yoi/.yoi/tickets/closed/20260530-204045-webfetch-readable-extraction/item.md

id	slug	title	status	kind	priority	labels	created_at	updated_at	assignee	legacy_ticket
20260530-204045-webfetch-readable-extraction	webfetch-readable-extraction	WebFetch: extract main HTML content with lightweight readability	closed	task	P2	web tools html	2026-05-30T20:40:45Z	2026-05-30T20:55:13Z	null	null

Background

WebFetch currently returns bounded, safety-checked content but the HTML path is still close to raw page text: it strips tags with a small local formatter and does not try to isolate the article/main content. For LLM use, a reader-mode style extraction layer is more useful than raw boilerplate-heavy page text.

readability-js was investigated but brings QuickJS / bundled JavaScript dependency weight. The desired direction is a lightweight pure-Rust extraction backend with fallback to the current html_to_text behavior.

Reference implementations checked out with ghq under .worktree/ghq-root/ for planning:

• github.com/quambene/readability-rs — crate readability-rs, MIT, small arc90-style extractor (Readable { title, content, text }).
• github.com/theiskaa/readabilityrs — crate readabilityrs, Apache-2.0, larger Mozilla Readability port with metadata/markdown support.
• github.com/readable-app/readability.rs — crate readable-readability, MIT, kuchiki-based extractor but sparse docs and older maintenance surface.

Requirements

• Keep WebSearch and WebFetch as separate tools. Do not add an automatic summarization/research tool in this ticket.
• Add a reader-mode extraction path for HTML responses in WebFetch.
• Use a pure-Rust dependency or local extraction implementation; do not use readability-js, QuickJS, Node, Python, or subprocess-based extraction.
• Prefer the lightweight readability-rs crate if it builds cleanly and produces usable title + main text; escalate if the crate is incompatible or obviously too low-quality for the included fixtures.
• Preserve the current network safety behavior: configured provider requirement, private/local host rejection, bounded redirects, response size limits, binary rejection, output truncation, and untrusted-content warning semantics.
• Preserve fallback behavior. If readability extraction fails or returns empty/too-short text, return HTML text produced by the existing local fallback rather than failing the entire fetch.
• Structure HTML fetch output so the LLM can distinguish extraction metadata from document content. At minimum include:
  • extraction method (readability or fallback name)
  • fallback indicator / reason when applicable
  • title when available
  • main text
  • existing fetch metadata such as URL/final URL/status/content type/truncation
• Keep HTML returned to the LLM as text by default. Do not expose full extracted HTML unless there is a clear existing output field need.
• Add focused tests with small HTML fixtures covering:
  • article/main content is preferred over nav/footer/sidebar boilerplate
  • fallback is used when readability extraction is not useful
  • output remains bounded/truncated under the existing output limit

Non-goals

• Provider expansion or changes to WebSearch provider selection.
• LLM-generated summaries inside WebFetch.
• Browser rendering, JavaScript execution, or dynamic page support.
• Large benchmark suite or exhaustive readability quality comparison.
• Public API/protocol changes beyond the tool result JSON shape.

Implementation plan

1. Add the selected pure-Rust readability dependency to crates/tools.
2. Introduce a small internal HTML extraction helper, e.g. extract_html_document(html, base_url, output_limit), wrapping readability success and fallback.
3. Update the ContentKind::Html branch in WebFetch rendering to use the helper.
4. Keep existing html_to_text as fallback and testable utility.
5. Update tests in crates/tools/src/web.rs or a focused tools test module.
6. Validate with formatting, focused tools tests, and broader checks appropriate to the dependency change.

Acceptance criteria

• WebFetch HTML responses prefer extracted main content over navigation/footer/sidebar boilerplate in tests.
• WebFetch still returns useful bounded text when readability extraction fails or is empty.
• Tool output clearly reports extraction method and fallback status.
• No JavaScript engine/runtime dependency is introduced.
• Cargo.lock and Nix cargo hash implications are handled or explicitly reported.
• cargo fmt --check, focused tools tests, cargo check -p tools, and ./tickets.sh doctor pass or any failure is clearly reported as unrelated/pre-existing.