performance-reviewer

Hot-path analysis agent that identifies algorithmic complexity issues, memory patterns, and I/O bottlenecks with quantified impact estimates.

The performance-reviewer agent focuses on code that runs frequently or on large data sets. It does not flag micro-optimizations in cold code — instead, it quantifies findings by complexity class and estimated real-world impact, distinguishes "measure first" recommendations from "obviously fix now" ones, and notes explicitly when current performance is acceptable. An O(n²) loop in a hot path is a finding; a 5 ms startup cost is not.

Role

Identify hot paths and analyze their algorithmic complexity (time and space)
Detect memory anti-patterns: allocations in hot loops, large object lifetimes, string concatenation in tight loops, unintended closure captures
Review I/O patterns: blocking calls on hot paths, N+1 query patterns, unbatched network requests, unnecessary serialization
Surface caching and parallelism opportunities, and provide a profiling plan for non-obvious concerns

When invoked

When load-test regressions are reported and the bottleneck has not yet been located
On changes to hot-path code where performance is a stated requirement
When a quality-reviewer pass notes high cyclomatic complexity that may also carry a performance cost
By $review when the changed code involves large data processing, database queries, or network I/O

Inputs

Source files under review, read via Read and Grep
ast_grep_search results for structural performance anti-patterns (nested loops, Array.includes in hot loops, etc.)
Optional: benchmark results, profiling output, or load-test logs

Outputs

A performance review report with overall status, critical hotspots with complexity class and estimated impact at realistic data volumes, optimization opportunities with expected improvement, a profiling plan for non-obvious concerns, and explicit confirmation of areas where current performance is acceptable

Limits

Does not flag cold code (startup paths, rarely-called utilities) unless the cost exceeds a clearly material threshold
Does not recommend optimization without quantifying the impact; "slow" is not a finding
Does not profile or benchmark code directly; provides a profiling plan and hands execution to executor or qa-tester

quality-reviewer — logic and maintainability review that can surface complexity issues
debugger — investigates confirmed performance regressions when a hot path is already suspected