Irclaw Research Team

irclaw Research Team

Overview

Spawn N parallel claude workers in Zellij panes. Each agent owns one research domain, fetches URLs, optionally parses docs via liteparse, and writes findings to research/<domain>.md. You watch them live and ctrl-c any runaway pane.

Quick Start

# Launch 4 research agents in a new Zellij tab
nu ~/.claude/skills/zellij-team/zellij-team.nu \
  --files [research/tasks/deep-dive-gastown.md \
           research/tasks/deep-dive-elixir.md \
           research/tasks/deep-dive-llm.md \
           research/tasks/deep-dive-pipeline.md]

Task files live in research/tasks/. Each is a self-contained prompt with context, URLs, and output path.

liteparse Pipeline

liteparse does NOT fetch URLs itself — pipe curl output in:

# Parse a remote PDF
curl -sL "https://example.com/doc.pdf" -o /tmp/doc.pdf
npx @llamaindex/liteparse /tmp/doc.pdf --output json > research/doc.json

# Parse an HTML page (save first)
curl -sL "https://example.com/page" -o /tmp/page.html
npx @llamaindex/liteparse /tmp/page.html --output json > research/page.json

Ollama Headless Query (no UI needed)

# Query local Ollama directly — skip vibecode-webgui for agent use
curl -s http://localhost:11434/api/generate \
  -d '{"model":"llama3","prompt":"Summarize: <paste content>","stream":false}' \
  | jq -r .response

For fleet nodes: Ollama runs on each node at localhost:11434. No central coordinator needed for per-node research.

Full Nushell Pipeline

# curl → liteparse → ollama → append to research file
def research-url [url: string, output_file: string, model: string = "llama3"] {
    let tmp = $"/tmp/(random chars -l 8).html"
    ^curl -sL $url -o $tmp
    let parsed = (^npx @llamaindex/liteparse $tmp --output json | from json)
    let text = ($parsed.pages | each { |p| $p.text } | str join "\n")
    let summary = (^curl -s http://localhost:11434/api/generate
        -d ($'{"model":"($model)","prompt":"Summarize for technical research: ($text | str substring 0..3000)","stream":false}')
        | from json | get response)
    $"\n## ($url)\n($summary)\n" | save --append $output_file
    rm $tmp
}

Task File Format

Each file in research/tasks/ follows this template:

Working dir: /path/to/project

Context: [what's already known — point to existing research/*.md]

URLs to fetch:
1. https://... — what to extract
2. https://... — what to extract

Output: Append findings to research/<domain>.md under ## Deep Dive

Key questions:
- Question 1?
- Question 2?

Research Output Structure

research/
  tasks/           ← agent task prompts (input)
  gastown-irclaw.md
  elixir-beam.md
  llm-ollama-webllm.md
  datadog-fleet.md
  tools-vibecode-liteparse.md
  pipeline-design.md
  SUMMARY.md       ← synthesized by leader after all agents complete

Zellij Controls

Fleet Distribution

Skills live at ~/.claude/skills/ on each node. Ansible playbook to sync:

ansible-playbook -i ansible/inventory ansible/sync-skills.yml

Key Constraints

• liteparse needs a local file — always curl first
• irclaw is private — skip github.com/irclaw-labs/irclaw, use local ansible playbooks for context
• Ollama port — always localhost:11434, not a fleet-wide coordinator
• ARM64 nodes — liteparse runs fine via npx on aarch64; no binary issues