caretaker

name	caretaker
description	Run the bounded harness caretaker — deterministic state gather (predictions, verifier gain, agent health, sensors, drift), then triage ONLY what's flagged under a hard authority boundary (files work, never ships changes). Triggers on /caretaker, "run the caretaker", "harness health check".
when_to_use	Daily (schedule it), after a heavy multi-session day, or whenever you want a fleet-health snapshot without spending tokens on a full audit. NOT for fixing things — the caretaker files work; humans (or explicit follow-up commands) ship it.
allowed-tools	Bash, Read, Edit

/caretaker — bounded harness caretaker

Deterministic-first (LOOP pattern): all state gathering is compiled into a script; the LLM reasons only about what's flagged.

Step 1 — gather

~/.claude/scripts/caretaker-daily.sh

• Exit 0: output its one-liner verbatim and STOP. No further tool calls.
• Exit 1: triage the listed action items under the authority boundary.

Authority boundary (hard rules)

The caretaker files work, it never ships:

MAY:

• Read anything under ~/.claude (telemetry, rl, sensor-triage notes, lessons-inbox)
• Run read-only diagnostics (rg, git status/log, gh run view, launchctl print)
• Assemble evidence for a WRONG prediction or degrading agent (≥2 items) into ~/.claude/rl/tuning-evidence.md and recommend the exact /tune-agent or revert command
• Consolidate sensor-triage notes into one diagnosis
• Move processed tuning-evidence entries under a ## Processed <date> heading (that ONE file is the only thing it edits)
• Post one notification: osascript -e 'display notification "..." with title "caretaker"'

MUST NEVER:

• Edit rules/, CLAUDE.md, agents/, skills/, hooks/, settings, or any code
• Promote candidates, apply mining patches, or run /tune-agent itself
• Restart services, push git, delete files, spawn agent fleets
• Exceed ~20 tool calls; if triage is bigger, say so and stop

Interpretation guidance

• A cost anomaly on a day of heavy interactive use is expected — flag it prominently only if cache hit rate also fell below 85%.
• One sensor anomaly = read its triage note; repeated anomalies from the same sensor within the cooldown window = the underlying issue persists, escalate in the notification.
• A degrading agent needs ≥2 concrete evidence items before it becomes a /tune-agent recommendation (one bad day is noise).
• A starving loop item (from check-loop-liveness.sh) means a closed loop's output artifact is stale/missing even though its process may be ticking — the highest-severity class here, because it is invisible in normal logs (2026-07 rating-drip outage: two weeks of healthy-looking ticks, zero delivered questions). Diagnose which stage of the loop starved (producer, transport, or consumer) and file the evidence + exact fix command; do NOT restart services yourself. A loop that is intentionally finished (e.g. rating sheet complete) is auto-exempted by the script — if it flags anyway, the exemption logic is the bug to report.

Output

End with: what was found, what was filed where, and the exact next commands for a human. One notification, no essays.

Scheduling

Best run automatically once a day. Claude Code desktop: create a scheduled task with this skill as the prompt (see templates/caretaker-scheduled-task.md in the claude-agent-os repo). Headless: cron a budget-capped session:

47 6 * * * ~/.local/bin/claude -p --max-budget-usd 2 --max-turns 25 "/caretaker"