system-health

# /system-health — Claude Code config hygiene Keeps the setup "in an optimization flow": a cheap regular measurement catches degradation (bloated memory/context files, broken links, stale entries, excessive permissions, junk) BEFORE it breaks. Like a server health check, but for your Claude Code configuration rather than servers. ## When to use - On command or the trigger phrases above. - Once a week (can be scheduled). - After a large session with heavy writes to memory. - When a built-in hook warns "memory index over limit" (that's the signal it's time). ## Process ### Step 1 — deterministic measurement (0 LLM tokens) A read-only shell pass that changes nothing and emits metrics with `✓OK` / `⚠️WARN` flags. Measure: - Size of the memory index and each topic file vs the model's read limit. - Broken `[[wiki-links]]` / pointers in the active index (a link to a non-existent file is a distractor). - Stale "next-session" / TODO carryovers that are actually finished. - Orphan files (referenced by no index — invisible to recall, dead weight). - Project context file size vs target. - Excessive permission entries (specific grants subsumed by a wildcard). - `.bak`/junk files lying around. ### Step 2 — interpretation (LLM) Read the report. Focus on `⚠️WARN` lines. Translate each into a concrete diagnosis: what grew/went stale and why it hurts (recall / tokens / security). ### Step 3 — targeted fixes BY AGREEMENT Propose minimal surgical fixes per WARN. **Don't act silently** (exc...