fix(cli): release the old agent's clients before dropping it for a rebuild

Interactive CLI sites that set self.agent = None (/personality, /reasoning, /fast,
model/route/credential change, first-run setup, MoA one-shot) now go through
_retire_agent(), which calls agent.release_clients() first: on the codex_app_server
route the app-server child belongs to the instance and was orphaned for the CLI
process lifetime (#72548).

hermes_cli/cli_agent_setup_mixin.py

@@ -212,6 +212,16 @@ def _resume_panel_colors() -> tuple:
         return tuple(default for _, default in _RESUME_SKIN_COLORS)
 
 
+def _retire_agent(cli) -> None:
+    """Drop ``cli.agent`` so the next turn rebuilds it, releasing its LLM clients first: the Codex
+    app-server child (and MCP descendants) belongs to the instance, so ``self.agent = None`` alone
+    orphans it for the CLI process lifetime (#72548). Session tool state is kept (soft release)."""
+    agent = cli.agent
+    if agent is not None and hasattr(agent, "release_clients"):
+        agent.release_clients()
+    cli.agent = None
+
+
 class CLIAgentSetupMixin:
     """Agent construction + session-resume display methods for ``HermesCLI``."""
 
@@ -324,7 +334,7 @@ class CLIAgentSetupMixin:
 
         # AIAgent/OpenAI client holds auth at init, so rebuild on key/routing/model change.
         if (credentials_changed or routing_changed or model_changed) and self.agent is not None:
-            self.agent = None
+            _retire_agent(self)
             self._active_agent_route_signature = None
         return True
 
@@ -492,7 +502,7 @@ class CLIAgentSetupMixin:
         except Exception as exc:
             logger.debug("first-run config re-sync failed: %s", exc)
         # Force credential re-resolution + agent rebuild on next use.
-        self.agent = None
+        _retire_agent(self)
         self._active_agent_route_signature = None
         if self._runtime_credentials_ready():
             _cprint("  ✓ Provider configured — you're ready to chat.")

hermes_cli/cli_chat_turn_mixin.py

@@ -18,6 +18,8 @@ from rich import box as rich_box
 from rich.panel import Panel
 from typing import Optional
 
+from hermes_cli.cli_agent_setup_mixin import _retire_agent
+
 
 class CLIChatTurnMixin:
     """chat() and its per-turn phase helpers."""
@@ -51,7 +53,7 @@ class CLIChatTurnMixin:
 
         turn_route = self._resolve_turn_agent_config(message)
         if turn_route["signature"] != self._active_agent_route_signature:
-            self.agent = None
+            _retire_agent(self)
         if self.agent is None:
             _cprint(f"{_DIM}Initializing agent...{_RST}")
         if not self._init_agent(model_override=turn_route["model"], runtime_override=turn_route["runtime"],
@@ -341,7 +343,7 @@ class CLIChatTurnMixin:
                 for _key, _value in _restore.items():
                     if _value is not None:
                         setattr(self, _key, _value)
-                self.agent = None
+                _retire_agent(self)
                 self._pending_moa_restore_model = None
                 self._pending_moa_disable_after_turn = False
         except Exception as exc:

hermes_cli/cli_commands_mixin.py

@@ -31,6 +31,7 @@ from rich.panel import Panel
 from hermes_constants import display_hermes_home, is_termux as _is_termux_environment
 from hermes_state_ids import new_session_id as mint_session_id
 from agent.turn_context import extract_api_content_sidecar
+from hermes_cli.cli_agent_setup_mixin import _retire_agent
 from hermes_cli.browser_connect import (
     DEFAULT_BROWSER_CDP_URL, discover_local_cdp_url, find_free_debug_port, is_browser_debug_ready,
     launch_chrome_debug, local_port_in_use, manual_chrome_debug_command)
@@ -1555,12 +1556,12 @@ class CLICommandsMixin:
                     cfg_get(read_raw_config(), "agent", "system_prompt", default=""))
             except Exception:
                 self.system_prompt = ""
-            self.agent = None  # Force re-init
+            _retire_agent(self)  # Force re-init
             _pr(f"{face} Personality cleared {scope}",
                 "  No personality overlay — using base agent behavior.")
         else:
             self.system_prompt = personality_prompt
-            self.agent = None  # Force re-init
+            _retire_agent(self)  # Force re-init
             _pr(f"{face} Personality set to '{name}' {scope}",
                 f"  \"{_ellipsize(personality_prompt, 60)}\"")
 
@@ -2595,7 +2596,7 @@ class CLICommandsMixin:
                        _dim_line('Display:      show, hide'),
                        _dim_line('Scope:        session-scoped by default, --global to persist'))
         self.reasoning_config = parsed
-        self.agent = None  # Force agent re-init with new reasoning config
+        _retire_agent(self)  # Force agent re-init with new reasoning config
         saved = explicit_global and _save("agent.reasoning_effort", arg)
         if saved:
             if not isinstance(CLI_CONFIG.get("agent"), dict):
@@ -2652,7 +2653,7 @@ class CLICommandsMixin:
         if arg not in _FAST_TIERS:
             return _cp(_dim_line(f'(._.) Unknown argument: {arg}'), usage)
         self.service_tier, saved_value = _FAST_TIERS[arg]
-        self.agent = None  # Force agent re-init with new service-tier config
+        _retire_agent(self)  # Force agent re-init with new service-tier config
         saved = explicit_global and _save("agent.service_tier", saved_value)
         outcome = _scope_outcome(explicit_global, saved)
         _cp(_accent_line(f"✓ {feature_name} set to {saved_value.upper()} {outcome}"))

hermes_cli/cli_model_switch_mixin.py

@@ -16,6 +16,7 @@ import threading
 
 from rich.markup import escape as _escape
 from utils import base_url_host_matches
+from hermes_cli.cli_agent_setup_mixin import _retire_agent
 
 # CLI-level fields describing the active model route; snapshotted before a switch / one-turn
 # override and restored wholesale on rollback. ``reasoning_config`` rides along because it is
@@ -910,7 +911,7 @@ class CLIModelSwitchMixin:
         self.api_key = "moa-virtual-provider"
         self.base_url = "moa://local"
         self.api_mode = "chat_completions"
-        self.agent = None
+        _retire_agent(self)
         self._pending_moa_disable_after_turn = True
         self._pending_agent_seed = payload
         _cprint(f"  MoA one-shot queued with preset {preset}; previous model will be restored after this turn.")

tests/hermes_cli/test_cli_retire_agent.py

@@ -0,0 +1,27 @@
+"""The interactive CLI must release the old AIAgent's LLM clients when it drops the instance
+for a rebuild (/personality, /reasoning, /fast, model/route/credential change, MoA one-shot):
+on the codex_app_server route the app-server child belongs to that instance and ``self.agent =
+None`` alone orphans it for the CLI process lifetime (#72548)."""
+
+from types import SimpleNamespace
+from unittest.mock import patch
+
+from hermes_cli.cli_commands_mixin import CLICommandsMixin
+
+
+class _FakeAgent:
+    def __init__(self):
+        self.release_calls = 0
+
+    def release_clients(self):
+        self.release_calls += 1
+
+
+def test_reasoning_command_releases_old_agent_clients_before_rebuild():
+    agent = _FakeAgent()
+    stub = SimpleNamespace(reasoning_config={"enabled": True, "effort": "medium"}, show_reasoning=False, agent=agent)
+    with patch("cli.save_config_value"), patch("cli._cprint"):
+        CLICommandsMixin._handle_reasoning_command(stub, "/reasoning high")
+    assert stub.reasoning_config == {"enabled": True, "effort": "high"}
+    assert stub.agent is None
+    assert agent.release_calls == 1