fix(gateway): /stop and /new bypass Level 1 active-session guard (#5765)

* fix(gateway): /stop and /new bypass Level 1 active-session guard

The base adapter's Level 1 guard intercepted ALL messages while an
agent was running, including /stop and /new. These commands were queued
as pending messages instead of being dispatched to the gateway runner's
Level 2 handler. When the agent eventually stopped (via the interrupt
mechanism), the command text leaked into the conversation as a user
message — the model would receive '/stop' as input and respond to it.

Fix: Add /stop, /new, and /reset to the bypass set in base.py alongside
/approve, /deny, and /status. Consolidate the three separate bypass
blocks into one. Commands in the bypass set are dispatched inline to the
gateway runner, where Level 2 handles them correctly (hard-kill for
/stop, session reset for /new).

Also add a safety net in _run_agent's pending-message processing: if the
pending text resolves to a known slash command, discard it instead of
passing it to the agent. This catches edge cases where command text
leaks through the interrupt_message fallback.

Refs: #5244

* test: regression tests for command bypass of active-session guard

17 tests covering:
- /stop, /new, /reset bypass the Level 1 guard when agent is running
- /approve, /deny, /status bypass (existing behavior, now tested)
- Regular text and unknown commands still queued (not bypassed)
- File paths like '/path/to/file' not treated as commands
- Telegram @botname suffix handled correctly
- Safety net command resolution (resolve_command detects known commands)

gateway/run.py

@@ -7131,6 +7131,27 @@ class GatewayRunner:
                     if pending:
                         logger.debug("Processing queued message after agent completion: '%s...'", pending[:40])
             
+            # Safety net: if the pending text is a slash command (e.g. "/stop",
+            # "/new"), discard it — commands should never be passed to the agent
+            # as user input.  The primary fix is in base.py (commands bypass the
+            # active-session guard), but this catches edge cases where command
+            # text leaks through the interrupt_message fallback.
+            if pending and pending.strip().startswith("/"):
+                _pending_parts = pending.strip().split(None, 1)
+                _pending_cmd_word = _pending_parts[0][1:].lower() if _pending_parts else ""
+                if _pending_cmd_word:
+                    try:
+                        from hermes_cli.commands import resolve_command as _rc_pending
+                        if _rc_pending(_pending_cmd_word):
+                            logger.info(
+                                "Discarding command '/%s' from pending queue — "
+                                "commands must not be passed as agent input",
+                                _pending_cmd_word,
+                            )
+                            pending = None
+                    except Exception:
+                        pass
+
             if pending:
                 logger.debug("Processing pending message: '%s...'", pending[:40])