass". Do NOT use for: (1) goals that cannot be expressed as observable evidence (ask user to refine first), (2) trivial one-shot tasks, (3) open-ended exploration with no clear done-state.
Agents Checking Agents — multi-agent orchestration with cross-validation using Claude Code Agent Teams and a cross-provider Codex/GPT-5 judge for bias-free final verdicts. Fresh agents for each phase — no agent reuse. Each worker/reviewer/fixer is spawned, does its work, and is shut down. The review loop iterates until the reviewer APPROVES or max rounds are reached. The lead operates in delegate mode: coordinates only, never implements. Use when: (1) user says /insistir, (2) complex tasks that benefit from parallel execution with quality gates, (3) "agents checking agents", "multi-agent", "swarm", or "orchestrate". Do NOT use for: (1) simple single-file changes, (2) sequential tasks with tight dependencies, (3) quick bug fixes or typos, (4) tasks where coordination overhead exceeds the benefit. Full pipeline: intake → plan → spawn team → [execute wave → adversarial review → codex judge] → cleanup.