SIGN IN SIGN UP

feat(evals): Add web-fetch and web-tools-selection eval suites (#1300)

Stacked on #1299 (which stacks on #1294). Docs for four new Langfuse
datasets — the snapshots stay out of version control per review
(Langfuse is the only copy, see README Design Decisions); the tool
changes they validated are in #1298.

- `web-fetch-evals` (8) / `web-fetch-evals-errors` (3): the
`apify/web-fetch` default tool
- `web-selection-evals` (7) / `web-selection-evals-errors` (2): the
rag-web-browser vs web-fetch vs search-actors clash,
specialized-Actor-over-generic-fetch

Final ladder on the current tool descriptions (#1298): claude-opus-5,
claude-sonnet-5, and claude-haiku-4-5 all green on both families, except
one documented residual — Haiku reproducibly rewrites ftp:// to https://
silently on `web-fetch-unsupported-protocol` (3x, despite the scheme
note in the description and on the `url` parameter); the case keeps
documenting it on purpose.

---------

Co-authored-by: Claude <noreply@anthropic.com>
J
Jiří Spilka committed
23b94c07f7cc373ac2d1276f4e4fa67d6658fecd
Parent: f572f12
Committed by GitHub <noreply@github.com> on 9/4/2026, 2:14:11 PM