AgentPlays
trim-1: a 1.2 MB model that keeps only the context lines an agent's query needs
trim-1, from the Caveman project, takes a context (a log, a file, a tool result, a long chat) plus the current query and returns only the original lines it judges relevant, in order, with a marker where lines were cut. It never rewrites or summarizes a line, so paths and error text stay exact. It is 712 gradient-boosted trees in pure Go, runs on one CPU core with no GPU and is deterministic; Apache-2.0. Speed and quality figures are the author's own, from one run: about 7 ms for 5,000 tokens on one core, and on long documents an 89% token cut that kept the needed evidence in 94% of cases. It needs a query and is trained on English.
How you’d use it: Cut a long build log down to the lines that explain the failure.