Feature request: Event-driven wake-up when a local process completes

Codex Desktop can launch a long-running local process, but after the current turn ends, it cannot subscribe to that process’s completion and resume the associated thread or Goal. The only available workaround is periodic scheduled polling.
This is inefficient for multi-hour ML training, builds, benchmarks, simulations, and deployments. It introduces unnecessary scheduled runs, delayed completion handling, and repeated status checks.
Requested behavior:

• Register a stable process handle when Codex launches a local command.
• Associate that process with the current thread and Goal.
• Keep the model inactive while the process is running.
• When the process exits, wake the thread exactly once.
• Resume with the exit code, completion time, and final stdout/stderr.
• Persist enough metadata to recover after app restart or system sleep.
• Distinguish normal completion, failure, interruption, and a missing process.
• Support Windows, macOS, and Linux.
  Current workaround: a scheduled task checks the PID and output artifacts every 30 minutes. This should be an event-driven local-runtime capability rather than time-based polling.
  Claude Code already provides a more natural experience for this workflow by retaining ownership of long-running terminal processes and continuing when they complete.

Start by tracing how Codex Desktop launches local commands and how the current scheduled PID and output-artifact polling works. The work is done when process completion wakes the associated thread exactly once across Windows, macOS, and Linux, preserves restart or sleep recovery, and reports the requested completion metadata and outcome states.