aux

Growth-weighted complexity hotspots per file (Tornhill "Your Code as a Crime Scene", 2015). Composes `ccx` (complexity) with `git log --numstat` (LOC change volume) and ranks files by hotspot score = sum_ccx × max(0, loc_delta). Files are classified into four quadrants — hotspot, stable_complex, churning_simple, calm — based on 75th-percentile cutoffs on both axes. Default time window is 14 days — tuned for agentic code generation where architectural damage from file growth accumulates in days, not months. Read-only. Requires git. Language coverage inherits from ccx.

Meta-skill for agent discovery and skill routing. Returns the structured capability registry for all aux skills so agents can select the right skill, understand composition patterns, and bootstrap with no prior context.

Cyclomatic Complexity (McCabe 1976) and Cognitive Complexity (Campbell 2018) per function across a codebase, computed in a single tree-sitter AST traversal per file. Language-agnostic — supports Python, JavaScript, TypeScript, Go, Rust, and Java on day one. Classifies functions into zones: simple, moderate, complex, untestable. Read-only. No regex text-tier fallback — files that fail to parse are reported as errors and skipped.

Semantic git diff skill. Surfaces what has changed since a git ref — files modified, line counts, and symbols added/removed (requires tree-sitter). Attacks accumulated session drift: shows exactly what changed and what symbol-level API surface was affected. Read-only.

Module dependency graph skill. Builds an import graph from source files, computes per-file coupling metrics (afferent Ca, efferent Ce, instability), and detects import cycles. Read-only. AST extraction via tree-sitter is the default; a text-tier regex fallback runs when AST extraction is unavailable for a specific file.

Agent-assisted read-only structural search. Executes tree-sitter AST queries over a codebase and returns matched captures without modifying any files.

Tiered dead code candidate audit. Enumerates symbols and modules with zero detected external references and emits advisory output for human verification. Never acts autonomously — all candidates require deeper-dive investigation via `aux usages` before any action is taken.

Robert C. Martin package design metrics skill. Computes coupling (Ca/Ce/I), abstractness (Na/Nc/A), and distance from the main sequence (D') per package. Classifies packages into zones: pain, uselessness, warning, clean, ok. Read-only. AST abstractness detection via tree-sitter is the default; a text-tier regex fallback runs when AST extraction fails for a specific file.

Hierarchical file-discovery + content-search + AST-structure pipeline. Enumerates files, searches their contents, and optionally filters to structural AST matches — all in a single deterministic invocation.

Symbol cross-reference skill. Returns all definition locations and all reference locations for a given symbol in one structured call. Uses ripgrep for exhaustive text matches enriched by tree-sitter for semantic definition tagging.

Move or rename files and directories with a dry-run gate. Single-call filesystem rename that avoids read-write-delete token waste. Batch-capable via plan mode.

Focused fixed-string text replacement across a codebase. Simpler than sed: one old string, one new string, one scope. Designed for Haiku-tier agents — no regex, no AST, no mode selection.