agent: a floor under compaction, /compact, and a way to turn it off

Compaction has folded the older half of the conversation at 70% since it
was built, and the threshold was chosen to leave headroom: the fold is
itself a model call with the history in it, and it is the one thing that
reliably breaks prompt caching. That argument is still right, and the
70% default has not moved.

What it does not cover is a history the split refuses to touch. The cut
keeps the last six messages, so a conversation of four enormous tool
results is too short to cut six off the end of — compactionSplit returns
-1, nothing is folded, and the window goes on filling with the only
mechanism for emptying it declining every turn. Below 70% that restraint
is correct. At 97% it is not: a verbatim tail is worth nothing if the
next request has no room to be sent. So the split now takes the tail
length as a parameter, and past -compact-floor it is tried again keeping
one message instead of six. For a conversation of any real length this
changes nothing, because the cut lands on the midpoint either way; its
only job is to make a short history of huge messages foldable at all.
When even that finds nothing, the loop says so through a new
EventNotice, which beats letting the next request come back as an API
error about a number the user cannot see.

/compact is the same fold on demand. It refuses while a turn is running,
because it rewrites the history that turn is building on, and it takes
the smaller tail past the floor for the same reason the floor does.
Afterwards the context gauge goes blank rather than keep reporting what
the model read before the conversation was replaced; the next turn
measures the window again and it returns.

auto_compact: "off" in the config, or -no-auto-compact for one run, is
for anyone who would rather lose the turn than the transcript. The
window then stops the turn instead of folding it, and the check happens
before the prompt joins the history — a turn that never ran must not
leave its prompt behind, or compacting and asking again would send it
twice. The flag only turns the fold off, never on, which is the
direction -yolo already runs in: a file should not quietly grant more
than the command line did. An absent key means on, so a config written
before this existed does not start stopping the agent.

Two things only running it would have shown. A successful fold left
lastInputTokens describing the history it had just thrown away, so with
auto-compaction off the stop kept firing on a number that had stopped
being true and /compact could never clear it — the count is now reset
with the history it described. And the mock API reported a fixed 9.2k
per turn, which no amount of talking would move; it now scales with the
request, so the gauge climbs the way a real conversation makes it climb
and a threshold can be reached by typing rather than by shrinking the
window until it is already crossed.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>