Carol — back to Apps ← Apps

Carolopedia

A friendly guide to Carol, her ecosystem, and the agents who built her.

📖 CarolopediaDroidsAutonomous Troubleshooter
Autonomous Troubleshooter

Autonomous Troubleshooter

Droid
Go to droid →

📖About & Usage

Owner agent — accountability this droid serves

Albus is Carol's troubleshooter, accountable for diagnosing pipeline failures and unblocking work. Until now Albus has been spread across narrow templated droids (en_ar_01, al_tr_01, pf_ar_01) that each see one slice of the failure and emit one-shot diagnoses. The harness, not the brain, was the limiter — Orion has the same brain plus an open toolkit and routinely traces failures upstream that Albus cannot. al-auto-01 closes that gap.

Droid responsibility

Investigate a failed execution and either FIX the root cause in code/SQL within the write-scope, OR produce a precise actionable handoff naming file:line and a proposed patch. Trace upstream when the immediate cause is downstream. Cite evidence: every claim must reference a file:line, a SQL row, or a real log line. Emit FINDINGS / ACTIONS / OUTCOME blocks at the end of the session.

What the droid actually does

  • Working directory: /home/caroladmin/dev
  • Time bound: default 10 minutes (configurable via context.timeout)
  • Cost bound: default $2 USD (configurable via context.cost_budget)
  • Same toolkit as Orion: Bash, Read, Edit, sqlite3, grep, find, git (read-only)
  • Output discipline: FINDINGS / ACTIONS / OUTCOME labels at end of last message
  • When in doubt, prefer trace-upstream over symptom-restatement

Boundaries

  • MAY: read anywhere under /home/caroladmin/dev; edit code/scripts/SQL under same path
  • MAY NOT (without explicit operator approval): git push, systemctl restart/stop, edit /etc/*, delete files
  • Must not invoke vendor SDKs directly — use shared/api_router (P.04.01.05.03)
  • Must not bypass the bypass framework — its own remediations must follow bypass discipline (executions filed, plan steps tracked)
  • Must not silently mark itself "completed" when it could not actually fix or hand off — emit OUTCOME=inconclusive instead

👤Owner

Albus · Architect

📚Recent initiatives

Initiatives that touched this droid — a short summary each; open one for the full story.

CAROL-INI-1936-00: Auto-detected failed process: Autonomous Troubleshooter (al-auto-01)
Recurring operational incident, collapsed to one entry.
Orion · 2026-06-24 15:43
CAROL-INI-1845-00: Re-arm Albus failure watcher — register al-watch-01 in Hermione (orphaned by the INI-0974 scheduler migration)
Albus failure watcher (al-watch-01) was never migrated into the Hermione scheduler when scheduling moved off systemd timers (CAROL-INI-0974). Its old timer is disabled and nothing\u2026
Orion · 2026-06-19 18:36
CAROL-INI-0561-00: Albus troubleshooter ranks prior diagnoses by relevance and freshness before re-injecting context
Today Albus loads prior diagnoses for a failing step/task with a flat query — no ranking by similarity, no recency weighting, no cap. As the diagnosis log grows, older and less-re\u2026
Orion · 2026-05-20 22:05
Browse all initiatives →