Files
oc/tests
only-cli f84074701d perf: spend one command where the tool used to need two
A tool call inside an agent session costs 23,000 to 33,000 tokens of
overhead whatever it prints, so the page-view win only reaches the
session total if answering a task takes fewer commands. Three places
were charging a command to say what the next command should be, each
found by capturing the command stream of a real agent run rather than
by reading the code.

A search result title is a link. Every engine puts it in an anchor
filling an <h2>, and the walk took the heading's text and returned,
dropping the href, so `do` on the most obvious number on a results page
printed the title back. The agent then spent a second command finding
the number that navigates. The href now rides along when the anchor is
the whole heading, which is the test documentation fails on purpose:
every heading in the Rust book and on an AWS CLI reference page carries
a permalink to its own id, and following one would refetch the page the
agent is already reading.

`find` pointed at its answer. With a single match it printed the block
and a number, and the agent's next command was always the `read` on
that number, so it now prints the region. With several matches it
showed a 200 character snippet of each even when the budget had room
for them whole, so it spends that room, on the same terms `FINISH`
already documents for a page that nearly fits.

A truncated block ended mid sentence. Asked for the first sentence of a
page, an agent was handed it complete, followed by a marker saying 302
characters were cut, and spent a command on `read` to find out whether
the sentence went on. The cut now falls on the last sentence that
finished inside the cap, and measured across five real pages it costs
nothing: four came out within three characters of before.

The package-lock name field catches up with the scoped package name,
which npm rewrites on any install.
2026-08-23 08:46:10 -04:00
..