arnonmoscona

Conduct a thorough code review in an isolated subagent context. Use when asked to review code, check for bugs, audit changes, or assess code quality. Accepts an optional scope argument (staged, changed, all-python, a specific file/package path, or an explicit space-separated list of files). An optional ticket ID can be appended to help the subagent load task context (e.g. "changed TOO-14").

Reduce cyclomatic and cognitive complexity in Python code. Break down complex functions, simplify control flow, and track complexity trends over time.

Search, read, create, and edit notes in the user's Obsidian vault. Use when asked to look up, write, or update Obsidian notes or memories — especially in claude-global-memory.

Audit the active toolguard permission configuration for security risks and flag dangerous allow rules and takeover-mode misconfigurations across the whole config hierarchy. Use when asked to security-check, audit, or review the toolguard / Claude Code permission setup, to find risky or over-broad allow rules, to verify takeover mode is sound, or as a start-of-session safety check. Runs a deterministic analyzer first, then optionally offers a deeper, judgement-based AI-assisted assessment. Read-only: it reports findings and proposes fixes but never edits the config.

Curate the active toolguard permission configuration through a guided, conversational review: find redundant, consolidatable, mis-levelled, and confusing rules across the whole config hierarchy, present them grouped by command family with a plain-English before/after, and -- only on explicit, per-item consent -- apply the changes the user approves. Use when asked to clean up, tidy, consolidate, simplify, de-duplicate, or maintain the toolguard / Claude Code permission rules, to shrink an over-grown allow list, or as a periodic config-curation checkpoint (e.g. before a push). Runs a deterministic analyzer for evidence, reasons about it in passes, and never applies anything automatically.