Do the subtraction and the shape of the failure is clear: a 240-second suite cancelled at second 60 leaves 180 seconds of compilation happening where nobody is watching, three quarters of the run [1]. Cancellation is a message to the client, not a signal to the child, so the process kept going while the model sat holding an error where its results should have been [2]. The workaround the model invented later, `sleep 240`, matched the suite's runtime exactly, which is the closest thing to a spec that anyone had given it [4][4].
The first patch treated the silence. Release 0.10.2 emits a progress notification every ten seconds while a command runs in the foreground [6]. Six of those intervals fit inside a sixty-second window, so the call tolerates five consecutive dropped beats before the silence adds up to a cancellation [2]. That fixed the killings without touching the blocking [7].
What happens now at second ten is worth reading closely. The command is not killed and not restarted; the same process is promoted to a background job [11]. Output capture is durable from the first byte, so nothing printed during its foreground life is lost at the boundary [12]. The tool call returns immediately with a resource link whose name is the command itself, which is what lets a client still render "make test ... still running" after a context compaction has discarded the original request [13]. Each job is an MCP resource with a URI like `octofs://jobs/17342-1` [17]. On exit the server emits `notifications/resources/updated` for that URI, and one read hands back the exit code and the output tail [14]. The longest blocking call in that path goes from 240 seconds to about 10, a factor of 24 [3].
The flag is the more portable lesson. Background execution shipped in 0.11.0 behind a flag on the shell tool, and 0.13.0 deleted it, because the flag asked the model to predict a duration it cannot know: `cargo build` is instant on a warm cache and six minutes cold [8][10][11]. The author traces the reasoning to 0.9.0, when a `--line-mode` switch was removed on the grounds that safety shipped behind a flag is safety most people never enable [9]. It is a rare release note that describes its own feature as the same bug in a different costume [8].
Two things have to be true before that ten-second threshold transfers. Your workload has to be bimodal the way builds and test suites are, with fast calls well under the line and slow ones well over it; if most of your calls land just past it, you promote nearly everything and pay a round trip on each. And your client has to render resource links and act on `resources/updated` [13][14]. A client that ignores both leaves you back at polling with a tidier URI [4].
One default I would copy outright: a resource read returns at most the last 30 KB [15]. The verdict lives at the end of a build log, and feeding a model the first 30 KB of a log whose final line reads FAILED is how you get a confident report that everything passed [16].
