- 0.71.1: browser_evaluate says it is the last rung, and which tool reads what it is not for
A host's model (invisible_dots, z-ai/glm-5.3-flash, four tasks on 0.71.0) made 13 of its 63 calls with
browser_evaluate, most of them to read a heading that browser_read_text reads: the description named what evaluate
is for and nothing of what it is not, and the ladder's fourth rung said "what none of the above can see", which a
model did not apply to a heading. A script runs in the page and the readers do not, so reading text that way is the
less clean choice for this browser.
The description now says it comes last, and that text, a heading or a price is browser_read_text's, markup
browser_read_html's and a control browser_snapshot's, because a script runs in the page and they do not. The
fourth rung of the page rules says "last". The description is 1006 characters, inside the 1008 of headroom its gate
holds.
- docs: the tool-surface figures of 0.71.1 in the fourteen pages that publish them
browser_evaluate's description grew by 163 characters, so the server declares 9,081 characters of description and
the content gate refused the 8,918 the pages said. Measured on 2026-10-11 the way the pages say (tiktoken
o200k_base): the descriptions alone are 2,096 tokens, 42 more than the 2,054 published, and the median per tool is
unchanged. The serialization behind the published 3,464 for the complete definitions is not recorded and four
forms tried give 3,422 to 3,675, every one of them 42 more on 0.71.1, so the pages say 3,506; forty turns of it are
about 140,000 tokens.