Orchestrate parallel research agents and write a cited report to disk.
npx skills add https://github.com/Lawrence908/chiron --skill deep-research-lawrence908
Please help me install this Agent Skill. Skill: deep-research Source: https://github.com/Lawrence908/chiron/tree/main/plugins/meta/skills/deep-research Command: npx skills add https://github.com/Lawrence908/chiron --skill deep-research-lawrence908
/deep-research "current state of small open-weight LLMs for on-device coding assistants" --depth deep to get a referenced markdown report saved under your research directory. ## Quick Start Ask the assistant to run /deep-research on your topic, optionally specifying depth, format, output path, and source lanes.High-intent search queries and answers about installing and using this skill.
Invoke /deep-research with your topic and optional flags for depth, format, output path, and sources. The orchestrator states a short plan, spawns parallel research agents, then writes a cited markdown report to disk.
Three tiers exist: quick uses 3 subquestions in about 5 minutes, standard uses 5-6 subquestions in about 15 minutes, and deep uses 8-10 subquestions with a mandatory gap pass taking 30 or more minutes.
Yes, the workflow is designed for unattended execution. After the initial plan statement it never asks questions mid-run, notes agent failures in the Method section, and always writes a report file even if results are thin.
Agents can search the web, documentation via Context7, arXiv and Hugging Face papers, the Hugging Face hub, and a local codebase through a read-only Explore agent. Lanes are auto-picked from the topic or set with the sources flag.
Each agent returns locally numbered citations in a fixed packet format. The orchestrator renumbers citations globally during synthesis and deduplicates sources by URL so every claim traces to a unique source entry.
Fan-out is capped at 12 initial agents plus 4 gap-pass follow-ups, and every claim must trace to an agent-cited source or it is dropped. Narrow factual questions waste compute on deep tier and should use quick instead.
blockscoutClarify codebase topics through guided discussion and propose a research-note outline.
UniquecreteConvert Markdown files into formatted PDFs using pandoc and xelatex.
vanyastaffVerifies claims against available evidence and data before responding.
Geralt1983Interact with Google Gemini models and NotebookLM via command line.