bash.ts: detached=true leaves zombie processes when command times out

[LifetimeVip]

Description

File: packages/core/src/tool/bash.ts:143-148

const command = ChildProcess.make(parameters.command, [], {
  cwd: target.canonical,
  shell,
  stdin: "ignore",
  detached: process.platform !== "win32",
  forceKillAfter: Duration.seconds(3),
})

detached: true on non-Windows puts the command in its own process group. When Effect.timeoutOrElse fires (L157), the Effect's scope cleanup only kills the direct child — grandchild processes survive. Meanwhile bash.ts:162-170 treats timeout as a SUCCESS (timedOut: true) and returns it to the AI, so the AI may retry while the original process is still running.

Repro: AI runs a long command (e.g., npm install). It times out. AI gets "Command exceeded timeout" (not an error). AI retries. Original npm is still writing to node_modules. Race condition → corruption.

Fix: Remove detached or add explicit process-group cleanup on timeout.

OpenCode version

v1.16.0 (commit 3cf1cef)

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• fix(opencode): zombie processes from detached shell children on Linux #29506: Same root cause — detached: true in the bash/shell tool on Linux causes zombie processes, with the same proposed fix (detached: false)
• Bash tool: timeout-killed processes not detected as completed; user must manually abort #28654: Related — bash tool timeout-killed processes not properly cleaned up, leaving orphan child processes
• bash tool hangs when command spawns background child processes #20902: Related — bash tool hangs when commands spawn background child processes (same underlying stdio/process-group behavior)

[LifetimeVip]

Reproduction: detached=true creates zombie processes

bash.ts line 147:

detached: process.platform !== "win32",

On macOS/Linux, detached: true spawns processes in a new process group. When the bash tool's timeout fires and kills the main shell process, orphaned background processes survive. The TODO at line 95 acknowledges this:

TODO: Revisit process-group cleanup and platform coverage with shell-specific tests if current AppProcess semantics do not fully cover it.

The fix should use process group termination (kill(-pid)) instead of just killing the parent shell process.

[LifetimeVip]

已提交此问题的修复 PR #31893（LLM runner last-step signal, MAX_STEPS configurable, shell timeout fixes）