A tool's result is cut at 32 KiB by default, not 60,000 bytes: what is cut is one read_output call away, and what is sent goes out again with every call after it¶
Decision 781 · accepted · 2026-10-05 · issue #407
- Why 32 KiB (32,768 bytes). The maintainer's choice between the old 60,000 and
the issue's candidate, 16 KiB, the blob inline limit where 771's stub floor sits.
32 KiB still sends whole what an agent ordinarily asks for: a
grepof 200 matching lines of up to about 160 bytes, and a file ofprecise_edit's size (1,560 lines, 31,042 bytes asread_filenumbers them) in one read. At 16 KiB that file is cut, and so is such agreppast about 80 bytes a line, and each cut the model needs the rest of is a round trip; a number below 32 KiB needs measurements of its own. - What a cut costs. The model is sent the first 32 KiB and the marker, and
read_outputpages the rest back, 200 lines a call: one round trip, the whole prompt again with the page, when the model needs the rest, and nothing when it does not. Against that, every later call carries up to 27,232 bytes less, until a compaction summarises the result away; in #407's dear run, four calls. Offline,cut_output, whose script does read the rest back, sends 39,649 tokens (four bytes each, 772) over its three calls where it sent 46,422: the call with the page carries the whole result either way, and the call before it 27 kB less. - One default.
grep,git_read,read_file,shellandweb_fetcheach kept a copy of 60,000 for a context with no config; they read%Config{}'s now, so the next move is one line. The key table saysread_outputpages the rest back, where it said the rest is kept as a blob.
Governs: apps/troupe_core/lib/troupe/config.ex