paper-banana-orchestration

PaperBanana Orchestration Skill

You are the orchestrator for the PaperBanana multi-agent pipeline. Your job is to coordinate 5 sub-agents to generate publication-quality academic illustrations.

Pipeline Overview

Retriever → Planner → Stylist → Visualizer → [Critic → Visualizer] ×N

File-Based Storage Convention

Long text content (descriptions, critic suggestions) is stored as separate files in subdirectories, NOT inline in pipeline_state.json. This keeps the state file lightweight and prevents issues with large JSON values.

Directory layout inside the output directory:

{output_dir}/
├── pipeline_state.json       # lightweight metadata + file path references
├── descriptions/             # all text descriptions and critic suggestions (.txt)
├── images/                   # all generated images (.jpg)
└── code/                     # matplotlib code for plot tasks (.py)

Convention: When a sub-agent writes a description, it writes the full text to a file (e.g., descriptions/desc0.txt) and stores only the relative file path in pipeline_state.json (e.g., "target_diagram_desc0": "descriptions/desc0.txt"). To read a description, construct the absolute path: {output_dir}/{relative_path}.

This convention applies to all description keys (target_*_desc*), critic suggestion keys (target_*_critic_suggestions*), image path keys (target_*_image_path), and code keys (target_*_code). All paths in pipeline_state.json are relative to output_dir.

Step 0: Parse User Input

Parse the user's input to determine:

1. task_type: "diagram" or "plot"

   • If the user provides a method description, methodology section, or asks for architecture/pipeline/framework diagrams → "diagram"
   • If the user provides raw data (tabular, JSON) or asks for charts/plots/visualizations → "plot"
   • If --type diagram or --type plot is explicitly specified, use that
   • If ambiguous, ask the user via AskUserQuestion

2. content: The main input content

   • For diagrams: methodology section text
   • For plots: raw data (tabular, JSON, or text)
   • If the user provides a file path, read the file content