@cratexnet

Summary

After updating to Workmux 0.1.211, Codex status tracking can leave stale active run IDs in the Workmux Codex runtime state. When this happens, the sidebar/pane remains stuck on working even though the visible Codex work has already completed or been interrupted.

This looks like a remaining edge case in the active-run tracking added for nested Codex agents, related to #154.

Environment

• Workmux: 0.1.211
• Codex CLI: 0.133.0
• Multiplexer: tmux
• Workmux Codex hooks installed from the bundled template

Observed behavior

A Workmux sidebar entry stayed in working indefinitely.

Inspecting the Workmux runtime state showed a non-empty active set under the Codex status runtime directory for that tmux pane. The corresponding Codex transcript metadata showed those turn IDs were no longer active:

• some had ended with turn_aborted / interrupted
• some were thread-spawned subagent turns that ended with task_complete

There were no visible active Codex turns corresponding to those stale active IDs, but Workmux continued to derive working from the non-empty active set.

Why this may leak active IDs

From Workmux 0.1.211 source:

• src/state/codex_status.rs tracks active Codex runs by session_id:turn_id.
• It accepts these events for run-id detection: UserPromptSubmit, PreToolUse, PermissionRequest, PostToolUse, and Stop.
• The bundled Codex hook template registers UserPromptSubmit, PostToolUse, and Stop.

From Codex 0.133.0 source/behavior:

• thread-spawned child agents complete with SubagentStop, not normal Stop
• SubagentStop payloads include agent_id / agent_type
• interrupted turns can end as turn_aborted without running the normal stop hook path

This means Workmux can add a run ID on UserPromptSubmit/PostToolUse, but miss the corresponding cleanup when the turn ends through SubagentStop or turn_aborted. Also, adding a SubagentStop hook alone may not be sufficient in the current implementation because codex_status.rs rejects payloads containing agent_id or agent_type.

Expected behavior

Any run ID added to the pane active set should eventually be removed when that Codex turn completes, the subagent completes, or the turn is interrupted/aborted. The pane/sidebar should not remain working when all tracked Codex turns are done.

Possible fix direction

A few possible approaches:

• handle SubagentStop as a cleanup event for session_id:turn_id
• include a bundled SubagentStop Workmux hook if needed
• allow SubagentStop payloads with agent_id / agent_type for cleanup
• add cleanup/reconciliation for turn_aborted or stale active IDs

I can provide more anonymized runtime-state details if helpful.