| name | agent-usage-audit |
| description | Audit installed Claude/Codex agents, skills, commands, metadata-level reference counts, and source-level token pressure to decide what to prune, keep, or route. This inventory skill does not mine historical correction content or decide the current session's handoff state. |
Agent Usage Audit
Prefer deterministic local audit scripts. Do not paste transcript contents, prompts,
raw command output, private paths, or project text.
Workflows
Resolve the dotfiles checkout explicitly before running its repo-backed helpers:
DOTFILES_DIR="${DOTFILES_DIR:-$HOME/work/dotfiles}"
test -f "$DOTFILES_DIR/scripts/agent_usage_audit.py"
Install/usage inventory
python3 "$DOTFILES_DIR/scripts/agent_usage_audit.py"
Use this to separate installed local assets, recent usage signals, and gaps where an
install exists but usage is absent or unproven. Counts are references, not successful
outcomes.
Source-level token pressure
python3 "$DOTFILES_DIR/scripts/agent_usage_audit.py" session-report --since 7d --format markdown --redact
Use this when deciding where token pressure is coming from across RTK, Claude local
logs, and Codex local logs. The report separates exact counters from estimated
tool-result pressure.
RTK safety signals
python3 "$DOTFILES_DIR/scripts/rtk_safety_report.py" --since 30d
Use this to inspect fallback, explicit bypass, and repeat-after-compression candidate
counts. Repeat candidates are correlation only: same hashed command and cwd within the
repeat window after a high-compression event.
Recommendation rules
- Recommend additions or removals only from observed patterns plus explicit user goals.
- Treat exact provider/local counters separately from heuristic token estimates.
- Treat RTK compression as command-output compression, not total-cost reduction.
- If data-source notes say a log is missing, report the gap instead of inferring from it.
