lab service --start kills tmux session when login shell is ~/hero/bin/nu #310

New issue

Open

opened 2026-06-03 17:04:44 +00:00 by thabeta · 0 comments

thabeta commented

2026-06-03 17:04:44 +00:00

Owner

Problem

Running lab service --start (or lab service hero_proc --start) causes the caller's tmux session to exit. This happens when the user's login shell is /home/driver/hero/bin/nu.

Root Cause

start_hero_proc() in crates/lab/src/service/hero_proc_exception.rs calls kill_all_hero_processes_except_self() as a pre-start cleanup sweep. This function walks /proc/<pid>/exe and SIGKILLs every process whose executable lives under ~/hero/bin/ — except the calling lab binary itself.

The exclusion logic only skips processes whose exe matches lab. It does not exclude:

The user's login shell (when it is a hero binary like nu)
Processes that are ancestors of the calling lab process
Session leaders of active terminal sessions

When the login shell is ~/hero/bin/nu, the tmux session's shell process matches the kill pattern. sigkill_with_group() then kills the entire process group, taking down tmux and everything in it.

Call Chain

lab service hero_proc --start
  → start_hero_proc()  [hero_proc_exception.rs]
    → kill_all_hero_processes_except_self()  [line ~413]
      → walks /proc/<pid>/exe
      → finds nu at /home/driver/hero/bin/nu (tmux shell)
      → sigkill_with_group(pid)  [line ~444]
        → kills PID + process group
          → tmux loses its shell → session exits

Relevant Code

crates/lab/src/service/hero_proc_exception.rs:

pub fn kill_all_hero_processes_except_self() {
    let hero_bin = crate::repo::paths::hero_bin_dir();
    let my_pid = std::process::id();
    let my_exe = std::fs::read_link("/proc/self/exe").ok();

    for entry in std::fs::read_dir("/proc").flatten() {
        // ... skips self
        let exe = std::fs::read_link(entry.path().join("exe"));
        if !exe.starts_with(&hero_bin) { continue; }
        if my_exe.as_ref().is_some_and(|me| me == &exe) { continue; }
        sigkill_with_group(pid);  // ← kills tmux shell
    }
}

Evidence

# tmux session tree:
tmux (457614)
  └── nu -c /home/driver/hero/bin/nu (457615) ← exe = ~/hero/bin/nu → MATCHES → KILLED
        └── nu (457618) ← also killed

# User's shell:
SHELL=/home/driver/hero/bin/nu

Suggested Fix

kill_all_hero_processes_except_self() should exclude ancestor processes of the calling lab process. Walk the PPID chain from my_pid up to init, collect all ancestor PIDs, and skip them in the kill loop.

This is the safest approach because:

The shell that launched lab is always an ancestor and must be preserved
It does not rely on fragile heuristics about session leaders or multiplexer detection
It handles arbitrary nesting (SSH → tmux → nu → lab)

Impact

Anyone whose login shell is ~/hero/bin/nu (or any other hero binary) will lose their terminal session every time they run lab service --start
This also affects any wrapper scripts or SSH sessions that use hero binaries as their shell

## Problem Running `lab service --start` (or `lab service hero_proc --start`) causes the caller's tmux session to exit. This happens when the user's login shell is `/home/driver/hero/bin/nu`. ## Root Cause `start_hero_proc()` in `crates/lab/src/service/hero_proc_exception.rs` calls `kill_all_hero_processes_except_self()` as a pre-start cleanup sweep. This function walks `/proc/<pid>/exe` and SIGKILLs every process whose executable lives under `~/hero/bin/` — except the calling `lab` binary itself. The exclusion logic only skips processes whose exe matches `lab`. It does not exclude: - The user's login shell (when it is a hero binary like `nu`) - Processes that are ancestors of the calling `lab` process - Session leaders of active terminal sessions When the login shell is `~/hero/bin/nu`, the tmux session's shell process matches the kill pattern. `sigkill_with_group()` then kills the entire process group, taking down tmux and everything in it. ## Call Chain ``` lab service hero_proc --start → start_hero_proc() [hero_proc_exception.rs] → kill_all_hero_processes_except_self() [line ~413] → walks /proc/<pid>/exe → finds nu at /home/driver/hero/bin/nu (tmux shell) → sigkill_with_group(pid) [line ~444] → kills PID + process group → tmux loses its shell → session exits ``` ## Relevant Code **`crates/lab/src/service/hero_proc_exception.rs`:** ```rust pub fn kill_all_hero_processes_except_self() { let hero_bin = crate::repo::paths::hero_bin_dir(); let my_pid = std::process::id(); let my_exe = std::fs::read_link("/proc/self/exe").ok(); for entry in std::fs::read_dir("/proc").flatten() { // ... skips self let exe = std::fs::read_link(entry.path().join("exe")); if !exe.starts_with(&hero_bin) { continue; } if my_exe.as_ref().is_some_and(|me| me == &exe) { continue; } sigkill_with_group(pid); // ← kills tmux shell } } ``` ## Evidence ``` # tmux session tree: tmux (457614) └── nu -c /home/driver/hero/bin/nu (457615) ← exe = ~/hero/bin/nu → MATCHES → KILLED └── nu (457618) ← also killed # User's shell: SHELL=/home/driver/hero/bin/nu ``` ## Suggested Fix `kill_all_hero_processes_except_self()` should exclude ancestor processes of the calling `lab` process. Walk the PPID chain from `my_pid` up to init, collect all ancestor PIDs, and skip them in the kill loop. This is the safest approach because: - The shell that launched `lab` is always an ancestor and must be preserved - It does not rely on fragile heuristics about session leaders or multiplexer detection - It handles arbitrary nesting (SSH → tmux → nu → lab) ## Impact - Anyone whose login shell is `~/hero/bin/nu` (or any other hero binary) will lose their terminal session every time they run `lab service --start` - This also affects any wrapper scripts or SSH sessions that use hero binaries as their shell