diff --git a/hermes_cli/models.py b/hermes_cli/models.py
index fc70420a09765..7f57b39b3549b 100644
--- a/hermes_cli/models.py
+++ b/hermes_cli/models.py
@@ -1714,7 +1714,8 @@ def _fetch_anthropic_models(
         return None
 
     resolved_base_url = base_url
-    token = (api_key or "").strip() or resolve_anthropic_token()
+    from agent.command_token_source import materialize_probe_api_key
+    token = materialize_probe_api_key(api_key) if api_key else resolve_anthropic_token()
     if not token:
         # A pool credential and its endpoint are one security boundary — never pair the pool key
         # with a caller-provided endpoint.
@@ -2236,6 +2237,8 @@ def probe_api_models(
     """Probe a ``/models`` endpoint with light URL heuristics (``base`` then ``base±/v1``).
     ``anthropic_messages`` mode sends ``x-api-key`` + ``anthropic-version`` instead of a bearer; the
     ``data[].id`` response shape is identical. ``models`` is None when no candidate answered."""
+    from agent.command_token_source import materialize_probe_api_key
+    api_key = materialize_probe_api_key(api_key)
     normalized = (base_url or "").strip().rstrip("/")
     if not normalized:
         return _probe_result(None, None, "")
diff --git a/hermes_cli/runtime_provider_backends.py b/hermes_cli/runtime_provider_backends.py
index 2ce619a3ab21a..9508ee0c6b28c 100644
--- a/hermes_cli/runtime_provider_backends.py
+++ b/hermes_cli/runtime_provider_backends.py
@@ -68,7 +68,8 @@ def _resolve_azure_foundry_runtime(*, requested_provider: str, model_cfg: Dict[s
     ``.env``/env or a per-request Entra ID token, trailing ``/v1`` stripped for Anthropic-style
     endpoints (the Anthropic SDK appends /v1/messages itself)."""
     rp = _rp()
-    explicit_api_key = str(explicit_api_key or "").strip()
+    explicit_key_is_callable = callable(explicit_api_key) and not isinstance(explicit_api_key, str)
+    explicit_api_key = explicit_api_key if explicit_key_is_callable else str(explicit_api_key or "").strip()
     explicit_base_url_clean = str(explicit_base_url or "").strip().rstrip("/")
     cfg_base_url, cfg_api_mode, cfg_auth_mode, cfg_entra = "", "chat_completions", "api_key", {}
     if rp._cfg_provider(model_cfg) == "azure-foundry":
@@ -90,12 +91,16 @@ def _resolve_azure_foundry_runtime(*, requested_provider: str, model_cfg: Dict[s
     if cfg_api_mode == "anthropic_messages":
         base_url = re.sub(r"/v1/?$", "", base_url)
     if cfg_auth_mode == "entra_id":
+        scope = str(cfg_entra.get("scope") or "").strip()
         # --api-key on the CLI while config says entra_id: honour the explicit string (escape hatch
         # for one-off testing).
-        if explicit_api_key:
+        if explicit_key_is_callable:
+            api_key, source, auth_mode, entra = explicit_api_key, "entra_id", "entra_id", (
+                {"scope": scope} if scope else {}
+            )
+        elif explicit_api_key:
             api_key, source, auth_mode, entra = explicit_api_key, "explicit", "api_key", {}
         else:
-            scope = str(cfg_entra.get("scope") or "").strip()
             api_key, source, auth_mode, entra = _azure_entra_credentials(cfg_entra), "entra_id", "entra_id", (
                 {"scope": scope} if scope else {}
             )
diff --git a/hermes_cli/runtime_provider_custom.py b/hermes_cli/runtime_provider_custom.py
index b42d6537a4aed..06001346e1680 100644
--- a/hermes_cli/runtime_provider_custom.py
+++ b/hermes_cli/runtime_provider_custom.py
@@ -478,19 +478,22 @@ def _resolve_named_custom_runtime(*, requested_provider: str, explicit_api_key:
         # The pool doesn't know the custom_providers fields — propagate them here too.
         _apply_custom_provider_extras(custom_provider, target_model, pool_result)
         return pool_result
-    explicit_key = (explicit_api_key or "").strip()
+    explicit_key_is_callable = callable(explicit_api_key) and not isinstance(explicit_api_key, str)
+    explicit_key = explicit_api_key if explicit_key_is_callable else str(explicit_api_key or "").strip()
     candidates = [
         explicit_key,
         _clean(custom_provider.get("api_key", "")),
         rp._getenv(_clean(custom_provider.get("key_env", "")), "").strip(),
         *rp._host_gated_env_key_candidates(base_url, ollama=False),
     ]
-    api_key: Any = next((c for c in candidates if rp.has_usable_secret(c)), "")
+    api_key: Any = explicit_key if explicit_key_is_callable else next(
+        (c for c in candidates if rp.has_usable_secret(c)), ""
+    )
     # ``key_cmd`` credentials are minted per request (short-lived bearers would go stale
     # mid-session); both wire clients accept a callable api_key (the Entra ID contract). An
     # explicit --api-key still wins as the one-off recovery escape hatch.
     key_cmd = _clean(custom_provider.get("key_cmd", ""))
-    if key_cmd and not rp.has_usable_secret(explicit_key):
+    if key_cmd and not explicit_key_is_callable and not rp.has_usable_secret(explicit_key):
         from agent.command_token_source import build_command_token_provider
         token_provider = build_command_token_provider(key_cmd, str(custom_provider.get("name", requested_provider) or "custom"))
         if token_provider is not None:
diff --git a/tests/agent/test_command_token_source.py b/tests/agent/test_command_token_source.py
index e7210480ef4ca..ab7f4716be705 100644
--- a/tests/agent/test_command_token_source.py
+++ b/tests/agent/test_command_token_source.py
@@ -200,6 +200,33 @@ def test_explicit_api_key_still_wins(self, monkeypatch):
         )
         assert runtime["api_key"] == "sk-explicit-override"
 
+    def test_explicit_callable_survives_named_custom_reresolution(self, monkeypatch):
+        """A live key_cmd source passed back through the main resolver must
+        remain callable instead of raising or becoming an object repr."""
+        from hermes_cli import runtime_provider as rp
+
+        config = {
+            "providers": {
+                "dbx": {
+                    "base_url": "https://example.invalid/v1",
+                    "api_mode": "chat_completions",
+                    "model": "m1",
+                    "key_cmd": "printf minted-token",
+                }
+            }
+        }
+        monkeypatch.setattr(rp, "load_config", lambda *a, **k: config)
+        monkeypatch.setattr("hermes_cli.config.load_config", lambda *a, **k: config)
+
+        token_provider = CommandTokenSource("printf forwarded-token", "dbx")
+        runtime = rp.resolve_runtime_provider(
+            requested="custom:dbx",
+            explicit_api_key=token_provider,
+            target_model="m1",
+        )
+
+        assert runtime["api_key"] is token_provider
+
 
 class TestCallableKeyGetsBearerAuth:
     """A callable api_key must reach the Anthropic bearer-hook client path.
@@ -291,7 +318,7 @@ def test_the_token_actually_gets_re_minted(self, tmp_path):
         assert src._expires_at is not None, "cache must carry a deadline"
         src._expires_at = time.monotonic() - 1  # simulate crossing it
         src()
-        assert len(counter.read_text()) == 2, "expired cache must re-run the helper"
+        assert len(counter.read_text(encoding="utf-8")) == 2, "expired cache must re-run the helper"
 
 
 class TestAuxiliaryResolverHonoursKeyCmd:
diff --git a/tests/hermes_cli/test_azure_foundry_entra.py b/tests/hermes_cli/test_azure_foundry_entra.py
index 2205a66cbdda3..20786e8ba24a4 100644
--- a/tests/hermes_cli/test_azure_foundry_entra.py
+++ b/tests/hermes_cli/test_azure_foundry_entra.py
@@ -143,6 +143,30 @@ def test_entra_with_explicit_api_key_uses_string_escape_hatch(self, fake_azure_i
         assert runtime["auth_mode"] == "api_key"
         assert runtime["source"] == "explicit"
 
+    def test_forwarded_entra_callable_preserves_identity_and_metadata(self, fake_azure_identity):
+        """A live provider re-resolution must not stringify its token source or
+        relabel Entra authentication as a static-key override."""
+        from hermes_cli.runtime_provider import _resolve_azure_foundry_runtime
+
+        token_provider = lambda: "forwarded-jwt"
+        runtime = _resolve_azure_foundry_runtime(
+            requested_provider="azure-foundry",
+            model_cfg={
+                "provider": "azure-foundry",
+                "base_url": "https://r.services.ai.azure.com/openai/v1",
+                "api_mode": "chat_completions",
+                "auth_mode": "entra_id",
+                "entra": {"scope": "https://ai.azure.com/.default"},
+                "default": "gpt-4o",
+            },
+            explicit_api_key=token_provider,
+        )
+
+        assert runtime["api_key"] is token_provider
+        assert runtime["auth_mode"] == "entra_id"
+        assert runtime["source"] == "entra_id"
+        assert runtime["entra"] == {"scope": "https://ai.azure.com/.default"}
+
 
 # ---------------------------------------------------------------------------
 # _resolve_azure_foundry_runtime: legacy api_key branch (regression)
diff --git a/tests/hermes_cli/test_model_validation.py b/tests/hermes_cli/test_model_validation.py
index ec3e26cdfede2..aad868b93f0b0 100644
--- a/tests/hermes_cli/test_model_validation.py
+++ b/tests/hermes_cli/test_model_validation.py
@@ -587,6 +587,22 @@ def test_probe_user_agent_sent_without_api_key(self):
         # No Authorization was set, but UA must still be present.
         assert req.get_header("Authorization") is None
 
+    def test_probe_materializes_callable_credentials(self):
+        """Capability probes must mint a callable before constructing auth headers."""
+        from unittest.mock import patch
+
+        body = b'{"data":[{"id":"gpt-4o"}]}'
+        token_provider = lambda: "minted-probe-token"
+        with patch(
+            "hermes_cli.models._urlopen_model_catalog_request",
+            return_value=self._make_mock_response(body),
+        ) as mock_urlopen:
+            result = probe_api_models(token_provider, "https://example.com/v1")
+
+        assert result["models"] == ["gpt-4o"]
+        req = mock_urlopen.call_args[0][0]
+        assert req.get_header("Authorization") == "Bearer minted-probe-token"
+
 
 
 
diff --git a/tests/tui_gateway/test_make_agent_provider.py b/tests/tui_gateway/test_make_agent_provider.py
index ea760e8671a0e..36f77226b469c 100644
--- a/tests/tui_gateway/test_make_agent_provider.py
+++ b/tests/tui_gateway/test_make_agent_provider.py
@@ -135,6 +135,7 @@ def switch_model(self, **kw):
     # Target session recorded a per-session override.
     assert sess_b["model_override"]["model"] == "zai/glm-5.1"
     assert sess_b["model_override"]["provider"] == "zai"
+    assert "api_key" not in sess_b["model_override"]
     # The switched agent mutated in place.
     assert sess_b["agent"].model == "zai/glm-5.1"
     # Sibling session is completely untouched.
diff --git a/tui_gateway/model_switch.py b/tui_gateway/model_switch.py
index 770732b6514f4..57bb32c4e4ae4 100644
--- a/tui_gateway/model_switch.py
+++ b/tui_gateway/model_switch.py
@@ -238,7 +238,7 @@ def _apply_model_switch(
     if pin_session_override and isinstance(session, dict) and not one_turn:
         session["model_override"] = {
             "model": result.new_model, "provider": result.target_provider,
-            "base_url": result.base_url, "api_key": result.api_key, "api_mode": result.api_mode}
+            "base_url": result.base_url, "api_mode": result.api_mode}
     if persist_global:
         _persist_model_switch(result)
     return {