diff --git a/src/app/endpoints/rlsapi_v1.py b/src/app/endpoints/rlsapi_v1.py
index 5d23b33f6..c6fd2b67e 100644
--- a/src/app/endpoints/rlsapi_v1.py
+++ b/src/app/endpoints/rlsapi_v1.py
@@ -51,6 +51,7 @@
     extract_provider_and_model_from_model_id,
     handle_known_apistatus_errors,
     is_context_length_error,
+    normalize_vertex_ai_model_id,
 )
 from utils.quota import check_tokens_available
 from utils.responses import (
@@ -343,9 +344,12 @@ async def _call_llm(
 
     logger.debug("Using model %s for rlsapi v1 inference", resolved_model_id)
 
+    # Normalize Vertex AI model IDs to work around llama-stack 0.6.x bug
+    normalized_model = normalize_vertex_ai_model_id(resolved_model_id)
+
     response = await client.responses.create(
         input=question,
-        model=resolved_model_id,
+        model=normalized_model,
         instructions=instructions,
         tools=tools or [],
         stream=False,
diff --git a/src/utils/compaction.py b/src/utils/compaction.py
index b87085ff0..de3603e1d 100644
--- a/src/utils/compaction.py
+++ b/src/utils/compaction.py
@@ -34,6 +34,7 @@
 
 from log import get_logger
 from models.compaction import ConversationSummary
+from utils.query import normalize_vertex_ai_model_id
 from utils.token_estimator import (
     estimate_conversation_tokens,
     estimate_tokens,
@@ -266,10 +267,14 @@ async def summarize_chunk(
     # by utils.responses.get_topic_summary and protects the directives from
     # prompt-injection via user message content that ends up in the
     # transcript.
+
+    # Normalize Vertex AI model IDs to work around llama-stack 0.6.x bug
+    normalized_model = normalize_vertex_ai_model_id(model)
+
     response = await client.responses.create(
         input=f"Conversation:\n{transcript}",
         instructions=SUMMARIZATION_PROMPT,
-        model=model,
+        model=normalized_model,
         stream=False,
         store=False,
     )
@@ -374,10 +379,14 @@ async def recursively_resummarize(
         model,
     )
     # Same instructions/input split as summarize_chunk — see comment there.
+
+    # Normalize Vertex AI model IDs to work around llama-stack 0.6.x bug
+    normalized_model = normalize_vertex_ai_model_id(model)
+
     response = await client.responses.create(
         input=transcript,
         instructions=RECURSIVE_RESUMMARIZATION_PROMPT,
-        model=model,
+        model=normalized_model,
         stream=False,
         store=False,
     )
diff --git a/src/utils/query.py b/src/utils/query.py
index 32b9673f0..51e246d91 100644
--- a/src/utils/query.py
+++ b/src/utils/query.py
@@ -534,6 +534,34 @@ def extract_provider_and_model_from_model_id(model_id: str) -> tuple[str, str]:
     return "", model_id
 
 
+def normalize_vertex_ai_model_id(model_id: str) -> str:
+    """Normalize Vertex AI model ID to work around llama-stack 0.6.x bug.
+
+    llama-stack 0.6.x has a bug in the inline::meta-reference responses provider
+    where it normalizes model IDs before checking against allowed_models, but doesn't
+    normalize the allowed_models list itself. This causes Vertex AI models to fail
+    validation because:
+    - Model is registered as: publishers/google/models/gemini-2.5-flash
+    - llama-stack strips to: google/gemini-2.5-flash internally
+    - Checks against allowed list: ['publishers/google/models/gemini-2.5-flash']
+    - Mismatch → 500 error
+
+    This workaround strips the publishers/google/models/ prefix to match what
+    llama-stack expects internally.
+
+    Fixed in llama-stack 0.7.0 via https://github.com/ogx-ai/ogx/pull/5169
+
+    Args:
+        model_id: The model ID, possibly in Vertex AI format
+
+    Returns:
+        Normalized model ID with Vertex AI prefix stripped if present
+    """
+    if model_id.startswith("publishers/google/models/"):
+        return model_id.replace("publishers/google/models/", "google/", 1)
+    return model_id
+
+
 def handle_known_apistatus_errors(
     error: LLSApiStatusError | OpenAIAPIStatusError, model_id: str
 ) -> AbstractErrorResponse:
diff --git a/src/utils/responses.py b/src/utils/responses.py
index 6c06e8cbb..a6d098199 100644
--- a/src/utils/responses.py
+++ b/src/utils/responses.py
@@ -118,6 +118,7 @@
 from utils.query import (
     extract_provider_and_model_from_model_id,
     handle_known_apistatus_errors,
+    normalize_vertex_ai_model_id,
     prepare_input,
 )
 from utils.suid import to_llama_stack_conversation_id
@@ -178,11 +179,14 @@ async def get_topic_summary(  # pylint: disable=too-many-nested-blocks
         The topic summary for the question
     """
     try:
+        # Normalize Vertex AI model IDs to work around llama-stack 0.6.x bug
+        normalized_model = normalize_vertex_ai_model_id(model_id)
+
         response = cast(
             ResponseObject,
             await client.responses.create(
                 input=question,
-                model=model_id,
+                model=normalized_model,
                 instructions=get_topic_summary_system_prompt(),
                 stream=False,
                 store=False,  # Don't store topic summary requests
@@ -389,9 +393,13 @@ async def prepare_responses_params(  # pylint: disable=too-many-arguments,too-ma
 
     # Build x-llamastack-provider-data header from MCP tool headers
     extra_headers = _build_provider_data_headers(tools)
+
+    # Normalize Vertex AI model IDs to work around llama-stack 0.6.x bug
+    normalized_model = normalize_vertex_ai_model_id(model)
+
     return ResponsesApiParams(
         input=input_text,
-        model=model,
+        model=normalized_model,
         instructions=system_prompt,
         tools=tools,
         conversation=llama_stack_conv_id,
diff --git a/tests/unit/utils/test_responses.py b/tests/unit/utils/test_responses.py
index 3ca93a907..8666cfd93 100644
--- a/tests/unit/utils/test_responses.py
+++ b/tests/unit/utils/test_responses.py
@@ -57,6 +57,7 @@
 from models.api.requests import QueryRequest
 from models.common.responses.types import InputTool, InputToolMCP
 from models.config import ApprovalFilter, ByokRag, ModelContextProtocolServer
+from utils.query import normalize_vertex_ai_model_id
 from utils.responses import (
     _build_chunk_attributes,
     _merge_tools,
@@ -3577,3 +3578,32 @@ async def test_merge_header_no_server_tools_returns_client_only(
         )
         assert tools is not None
         assert len(tools) == 1
+
+
+class TestNormalizeVertexAIModelId:
+    """Tests for normalize_vertex_ai_model_id function."""
+
+    def test_normalizes_vertex_ai_model_id(self) -> None:
+        """Test that Vertex AI model IDs are normalized correctly."""
+        input_model = "publishers/google/models/gemini-2.5-flash"
+        expected = "google/gemini-2.5-flash"
+        assert normalize_vertex_ai_model_id(input_model) == expected
+
+    def test_normalizes_vertex_ai_model_id_with_version(self) -> None:
+        """Test normalization with versioned Vertex AI model ID."""
+        input_model = "publishers/google/models/gemini-1.5-pro-001"
+        expected = "google/gemini-1.5-pro-001"
+        assert normalize_vertex_ai_model_id(input_model) == expected
+
+    def test_preserves_non_vertex_ai_model_ids(self) -> None:
+        """Test that non-Vertex AI model IDs are returned unchanged."""
+        # Regular model IDs should pass through
+        assert normalize_vertex_ai_model_id("gpt-4") == "gpt-4"
+        assert normalize_vertex_ai_model_id("openai/gpt-4") == "openai/gpt-4"
+        assert normalize_vertex_ai_model_id("watsonx/model") == "watsonx/model"
+
+    def test_preserves_gemini_api_format(self) -> None:
+        """Test that Gemini API format (models/...) is preserved."""
+        # Gemini API format doesn't have the publishers prefix
+        gemini_api_format = "models/gemini-2.5-flash"
+        assert normalize_vertex_ai_model_id(gemini_api_format) == gemini_api_format