diff --git a/agent/agent_init.py b/agent/agent_init.py
index 5b78192587..e482570b8d 100644
--- a/agent/agent_init.py
+++ b/agent/agent_init.py
@@ -2830,6 +2830,12 @@ def init_agent(
     agent.codex_app_server_auto_compaction = codex_app_server_auto_compaction
     agent.codex_responses_native_compaction = codex_responses_native_compaction
     agent.codex_responses_compact_threshold = codex_responses_compact_threshold
+    from agent.native_compaction import resolve_native_compaction_capabilities
+    agent.capabilities = resolve_native_compaction_capabilities(
+        model=agent.model,
+        base_url=agent.base_url,
+        is_codex_backend=(agent.provider or "").strip().lower() == "openai-codex",
+    )
     agent.max_compression_attempts = compression_max_attempts
     agent.compression_idle_compact_after_seconds = (
         compression_idle_compact_after_seconds
diff --git a/agent/agent_runtime_helpers.py b/agent/agent_runtime_helpers.py
index 74dc486da2..fabfe751e7 100644
--- a/agent/agent_runtime_helpers.py
+++ b/agent/agent_runtime_helpers.py
@@ -1748,6 +1748,8 @@ def restore_primary_runtime(agent) -> bool:
         if hasattr(agent, "_transport_cache"):
             agent._transport_cache.clear()
         agent.api_key = rt["api_key"]
+        if "capabilities" in rt:
+            agent.capabilities = dict(rt.get("capabilities") or {})
         agent._reasoning_echo_flag = rt.get("reasoning_echo_flag", False)
         agent.request_overrides = dict(rt.get("request_overrides") or {})
         agent._client_kwargs = dict(rt["client_kwargs"])
@@ -2895,7 +2897,15 @@ def _apply_switched_provider_request_overrides(agent, new_provider):
     agent.request_overrides = overrides
 
 
-def switch_model(agent, new_model, new_provider, api_key='', base_url='', api_mode=''):
+def switch_model(
+    agent,
+    new_model,
+    new_provider,
+    api_key='',
+    base_url='',
+    api_mode='',
+    capabilities=None,
+):
     """Switch the model/provider in-place for a live agent.
 
     Called by the /model command handlers (CLI and gateway) after
@@ -2910,6 +2920,7 @@ def switch_model(agent, new_model, new_provider, api_key='', base_url='', api_mo
     turn-scoped).
     """
     from hermes_cli.providers import determine_api_mode
+    from agent.native_compaction import resolve_native_compaction_capabilities
 
     # ── Determine api_mode if not provided ──
     # Pass model so dual-wire providers (Nous Portal anthropic/* → Messages)
@@ -2918,6 +2929,16 @@ def switch_model(agent, new_model, new_provider, api_key='', base_url='', api_mo
     if not api_mode:
         api_mode = determine_api_mode(new_provider, base_url, model=new_model)
 
+    destination_capabilities = (
+        dict(capabilities)
+        if isinstance(capabilities, dict)
+        else resolve_native_compaction_capabilities(
+            model=new_model,
+            base_url=base_url,
+            is_codex_backend=(new_provider or '').strip().lower() == 'openai-codex',
+        )
+    )
+
     # Defense-in-depth: ensure OpenCode base_url doesn't carry a trailing
     # /v1 into the anthropic_messages client, which would cause the SDK to
     # hit /v1/v1/messages.  `model_switch.switch_model()` already strips
@@ -2964,6 +2985,7 @@ def switch_model(agent, new_model, new_provider, api_key='', base_url='', api_mo
             "_is_anthropic_oauth",
             "_config_context_length",
             "_reasoning_echo_flag",
+            "capabilities",
         )
     }
     # _client_kwargs is a dict — snapshot a shallow copy so mutating the
@@ -3180,9 +3202,13 @@ def switch_model(agent, new_model, new_provider, api_key='', base_url='', api_mo
     except Exception:
         _destination_context_intent = None
     agent._config_context_length = _destination_context_intent
-    _runtime_context_length = agent._ensure_lmstudio_runtime_loaded(
-        _destination_context_intent
-    )
+    try:
+        _runtime_context_length = agent._ensure_lmstudio_runtime_loaded(
+            _destination_context_intent
+        )
+    except Exception:
+        _restore_snapshot()
+        raise
     if agent._lmstudio_load_was_unverified(_runtime_context_length):
         logger.warning(
             "LM Studio model activation was rejected or completed without a "
@@ -3226,22 +3252,26 @@ def switch_model(agent, new_model, new_provider, api_key='', base_url='', api_mo
         # length normally resolves via config or static catalogs and
         # never hits a probe, but coerce to empty string defensively.
         _ctx_api_key = agent.api_key if isinstance(agent.api_key, str) else ""
-        new_context_length = get_model_context_length(
-            agent.model,
-            base_url=agent.base_url,
-            api_key=_ctx_api_key,
-            provider=agent.provider,
-            config_context_length=_effective_context_length,
-            custom_providers=_sm_custom_providers,
-        )
-        agent.context_compressor.update_model(
-            model=agent.model,
-            context_length=new_context_length,
-            base_url=agent.base_url,
-            api_key=agent.api_key,  # context_compressor forwards to call_llm; callable preserved
-            provider=agent.provider,
-            api_mode=agent.api_mode,
-        )
+        try:
+            new_context_length = get_model_context_length(
+                agent.model,
+                base_url=agent.base_url,
+                api_key=_ctx_api_key,
+                provider=agent.provider,
+                config_context_length=_effective_context_length,
+                custom_providers=_sm_custom_providers,
+            )
+            agent.context_compressor.update_model(
+                model=agent.model,
+                context_length=new_context_length,
+                base_url=agent.base_url,
+                api_key=agent.api_key,  # context_compressor forwards to call_llm; callable preserved
+                provider=agent.provider,
+                api_mode=agent.api_mode,
+            )
+        except Exception:
+            _restore_snapshot()
+            raise
 
     # ── Re-resolve reasoning_config from per-model override ──
     # The new model may have a different reasoning_effort override. Re-read
@@ -3264,6 +3294,10 @@ def switch_model(agent, new_model, new_provider, api_key='', base_url='', api_mo
     # ── Invalidate cached system prompt so it rebuilds next turn ──
     agent._cached_system_prompt = None
 
+    # Publish the destination capability map only after every runtime setup
+    # above has succeeded. Failed switches must leave the old map intact.
+    agent.capabilities = destination_capabilities
+
     # ── Reset the cross-turn stale-call circuit breaker (#58962) ──
     # The breaker's error text tells the user to "switch models ... then
     # retry"; without this reset the streak stays latched and the freshly
@@ -3291,6 +3325,7 @@ def switch_model(agent, new_model, new_provider, api_key='', base_url='', api_mo
         # recovery or fallback restore would resurrect the PRE-switch
         # overrides via the stale init-time snapshot (#75091 seam).
         "request_overrides": dict(getattr(agent, "request_overrides", {}) or {}),
+        "capabilities": dict(getattr(agent, "capabilities", {}) or {}),
         "compressor_model": getattr(_cc, "model", agent.model) if _cc else agent.model,
         "compressor_base_url": getattr(_cc, "base_url", agent.base_url) if _cc else agent.base_url,
         "compressor_api_key": getattr(_cc, "api_key", "") if _cc else "",
diff --git a/agent/chat_completion_helpers.py b/agent/chat_completion_helpers.py
index 42f186184c..009f6a5690 100644
--- a/agent/chat_completion_helpers.py
+++ b/agent/chat_completion_helpers.py
@@ -2932,6 +2932,12 @@ def try_activate_fallback(agent, reason: "FailoverReason | None" = None) -> bool
         # short-circuit the freshly activated fallback before it gets a
         # single stream attempt.
         _reset_stale_streak(agent)
+        from agent.native_compaction import resolve_native_compaction_capabilities
+        agent.capabilities = resolve_native_compaction_capabilities(
+            model=agent.model,
+            base_url=agent.base_url,
+            is_codex_backend=fb_provider == "openai-codex",
+        )
         return True
     except Exception as e:
         if fb_provider == "nous":
diff --git a/agent/native_compaction.py b/agent/native_compaction.py
index ba98ea2a48..8d098adf4f 100644
--- a/agent/native_compaction.py
+++ b/agent/native_compaction.py
@@ -68,6 +68,24 @@ def is_native_compaction_model(model: Optional[str]) -> bool:
     return _ELIGIBLE_MODEL_MARKER in (model or "").lower()
 
 
+def resolve_native_compaction_capabilities(
+    *,
+    model: Optional[str],
+    base_url: Optional[str],
+    is_codex_backend: bool = False,
+) -> Dict[str, bool]:
+    """Resolve the native-compaction capability for a runtime destination.
+
+    The result is deliberately explicit: a resolved ``False`` is different
+    from an unresolved capability and must survive model switches unchanged.
+    """
+    eligible = is_native_compaction_model(model) and is_direct_openai_route(
+        base_url,
+        is_codex_backend=is_codex_backend,
+    )
+    return {"native_compaction": eligible}
+
+
 def is_direct_openai_route(
     base_url: Optional[str],
     *,
@@ -160,6 +178,10 @@ def native_compaction_context_management(
     (``agent.codex_responses_native_compaction = False``, set by the
     conversation loop's rejection recovery) takes effect on the next call.
     """
+    capabilities = getattr(agent, "capabilities", None)
+    if isinstance(capabilities, dict):
+        if not bool(capabilities.get("native_compaction", False)):
+            return None
     if not bool(getattr(agent, "codex_responses_native_compaction", False)):
         return None
     # compression.enabled: false disables ALL automatic compaction, native
diff --git a/cli.py b/cli.py
index bcf71903eb..f92fd5d4cb 100644
--- a/cli.py
+++ b/cli.py
@@ -10306,6 +10306,7 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
                             api_key=_reset_result.api_key,
                             base_url=_reset_result.base_url,
                             api_mode=_reset_result.api_mode,
+                            capabilities=_reset_result.runtime_capabilities,
                         )
                     self.model = _reset_result.new_model
                     self.provider = _reset_result.target_provider
@@ -11480,6 +11481,7 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
                     api_key=snapshot.get("api_key", ""),
                     base_url=snapshot.get("base_url", ""),
                     api_mode=snapshot.get("api_mode", ""),
+                    capabilities=snapshot.get("capabilities"),
                 )
             except Exception as exc:
                 logger.warning("CLI one-turn model restore failed: %s", exc)
@@ -11624,6 +11626,7 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
                     api_key=result.api_key,
                     base_url=result.base_url,
                     api_mode=result.api_mode,
+                    capabilities=result.runtime_capabilities,
                 )
             except Exception as exc:
                 # The agent rolled itself back to the old working model/client.
@@ -12014,6 +12017,7 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
                     api_key=result.api_key,
                     base_url=result.base_url,
                     api_mode=result.api_mode,
+                    capabilities=result.runtime_capabilities,
                 )
             except Exception as exc:
                 # Agent rolled itself back; roll the CLI back too and abort so a
diff --git a/gateway/slash_commands.py b/gateway/slash_commands.py
index 3279f40439..b8ce7cc616 100644
--- a/gateway/slash_commands.py
+++ b/gateway/slash_commands.py
@@ -1950,6 +1950,7 @@ class GatewaySlashCommandsMixin:
                                     api_key=result.api_key,
                                     base_url=result.base_url,
                                     api_mode=result.api_mode,
+                                    capabilities=result.runtime_capabilities,
                                 )
                             except Exception as exc:
                                 # The in-place swap rolled the agent back to the
@@ -2264,6 +2265,7 @@ class GatewaySlashCommandsMixin:
                         api_key=result.api_key,
                         base_url=result.base_url,
                         api_mode=result.api_mode,
+                        capabilities=result.runtime_capabilities,
                     )
                 except Exception as exc:
                     # In-place swap rolled the agent back to the OLD working
diff --git a/hermes_cli/model_switch.py b/hermes_cli/model_switch.py
index c50df9e104..e5db94f47a 100644
--- a/hermes_cli/model_switch.py
+++ b/hermes_cli/model_switch.py
@@ -626,6 +626,7 @@ class ModelSwitchResult:
     provider_label: str = ""
     resolved_via_alias: str = ""
     capabilities: Optional[ModelCapabilities] = None
+    runtime_capabilities: Optional[dict[str, bool]] = None
     model_info: Optional[ModelInfo] = None
     is_global: bool = False
 
@@ -2173,6 +2174,12 @@ def switch_model(
 
     # --- Get capabilities (legacy) ---
     capabilities = get_model_capabilities(target_provider, new_model, allow_network=True)
+    from agent.native_compaction import resolve_native_compaction_capabilities
+    runtime_capabilities = resolve_native_compaction_capabilities(
+        model=new_model,
+        base_url=base_url,
+        is_codex_backend=target_provider.strip().lower() == "openai-codex",
+    )
 
     # --- Get full model info from models.dev ---
     model_info = get_model_info(target_provider, new_model, allow_network=True)
@@ -2215,6 +2222,7 @@ def switch_model(
         provider_label=provider_label,
         resolved_via_alias=resolved_alias,
         capabilities=capabilities,
+        runtime_capabilities=runtime_capabilities,
         model_info=model_info,
         is_global=is_global,
     )
diff --git a/run_agent.py b/run_agent.py
index 1b87729159..ecec7b9b82 100644
--- a/run_agent.py
+++ b/run_agent.py
@@ -899,10 +899,26 @@ class AIAgent:
             return_load_result=True,
         )
 
-    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='',
+        capabilities=None,
+    ):
         """Forwarder — see ``agent.agent_runtime_helpers.switch_model``."""
         from agent.agent_runtime_helpers import switch_model
-        return switch_model(self, new_model, new_provider, api_key, base_url, api_mode)
+        return switch_model(
+            self,
+            new_model,
+            new_provider,
+            api_key,
+            base_url,
+            api_mode,
+            capabilities,
+        )
 
     def _safe_print(self, *args, **kwargs):
         """Print that silently handles broken pipes / closed stdout.
diff --git a/tests/run_agent/test_fallback_reasoning_override.py b/tests/run_agent/test_fallback_reasoning_override.py
index 911e8205d6..349508942d 100644
--- a/tests/run_agent/test_fallback_reasoning_override.py
+++ b/tests/run_agent/test_fallback_reasoning_override.py
@@ -75,6 +75,7 @@ class TestFallbackReasoningOverride:
             "client_kwargs": {},
             "use_prompt_caching": False,
             "use_native_cache_layout": False,
+            "capabilities": {"native_compaction": True},
             "reasoning_config": {"enabled": True, "effort": "medium"},
             "compressor_model": "gemini-flash",
             "compressor_base_url": "",
@@ -95,6 +96,7 @@ class TestFallbackReasoningOverride:
         agent.model = "claude-opus-4.5"
         agent.provider = "anthropic"
         agent.reasoning_config = {"enabled": True, "effort": "xhigh"}
+        agent.capabilities = {"native_compaction": False}
         agent.context_compressor = MagicMock()
         agent.base_url = ""
         agent._anthropic_prompt_cache_policy = MagicMock(return_value=(False, False))
@@ -105,6 +107,7 @@ class TestFallbackReasoningOverride:
         assert result is True
         # reasoning_config should be restored to primary's value (medium)
         assert agent.reasoning_config == {"enabled": True, "effort": "medium"}
+        assert agent.capabilities == {"native_compaction": True}
 
     def test_fallback_global_fallback_with_yaml_false(self):
         """Fallback global fallback must not coerce YAML boolean False.
diff --git a/tests/run_agent/test_native_compaction_switch_capabilities.py b/tests/run_agent/test_native_compaction_switch_capabilities.py
new file mode 100644
index 0000000000..dde00c4bea
--- /dev/null
+++ b/tests/run_agent/test_native_compaction_switch_capabilities.py
@@ -0,0 +1,48 @@
+from types import SimpleNamespace
+
+from agent.native_compaction import (
+    native_compaction_context_management,
+    resolve_native_compaction_capabilities,
+)
+
+
+def _agent(capabilities):
+    return SimpleNamespace(
+        model="gpt-5.6",
+        base_url="https://proxy.example/v1",
+        codex_responses_native_compaction=True,
+        compression_enabled=True,
+        codex_responses_compact_threshold=200_000,
+        context_compressor=None,
+        capabilities=capabilities,
+    )
+
+
+def test_trusted_destination_capabilities_are_explicitly_enabled():
+    capabilities = resolve_native_compaction_capabilities(
+        model="gpt-5.6",
+        base_url="https://api.openai.com/v1",
+    )
+
+    assert capabilities == {"native_compaction": True}
+
+
+def test_untrusted_destination_capabilities_are_explicitly_denied():
+    capabilities = resolve_native_compaction_capabilities(
+        model="gpt-5.6",
+        base_url="https://openrouter.ai/api/v1",
+    )
+
+    assert capabilities == {"native_compaction": False}
+
+
+def test_explicit_false_capability_denies_native_payload():
+    agent = _agent({"native_compaction": False})
+
+    assert native_compaction_context_management(agent, is_codex_backend=False) is None
+
+
+def test_missing_capability_keeps_default_deny():
+    agent = _agent({})
+
+    assert native_compaction_context_management(agent, is_codex_backend=False) is None
diff --git a/tests/run_agent/test_switch_model_context.py b/tests/run_agent/test_switch_model_context.py
index 22321fdc00..8684baa700 100644
--- a/tests/run_agent/test_switch_model_context.py
+++ b/tests/run_agent/test_switch_model_context.py
@@ -284,3 +284,30 @@ def test_lmstudio_switch_uses_destination_context_and_verified_runtime(monkeypat
     assert call_kwargs.get("config_context_length") == 100_000
     assert agent._config_context_length == 120_000
     assert agent.context_compressor.context_length == 100_000
+
+
+def test_later_lmstudio_failure_restores_runtime_capabilities(monkeypatch):
+    agent = _make_agent_with_compressor(config_context_length=32_768)
+    agent.capabilities = {"native_compaction": True}
+    original_client = agent.client
+
+    monkeypatch.setattr(
+        AIAgent,
+        "_ensure_lmstudio_runtime_loaded",
+        lambda *_args, **_kwargs: (_ for _ in ()).throw(
+            RuntimeError("simulated LM Studio failure")
+        ),
+    )
+
+    with pytest.raises(RuntimeError, match="simulated LM Studio failure"):
+        agent.switch_model(
+            "new-model",
+            "openrouter",
+            api_key="sk-new",
+            base_url="https://openrouter.ai/api/v1",
+        )
+
+    assert agent.model == "primary-model"
+    assert agent.provider == "openrouter"
+    assert agent.client is original_client
+    assert agent.capabilities == {"native_compaction": True}
diff --git a/tests/run_agent/test_switch_model_rollback.py b/tests/run_agent/test_switch_model_rollback.py
index efedad9895..8f8f7686a9 100644
--- a/tests/run_agent/test_switch_model_rollback.py
+++ b/tests/run_agent/test_switch_model_rollback.py
@@ -46,6 +46,7 @@ def _make_agent_openrouter():
     agent._fallback_chain = []
     agent._fallback_model = None
     agent._config_context_length = None
+    agent.capabilities = {"native_compaction": False}
 
     return agent
 
@@ -73,6 +74,7 @@ def _make_agent_anthropic():
     agent._fallback_chain = []
     agent._fallback_model = None
     agent._config_context_length = None
+    agent.capabilities = {"native_compaction": False}
 
     return agent
 
@@ -108,6 +110,7 @@ def test_openai_client_rebuild_failure_rolls_back_to_original_state():
     assert agent.api_key == "or-key-original"
     assert agent.client is original_client
     assert agent._client_kwargs == original_kwargs
+    assert agent.capabilities == {"native_compaction": False}
 
 
 def test_anthropic_client_rebuild_failure_rolls_back_to_original_state():
diff --git a/tui_gateway/server.py b/tui_gateway/server.py
index 576d6db45f..e4b70649ea 100644
--- a/tui_gateway/server.py
+++ b/tui_gateway/server.py
@@ -6135,6 +6135,7 @@ def _restore_agent_model_runtime(agent, snapshot: dict | None) -> None:
             api_key=snapshot.get("api_key", ""),
             base_url=snapshot.get("base_url", ""),
             api_mode=snapshot.get("api_mode", ""),
+            capabilities=snapshot.get("capabilities"),
         )
 
 
@@ -6347,6 +6348,7 @@ def _apply_model_switch(
                 api_key=result.api_key,
                 base_url=result.base_url,
                 api_mode=result.api_mode,
+                capabilities=result.runtime_capabilities,
             )
         except Exception as exc:
             # The in-place swap rolled the agent back to the old working