Simplify recently-changed code

One quality pass over the code changed this session: read-only sub-agents surface code smells, you apply the ones worth applying. Not a loop: one pass, done. Correctness (crashes, data loss, contract changes) is harden-code's job; this skill is simplicity only. If nothing changed, ask which scope to use.

The bar is simplicity. The goal is clear, clean-cut features built on the codebase's established patterns, not bells and whistles, and no scaffolding for features that don't exist yet. If a simpler shape does the same job, or an existing pattern already covers it, that's the finding.

Smells to surface

• Dead code: unused functions, unreachable branches, orphaned exports, commented-out blocks.
• Duplicate code: the same logic in two places, near-duplicate functions, repeated validation.
• Data clumps: the same field group or shape redeclared inline across files that should be one shared type/schema.
• Misplaced responsibility: logic in the wrong layer (business rules in transport, validation in storage, presentation in domain models); anything that violates the separation of concerns CLAUDE.md defines.
• Speculative generality: abstraction, indirection, or config built for a future that hasn't arrived (per CLAUDE.md's restraint rule).

Leave alone: bugs, crashes, data loss (note them for harden-code, don't fix here); behavior-changing performance tweaks; and nits the formatter already handles.

Process

1. Scope: the files changed this session (ask if you can't tell reliably).
2. Partition: group them into cohesive slices, one per import neighborhood or feature path; one sub-agent per slice.
3. Dispatch in parallel: read-only sub-agents, each given its file list and slice name, each returning findings as file:line · smell · suggested fix. They advise; they never edit.
   • Adversarial variant (agent team): when agent teams are enabled, run this as a team of QA skeptics instead. Each teammate owns a slice, argues that its code isn't earning its keep (is this abstraction premature, does this deviate from an established pattern, is this bell-and-whistle wanted by an actual feature), and teammates cross-examine each other before reporting. A smell that survives the challenge is worth applying.
4. Triage: apply a finding only if it simplifies and preserves behavior. Skip anything that changes behavior, contradicts CLAUDE.md conventions, or doesn't survive your own scrutiny, each skip with a one-line reason.
5. Apply the survivors yourself: you have the context, don't delegate to a fixer.
6. Report, per slice: audited, found, applied, skipped and why.