[ayman3000]

Bug (seen in a Naseem benchmark run): the model ran … python3 -m http.server 8080 & … in run_shell. The background server held the stdout pipe open, run_shell waited for EOF until its 120 s timeout, then killed the process group — server included — so the next chrome_devtools__new_page got ERR_CONNECTION_REFUSED.

Fix: the run ends when zsh exits, not when the pipe closes.

• Output already written is collected (0.3 s drain). A child still holding the pipe is left running; its later output is drained and dropped, and the result says so.
• Timeout and Stop still SIGKILL the whole process group.
• PipeCollector reads with a dispatch readability handler and keeps the FileHandle alive until EOF (otherwise the read end closes and the server dies of SIGPIPE on its next log line — caught in the live repro).
• Lossy UTF-8 decode so the byte cap can't blank output; python3 -m http.server added to the auto-background markers.

Tests

• Replaced shellTimesOutEvenWhenOrphanHoldsStdout (it asserted the old behavior) with shellReturnsWhenShellExitsAndLeavesBackgroundChildRunning — failed before the fix (21 s, killed child), passes after (<1 s, child alive).
• shellTimeoutKillsTheWholeGroup, shellDetectsPython3HTTPServer.
• Live repro (not committed): python3 -m http.server 8799 & sleep 1; echo up returns in 1.5 s and three later curls get the page.
• Full suite: 665 tests pass (1 pre-existing known issue).

🤖 Generated with Claude Code