[add] Windsurf-Cascade KPL (Issue #25)

[rhyannonjoy]

• Resolves Issue platform data (known platform limits): add Cascade web fetch notes #25
• Tested on Hybrid Arena Mode to widen LLM pool
• Routing and toolchain deterministic with read_url_content->view_content_chunk calls; curl bypass when agent acknowledges that Cascade chunking pipeline includes a type of truncation by design (chunk index with summaries, sometimes populated sometimes not, metadata, sometimes per-chunk truncation include byte-loss notes)
• No retrieval limit observed, but agent-reasoned write ceiling. Agents tend to claim to perform full chunk analysis, but struggle to prove it by completing a write task. With that said, it’s a pipeline designed to emulate “human-skimming” and doesn’t claim to be crawl on-demand. The bottom line is that this pattern creates complications for anyone building agentic infrastructure, just asking agents to read stuff from the web, or using agents to maintain documentation workflows.

—

Cascade Test Details

Like Copilot and Cursor, Cascade testing uses manual chat sessions in the Windsurf IDE.
The framework generates prompts, but execution requires copy-paste into the Cascade chat
panel. Cascade is the first platform in this collection with a user-invocable web directive,
@web, making it the primary variable under test alongside the standard truncation questions.
All 11 URLs run across three tracks; interpreted and explicit track results compared
directly to isolate the @web effect.

Category	Question	Source
Baseline	What does Cascade retrieve by default? Does @web change the ceiling/output size on the same URL?	MongoDB docs ~20KB–256KB; HTML, Markdown URL variants
Structured Content	How does Cascade handle tables, code blocks, nested headings, and JavaScript-rendered pages?	Wikipedia, Anthropic API docs, Markdown Guide, Gemini docs
Offset/ Pagination	Does view_content_chunk auto-paginate after truncation, or only when prompted?	256KB MongoDB tutorial; fragment navigation via URL # identifier
Edge Cases	How does Cascade handle redirect chains, SPAs, raw Markdown files, and JSON endpoints?	5-hop redirect chain, Gemini landing page, GitHub raw .md

Cascade self-reported three tools: read_url_content for
URL fetch, view_content_chunk for paginating documents via DocumentId, and search_web
for query-based lookup. read_url_content requires approval before execution,
a behavior with no Copilot or Cursor equivalent.

[dacharyc]

Arena mode sounds interesting! But dang, the retrieval details here are - not good news for folks who are asking Windsurf Cascade to read things.

Such a fascinating read - great work, and thank you for continuing your exploration!