mirror of https://github.com/sbt/sbt.git
0c03718218
Since #8678 (sbt 1.12.2), Fork.apply has routed the `connectInput && outputStrategy == StdoutOutput` case through a new JProcess path that calls `jpb.inheritIO()`. `inheritIO` ties the child JVM's stdio to the sbt server JVM's OS-level fd 1/2, bypassing the `Terminal` indirection. In server mode (which `sbt --client` uses), that indirection is how task output reaches the client terminal — so the child's stdout/stderr went nowhere observable to the client. Keep `redirectInput(Redirect.INHERIT)` so REPL/raw-mode keystrokes still reach the child (preserving the use case #8678 added), but leave stdout/stderr as PIPE and pump them to the active terminal's output/error streams in two daemon threads. The pump threads are joined after `p.waitFor()` so all buffered output drains before the task completes. The active terminal is captured once on the task thread rather than re-resolved per-write via the `System.out`/`err` proxies, otherwise a concurrent terminal swap can route mid-pump bytes to the wrong client (observed on the CI JDK 8 matrix; the proxy's `activeTerminal.get()` is evaluated at every write). Forced scala/scala3 CI to downgrade from sbt 1.12.10 to 1.12.1 (see scala/scala3#25995). Diagnosis credit: @mbovel in #9185. Fixes sbt/sbt#9185. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|---|---|---|
| .. | ||
| src | ||
| NOTICE | ||