Deepen

Goal

Surface architectural friction in an existing codebase and propose deepening opportunities — changes that turn shallow modules into deep ones. Aim at testability and navigability, not tidiness.

The run is: scope → parallel explorers → a visual HTML report of the candidates → a grill through the one the user picks → a brainstorm brief the crank pipeline routes to its spec phase. The report is what the user reads; the brief is what the pipeline reads.

Hard Rules

• The target repo is read-only. This run scans, it doesn't refactor. The report goes to the OS temp dir; the brief goes to the ignored .crank/ directory at the repository root; the only things that may land in the tracked tree are the two deliberate side effects the user approves during the grill — a term written into CONTEXT.md, and an offered ADR. Nothing else.
• The report is presentation only. It exists to let the user see the candidates and pick one. It is never re-read as a source of truth. Everything downstream needs travels in the brief, or in the effort's grounding file (ARTIFACT-HOME.md → Grounding); the brief stays sufficient alone.
• One candidate per grill loop. The user picks one card; the grill walks that one to ground. Re-entry is cheap for as long as the report file survives — when a loop ends, offer to reopen the report and pick another card rather than rescanning by reflex.
• No interfaces before the grill. Explorers don't propose them and cards don't commit to them. The single permitted exception is a card's structural pseudo-code panel — what a module exposes and what it hides — with no parameter lists, no return types, and no names anyone will be held to. Signatures are earned in the grill, not before it.
• Explorers are read-only and standard tier. They read, classify, and report; they change nothing.

Flow

1. Scope