Search the web and fetch content with zero-config fallbacks

Gives Pi Agent its own web-access package (pi-web-access, installed globally) for current web information, page fetches, PDFs, YouTube, and GitHub content - zero-config via Exa MCP with no API key required, falling back from Exa to Perplexity to Gemini if needed.

Use it when a Pi Agent task needs current web information, page fetches, PDFs, YouTube, or GitHub content, or specifically when Pi should use its own web-access package instead of another agent's browser tool. Every web_search call must include workflow: "none" with no exceptions, for a single query or a batched list of queries, since this skips the interactive browser curator popup the user does not want opening:

web_search({ queries: ["query 1", "query 2"], workflow: "none" })

Four tools are available: web_search returns synthesized answers with citations and can be called many times per turn (always with workflow: "none"); code_search is zero-key Exa code-context search, preferred over generic web_search for library, API, or code lookups; fetch_content fetches one or more URLs to markdown and handles PDFs, YouTube, and GitHub; and get_search_content pulls the full text of big pages (over 30k characters) that get truncated in normal responses, so they don't blow context. fetch_content behaves differently by source: GitHub URLs are cloned rather than scraped, giving real files plus a local path to explore with read/bash (private repos need the gh CLI), which is the right choice for dev work; PDFs are auto-extracted to markdown in ~/Downloads/, readable in sections, text-only with no OCR; and YouTube or video URLs return full raw transcripts plus frame extraction, which needs a GEMINI_API_KEY (not zero-config) and, for frame extraction specifically, ffmpeg and yt-dlp.

Query counts are hard minimums keyed to the user's phrasing, and must be counted before answering rather than stopped short: "web search" needs at least 2 varied queries before synthesizing; "extensive web research" needs at least 4 queries across totally different keywords and angles; and "deep research" needs at least 8 queries across totally different keywords and angles, run in 2-3 successive batches that refine angles after each one. A single batched web_search call counts every query inside queries[] toward that total, and a first batch under the minimum means firing another batch before synthesizing. If the Exa-Perplexity-Gemini chain fails, or ranked results with URLs are specifically needed, DeepAPI web search is the fallback:

test -n "$DEEPAPI_API_KEY" || { echo "DEEPAPI_API_KEY is not set"; exit 1; }
curl -s --max-time 60 "https://deepapi.co/v1/search/web" \
  -H "Authorization: Bearer $DEEPAPI_API_KEY" -H "Content-Type: application/json" \
  -H "Idempotency-Key: $(uuidgen)" \
  -d '{"query": "your search terms", "maxResults": 5, "maxCostUsd": "0.05"}'

Results land in .output (title, url, snippet per item), the query must stay under 500 characters, and full details live in the separate deepapi skill.

Pi Agent tasks that need live web information, code-context lookups, or content fetched from PDFs, YouTube, or GitHub, without opening the interactive browser curator popup.