fix(ci): bedrock recordings hash stability improvements

- Skip embedding model validation for bedrock setup (empty string forces override)
- Exclude stream_options from hash computation for consistent recordings

The bedrock provider adds stream_options for telemetry but this was causing
hash mismatches since recordings were made before this was added. Now
infrastructure/telemetry fields are excluded from the hash computation.

Author: skamenan7

src/llama_stack/testing/api_recorder.py

@@ -77,15 +77,29 @@ def _replace(match: re.Match[str]) -> str:
     return _FLOAT_IN_STRING_PATTERN.sub(_replace, value)
 
 
-def _normalize_body_for_hash(value: Any) -> Any:
-    """Recursively normalize a JSON-like value to improve hash stability."""
+# Keys to exclude from hash computation - these are infrastructure/telemetry fields
+# that don't affect the logical request identity
+_HASH_EXCLUDED_KEYS = {"stream_options"}
+
+
+def _normalize_body_for_hash(value: Any, is_top_level: bool = True) -> Any:
+    """Recursively normalize a JSON-like value to improve hash stability.
+
+    Args:
+        value: The value to normalize
+        is_top_level: Whether this is the top-level body dict (for key exclusion)
+    """
 
     if isinstance(value, dict):
-        return {key: _normalize_body_for_hash(item) for key, item in value.items()}
+        return {
+            key: _normalize_body_for_hash(item, is_top_level=False)
+            for key, item in value.items()
+            if not (is_top_level and key in _HASH_EXCLUDED_KEYS)
+        }
     if isinstance(value, list):
-        return [_normalize_body_for_hash(item) for item in value]
+        return [_normalize_body_for_hash(item, is_top_level=False) for item in value]
     if isinstance(value, tuple):
-        return tuple(_normalize_body_for_hash(item) for item in value)
+        return tuple(_normalize_body_for_hash(item, is_top_level=False) for item in value)
     if isinstance(value, float):
         return round(value, 5)
     if isinstance(value, str):
@@ -162,14 +176,6 @@ def normalize_inference_request(method: str, url: str, headers: dict[str, Any],
     normalized_json = json.dumps(normalized, sort_keys=True)
     request_hash = hashlib.sha256(normalized_json.encode()).hexdigest()
 
-    if is_debug_mode():
-        logger.info("[RECORDING DEBUG] Hash computation:")
-        logger.info(f"  Test ID: {test_id}")
-        logger.info(f"  Method: {method.upper()}")
-        logger.info(f"  Endpoint: {parsed.path}")
-        logger.info(f"  Model: {body.get('model', 'N/A')}")
-        logger.info(f"  Computed hash: {request_hash}")
-
     return request_hash
 
 

tests/integration/conftest.py

@@ -134,9 +134,13 @@ def pytest_configure(config):
             if k not in os.environ:
                 os.environ[k] = str(v)
         # Apply defaults if not provided explicitly
+        # Empty string ("") is a special case that means "force override to disable"
+        # This allows setups like bedrock to skip embedding validation
         for dest, value in setup_obj.defaults.items():
-            current = getattr(config.option, dest, None)
-            if current is None:
+            if value == "":
+                # Empty string forces override - used to disable features
+                setattr(config.option, dest, value)
+            elif getattr(config.option, dest, None) is None:
                 setattr(config.option, dest, value)
 
     # Apply global fallback for embedding_dimension if still not set

tests/integration/suites.py

@@ -107,6 +107,7 @@ class Setup(BaseModel):
         description="AWS Bedrock provider with OpenAI GPT-OSS model (us-west-2)",
         defaults={
             "text_model": "bedrock/openai.gpt-oss-20b-1:0",
+            "embedding_model": "",  # Bedrock doesn't support embeddings
         },
     ),
     "gpt": Setup(