OpenCode Batch Processor

A Node.js package for batch processing folders with OpenCode AI SDK.

For detailed configuration documentation, see CONFIG_GUIDE.md.

Features

• Extract compressed files (zip, tar, tar.gz, rar, 7z)
• Process each folder with configurable agent and prompts
• Structured output scoring with JSON Schema
• Export results to CSV

Installation

npm install

Usage

CLI

# Validate configuration
node bin/opencode-batch.js --config config.example.md --validate

# Process folders (point to the directory containing archives/subfolders)
node bin/opencode-batch.js --config config.example.md --input "./Example/files" --output ./results.csv

# Enable debug output for troubleshooting
node bin/opencode-batch.js --config config.example.md --input "./Example/files" --debug

Programmatic

const { BatchProcessor } = require('./src');

const processor = BatchProcessor.fromConfigFile('config.md');
await processor.run('./Example/files', './results.csv');

Configuration File Format

The configuration file uses YAML frontmatter with markdown content:

---
agent: Build
model: anthropic/claude-3-5-sonnet-20241022
base_url: http://localhost:4096
extract: true
output_csv: results.csv
structured_output:
  type: json_schema
  schema:
    type: object
    properties:
      format_score:
        type: number
        description: Format score (0-100)
      prompt_effort_score:
        type: number
        description: Prompt effort score (0-100)
      core_prediction:
        type: number
        description: Core prediction result score (0-100)
      report_score:
        type: number
        description: Report score (0-100)
    required: [format_score, prompt_effort_score, core_prediction, report_score]
---

# Prompts

Please grade the student's assignment based on the following criteria:

1. Format: Check if the report follows the required format
2. Prompt Effort: Evaluate the effort put into prompt engineering
3. Core Prediction: Assess the accuracy of core predictions
4. Report Quality: Evaluate overall report quality

Folder contents:
{folder_contents}

Please provide scores for each dimension.

Configuration Options

Option	Type	Default	Description
agent	string	Build	Agent name
model	string	null	Model identifier (provider/model)
base_url	string	http://localhost:4096	OpenCode server URL
username	string	opencode	OpenCode server username (for basic auth)
password	string	null	OpenCode server password (for basic auth)
extract	boolean	true	Auto-extract archives
extract_extensions	array	[.zip, .tar, .tar.gz, .tgz, .rar, .7z]	Extensions to extract
auto_approve	boolean	true	Auto-approve Agent permission requests
timeout	number	300	Prompt timeout in seconds
delay_between_folders	number	0	Delay in seconds between processing folders (0 = no delay)
retry_count	number	0	Number of retries on failure (0 = no retry). Waits 10s between retries.
output_csv	string	results.csv	Output CSV filename
structured_output	object	null	JSON Schema for structured output
csv_columns	object	null	Column name mapping

Check Available Models

Before running, check what models are available on your OpenCode server:

# List available models
curl http://localhost:4096/config/providers

# Or with authentication
curl -u opencode:yourpassword http://localhost:4096/config/providers

CLI Options

Option	Description
-c, --config <path>	Path to configuration markdown file (required)
-i, --input <path>	Input directory containing archives/subfolders
-o, --output <path>	Output CSV file path
--validate	Validate configuration file only
--debug	Enable debug output for troubleshooting
--filter <file>	Only process folders listed in the filter file

Retry Failed Folders

If some folders fail processing, they are automatically saved to failed.txt in the input directory. To retry only those folders:

# First run - some may fail
node bin/opencode-batch.js --config config.example.md --input "./Example/files"

# Retry only failed folders
node bin/opencode-batch.js --config config.example.md --input "./Example/files" --filter ./Example/files/failed.txt

Permission Handling

When auto_approve: true (default), the tool will:

1. Subscribe to OpenCode server events
2. Listen for permission requests from the Agent
3. Automatically approve all permissions with "always" response
4. This prevents the Agent from stopping and waiting for user input

If you want manual control, set auto_approve: false.