Release v1.27.1

.github/workflows/run-eval.yml

@@ -26,7 +26,8 @@ on:
             sdk_ref:
                 description: SDK commit/ref to evaluate (must be a semantic version like v1.0.0 unless 'Allow unreleased branches' is checked)
                 required: true
-                default: v1.27.0
+                default: v1.27.1
+
 
 
 

openhands-agent-server/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "openhands-agent-server"
-version = "1.27.0"
+version = "1.27.1"
 description = "OpenHands Agent Server - REST/WebSocket interface for OpenHands AI Agent"
 
 requires-python = ">=3.12"

openhands-sdk/openhands/sdk/llm/llm.py

@@ -2100,8 +2100,9 @@ def _apply_prompt_caching(self, messages: list[Message]) -> None:
                 # Single block: mark it for caching
                 sys_content[0].cache_prompt = True
 
-        # Anthropic and Gemini both use these cache_control markers. LiteLLM
-        # performs the provider-specific cache setup for Gemini downstream.
+        # Second breakpoint: mark the last user/tool message so the cached prefix
+        # extends every turn. Anthropic-only; Gemini is excluded from
+        # PROMPT_CACHE_MODELS because its cache can't extend this way.
         for message in reversed(messages):
             if message.role in ("user", "tool"):
                 message.content[

openhands-sdk/openhands/sdk/llm/utils/model_features.py

@@ -117,10 +117,9 @@ def _supports_reasoning_effort(model: str | None) -> bool:
     "claude-opus-4-7",
     "claude-opus-4-8",
     "claude-sonnet-4-6",
-    # Gemini uses the same cache_control marker format. LiteLLM handles
-    # Vertex/Gemini context-cache creation when these markers are present.
-    "gemini-2.5",
-    "gemini-3",
+    # Do NOT add Gemini: explicit cache_control markers freeze its cache at the
+    # static prefix and disable Google's implicit caching on the growing body
+    # (~6-14x cost). Gemini uses implicit prefix caching instead.
 ]
 
 # Models that support a top-level prompt_cache_retention parameter

openhands-sdk/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "openhands-sdk"
-version = "1.27.0"
+version = "1.27.1"
 description = "OpenHands SDK - Core functionality for building AI agents"
 
 requires-python = ">=3.12"

openhands-tools/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "openhands-tools"
-version = "1.27.0"
+version = "1.27.1"
 description = "OpenHands Tools - Runtime tools for AI agents"
 
 requires-python = ">=3.12"

openhands-workspace/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "openhands-workspace"
-version = "1.27.0"
+version = "1.27.1"
 description = "OpenHands Workspace - Docker and container-based workspace implementations"
 
 requires-python = ">=3.12"

tests/sdk/llm/test_model_features.py

@@ -123,10 +123,11 @@ def test_extended_thinking_support(model, expected_extended_thinking):
         ("anthropic.claude-3-5-sonnet-20241022", True),
         ("anthropic.claude-3-haiku-20240307", True),
         ("anthropic.claude-3-opus-20240229", True),
-        # Gemini explicit context caching through LiteLLM.
-        ("gemini-2.5-pro", True),
-        ("gemini-3.1-pro-preview", True),
-        ("litellm_proxy/gemini-3.1-pro-preview", True),
+        # Gemini must NOT use explicit cache_control markers: they freeze the
+        # cache at the static prefix and disable Google's implicit caching.
+        ("gemini-2.5-pro", False),
+        ("gemini-3.1-pro-preview", False),
+        ("litellm_proxy/gemini-3.1-pro-preview", False),
         ("gpt-4o", False),  # OpenAI doesn't support explicit prompt caching
         ("gemini-1.5-pro", False),
         ("unknown-model", False),

tests/sdk/llm/test_prompt_caching_cross_conversation.py

@@ -144,13 +144,24 @@ def on_event(event):
     assert messages[1].content[-1].cache_prompt is True
 
 
-def test_gemini_prompt_caching_marks_formatted_messages():
-    """Gemini models should emit cache_control markers when caching is enabled."""
+def test_gemini_prompt_caching_emits_no_markers():
+    """REGRESSION: Gemini must not emit explicit cache_control markers.
+
+    Explicit markers freeze Gemini's cache at the static prefix and disable
+    Google's implicit caching on the growing body (~6-14x cost). No markers keeps
+    Gemini on the implicit-caching path, where the cached prefix grows.
+    """
     llm = LLM(
         model="litellm_proxy/gemini-3.1-pro-preview",
         usage_id="test",
         caching_prompt=True,
     )
+
+    # Explicit-breakpoint caching must be inactive for Gemini.
+    assert llm.is_caching_prompt_active() is False
+
+    # System (index 0) and last user message (index 3) are non-adjacent — the
+    # case that froze the LiteLLM/Vertex cache at the static prefix.
     messages = [
         Message(
             role="system",
@@ -159,19 +170,21 @@ def test_gemini_prompt_caching_marks_formatted_messages():
                 TextContent(text="Dynamic context"),
             ],
         ),
-        Message(
-            role="user",
-            content=[TextContent(text="Hello")],
-        ),
+        Message(role="user", content=[TextContent(text="First question")]),
+        Message(role="assistant", content=[TextContent(text="First answer")]),
+        Message(role="user", content=[TextContent(text="Second question")]),
     ]
 
     formatted_messages = llm.format_messages_for_llm(messages)
 
-    system_content = formatted_messages[0]["content"]
-    user_content = formatted_messages[1]["content"]
-    assert system_content[0]["cache_control"] == {"type": "ephemeral"}
-    assert "cache_control" not in system_content[1]
-    assert user_content[-1]["cache_control"] == {"type": "ephemeral"}
+    # No cache_control marker anywhere in the formatted payload.
+    for message in formatted_messages:
+        assert "cache_control" not in message
+        content = message.get("content")
+        if isinstance(content, list):
+            for block in content:
+                if isinstance(block, dict):
+                    assert "cache_control" not in block
 
 
 @pytest.mark.parametrize(

tests/sdk/llm/test_responses_parsing_and_kwargs.py

@@ -338,8 +338,9 @@ def test_responses_retries_without_caching_on_prompt_cache_too_small(mock_respon
 
     # Pick a model that supports prompt caching so is_caching_prompt_active()
     # is True and the retry branch is reachable on the responses() path.
+    # (Gemini no longer uses explicit caching, so use an Anthropic model here.)
     llm = LLM(
-        model="gemini-3-flash",
+        model="claude-sonnet-4-20250514",
         api_key=SecretStr("test_key"),
         usage_id="test-llm",
         caching_prompt=True,
@@ -417,8 +418,10 @@ async def test_aresponses_retries_without_caching_on_prompt_cache_too_small(
     )
     mock_aresponses.side_effect = [cache_error, success_resp]
 
+    # Anthropic model so is_caching_prompt_active() is True (Gemini no longer
+    # uses explicit caching); mirrors the sync test above.
     llm = LLM(
-        model="gemini-3-flash",
+        model="claude-sonnet-4-20250514",
         api_key=SecretStr("test_key"),
         usage_id="test-llm",
         caching_prompt=True,