fix: preserve credential pool through /model session override

cli.py

@@ -5475,12 +5475,15 @@ def _apply_model_switch_result(self, result, persist_global: bool) -> None:
 
         if self.agent is not None:
             try:
+                if result.credential_pool is not None:
+                    self._credential_pool = result.credential_pool
                 self.agent.switch_model(
                     new_model=result.new_model,
                     new_provider=result.target_provider,
                     api_key=result.api_key,
                     base_url=result.base_url,
                     api_mode=result.api_mode,
+                    credential_pool=result.credential_pool,
                 )
             except Exception as exc:
                 _cprint(f"  ⚠ Agent swap failed ({exc}); change applied to next session.")
@@ -5699,12 +5702,15 @@ def _handle_model_switch(self, cmd_original: str):
         # Apply to running agent (in-place swap)
         if self.agent is not None:
             try:
+                if result.credential_pool is not None:
+                    self._credential_pool = result.credential_pool
                 self.agent.switch_model(
                     new_model=result.new_model,
                     new_provider=result.target_provider,
                     api_key=result.api_key,
                     base_url=result.base_url,
                     api_mode=result.api_mode,
+                    credential_pool=result.credential_pool,
                 )
             except Exception as exc:
                 _cprint(f"  ⚠ Agent swap failed ({exc}); change applied to next session.")

gateway/run.py

@@ -959,7 +959,8 @@ class GatewayRunner:
     _restart_detached: bool = False
     _restart_via_service: bool = False
     _stop_task: Optional[asyncio.Task] = None
-    _session_model_overrides: Dict[str, Dict[str, str]] = {}
+    # Values are mostly str but credential_pool is a CredentialPool instance.
+    _session_model_overrides: Dict[str, Dict[str, Any]] = {}
     _session_reasoning_overrides: Dict[str, Dict[str, Any]] = {}
     # Stale-code self-check defaults (see _detect_stale_code()).  Class-level
     # so tests that construct GatewayRunner via ``object.__new__`` without
@@ -1058,7 +1059,7 @@ def __init__(self, config: Optional[GatewayConfig] = None):
 
         # Per-session model overrides from /model command.
         # Key: session_key, Value: dict with model/provider/api_key/base_url/api_mode
-        self._session_model_overrides: Dict[str, Dict[str, str]] = {}
+        self._session_model_overrides: Dict[str, Dict[str, Any]] = {}
         # Per-session reasoning effort overrides from /reasoning.
         # Key: session_key, Value: parsed reasoning config dict.
         self._session_reasoning_overrides: Dict[str, Dict[str, Any]] = {}
@@ -1441,7 +1442,23 @@ def _resolve_session_agent_runtime(
                 "base_url": override.get("base_url"),
                 "api_mode": override.get("api_mode"),
             }
+            # Carry credential_pool from override if stored (new behavior)
+            if override.get("credential_pool") is not None:
+                override_runtime["credential_pool"] = override["credential_pool"]
             if override_runtime.get("api_key"):
+                # Credential pool fix: /model stores api_key but may lack
+                # credential_pool (ModelSwitchResult has no such field).
+                # When the provider is a pool name (e.g. "custom:mimo-sgp-friend"),
+                # resolve the pool here so the agent can rotate on 429/402.
+                _prov = override.get("provider", "")
+                if _prov.startswith("custom:") and "credential_pool" not in override_runtime:
+                    try:
+                        from agent.credential_pool import load_pool
+                        _pool = load_pool(_prov)
+                        if _pool and _pool.has_credentials():
+                            override_runtime["credential_pool"] = _pool
+                    except Exception:
+                        pass
                 logger.debug(
                     "Session model override (fast): session=%s config_model=%s -> override_model=%s provider=%s",
                     resolved_session_key or "", model, override_model,
@@ -7532,6 +7549,7 @@ async def _on_model_selected(
                     api_key=result.api_key,
                     base_url=result.base_url,
                     api_mode=result.api_mode,
+                    credential_pool=result.credential_pool,
                 )
             except Exception as exc:
                 logger.warning("In-place model switch failed for cached agent: %s", exc)
@@ -7553,6 +7571,7 @@ async def _on_model_selected(
             "api_key": result.api_key,
             "base_url": result.base_url,
             "api_mode": result.api_mode,
+            "credential_pool": result.credential_pool,
         }
 
         # Evict cached agent so the next turn creates a fresh agent from the

hermes_cli/model_switch.py

@@ -23,7 +23,10 @@
 import logging
 import re
 from dataclasses import dataclass
-from typing import List, NamedTuple, Optional
+from typing import TYPE_CHECKING, List, NamedTuple, Optional
+
+if TYPE_CHECKING:
+    from agent.credential_pool import CredentialPool
 
 from hermes_cli.providers import (
     custom_provider_slug,
@@ -246,6 +249,10 @@ class ModelSwitchResult:
     capabilities: Optional[ModelCapabilities] = None
     model_info: Optional[ModelInfo] = None
     is_global: bool = False
+    # CredentialPool for the resolved target provider, when one was loaded.
+    # Callers must propagate this into runtime kwargs / session overrides so
+    # per-provider 429 rotation continues to work after a /model switch.
+    credential_pool: Optional["CredentialPool"] = None
 
 
 @dataclass
@@ -818,6 +825,7 @@ def switch_model(
     api_key = current_api_key
     base_url = current_base_url
     api_mode = ""
+    _resolved_pool = None
 
     if provider_changed or explicit_provider:
         try:
@@ -828,6 +836,7 @@ def switch_model(
             api_key = runtime.get("api_key", "")
             base_url = runtime.get("base_url", "")
             api_mode = runtime.get("api_mode", "")
+            _resolved_pool = runtime.get("credential_pool")
         except Exception as e:
             return ModelSwitchResult(
                 success=False,
@@ -853,6 +862,7 @@ def switch_model(
                 api_key = runtime.get("api_key", "")
                 base_url = runtime.get("base_url", "")
                 api_mode = runtime.get("api_mode", "")
+                _resolved_pool = runtime.get("credential_pool")
         except Exception:
             pass
 
@@ -978,6 +988,7 @@ def switch_model(
         capabilities=capabilities,
         model_info=model_info,
         is_global=is_global,
+        credential_pool=_resolved_pool,
     )
 
 

run_agent.py

@@ -2278,7 +2278,7 @@ def _ensure_lmstudio_runtime_loaded(self, config_context_length: Optional[int] =
         except Exception as err:
             logger.debug("LM Studio preload skipped: %s", err)
 
-    def switch_model(self, new_model, new_provider, api_key='', base_url='', api_mode=''):
+    def switch_model(self, new_model, new_provider, api_key='', base_url='', api_mode='', credential_pool=None):
         """Switch the model/provider in-place for a live agent.
 
         Called by the /model command handlers (CLI and gateway) after
@@ -2325,6 +2325,15 @@ def switch_model(self, new_model, new_provider, api_key='', base_url='', api_mod
         if api_key:
             self.api_key = api_key
 
+        # ── Swap credential pool if provided ──
+        # The /model pipeline resolves pools via load_pool() but the result
+        # was historically lost in the ModelSwitchResult → session-override
+        # chain.  Callers that resolve the pool themselves can now pass it
+        # through so the agent can rotate on 429/402 after a mid-session
+        # model switch.
+        if credential_pool is not None:
+            self._credential_pool = credential_pool
+
         # ── Build new client ──
         if api_mode == "anthropic_messages":
             from agent.anthropic_adapter import (