503 Service temporarily unavailable causes Command Code to stop after /compact and prevents queued message from continuing

[BossCandy123]

Summary

Command Code becomes stuck after a 503 Service temporarily unavailable error and I cannot reliably continue the queued task/message.

This occurred while working in an active coding session. I attempted to use /compact-mode, received the message that Command Code was still working, interrupted the current turn, and then gave the next instruction.

Command Code started processing the request and loading skills/tastes, but the service then returned repeated 503 errors. After the error, my following queued instruction was not processed normally.

This interrupts long-running development work and makes it difficult to resume the session without losing context.

Expected Behavior

Command Code should recover gracefully from a temporary 503 service error without stopping or losing the current task.

If a secondary service such as taste learning fails, the main coding agent should continue working instead of terminating the entire turn.

Expected recovery behavior:

1. Preserve the current conversation and task context.
2. Preserve any queued user message or instruction.
3. Allow "continue" to retry the failed operation without losing progress.
4. Retry only the failed service when possible instead of restarting the entire agent turn.
5. If taste learning is temporarily unavailable, skip or defer it and continue the main coding task.
6. After interrupting a turn to use /compact-mode, the session should remain stable and accept the next instruction normally.
7. A temporary backend failure should not require restarting a long development session or re-entering previous instructions.

Actual Behavior

Command Code returned:

"taste learning failed: Service temporarily unavailable. Please try again shortly."

followed by:

"Error: 503 Service temporarily unavailable. Please try again shortly."

It then instructed me to type "continue".

The error occurred more than once with different Trace IDs, and the next instruction I entered was not processed normally.

The session had already completed significant work, so restarting the conversation is undesirable because it risks losing working context.

Steps to reproduce the issue

1. Start a Command Code session in an existing repository with a long-running coding task.
2. Allow Command Code to perform several tool operations and accumulate conversation context.
3. Enter /compact-mode while Command Code is still working.
4. Command Code reports:
   "Cannot run /compact-mode while Command Code is working. Press Esc or Ctrl+C to interrupt the current turn first."
5. Interrupt the current turn.
6. Give another development instruction.
7. Command Code begins processing the instruction and loads skills/taste information.
8. A taste-learning request fails with HTTP 503.
9. Command Code displays:
   "Service temporarily unavailable. Please try again shortly."
10. Send the next instruction or attempt to continue.
11. Command Code fails again or does not reliably process the queued instruction.

Command Code Version

1.4.5

Operating System

Windows

Terminal/IDE

Unknown

Shell

cmd.exe

Additional context

Environment:

• Windows 11
• Command Code version: 1.4.5
• Shell: powershe7.exe
• Repository: existing Git project with a long-running development session
• The session contained substantial context, tool usage, tests, and checkpoint work.

The failure appears to have occurred during "taste learning". Ideally, failure of that optional/secondary operation should not terminate or block the main coding task.

Errors observed:

Error 1:
taste learning failed: Service temporarily unavailable. Please try again shortly.
Error: 503 Service temporarily unavailable. Please try again shortly.
Trace ID: f9605718a804fe5f7b748a5461837565

Error 2:
taste learning failed: Service temporarily unavailable. Please try again shortly.
Error: 503 Service temporarily unavailable. Please try again shortly.
Trace ID: 031746f653df77f62dd54cfa3ff0f683

The UI instructed me to type "continue" to retry.

A useful recovery behavior would be:

• preserve the failed/queued user instruction,
• retry only the failed backend request,
• allow the coding agent to continue if taste learning is unavailable,
• and avoid requiring the user to recreate or restart a long development session.

I can provide additional logs/screenshots if needed.