Audit with an agent team

Audit whatever the user names: a change, file, feature, or refactor. If they name no target, audit the current uncommitted changes.

Treat the area under audit as an import graph: the files in scope, plus (one hop out) the neighboring files they import or that import them. Split it into connected subgraphs, one per teammate, so each traces its own edges and confirms data flows cleanly across every boundary, not just where it lands. Aim for 3–5 teammates; read a subgraph directly instead of spawning when it's small enough to verify by eye.

Spawn an agent team and, in the spawn instruction, tell the lead to:

• give each teammate a self-contained brief: the target, the subgraph it owns, and the neighboring files to trace into (teammates don't inherit this conversation, so spell out the change);
• have each critique adversarially (hunt for what breaks, don't rubber-stamp) and report findings as category · severity · concrete failure scenario;
• have teammates message each other to challenge findings before reporting, like a scientific debate. A finding that survives peer cross-examination is credible; one that doesn't is dropped.

If agent teams aren't enabled (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS), fall back to parallel subagents in a single message, with the same briefs, minus the cross-examination.

Collect the surviving audits and judge what actually holds. You are the final filter, not a pass-through.

Presenting the output: mandatory

Group the surviving findings into categories (e.g. reliability, type/SoC consolidation, config cleanup, test coverage, dead code) and present them with the AskUserQuestion tool as a multi-select question, one option per category, labeled with its severity, described by what acting on it entails. Do not apply anything until the user has chosen.

• Lead with a tight written synthesis (verdict + per-category findings) so the choice is informed, not blind.
• Set multiSelect: true; order options by severity, most important first, appending "(Recommended)" where apt.
• Act on exactly the selected categories and skip the rest. Even when the findings look obvious, the user decides scope.