diff --git a/agent/agent_runtime_helpers.py b/agent/agent_runtime_helpers.py
index 1cde73419a449..47d5fe8ac1872 100644
--- a/agent/agent_runtime_helpers.py
+++ b/agent/agent_runtime_helpers.py
@@ -1007,6 +1007,10 @@ def try_recover_primary_transport(
         if hasattr(agent, "_transport_cache"):
             agent._transport_cache.clear()
         agent.api_key = rt["api_key"]
+        # Restore the primary's reasoning effort if a fallback entry
+        # overrode it (#21256).  Guarded for snapshots predating the key.
+        if "reasoning_config" in rt:
+            agent.reasoning_config = rt["reasoning_config"]
 
         if agent.api_mode == "anthropic_messages":
             from agent.anthropic_adapter import build_anthropic_client
@@ -1172,6 +1176,12 @@ def restore_primary_runtime(agent) -> bool:
         agent.api_key = rt["api_key"]
         agent._client_kwargs = dict(rt["client_kwargs"])
         agent._use_prompt_caching = rt["use_prompt_caching"]
+        # Restore the primary's reasoning effort (a fallback entry may have
+        # overridden it via reasoning_effort).  Older snapshots predate this
+        # key — only restore when it was captured, to avoid clobbering the
+        # live value with None.  See #21256.
+        if "reasoning_config" in rt:
+            agent.reasoning_config = rt["reasoning_config"]
         # Default to native layout when the restored snapshot predates the
         # native-vs-proxy split (older sessions saved before this PR).
         agent._use_native_cache_layout = rt.get(
@@ -2094,6 +2104,11 @@ def switch_model(agent, new_model, new_provider, api_key='', base_url='', api_mo
         "compressor_context_length": _cc.context_length if _cc else 0,
         "compressor_api_mode": getattr(_cc, "api_mode", agent.api_mode) if _cc else agent.api_mode,
         "compressor_threshold_tokens": _cc.threshold_tokens if _cc else 0,
+        # Snapshot the primary's reasoning effort so a fallback entry that
+        # overrides it (fallback_model[N].reasoning_effort) can be reverted
+        # when the primary is restored.  See per-entry reasoning support in
+        # try_activate_fallback (#21256).
+        "reasoning_config": getattr(agent, "reasoning_config", None),
     }
     if api_mode == "anthropic_messages":
         agent._primary_runtime.update({
diff --git a/agent/chat_completion_helpers.py b/agent/chat_completion_helpers.py
index b93dcf661717f..f6115cc2eaad6 100644
--- a/agent/chat_completion_helpers.py
+++ b/agent/chat_completion_helpers.py
@@ -1511,12 +1511,49 @@ def try_activate_fallback(agent, reason: "FailoverReason | None" = None) -> bool
             agent._transport_cache.clear()
         agent._fallback_activated = True
 
+        # Per-entry reasoning effort override (#21256).  A fallback_model
+        # entry may carry ``reasoning_effort`` (none through ultra) to run that
+        # tier at a different thinking depth than the
+        # global ``agent.reasoning_effort``. The primary's reasoning_config is
+        # snapshotted in ``_primary_runtime`` and restored on primary recovery,
+        # so this only affects the fallback turn(s). Absent/blank keeps the
+        # current reasoning effort unchanged.
+        _fb_reasoning_effort = str(fb.get("reasoning_effort") or "").strip()
+        if _fb_reasoning_effort:
+            try:
+                from hermes_constants import parse_reasoning_effort
+
+                _fb_reasoning_config = parse_reasoning_effort(_fb_reasoning_effort)
+                if _fb_reasoning_config is not None:
+                    agent.reasoning_config = _fb_reasoning_config
+                    logger.info(
+                        "Fallback %s/%s: reasoning effort override → %s",
+                        fb_provider,
+                        fb_model,
+                        _fb_reasoning_effort,
+                    )
+                else:
+                    logger.warning(
+                        "Fallback %s/%s: unknown reasoning_effort '%s' — "
+                        "keeping current effort",
+                        fb_provider,
+                        fb_model,
+                        _fb_reasoning_effort,
+                    )
+            except Exception as _re_exc:  # noqa: BLE001
+                logger.warning(
+                    "Fallback %s/%s: failed to apply reasoning_effort '%s': %s",
+                    fb_provider,
+                    fb_model,
+                    _fb_reasoning_effort,
+                    _re_exc,
+                )
+
         # Rebind the credential pool to the fallback provider when the provider
         # changes.  Keeping the primary pool attached would make downstream
         # recovery (rate_limit / billing / auth) mutate the wrong credential
         # set and can overwrite the fallback's base_url back to the primary
         # endpoint.  See #33163.
-        #
         # When the fallback shares the pool's provider (e.g. both openrouter
         # entries with different routing) the pool is preserved.  When the
         # providers differ, load the fallback provider's own pool if one exists
diff --git a/hermes_cli/config.py b/hermes_cli/config.py
index 03570bceba498..6da5df05f1bd7 100644
--- a/hermes_cli/config.py
+++ b/hermes_cli/config.py
@@ -5327,6 +5327,19 @@ def validate_config_structure(config: Optional[Dict[str, Any]] = None) -> List["
                             f"fallback_model[{i}] is missing 'model' field",
                             "Add: model: <model-name>",
                         ))
+                    # Optional per-entry reasoning_effort override (#21256).
+                    _re = entry.get("reasoning_effort")
+                    if _re is not None and str(_re).strip():
+                        from hermes_constants import VALID_REASONING_EFFORTS, parse_reasoning_effort
+
+                        if parse_reasoning_effort(_re) is None:
+                            _valid = ", ".join(("none", *VALID_REASONING_EFFORTS))
+                            issues.append(ConfigIssue(
+                                "warning",
+                                f"fallback_model[{i}] has invalid reasoning_effort "
+                                f"'{_re}' — ignored (global effort used)",
+                                f"Use one of: {_valid}",
+                            ))
         elif not isinstance(fb, dict):
             issues.append(ConfigIssue(
                 "error",
diff --git a/tests/run_agent/test_primary_runtime_restore.py b/tests/run_agent/test_primary_runtime_restore.py
index 0d732277e977f..b7170dd2b8ea9 100644
--- a/tests/run_agent/test_primary_runtime_restore.py
+++ b/tests/run_agent/test_primary_runtime_restore.py
@@ -731,3 +731,105 @@ def test_cooldown_not_set_when_already_on_fallback(self):
 
         # second call should not have extended the cooldown
         assert second_cooldown == first_cooldown
+
+
+# =============================================================================
+# Per-entry reasoning_effort override (#21256)
+# =============================================================================
+
+class TestFallbackReasoningEffort:
+    """A fallback_model entry may carry reasoning_effort to run that tier at a
+    different thinking depth than the global agent.reasoning_effort, restored
+    when the primary comes back."""
+
+    def test_override_applied_on_fallback(self):
+        agent = _make_agent(
+            fallback_model={
+                "provider": "openai-codex",
+                "model": "gpt-5.5",
+                "reasoning_effort": "xhigh",
+            },
+        )
+        # Primary starts with no/medium reasoning override.
+        agent.reasoning_config = None
+
+        mock_client = _mock_resolve()
+        with patch("agent.auxiliary_client.resolve_provider_client", return_value=(mock_client, None)):
+            assert agent._try_activate_fallback() is True
+
+        # The fallback entry's xhigh effort is now active.
+        assert agent.reasoning_config == {"enabled": True, "effort": "xhigh"}
+
+    def test_override_restored_on_primary(self):
+        agent = _make_agent(
+            fallback_model={
+                "provider": "openai-codex",
+                "model": "gpt-5.5",
+                "reasoning_effort": "xhigh",
+            },
+        )
+        # Primary runs at an explicit medium effort.  Set it on both the live
+        # agent and the captured _primary_runtime snapshot (init captured the
+        # value present at construction time).
+        primary_cfg = {"enabled": True, "effort": "medium"}
+        agent.reasoning_config = primary_cfg
+        agent._primary_runtime["reasoning_config"] = primary_cfg
+
+        mock_client = _mock_resolve()
+        with patch("agent.auxiliary_client.resolve_provider_client", return_value=(mock_client, None)):
+            agent._try_activate_fallback()
+        assert agent.reasoning_config == {"enabled": True, "effort": "xhigh"}
+
+        # Restoring the primary brings back medium.
+        with patch("run_agent.OpenAI", return_value=MagicMock()):
+            assert agent._restore_primary_runtime() is True
+        assert agent.reasoning_config == primary_cfg
+
+    def test_absent_key_leaves_reasoning_unchanged(self):
+        agent = _make_agent(
+            fallback_model={"provider": "openrouter", "model": "anthropic/claude-sonnet-4"},
+        )
+        primary_cfg = {"enabled": True, "effort": "high"}
+        agent.reasoning_config = primary_cfg
+
+        mock_client = _mock_resolve()
+        with patch("agent.auxiliary_client.resolve_provider_client", return_value=(mock_client, None)):
+            agent._try_activate_fallback()
+
+        # No reasoning_effort on the entry → keep whatever was active.
+        assert agent.reasoning_config == primary_cfg
+
+    def test_invalid_level_ignored(self):
+        agent = _make_agent(
+            fallback_model={
+                "provider": "openrouter",
+                "model": "model-x",
+                "reasoning_effort": "extreme",  # not a valid level
+            },
+        )
+        primary_cfg = {"enabled": True, "effort": "medium"}
+        agent.reasoning_config = primary_cfg
+
+        mock_client = _mock_resolve()
+        with patch("agent.auxiliary_client.resolve_provider_client", return_value=(mock_client, None)):
+            agent._try_activate_fallback()
+
+        # Unknown level → no override, primary effort preserved.
+        assert agent.reasoning_config == primary_cfg
+
+    def test_none_disables_reasoning_on_fallback(self):
+        agent = _make_agent(
+            fallback_model={
+                "provider": "openrouter",
+                "model": "fast-model",
+                "reasoning_effort": "none",
+            },
+        )
+        agent.reasoning_config = {"enabled": True, "effort": "high"}
+
+        mock_client = _mock_resolve()
+        with patch("agent.auxiliary_client.resolve_provider_client", return_value=(mock_client, None)):
+            agent._try_activate_fallback()
+
+        # "none" → reasoning explicitly disabled for this tier.
+        assert agent.reasoning_config == {"enabled": False}