claude-code-config: Skill for Claude Code

Audit this project's durable memory store and present prune recommendations. Same discipline as /pm-clean: evidence for every recommendation, conservative on "stale," and nothing is removed without an explicit yes. All detection and mutation lives in the shared, tested script .claude/scripts/memory-audit.py — this file is the recommend-then-confirm layer.

Step 0: Locate the audit script

Prefer the global install; fall back to the in-repo copy when developing the skill itself.

SCRIPT=""
for candidate in \
  "$HOME/.claude/skills-worktree/.claude/scripts/memory-audit.py" \
  "$HOME/.claude/scripts/memory-audit.py" \
  ".claude/scripts/memory-audit.py"; do
  if [[ -f "$candidate" ]]; then SCRIPT="$candidate"; break; fi
done
if [[ -z "$SCRIPT" ]]; then
  echo "ERROR: memory-audit.py not found (checked ~/.claude/scripts/, skills-worktree, in-repo)" >&2
  exit 1
fi

Parse $ARGUMENTS: pass through an explicit --dir PATH (for a non-current project or a relocated store) and translate --budget-kb N to --budget-bytes $((N*1024)). With no arguments the script auto-detects the current project's memory dir — the ROOT project, not the worktree, since memory is pinned to the root project across worktree sessions.

Step 1: Scan (read-only)

python3 "$SCRIPT" --check --json [--dir PATH] [--budget-bytes N]

Exit 0 = scanned (a clean store is a valid result). Exit 3 = memory dir not found — tell the user and offer to re-run with --dir PATH. The JSON has memory_dir, index (bytes, entry_count, file_count, budget_bytes, over_budget), orphans, dangling, and advisory (duplicate_names, superseded). Read memory_dir back to the user so they know exactly which store is in scope.

Step 2: Present recommendations

Group findings into scannable sections. Omit any empty section. If nothing is flagged and the index is under budget, report "Memory store is clean — {entry_count} entries, {bytes} bytes, nothing to prune." and stop.