automation

Configure and explain unattended vault maintenance. Keep the separation clear: the skill conducts the interview and applies judgment; runtimes validate and persist settings; Claude's Scheduled page or another host owns scheduling.

Read references/claude-scheduling-guide.md before giving scheduling instructions. Read references/overnight-prompt-guide.md when customizing the generated prompt contract.

Feature Filter

Accept capabilities that amplify researcher judgment, expose patterns, or automate mechanical work within agreed boundaries. Do not automate away intellectual decisions. Draft feedback and gap analysis default off; reflection synthesis defaults on.

Automation Profile

Preferences live under automation in .carrel/environment.json:

{
  "enabled": false,
  "inbox_processing": true,
  "vault_health": true,
  "cross_linking_suggestions": true,
  "gap_analysis": false,
  "draft_feedback": false,
  "reflection_synthesis": true,
  "wiki_maintenance": false,
  "trust_level": "advisory",
  "model": "sonnet",
  "schedule": "daily",
  "review_cadence": "quarterly",
  "last_reviewed": null
}

Schedules are daily, weekdays, or weekly; review cadence is monthly, quarterly, or biannual. Update the vault's behavioral guidance and _meta/my-environment.md after profile changes.

Graduated Trust

Explain every intermediate level before raising trust. You MUST obtain the researcher's explicit confirmation before running configure with any trust upgrade — the runtime writes no approval artifact for the Advisory→Consultative bootstrap, so your interview confirmation is the approval of record (decision 2026-07-10). Only after that explicit confirmation, invoke configure once with --trust-level consultative; both runtimes permit exactly that Advisory-to-Consultative bootstrap and persist the approved configuration in one validated write. A direct Advisory-to-Delegated or Advisory-to-Partnership jump is rejected. Never treat a request to configure unrelated settings as permission to raise trust.

Run an action-specific trust check immediately before a guarded operation. Examples:

carrel trust check automation:propose --vault .
carrel trust check automation:execute --vault .
carrel trust check vault:reorganize --vault .

Stop on a failed check and explain the required level.

Unattended Contract

An unattended agent must:

• never ask a question or wait for input;
• defer ambiguous type, cloud-consent, sensitivity, OCR, or transcription-quality decisions;
• create/append _meta/pending-decisions.md only when a real item is deferred;
• create/append _meta/pending-approvals.md only when a real consultative proposal exists;
• continue safe independent items after a defer;
• log every delegated/partnership action with a path and usable revert;
• stop after writing the morning brief.

Use --unattended only in scheduled batch invocations. Interactive batches retain confirmations and inline questions.

Prompt Lifecycle

Automation configure commands persist only the profile and set last_reviewed; they do not create prompt or pending files. Generate the prompt as a separate explicit step:

carrel vault automation-prompt --vault .

Pass --force to replace an existing prompt. No command creates an automatic .prev backup. The generated prompt must locate the vault via .carrel/environment.json, not embed an absolute path. Prompt generation enforces automation:write-prompt, a Consultative action, so an Advisory profile must first complete an explicitly approved automation configuration step.

Configuration Interview

For a first setup, cover naturally:

1. Capabilities to enable. Offer wiki maintenance only when a field map exists.
2. Trust level and experimental status of Delegated/Partnership.
3. Sonnet or Opus preference.
4. Daily, weekdays, or weekly schedule.
5. Monthly, quarterly, or biannual review cadence.
6. Local-vault access requirements for the chosen scheduler.

For a returning researcher, summarize current values and change only requested fields. If a field map was activated since the last review, offer wiki maintenance without enabling it automatically.

Calling Patterns

This is the Claude Code plugin skill; its runtime is the typed carrel CLI, which uses boolean switch pairs. Do not pass true or false after these flags:

carrel automate configure --enabled --trust-level consultative \
  --model sonnet --schedule daily --review-cadence quarterly \
  --inbox-processing --vault-health --cross-linking \
  --no-gap-analysis --no-draft-feedback --reflection-synthesis \
  --no-wiki-maintenance --vault .

On hosts without this plugin, the portable pack's runtime takes explicit boolean values instead; see skills/carrel/references/workflows/automation.md.

After configure succeeds, generate the prompt separately, review it, then guide the researcher through the Scheduled page in Claude (or /schedule in a conversation). A local-vault task needs the connected folder and Desktop availability when it runs. The host's saved schedule is not modified by Carrel profile changes.

Do not quote fixed dollar estimates. Claude scheduled tasks use the paid plan's allocation; programmatic/API usage depends on the selected authentication and provider account.

Morning Brief

Write _meta/briefs/YYYY-MM-DD.md with:

• inbox processed/skipped/failed/deferred counts;
• vault health counts, broken links, orphans, and stale drafts;
• high-confidence suggestions only;
• active-plan next steps;
• actions and reverts only for Delegated/Partnership;
• field-map counts when enabled: new/updated pages, low-confidence, contested, contradictions, missing single-source confidence, source drift, orphans, and un-ingested sources;
• one substantive synthesis insight, not only counts.

Advisory and Consultative briefs omit “Actions Taken.” A brief may link to pending files only when those files/items actually exist.

Suggestion Confidence

• High: shared citation and concept, with substantial notes; include in brief.
• Medium: shared citation or concept; keep in suggestions only.
• Low: loose thematic similarity; keep in suggestions only.

Wiki Maintenance

Unattended field-map writes require Delegated trust and wiki:write. Consultative approval obtained in an earlier interactive session is applied through wiki:apply-approved, not converted into general autonomous authority. Follow the knowledge-wiki skill for provenance, source digests, quality signals, and lint counts.

Related

• /carrel-automate triggers this skill.
• convert and transcribe own interactive batch judgment.
• knowledge-wiki owns field-map write semantics.
• references/claude-scheduling-guide.md contains current Claude Scheduled and claude -p setup.
• references/overnight-prompt-guide.md contains prompt assembly and trust blocks.