nobu/research-agent-opencode
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
Research agent for Open Code which gives ability to process 50+ results for each request and search for domain knowledge data
131 commits 1 branch 0 tags 978 KiB
Find a file
Exact
Nobu b11d560db4 AGENTS.md: operator-output refinements (ISTJ register, inline uncertainty, sign-off blacklist, density rule)
2026-09-26 18:27:30 +05:00
.opencode
examples
.gitignore
AGENTS.md AGENTS.md: operator-output refinements (ISTJ register, inline uncertainty, sign-off blacklist, density rule) 2026-09-26 18:27:30 +05:00
opencode.json
README.md

README.md

Web Search Agent

Deep web search for OpenCode (v1 & v2). Fetches 30 search results and up to 20 pages per query — far beyond the typical 10-20 result limit of built-in tools — and delivers a compact digest plus a full report file you can grep.

Why use it

Most LLM search tools return 10-20 results, limiting research depth on complex questions. This tool digs deeper — more sources, smarter extraction, quality filtering, better answers. Works with any LLM through OpenCode.

Quick Start

git clone https://github.com/itohnobue/research-agent-opencode
cp -R research-agent-opencode/.opencode /path/to/your/project/
cp research-agent-opencode/AGENTS.md /path/to/your/project/
cp research-agent-opencode/opencode.json /path/to/your/project/   # skip if you already have one

If you already have an AGENTS.md or opencode.json, keep your own (append this AGENTS.md's instructions) instead of overwriting. This teaches OpenCode to route all web searches through this tool. Test it: "Search for the most performant Rust web frameworks"

On first run, uv is bootstrapped repo-local into <repo>/tmp/uv/ (never system-wide — no installers touch ~/.local/bin, PATH, or shell profiles) and all Python dependencies are resolved from PEP 723 inline metadata via uv run --no-project, so a stray pyproject.toml in your project can never hijack the run. No API keys required.

Usage

.opencode/tools/web_search.sh "React server components best practices" --tech
.opencode/tools/web_search.sh "CRISPR delivery methods" --sci --med
.opencode/tools/web_search.sh "Kalman filter implementations" --sci
.opencode/tools/web_search.sh --url https://example.com   # direct page fetch (pages only — never file downloads; use curl -L -o for files)
Flag Sources Best for
(none) DuckDuckGo General web
--tech + Hacker News, Stack Overflow, Dev.to, GitHub Software, DevOps
--sci + arXiv, OpenAlex CS, physics, math, engineering
--med + PubMed, Europe PMC, OpenAlex Medicine, clinical trials
--url direct fetch of one specific URL, skips search Known-page retrieval only (never file downloads), full page text saved raw to its own report file (no quality filters)

Key features

  • 30 results / up to 20 pages per query via DuckDuckGo
  • Anti-bot bypass — Scrapling with TLS fingerprinting, auto-fallback to httpx
  • Smart extraction — Trafilatura content-area detection (article body, not sidebars)
  • Quality filters — content-farm/syndication detection (F5), recency filtering on time-sensitive queries (F7), stub-page drop, cross-page deduplication
  • Non-English support — auto-detects Japanese, Chinese, Korean and sets the appropriate DDG region
  • Wayback Machine fallback for paywalled pages
  • PDF extraction via pdftotext (poppler)
  • Wall-clock watchdog — hard timeout kills hung fetches on all platforms

Full feature list and blocked/API-routed domain documentation in .opencode/agents/web-searcher.md.

Requirements

  • Python 3.11+ (managed by uv; downloaded into uv's cache if needed)
  • pdftotext (optional — from poppler, for PDF content extraction)

License

MIT