| name | research-vault-automation |
| description | Compose Codex automation prompts for Research PDF Vault digests and alerts that read generated reports or triage outputs and notify humans without running background vault workers. |
Research Vault Automation
When To Use
Use this skill when the user wants scheduled or event-style Codex automation prompts for Research PDF Vault. Automations should read generated reports, review queue triage, and citation-slot readiness output, then produce a notification or summary.
Safety Boundary
Codex automations are report readers and notification writers. They avoid direct OCR, direct PDF vectorization, and persistent worker service launch. If fresh processing is needed, tell the user which one-shot CLI command should be run outside the automation prompt.
Prompt Templates
Use the templates under plugins/research-pdf-vault/assets/automation-prompts/:
daily-digest.md
review-queue-alert.md
ingestion-failure-alert.md
citation-slot-readiness.md
Inputs
Use one or more of these generated inputs:
- CLI output from
rpv.py setup --check, rpv.py scan --once, or rpv.py ingest --once.
- Review queue output from
rpv.py review list and rpv.py review show.
- Discord dry-run payloads from
rpv.py notify discord --event review-queue --dry-run.
- MCP output from
list_review_queue and apply_review_decision when available.
- Worker report summaries from local report tools such as
list_reports and get_report.
- Citation output from
rpv.py citation-slots build.
- Literature-map output from
rpv.py literature-map report.
Automation Pattern
- Read the latest generated report, queue, or citation-slot artifact supplied by the user or scheduler.
- Extract counts, status changes, failures, and next actions.
- Write a short notification with no private file paths or real PDF names.
- Include the exact one-shot CLI command a human can run when fresh processing is required.
- Prefer Red-lane decision alerts over Amber alerts unless the config explicitly lists Amber in
manual_review_lanes.
Notification Rules
- Keep examples public-safe and synthetic.
- Never include secrets, personal paths, institutional strings, or real document titles.
- Prefer IDs, counts, status labels, and short remediation steps.
- Discord messages should contain counts and commands only, not document titles or source paths.
- Do not imply that an automation prompt is a persistent service.