fix: Disable prompt caching in default condenser (#7781)

Co-authored-by: openhands <openhands@all-hands.dev>
Co-authored-by: Calvin Smith <calvin@all-hands.dev>
Co-authored-by: Xingyao Wang <xingyao@all-hands.dev>

openhands/memory/condenser/impl/llm_attention_condenser.py

@@ -114,8 +114,14 @@ class LLMAttentionCondenser(RollingCondenser):
 
     @classmethod
     def from_config(cls, config: LLMAttentionCondenserConfig) -> LLMAttentionCondenser:
+        # This condenser cannot take advantage of prompt caching. If it happens
+        # to be set, we'll pay for the cache writes but never get a chance to
+        # save on a read.
+        llm_config = config.llm_config.model_copy()
+        llm_config.caching_prompt = False
+
         return LLMAttentionCondenser(
-            llm=LLM(config=config.llm_config),
+            llm=LLM(config=llm_config),
             max_size=config.max_size,
             keep_first=config.keep_first,
         )

openhands/memory/condenser/impl/llm_summarizing_condenser.py

@@ -155,8 +155,14 @@ CURRENT_STATE: Last flip: Heads, Haiku count: 15/20"""
     def from_config(
         cls, config: LLMSummarizingCondenserConfig
     ) -> LLMSummarizingCondenser:
+        # This condenser cannot take advantage of prompt caching. If it happens
+        # to be set, we'll pay for the cache writes but never get a chance to
+        # save on a read.
+        llm_config = config.llm_config.model_copy()
+        llm_config.caching_prompt = False
+
         return LLMSummarizingCondenser(
-            llm=LLM(config=config.llm_config),
+            llm=LLM(config=llm_config),
             max_size=config.max_size,
             keep_first=config.keep_first,
             max_event_length=config.max_event_length,

openhands/memory/condenser/impl/structured_summary_condenser.py

@@ -311,8 +311,14 @@ Capture all relevant information, especially:
     def from_config(
         cls, config: StructuredSummaryCondenserConfig
     ) -> StructuredSummaryCondenser:
+        # This condenser cannot take advantage of prompt caching. If it happens
+        # to be set, we'll pay for the cache writes but never get a chance to
+        # save on a read.
+        llm_config = config.llm_config.model_copy()
+        llm_config.caching_prompt = False
+
         return StructuredSummaryCondenser(
-            llm=LLM(config=config.llm_config),
+            llm=LLM(config=llm_config),
             max_size=config.max_size,
             keep_first=config.keep_first,
             max_event_length=config.max_event_length,

tests/unit/test_condenser.py

@@ -331,10 +331,7 @@ def test_llm_summarizing_condenser_from_config():
     config = LLMSummarizingCondenserConfig(
         max_size=50,
         keep_first=10,
-        llm_config=LLMConfig(
-            model='gpt-4o',
-            api_key='test_key',
-        ),
+        llm_config=LLMConfig(model='gpt-4o', api_key='test_key', caching_prompt=True),
     )
     condenser = Condenser.from_config(config)
 
@@ -344,6 +341,10 @@ def test_llm_summarizing_condenser_from_config():
     assert condenser.max_size == 50
     assert condenser.keep_first == 10
 
+    # Since this condenser can't take advantage of caching, we intercept the
+    # passed config and manually flip the caching prompt to False.
+    assert not condenser.llm.config.caching_prompt
+
 
 def test_llm_summarizing_condenser_invalid_config():
     """Test that LLMSummarizingCondenser raises error when keep_first > max_size."""
@@ -474,6 +475,7 @@ def test_llm_attention_condenser_from_config():
         llm_config=LLMConfig(
             model='gpt-4o',
             api_key='test_key',
+            caching_prompt=True,
         ),
     )
     condenser = Condenser.from_config(config)
@@ -484,6 +486,10 @@ def test_llm_attention_condenser_from_config():
     assert condenser.max_size == 50
     assert condenser.keep_first == 10
 
+    # Since this condenser can't take advantage of caching, we intercept the
+    # passed config and manually flip the caching prompt to False.
+    assert not condenser.llm.config.caching_prompt
+
 
 def test_llm_attention_condenser_invalid_config():
     """Test that LLMAttentionCondenser raises an error if the configured LLM doesn't support response schema."""
@@ -614,6 +620,7 @@ def test_structured_summary_condenser_from_config():
         llm_config=LLMConfig(
             model='gpt-4o',
             api_key='test_key',
+            caching_prompt=True,
         ),
     )
     condenser = Condenser.from_config(config)
@@ -624,6 +631,10 @@ def test_structured_summary_condenser_from_config():
     assert condenser.max_size == 50
     assert condenser.keep_first == 10
 
+    # Since this condenser can't take advantage of caching, we intercept the
+    # passed config and manually flip the caching prompt to False.
+    assert not condenser.llm.config.caching_prompt
+
 
 def test_structured_summary_condenser_invalid_config():
     """Test that StructuredSummaryCondenser raises error when keep_first > max_size."""