llm-multi

A plugin for LLM that combines file bundling and LLM mapping capabilities into two complementary commands: llm archive and llm map.

Commands

llm archive

Bundle files into JSONL format or extract files from a JSONL archive:

llm archive [files...]

Options:

• --create - Create an archive from files (default)
• --extract - Extract files from an archive (exclusive with --create)
• --basename - For create: strip path and use only filename. For extract: ignore paths, write only filenames to current directory (last copy wins on collisions)
• --basedir <str> - For create and extract: strip this prefix from paths, refuse to process files without this prefix, printing warnings instead
• --format - Output format: jsonl, json, jsonarr, xml, xmlish

llm map

Apply an LLM prompt to each item in a JSON structure or repeated content:

llm map data.jsonl "Summarize this: {item}"

The item will be inserted at the first {item} in the prompt. If no {item} exists, it will be inserted at the end, as \n{item}.

Options:

• --model, -m - Model to use (default: gpt-4o-mini)
• --temperature, -t - Temperature for generation (default: 0.0)
• --tokens - Max tokens (default: 15000)
• --concurrency, -c - Number of concurrent requests (default: 16)
• --output, -o - Write output to file
• --branches - Repeat content this many times, prefix the base filename with {i}_
• --content - Include original content in output (incompatible with --json/--jsonarr output formats)
• --format - Set both input and output format: jsonl (default), json, jsonarr
• --iformat - Input format: jsonl (default), json, jsonarr
• --oformat - Output format: jsonl (default), json, jsonarr

Format descriptions

JSON based formats, useful for piping into map -

• jsonl: One JSON object per line with path and content fields
• json: Single JSON object with filenames as keys and content as values
• jsonarr: JSON array containing only content (no filenames)

XML based formats, useful to pass an entire archive to an LLM (llm archive ... | llm prompt ...):

• xml: XML elements with properly escaped content (e.g., <foo.html>&lt;em&gt;escaped&lt;/em&gt;</foo.html>)
• xmlish: XML elements with raw content, no escaping (e.g., <foo.html><em>preserved</em></foo.html>)

Workflow Example

Process multiple files in parallel:

llm archive file1.py file2.js | llm map - "What programming language is this?"

Process and extract:

llm archive file1.py file2.js | llm map - "Add --reticulate-splines option" | llm archive --extract

Process multiple files with one call:

llm archive --format xmlish file1.py file2.js | llm prompt "Is the algorithm identical in both files"