harness-ultracodex

Use when uma tarefa de alto esforco ou dificuldade precisa de um juiz robusto e contraparte adversarial antes de concluir.

Use when delegating bounded subtasks to another agent or lighter model through Codex CLI, Hermes CLI, Antigravity CLI, or a local orchestration script to save tokens without losing control of scope and evidence.

Use when a task depends on library, framework, SDK, API, CLI, cloud service, version-specific behavior, setup instructions, migration guidance, or any external documentation that may be stale or mutable.

Use when debugging panes, errors, anomalies, regressions, flaky behavior, broken tests, incidents, unexpected UI behavior, or any failure where jumping straight to a fix could hide the real cause.

Use when a task changes UI, layout, frontend behavior, browser interaction, screenshots, visual regressions, E2E flows, accessibility-visible states, or anything that must be inspected in a browser.

Use when delivering a high-risk plan, architecture, audit, migration, security conclusion, refactor strategy, or any non-obvious technical recommendation that could be wrong in costly ways.

Use when choosing whether to answer briefly, write a technical plan, perform an audit, implement code, run incremental execution, or stop for confirmation based on ambiguity, risk, reversibility, and evidence needs.

Use when claiming something is done, fixed, safe, broken, vulnerable, tested, untested, production-ready, or when summarizing findings from files, code, tests, audits, logs, or inferred patterns.

Use when a user request may contain a gap between the literal ask and the real operational need, especially for refactors, fixes, audits, legacy work, risky commands, ambiguous scope, or requests where the best delivery format is not obvious.

Use when changing legacy code, refactoring behavior that lacks tests, reducing risk in old modules, separating concerns, migrating code incrementally, or when a requested cleanup may accidentally change production behavior.

Use when a task may affect data, production, security, billing, accounting, migrations, destructive commands, public contracts, legacy behavior, or when deciding whether to ask, infer, research, test, or act.