agentsclimarketplace
Sign in

Agentic search

Skill appautomaton/webmaton/skills/agentic-search

Grok-primary deep research skill for source-backed web work. Use when the task needs current web research, grounded citations, supplementary Tavily/Firecrawl source discovery, high-fidelity page-to-Markdown fetch, Tavily site mapping, verbatim quote extraction, source reranking, or reusable multi-step research sessions. Use as a supplement when ordinary search results feel incomplete, thinly sourced, or need more URLs, quotes, fetched pages, or session composition; not for simple one-off facts.From its SKILL.md

Install
npx -y skills add appautomaton/webmaton --skill agentic-search

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

3 things to look at

  • reads credentialsReads from 3 credential sources: `GROK_API_KEY` and 2 more.
  • 18 stars18 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
  • runs commandsInstructs the agent to run 6 commands, including `uv run scripts/agentic_search.py` and 5 more.

SKILL.md

3.3 KB, 711 tokens by cl100k_base, as published. Nobody here has run it

agentic-search

#ScriptIntent
1scripts/agentic_search.pyResearch a topic — AI-reasoned answer with citations
2scripts/agentic_fetch.pyFull page → Markdown, no summarization
3scripts/agentic_map.pyEnumerate URLs under a site (Tavily-only)
4scripts/agentic_extract.pyVerbatim title + 2–4 quotes from a URL
5scripts/agentic_rank.pyRerank session sources by a refined query
6scripts/agentic_get_sources.pyRetrieve or list cached sessions

How to invoke

All scripts use PEP 723 inline deps — always invoke via uv run. Allow up to 3 minutes per call.

uv run scripts/agentic_search.py --query "..." [--extra-sources N] [--auto-fetch-top N] [--platform "..."]
uv run scripts/agentic_fetch.py --url "..." [--engine auto|tavily|firecrawl|grok]
uv run scripts/agentic_map.py --url "..." [--instructions "..."] [--limit N]
uv run scripts/agentic_extract.py --url "..." [--session-id S]
uv run scripts/agentic_rank.py --query "..." --session-id S
uv run scripts/agentic_get_sources.py --list | --session-id S

Session workflow

agentic_search generates a session_id persisted to disk — the connective tissue for composing steps:

agentic_search → session_id
     ├─→ agentic_rank --session-id S --query Q     (rerank in place; mutates session)
     ├─→ agentic_extract --session-id S --url U    (append verbatim quotes)
     └─→ agentic_get_sources --session-id S        (inspect full session)

Sessions survive between invocations but may be cleared on OS reboot.

Operating discipline

Load the relevant reference before composing a non-trivial query — don't load all four upfront.

  • references/search-discipline.md — two-pass methodology, citation contract, time-context heuristic. Read for research tasks.
  • references/fetch-fidelity.md — engine trade-offs, fidelity guarantees, when not to fetch. Read before presenting extracted content.
  • references/extract-and-rank.md — extract vs fetch, when to rank, session composition patterns. Read for multi-step workflows.
  • references/provider-quirks.md — env vars, provider schemas, retry policy, debugging. Read when configuring or debugging.

Failure modes

  • No GROK_API_KEY / GROK_API_URL → config error on any Grok-using script.
  • agentic_fetch --engine auto total failure → retry with --engine grok or fall back to built-in WebFetch.
  • agentic_map without TAVILY_API_KEY → hard exit.
  • agentic_rank --session-id S session expired → re-run agentic_search.
  • sources_count: 0 with non-empty content → Grok used training data; check providers_used (grok-web-search = real citations, grok = heuristic fallback).

What ships with it: 15 files

115.1 KB alongside SKILL.md, 9 of them executable

references/

scripts/

Keep looking

Skills are one crate of 325,949. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.