oc-proposing-deepeninglisted

# Proposing deepening — find where the codebase went shallow, propose the fix An existing codebase was not written under the module-depth rule. This skill finds where that hurts most and proposes a bounded set of retrofits. It **analyses and proposes; it never changes code.** Announce at start (pt-br): "Vou varrer o código e trazer no máximo 5 candidatos a reforma." All reasoning and identifiers in English. Every message to the operator is **pt-br, product-language** — he decides on impact and risk, never on class names. --- ## Where it runs (OpenCode) — `build` Runs in the sole **`build` primary agent**, after normal entry routing identifies it as a no-delivery discovery request. The result is intentionally narrow: it may write the candidates document, but never source code, an issue, a delivery task, or a commit. Do not use a generic shell write path to work around that boundary. --- ## Hard precondition — LOCAL only, never headless The output of this skill is a **question to the operator** ("qual, se algum, vale reformar?"). In a headless run there is nobody to answer, ending a turn on a question kills the run, and an unanswered candidate list would be picked up by an autonomous engine as work to do. So: **if the session is not interactive, refuse immediately and say why.** Headless is active when **any** of these holds: - the trigger prompt says to run **autonomously** / "sem perguntar" / "without asking" — the VPS cron dispatcher always prepends a fixed pref