Dependency expansion

sift select, the picker’s generate/quit-and-render paths, and the MCP pack_context tool automatically pull in the dependencies of whatever was selected: after the budget optimizer picks the review set, sift walks the import graph outward and re-runs the optimizer over the union — first-pass picks keep their chosen modes as preferences, dependencies enter biased toward signatures. One budget, one optimization, so dependencies compete fairly instead of starving on post-upgrade crumbs.

How it works

During ranking, sift already parses every collected file’s imports for fan-in centrality. That same pass now retains a file → targets adjacency list, so dependency expansion costs no second walk. Language drivers resolve their own import syntax to confirmed, collected files:

• Go, JavaScript, TypeScript commit to real files (relative imports, @//~/ aliases, extension and index.* fallbacks; Go package imports expand to the package’s collected files).
• All other languages fall back to generic prefix expansion — correct enough, and nothing regresses.

Import cycles are broken with a visited set.

Tuning

Flag	Default	Meaning
--max-depth N	2	How many import hops to follow from the selected set. 0 disables expansion, -1 follows without limit (the token budget still caps spend).

Dependencies enter the optimizer biased toward signatures mode — they’re context, not the thing under review — and the optimizer decides the rest within the leftover budget.