MAGIC Agent Skills is now open source! Star on GitHub
MAGIC Agent SkillsMAGIC Agent Skills
Commands

/magic:deliver

/magic:deliver — Deliver processed data to a destination

Usage

/magic:deliver [destination]

Description

Deliver your processed data to an external destination. This command guides you through the delivery workflow with appropriate safety checks.

Destinations

Database (db)

Deliver processed data to a database table.

Read the magic-data-transformation SKILL.md ## Database Delivery section for delivery guidance.

Workflow:

  1. Check workspace state for current data (latest checkpoint or DataFrame)
  2. Ask user for delivery target:
    • Environment variable for connection (default: DATABASE_URL)
    • Target table name
    • Write mode: append (default) / replace
  3. Validate data with validate_transform.py (transformation skill, Tier A scriptable — call directly) before delivery
  4. Present write plan: row count, target table, mode, column mapping
  5. PAUSE (always): Confirm before writing — "Write {N} rows to {table}? [y/N]"
  6. Execute write via deliver_to_db.py
  7. Update workspace_state.md with delivery results

Safety: This command creates a NON-read-only connection. The PAUSE gate is mandatory and cannot be skipped even in autonomous mode.

HuggingFace Hub (hf)

Upload processed data to a HuggingFace dataset repository.

Workflow:

  1. Confirm repo name and visibility (private by default)
  2. Generate dataset card from data metadata
  3. Upload data folder with atomic commit
  4. Report Hub URL

Example:

/magic:deliver hf

Required: HF_TOKEN env var or huggingface-cli login

Scripts used:

  • generate_dataset_card.py (loading skill, callable tool) — creates README.md with schema and provenance
  • deliver_to_hf.py (transformation skill, callable tool) — uploads folder to HF Hub (has built-in confirmation prompt)

Local File (local)

Save processed data to a local output directory (default behavior without destination).

Example:

/magic:deliver local

Safety

  • All external deliveries require user confirmation (PAUSE gate)
  • Database writes use non-read-only connections — PAUSE is mandatory
  • Defaults to private visibility for HuggingFace repos
  • Credentials are never included in uploads or dataset cards
  • Use --create-pr for PR-based HF delivery when working in teams
Was this page helpful?
Edit on GitHub

Last updated on

On this page