Skip to content

A large tool result a compaction left behind is sent from then on as one line naming the read_output call that returns it, not again in full on every call

Decision 771 · accepted · 2026-10-04 · issue #389

Superseded in part

In part by 1024.

A compaction summarises the older part of the conversation and keeps the last few messages as they were, so a 60 kB grep the model read just before it went out again with every call after it, until a later compaction summarised it away. - What is behind. The messages a compaction kept, up to the model's last reply among them: State.compacted_through, set by the compaction and folded from its compacted event. A result after that reply is the one a compaction in the middle of a turn was about to send; the model has not read it, so it goes whole until the next compaction. Results since the compaction are never touched. - Which results. Those over the blob inline limit, 16 KiB: the log already keeps them as blobs, so the stub's id names one without a new write (Output.keep/2 makes sure). Below that a stub saves less than the read_output round trip it may cost, which sends the whole prompt again, and a floor of its own would be one more number to guess before slice 1 measures. - An error result by the same rule. The block keeps its error flag, so the model still knows the call failed. Most failures are a short message, under the floor; a large one is a failing command's output, the same bulk as any other. - The stub. [output of grep (48213 bytes) left out since a compaction. Call read_output(id: "sha256:…", offset: 1, limit: 200) to see it again.], the call a truncation marker names (650). The tool is named by the call just before the result, not by id across the conversation: a gateway that numbers its calls gives every response a call_0. - Only in what is sent. build_request stubs; the conversation in memory, the snapshot and the log keep the result whole, so a replay rebuilds the same conversation and the same boundary, and an audit sees everything. The summariser still reads what it summarises whole: it condenses those results rather than sending them again. - Only where read_output is offered. A profile without it (explore, answer, ask, librarian) keeps its results whole: a stub it cannot expand is a result lost. - A summary that fails leaves the conversation, and the boundary with it, as they were (774). - The prefix stays put. What is behind changes only when a compaction rewrites the head of the conversation anyway, so from one call to the next the prompt starts the same, which a prompt cache (slice 2) needs.

Governs: apps/troupe_core/lib/troupe/agent/server.ex, apps/troupe_core/lib/troupe/agent/state.ex, apps/troupe_core/lib/troupe/tools/output.ex, apps/troupe_core/lib/troupe/tools/read_output.ex