Bug: Terminal does not accept keyboard input after agent process exits

[jidohyun]

Bug Description

After an agent (e.g. Codex) finishes its session and exits back to the shell prompt, the terminal becomes unresponsive to keyboard input. The shell prompt is visible and the cursor is blinking, but no keystrokes are delivered to the terminal. This occurs in the multi-agent task view.

Steps to Reproduce

1. Create a multi-agent task with Codex (or any agent)
2. Let the agent run and complete its work, or exit the agent session (e.g. via /exit or the agent hitting a rate limit)
3. The agent process exits and the shell prompt appears in the terminal
4. Try to type any command in the terminal

Actual Behavior

The terminal does not accept any keyboard input. The shell prompt is displayed correctly but keystrokes are completely ignored. The only workaround is to restart the task or the app.

Expected Behavior

After the agent process exits, the terminal should return to a normal interactive shell where the user can type commands freely.

Possible Cause

The PTY session may not be re-enabling stdin after the agent process exits, or the xterm.js terminal instance may still have disableStdin set. It could also be a focus management issue where the terminal fails to reclaim keyboard focus after the agent exit event.

Related issues (similar symptoms but different triggers):

• GEN-432 — Terminal unresponsive on initial startup
• GEN-571 — Text input fields randomly stop accepting input

Environment

• emdash version: latest (built from main @ d8cb465)
• OS: macOS
• Agent: Codex (observed in multi-agent task view)

Screenshot

The screenshot below shows the terminal after the Codex agent exited. The shell prompt is visible at the bottom but keyboard input is not accepted:

Shell prompt is displayed (~/De/B/w/ai-image-local-codex-6m0 on emdash/ai-image-local-codex-6m0) but no keystrokes register.

[jidohyun]

Debugging Results

After extensive investigation, here are the findings:

What works

• xterm.js receives keyboard input — terminal.onData() fires correctly with each keystroke
• Renderer sends input to main process — ptyInput({ id, data }) IPC is called
• PTY record exists in main process — writePty() finds the record and calls proc.write(data)

What doesn't work

• The shell does not echo input back — despite proc.write(data) being called on the PTY, no output is returned to the terminal

Root cause analysis

When a direct-spawned agent (Codex) exits, the following sequence occurs:

1. spawnDirectAgent registers proc.onExit() in ptyManager.ts (line ~1237) which calls onDirectCliExitCallback(id, cwd) — an async function
2. pty:startDirect handler registers its own proc.onExit() in ptyIpc.ts (line ~1348)
3. When the agent exits, both onExit handlers fire sequentially (same EventEmitter)
4. The ptyManager handler starts the async shell respawn but doesn't await
5. The ptyIpc handler runs synchronously and calls removePtyRecord(id) (line ~1371) — deleting the PTY record
6. The async shell respawn completes and calls startPty({id, cwd}) which creates a new PTY and stores it via ptys.set(id, ...)
7. A new pty:started event is sent to the renderer

However, the ptyIpc handler's proc.onExit (step 5) also has a guard at line ~1359:

const current = getPty(id);
if (current && current !== proc) { return; }

The timing issue: if startPty completes before the ptyIpc onExit runs the guard, the guard will skip cleanup. But if the ptyIpc onExit runs first (before startPty resolves), it will removePtyRecord(id) and then startPty re-creates it.

The key finding: Even when the PTY record IS correctly re-created (confirmed by writePty not logging "no PTY record found"), proc.write(data) on the new shell PTY does not produce output. This suggests the new shell process's stdin/stdout pipe may not be properly connected, or the PTY process itself is in a bad state.

Event ordering (confirmed via logging)

In the renderer, events arrive in this order:

1. pty:started (from shell respawn) — arrives FIRST
2. pty:exit (from ptyIpc onExit handler) — arrives SECOND

This reversed ordering itself is a secondary issue, but the primary problem is that proc.write() on the respawned shell has no effect.

Recommendation

The issue appears to be in the main process PTY lifecycle management (ptyManager.ts / ptyIpc.ts), specifically in how startPty interacts with the cleanup from the previous PTY's exit handler. The removePtyRecord(id) + subsequent ptys.set(id, newProc) race may leave the PTY in an inconsistent state where the node-pty process handle is valid but the underlying pseudoterminal pipe is broken.

Files involved

• src/main/services/ptyManager.ts — spawnDirectAgent(), startPty(), writePty()
• src/main/services/ptyIpc.ts — setOnDirectCliExit(), pty:startDirect handler
• src/renderer/terminal/TerminalSessionManager.ts — setupPtyDataListener()

[andrassmuk]

Hey, I took a look at this and put together a fix.

The root cause is a race condition between two onExit handlers when a direct-spawned agent exits. The ptyManager handler
starts an async shell respawn (startPty), but the ptyIpc handler runs synchronously right after it — deleting the PTY record
and sending pty:exit to the renderer before the new shell is ready. The renderer sees the exit event and stops listening, so
the respawned shell becomes unresponsive to keyboard input.

The fix skips cleanup in the ptyIpc handler for direct spawns and lets the respawn callback manage the full lifecycle
instead. Also added proper cleanup for cases where the respawn itself fails.

Branch with the fix: https://github.com/andrassmuk/emdash/tree/fix/1519-terminal-input-after-agent-exit

If this looks good, I can open a PR.

[jidohyun]

Hey, I took a look at this and put together a fix.

The root cause is a race condition between two onExit handlers when a direct-spawned agent exits. The ptyManager handler
starts an async shell respawn (startPty), but the ptyIpc handler runs synchronously right after it — deleting the PTY record
and sending pty:exit to the renderer before the new shell is ready. The renderer sees the exit event and stops listening, so
the respawned shell becomes unresponsive to keyboard input.

The fix skips cleanup in the ptyIpc handler for direct spawns and lets the respawn callback manage the full lifecycle
instead. Also added proper cleanup for cases where the respawn itself fails.

Branch with the fix: https://github.com/andrassmuk/emdash/tree/fix/1519-terminal-input-after-agent-exit

If this looks good, I can open a PR.

Hi, @andrassmuk
That's looks great!