From f5b159e38d9f4326084982d5a83b44852c246f6b Mon Sep 17 00:00:00 2001
From: lei_lei <96427312+leilei3167@users.noreply.github.com>
Date: Mon, 31 Aug 2026 08:47:43 +0000
Subject: [PATCH 1/7] fix(health): drop model capability flags from
 health-check probes

supports_* and related model_info metadata were ending up on the probe
kwargs and Bedrock rejected them as Extra inputs. Copy params before
mutating and strip those keys before the call.

Fixes #38941
---
 litellm/proxy/health_check.py                 | 31 ++++++++++++++-
 .../litellm_utils_tests/test_health_check.py  | 39 +++++++++++++++++++
 2 files changed, 69 insertions(+), 1 deletion(-)

diff --git a/litellm/proxy/health_check.py b/litellm/proxy/health_check.py
index b1e4f6fd9c3..cad108f6ba1 100644
--- a/litellm/proxy/health_check.py
+++ b/litellm/proxy/health_check.py
@@ -112,6 +112,30 @@ def _should_inject_health_check_max_tokens(model_info: Mapping[str, object], mod
 # Health-check modes that forward `reasoning_effort` to the provider (chat-style calls).
 _HEALTH_CHECK_MODES_SUPPORTING_REASONING_EFFORT: Final = frozenset((None, "chat", "completion"))
 
+# Model-map / model_info capability metadata that must never ride on the health-check
+# probe as provider request fields. Bedrock (and similar) reject them with
+# ``supports_*: Extra inputs are not permitted`` (#38941). They reach the body via
+# ``add_provider_specific_params_to_optional_params`` → ``{**optional_params}``.
+_HEALTH_CHECK_MODEL_METADATA_KEYS: Final = frozenset(
+    {
+        "reasoning_effort_levels",
+        "default_reasoning_effort",
+        "bedrock_output_config_effort_ceiling",
+        "bedrock_converse_supports_strict_tools",
+        "thinking_always_on",
+    }
+)
+
+
+def _is_health_check_model_metadata_key(key: str) -> bool:
+    """True for model capability / catalog metadata that is not a request param."""
+    return key.startswith("supports_") or key in _HEALTH_CHECK_MODEL_METADATA_KEYS
+
+
+def _strip_model_metadata_from_health_params(litellm_params: dict) -> dict:
+    """Drop model_info capability keys that leaked onto the health-check probe params."""
+    return {k: v for k, v in litellm_params.items() if not _is_health_check_model_metadata_key(k)}
+
 
 def _get_process_rss_mb() -> float | None:
     """
@@ -757,6 +781,7 @@ def _update_litellm_params_for_health_check(model_info: dict, litellm_params: di
     """
     Update the litellm params for health check.
 
+    - copies `litellm_params` so the shared deployment dict is not mutated
     - merges `model_info.health_check_params` into the probe request, so a deployment whose provider
       requires a payload field litellm does not synthesize (e.g. `mediaSource` for Bedrock TwelveLabs
       Pegasus) can supply it. The dedicated knobs below are applied afterwards and win on conflict.
@@ -769,7 +794,11 @@ def _update_litellm_params_for_health_check(model_info: dict, litellm_params: di
     - updates the `model` param with the `health_check_model` if it exists Doc: https://docs.litellm.ai/docs/proxy/health#wildcard-routes
     - updates the `voice` param with the `health_check_voice` for `audio_speech` mode if it exists Doc: https://docs.litellm.ai/docs/proxy/health#text-to-speech-models
     - for Bedrock models with region routing (bedrock/region/model), strips the litellm routing prefix but preserves the model ID, and pins `custom_llm_provider` to `bedrock` (only when the deployment hasn't already set one, so an explicit `bedrock_converse` survives) so the bare model id still resolves to the provider (e.g. cross-region ids like `us.cohere.embed-v4:0`)
+    - strips model capability metadata (`supports_*`, effort ceilings, …) so it cannot leak into
+      the provider request body (#38941)
     """
+    # Copy first: callers pass the live deployment litellm_params dict.
+    litellm_params = dict(litellm_params)
     mode: Final = _resolve_health_check_mode(
         model_info,
         litellm_params,  # any-ok: untyped router config dict
@@ -843,7 +872,7 @@ def _update_litellm_params_for_health_check(model_info: dict, litellm_params: di
                 "bedrock"
             )
 
-    return litellm_params
+    return _strip_model_metadata_from_health_params(litellm_params)
 
 
 async def perform_health_check(
diff --git a/tests/litellm_utils_tests/test_health_check.py b/tests/litellm_utils_tests/test_health_check.py
index cfdddd20263..70f55944bef 100644
--- a/tests/litellm_utils_tests/test_health_check.py
+++ b/tests/litellm_utils_tests/test_health_check.py
@@ -832,3 +832,42 @@ async def test_health_check_with_custom_llm_provider():
         # Should succeed without "LLM Provider NOT provided" error
         assert "error" not in response
         assert isinstance(response, dict)
+
+
+def test_health_check_strips_model_capability_metadata():
+    """
+    Capability flags on litellm_params / health_check_params must not ride the probe.
+
+    Bedrock rejects them as request fields (#38941):
+    ``supports_max_reasoning_effort: Extra inputs are not permitted``.
+    Also ensure the shared deployment dict is not mutated.
+    """
+    from litellm.proxy.health_check import _update_litellm_params_for_health_check
+
+    original = {
+        "model": "bedrock/anthropic.claude-3-7-sonnet-20240620-v1:0",
+        "api_key": "fake_key",
+        "supports_max_reasoning_effort": True,
+        "supports_xhigh_reasoning_effort": None,
+        "bedrock_output_config_effort_ceiling": "xhigh",
+        "thinking": {"type": "enabled", "budget_tokens": 1024},
+    }
+    model_info = {
+        "health_check_params": {
+            "supports_max_reasoning_effort": True,
+            "mediaSource": {"s3Location": {"uri": "s3://bucket/key"}},
+        },
+    }
+
+    updated = _update_litellm_params_for_health_check(model_info, original)
+
+    assert "supports_max_reasoning_effort" not in updated
+    assert "supports_xhigh_reasoning_effort" not in updated
+    assert "bedrock_output_config_effort_ceiling" not in updated
+    # Legitimate probe / provider fields stay
+    assert "messages" in updated
+    assert updated["thinking"] == {"type": "enabled", "budget_tokens": 1024}
+    assert updated["mediaSource"] == {"s3Location": {"uri": "s3://bucket/key"}}
+    # Shared deployment dict must stay untouched
+    assert "messages" not in original
+    assert original["supports_max_reasoning_effort"] is True

From 2800120432c0f02ac72501d9a9cd1fddcafabdfc Mon Sep 17 00:00:00 2001
From: lei_lei <96427312+leilei3167@users.noreply.github.com>
Date: Mon, 31 Aug 2026 09:04:27 +0000
Subject: [PATCH 2/7] fix(health): satisfy type-discipline LIT001/LIT002 on
 probe strip

Annotate the required mutable dict copy/filter so the gate does not count new dict annotations or constructions.
---
 litellm/proxy/health_check.py | 12 +++++++++---
 1 file changed, 9 insertions(+), 3 deletions(-)

diff --git a/litellm/proxy/health_check.py b/litellm/proxy/health_check.py
index cad108f6ba1..02674b53357 100644
--- a/litellm/proxy/health_check.py
+++ b/litellm/proxy/health_check.py
@@ -132,9 +132,15 @@ def _is_health_check_model_metadata_key(key: str) -> bool:
     return key.startswith("supports_") or key in _HEALTH_CHECK_MODEL_METADATA_KEYS
 
 
-def _strip_model_metadata_from_health_params(litellm_params: dict) -> dict:
+def _strip_model_metadata_from_health_params(
+    litellm_params: Mapping[str, object],
+) -> dict:  # mutable-ok: ahealth_check kwargs and LiteLLM_Params(**probe) need a mutable dict
     """Drop model_info capability keys that leaked onto the health-check probe params."""
-    return {k: v for k, v in litellm_params.items() if not _is_health_check_model_metadata_key(k)}
+    return {  # mutable-ok: ahealth_check kwargs and LiteLLM_Params(**probe) need a mutable dict
+        k: v
+        for k, v in litellm_params.items()
+        if not _is_health_check_model_metadata_key(k)
+    }
 
 
 def _get_process_rss_mb() -> float | None:
@@ -798,7 +804,7 @@ def _update_litellm_params_for_health_check(model_info: dict, litellm_params: di
       the provider request body (#38941)
     """
     # Copy first: callers pass the live deployment litellm_params dict.
-    litellm_params = dict(litellm_params)
+    litellm_params = dict(litellm_params)  # mutable-ok: copy so probe mutations don't rewrite the shared deployment dict  # rebind-ok: local probe copy
     mode: Final = _resolve_health_check_mode(
         model_info,
         litellm_params,  # any-ok: untyped router config dict

From 954c850205ac0c0543f0bda035ffa765efc83e0c Mon Sep 17 00:00:00 2001
From: lei_lei <96427312+leilei3167@users.noreply.github.com>
Date: Mon, 31 Aug 2026 09:10:00 +0000
Subject: [PATCH 3/7] style(health_check): ruff format

---
 litellm/proxy/health_check.py | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/litellm/proxy/health_check.py b/litellm/proxy/health_check.py
index 02674b53357..be59b245e18 100644
--- a/litellm/proxy/health_check.py
+++ b/litellm/proxy/health_check.py
@@ -137,9 +137,7 @@ def _strip_model_metadata_from_health_params(
 ) -> dict:  # mutable-ok: ahealth_check kwargs and LiteLLM_Params(**probe) need a mutable dict
     """Drop model_info capability keys that leaked onto the health-check probe params."""
     return {  # mutable-ok: ahealth_check kwargs and LiteLLM_Params(**probe) need a mutable dict
-        k: v
-        for k, v in litellm_params.items()
-        if not _is_health_check_model_metadata_key(k)
+        k: v for k, v in litellm_params.items() if not _is_health_check_model_metadata_key(k)
     }
 
 
@@ -804,7 +802,9 @@ def _update_litellm_params_for_health_check(model_info: dict, litellm_params: di
       the provider request body (#38941)
     """
     # Copy first: callers pass the live deployment litellm_params dict.
-    litellm_params = dict(litellm_params)  # mutable-ok: copy so probe mutations don't rewrite the shared deployment dict  # rebind-ok: local probe copy
+    litellm_params = dict(
+        litellm_params
+    )  # mutable-ok: copy so probe mutations don't rewrite the shared deployment dict  # rebind-ok: local probe copy
     mode: Final = _resolve_health_check_mode(
         model_info,
         litellm_params,  # any-ok: untyped router config dict

From 77f3be043eb7a742bc6ded10ef8bef18397459cf Mon Sep 17 00:00:00 2001
From: lei_lei <96427312+leilei3167@users.noreply.github.com>
Date: Mon, 31 Aug 2026 09:17:02 +0000
Subject: [PATCH 4/7] style(health_check): put mutable-ok on dict()
 construction line

---
 litellm/proxy/health_check.py | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)

diff --git a/litellm/proxy/health_check.py b/litellm/proxy/health_check.py
index be59b245e18..99934726532 100644
--- a/litellm/proxy/health_check.py
+++ b/litellm/proxy/health_check.py
@@ -802,9 +802,7 @@ def _update_litellm_params_for_health_check(model_info: dict, litellm_params: di
       the provider request body (#38941)
     """
     # Copy first: callers pass the live deployment litellm_params dict.
-    litellm_params = dict(
-        litellm_params
-    )  # mutable-ok: copy so probe mutations don't rewrite the shared deployment dict  # rebind-ok: local probe copy
+    litellm_params = dict(litellm_params)  # mutable-ok: copy so probe mutations don't rewrite the shared deployment dict  # rebind-ok: local probe copy
     mode: Final = _resolve_health_check_mode(
         model_info,
         litellm_params,  # any-ok: untyped router config dict

From 5da94728b4ded429b496c1f5bcd66b53132056ad Mon Sep 17 00:00:00 2001
From: lei_lei <96427312+leilei3167@users.noreply.github.com>
Date: Mon, 31 Aug 2026 09:55:43 +0000
Subject: [PATCH 5/7] style(health_check): ruff format

---
 litellm/proxy/health_check.py | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/litellm/proxy/health_check.py b/litellm/proxy/health_check.py
index 99934726532..be59b245e18 100644
--- a/litellm/proxy/health_check.py
+++ b/litellm/proxy/health_check.py
@@ -802,7 +802,9 @@ def _update_litellm_params_for_health_check(model_info: dict, litellm_params: di
       the provider request body (#38941)
     """
     # Copy first: callers pass the live deployment litellm_params dict.
-    litellm_params = dict(litellm_params)  # mutable-ok: copy so probe mutations don't rewrite the shared deployment dict  # rebind-ok: local probe copy
+    litellm_params = dict(
+        litellm_params
+    )  # mutable-ok: copy so probe mutations don't rewrite the shared deployment dict  # rebind-ok: local probe copy
     mode: Final = _resolve_health_check_mode(
         model_info,
         litellm_params,  # any-ok: untyped router config dict

From 7206a4dd88dbc2213bd916ae5129bc305de51f6d Mon Sep 17 00:00:00 2001
From: lei_lei <96427312+leilei3167@users.noreply.github.com>
Date: Mon, 31 Aug 2026 10:37:20 +0000
Subject: [PATCH 6/7] style(health_check): keep mutable-ok on probe dict copy

Co-authored-by: lei_lei <leilei3167@users.noreply.github.com>
---
 litellm/proxy/health_check.py | 35 ++++++++++++++++-------------------
 1 file changed, 16 insertions(+), 19 deletions(-)

diff --git a/litellm/proxy/health_check.py b/litellm/proxy/health_check.py
index be59b245e18..39033cd6ebf 100644
--- a/litellm/proxy/health_check.py
+++ b/litellm/proxy/health_check.py
@@ -801,44 +801,41 @@ def _update_litellm_params_for_health_check(model_info: dict, litellm_params: di
     - strips model capability metadata (`supports_*`, effort ceilings, …) so it cannot leak into
       the provider request body (#38941)
     """
-    # Copy first: callers pass the live deployment litellm_params dict.
-    litellm_params = dict(
-        litellm_params
-    )  # mutable-ok: copy so probe mutations don't rewrite the shared deployment dict  # rebind-ok: local probe copy
+    probe = dict(litellm_params)  # mutable-ok: copy so probe mutations do not rewrite the shared deployment
     mode: Final = _resolve_health_check_mode(
         model_info,
-        litellm_params,  # any-ok: untyped router config dict
+        probe,  # any-ok: untyped router config dict
     )
     _health_check_params: Final = model_info.get("health_check_params", None)
     if isinstance(_health_check_params, dict):
-        litellm_params.update(_health_check_params)
+        probe.update(_health_check_params)
     elif _health_check_params is not None:
         logger.warning(
             "health_check_params for model %s is a %s, expected a dict. Ignoring it.",
-            litellm_params.get("model"),
+            probe.get("model"),
             type(_health_check_params).__name__,
         )
 
-    litellm_params["messages"] = _get_random_llm_message()
+    probe["messages"] = _get_random_llm_message()
     if _should_inject_health_check_max_tokens(
         model_info,
         mode,  # any-ok: untyped router config dict
     ):
-        _resolved_max_tokens: Final = _resolve_health_check_max_tokens(model_info, litellm_params)
+        _resolved_max_tokens: Final = _resolve_health_check_max_tokens(model_info, probe)
         if _resolved_max_tokens is not None:
-            litellm_params["max_tokens"] = _resolved_max_tokens
+            probe["max_tokens"] = _resolved_max_tokens
 
     # Per-model reasoning effort for health checks only (e.g. reasoning_effort=none).
     if mode in _HEALTH_CHECK_MODES_SUPPORTING_REASONING_EFFORT:
         _hc_reasoning_effort: Final = model_info.get("health_check_reasoning_effort", None)
         if _hc_reasoning_effort is not None:
-            litellm_params["reasoning_effort"] = _hc_reasoning_effort
+            probe["reasoning_effort"] = _hc_reasoning_effort
 
     _health_check_model: Final = model_info.get("health_check_model", None)
     if _health_check_model is not None:
-        litellm_params["model"] = _health_check_model
+        probe["model"] = _health_check_model
     if mode == "audio_speech":
-        litellm_params["voice"] = model_info.get("health_check_voice", "alloy")
+        probe["voice"] = model_info.get("health_check_voice", "alloy")
 
     # Handle Bedrock region routing format: bedrock/region/model
     # This is needed because health checks bypass get_llm_provider() for the model param
@@ -849,10 +846,10 @@ def _update_litellm_params_for_health_check(model_info: dict, litellm_params: di
     # Issue: Stripping these breaks AWS requirement for inference profile IDs
     #
     # Must also preserve route prefixes (converse/, invoke/) and handlers (llama/, deepseek_r1/, etc.)
-    if litellm_params["model"].startswith("bedrock/"):
+    if probe["model"].startswith("bedrock/"):
         from litellm.llms.bedrock.common_utils import BedrockModelInfo
 
-        model = litellm_params["model"]
+        model = probe["model"]
         # Strip only the bedrock/ prefix (preserve routes like converse/, invoke/)
         model = model.removeprefix("bedrock/")  # len("bedrock/") = 8
 
@@ -872,13 +869,13 @@ def _update_litellm_params_for_health_check(model_info: dict, litellm_params: di
                 filtered_parts.append(part)
 
         model = "/".join(filtered_parts)
-        litellm_params["model"] = model
-        if not litellm_params.get("custom_llm_provider"):  # any-ok: untyped router dict
-            litellm_params["custom_llm_provider"] = (  # any-ok: untyped router dict
+        probe["model"] = model
+        if not probe.get("custom_llm_provider"):  # any-ok: untyped router dict
+            probe["custom_llm_provider"] = (  # any-ok: untyped router dict
                 "bedrock"
             )
 
-    return _strip_model_metadata_from_health_params(litellm_params)
+    return _strip_model_metadata_from_health_params(probe)
 
 
 async def perform_health_check(

From 42b7c99414c8abcf2083eb4a8f28372d84aa7daf Mon Sep 17 00:00:00 2001
From: lei_lei <imleilei123@gmail.com>
Date: Sun, 13 Sep 2026 07:45:35 +0000
Subject: [PATCH 7/7] chore(ui): regenerate schema.d.ts for OpenAPI sync

---
 ui/litellm-dashboard/src/lib/http/schema.d.ts | 2 --
 1 file changed, 2 deletions(-)

diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts
index 7eadaa6c991..839aa52fa84 100644
--- a/ui/litellm-dashboard/src/lib/http/schema.d.ts
+++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts
@@ -16781,7 +16781,6 @@ export interface paths {
          *     - permissions: Optional[dict] - [Not Implemented Yet] User-specific permissions, eg. turning off pii masking.
          *     - metadata: Optional[dict] - Metadata for user, store information for user. Example metadata = {"team": "core-infra", "app": "app2", "email": "ishaan@berri.ai" }
          *     - max_parallel_requests: Optional[int] - Rate limit a user based on the number of parallel requests. Raises 429 error, if user's parallel requests > x.
-         *     - soft_budget: Optional[float] - Get alerts when user crosses given budget, doesn't block requests.
          *     - model_max_budget: Optional[dict] - Model-specific max budget for user. [Docs](https://docs.litellm.ai/docs/proxy/users#add-model-specific-budgets-to-keys)
          *     - budget_fallbacks: Optional[Dict[str, List[str]]] - Per-model fallback chain tried in order when that model's own `model_max_budget` is exceeded, e.g. {"gpt-4o": ["gpt-4o-mini"]}.
          *     - model_rpm_limit: Optional[float] - Model-specific rpm limit for user. [Docs](https://docs.litellm.ai/docs/proxy/users#add-model-specific-limits-to-keys)
@@ -16887,7 +16886,6 @@ export interface paths {
          *         - permissions: Optional[dict] - [Not Implemented Yet] User-specific permissions, eg. turning off pii masking.
          *         - metadata: Optional[dict] - Metadata for user, store information for user. Example metadata = {"team": "core-infra", "app": "app2", "email": "ishaan@berri.ai" }
          *         - max_parallel_requests: Optional[int] - Rate limit a user based on the number of parallel requests. Raises 429 error, if user's parallel requests > x.
-         *         - soft_budget: Optional[float] - Get alerts when user crosses given budget, doesn't block requests.
          *         - model_max_budget: Optional[dict] - Model-specific max budget for user. [Docs](https://docs.litellm.ai/docs/proxy/users#add-model-specific-budgets-to-keys)
          *         - budget_fallbacks: Optional[Dict[str, List[str]]] - Per-model fallback chain tried in order when that model's own `model_max_budget` is exceeded, e.g. {"gpt-4o": ["gpt-4o-mini"]}.
          *         - model_rpm_limit: Optional[float] - Model-specific rpm limit for user. [Docs](https://docs.litellm.ai/docs/proxy/users#add-model-specific-limits-to-keys)