You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Sol/Astra: Worker.ts wait defaults to 540 s, but Claude Code moves a subagent's foreground Bash call to the background after 180 s, and the codex run is then killed #2824
What.agents/Sol.md and agents/Astra.md start the codex helper in the background and then wait in the foreground with bun ~/.claude/LIFEOS/TOOLS/Worker.ts wait <lane> --slug <slug> (Bash timeout: 560000). With no --max-seconds, the wait uses WAIT_DEFAULT_SECONDS = 540 (LIFEOS/TOOLS/Worker.ts:179).
Claude Code 2.1.322 moves a subagent's foreground Bash call to the background once it has run for 180 s (the client's constant is 180000 ms, applied to async subagents unless the prompt cache setting is "off" or "1h").
What happens. Observed timeline of one failed Sol audit:
Time
Event
+0 s
The wrapper starts CodexProgress.ts in the background and waits in the foreground
+3 min
The client moves the wait to the background
+3 min 15 s
The wrapper has nothing left in the foreground, so it ends its turn
+8.5 min
Both background tasks end with "failed with exit code 144", and the helper logs failed to send SIGTERM ... EPERM
Result
verdict "error", exit_code null, duration 512,701 ms, no audit
Shorter runs sometimes survive, because codex can finish inside the window between the wrapper going idle and the kill (two earlier runs of 405 s and 464 s were backgrounded the same way and still finished). Longer runs are at risk; the one observed failure ended at 512 s.
Suggested fix. Keep every single wait under 180 s, so the wrapper stays in the foreground and simply loops on running, as both agent files already instruct:
Lower WAIT_DEFAULT_SECONDS to 150 (any value comfortably under 180 works). This fixes every caller at once.
Update the "blocks for up to 540 seconds" sentence in both agent files to match.
In practice, ten consecutive Sol audits that waited at most 150 s per call each made 2 to 4 waits, none was moved to the background, and all of them finished.
Open question for the maintainers. What ends the backgrounded tasks about five minutes after the wrapper goes idle? It may be tied to the subagent's lifecycle or to the prompt-cache window. If that kill is intended client behaviour, then any LifeOS wrapper that holds a long foreground call is exposed, not only these two.
Activity