fix: /queue now supports multiple queued messages via FIFO

[willy-scr]

Bug

/queue only stores one message per session. Sending /queue prompt A then /queue prompt B silently overwrites A — only B is ever processed.

Root Cause

_pending_messages is typed as Dict[str, MessageEvent] — a single value per session key. Every write (via /queue, interrupt follow-ups, or sentinel buffering) overwrites the previous entry:

adapter._pending_messages[_quick_key] = queued_event  # overwrites!

Fix

Change the data structure to a FIFO queue:

# Before
self._pending_messages: Dict[str, MessageEvent] = {}

# After  
self._pending_messages: Dict[str, list[MessageEvent]] = {}

Changes

gateway/platforms/base.py:

• _pending_messages type: Dict[str, MessageEvent] → Dict[str, list[MessageEvent]]
• merge_pending_message_event(): appends to list (photo bursts merge into last entry)
• get_pending_message(): pops first item (FIFO)
• Drain loop in _process_message_background: pops first from queue, processes it, then checks if more remain

gateway/run.py:

• /queue handler: adapter._pending_messages[key] = event → .setdefault(key, []).append(event)
• /stop and /new: drain entire queue instead of popping one
• Sentinel buffering: same append pattern

Testing

1. Send a long-running prompt to the agent
2. While processing, send /queue task A
3. Then /queue task B
4. Then /queue task C
5. Before fix: Only task C is processed (A and B silently lost)
6. After fix: task A → task B → task C processed in order, one at a time

Backward Compatibility

Photo burst merging still works — consecutive PHOTO events merge into the last queue entry instead of a single slot. No behavioral change for single-message queues.

[teknium1]

Thanks for the thorough write-up and the clean fix, @willy-scr! This automated hermes-sweeper review found that the same bug was independently addressed on main before this PR could be merged.

• Commit 1dfcc2ffc — fix(gateway): /queue is now a true FIFO — each invocation gets its own turn (fix(gateway): /queue is now a true FIFO — each invocation gets its own turn #16175) — landed the FIFO fix via _enqueue_fifo / _promote_queued_event / _queued_events in gateway/run.py.
• The /queue handler (line 3533) now calls self._enqueue_fifo(...) instead of the bare dict assignment you identified.
• _queue_depth() (line 1280) tracks slot + overflow depth and is surfaced to the user on subsequent /queue calls.

Main's implementation kept _pending_messages: Dict[str, MessageEvent] in base.py unchanged and added an overflow list at the GatewayServer level, rather than changing the base type — a slightly different approach, but the observable bug (silent overwrite of earlier queued messages) is fixed. Closing as implemented on main.