Audit My Automations

Operators set automations up and never look at them again — so they don't notice when one silently stops firing, starts erroring, or sends twice. This skill is the regular check-up.

Read first: knowledge/automation-method.md — especially §6 (safety dials) and §8 (reading the run log). The flags this skill raises map directly to those sections.

Step 1 — Fetch the health digest

python ~/.claude/bos-run.py audit-my-automations

One call: lists every automation (with triggers + actions inline), samples each one's recent runs, and computes per-automation flags plus cross-automation trigger overlaps. The shape is documented at the bottom of fetch.py.

If the script can't run (auth/network), fall back to mcp__trustpager__list_automations + mcp__trustpager__list_automation_runs per automation — but that's many calls; prefer the script.

Step 2 — Present the report, worst first

Lead with what's broken, then what's risky, then a one-line "healthy" tally. Never dump all automations as a flat list.

🔧 NEEDS ATTENTION (N)
  → "Renewal reminder" — 3 failed runs this week. Last error: <error_message>. → check it
  → "Old intake" — has a trigger but no actions (does nothing). → add actions or delete

⚠️ WORTH A LOOK (N)
  → "Quote follow-up" — enabled but hasn't fired in 41 days. Trigger may be wrong, or just quiet.
  → "Welcome email" — sends email but dedup is OFF — a double-submit double-emails. → turn dedup on
  → "Lead → quoting API" — calls a webhook with no daily cap. → set max_executions_per_day as a seatbelt
  → "Win celebration" — 9 of last 10 runs SKIPPED. Conditions may be too tight to ever pass.

🔁 OVERLAPS (N)
  → "Tag new lead" and "Email new lead" both fire on form_completed (Intake form).
    Fine if intentional; consolidate into one automation with multiple actions if not.

✅ HEALTHY: 12 automations firing normally.

What each flag means + the fix

Step 3 — Offer fixes (with approval)

For the safe, mechanical fixes, offer to apply them — one at a time, with a yes:

• Turn on dedup → mcp__trustpager__update_automation(automation_id, dedup_enabled=true, dedup_window_minutes=60)
• Set a daily cap → update_automation(automation_id, max_executions_per_day=N)
• Disable a redundant duplicate → mcp__trustpager__disable_automation(automation_id)

Never delete an automation without explicit confirmation naming it. For "why is it failing / skipping" deep-dives, hand to /why-didnt-it-fire. For building a missing automation, /automate-this.

What to never do

• ❌ Don't present automations alphabetically or as one flat list — rank by severity.
• ❌ Don't auto-apply fixes. Offer, get a yes, then apply — one at a time.
• ❌ Don't flag disabled as broken — staged/off is a valid state. Count it, don't alarm about it.
• ❌ Don't treat skipped runs as failures — that's conditions working as designed. Only mostly_skipped is worth a mention.

Output shape

Open with the headline tally ("18 automations: 12 healthy, 2 need attention, 4 worth a look"), then the prioritised sections, then the single most important fix to make first.