Skip to content

SWE-agent takes a GitHub issue and tries to automatically fix it, using your LM of choice. It can also be employed for offensive cybersecurity or competitive coding challenges. [NeurIPS 2024]

Doc: Anthropic, cache control and more (#970)

Kilian Lieret1a78be6fc706
diff --git a/docs/config/models.md b/docs/config/models.md
index a97c2adf..4517db90 100644
--- a/docs/config/models.md
+++ b/docs/config/models.md
@@ -18,8 +18,26 @@ Remember to unset spending limits and configure the action parser if you cannot
### Anthropic Claude
-Prompt caching makes SWE-agent several times more affordable. While this is done automatically for models like gpt-4o,
-care has to be taken for Anthropic Claude. Anthropic Claude gives you 4 cache break points per key.
+Prompt caching makes SWE-agent several times more affordable. While this is done automatically for models like `gpt-4o`,
+care has to be taken for Anthropic Claude, as you need to manually set the cache break points.
+
+For this, include the following history processor:
+
+```yaml
+agent:
+ history_processors:
+ - type: cache_control
+ last_n_messages: 2
+```
+
+!!! warning "Other history processors"
+
+ Other history processors might interfere with the prompt caching
+ if you are not careful.
+ However, if your history processor is only modifying the last observation,
+ you can combine as done [here](https://github.com/SWE-agent/SWE-agent/blob/main/config/sweagent_heavy.yaml).
+
+Anthropic Claude gives you 4 cache break points per key.
You need two of them for a single agent run (because the break points are both used to retrieve and set the cache).
Therefore, you can only run two parallel instances of SWE-agent with [`run-batch`](../usage/batch_mode.md) per key.
To support more parallel running instances, supply multiple keys as described below.
diff --git a/docs/reference/history_processor_config.md b/docs/reference/history_processor_config.md
index 59e65063..e3246dbd 100644
--- a/docs/reference/history_processor_config.md
+++ b/docs/reference/history_processor_config.md
@@ -1,8 +1,14 @@
# History processor configuration
+History processors can filter the history/trajectory to query the model.
+For example, a very simple history processor would be one that strips away old observations to reduce context when querying the model.
+
::: sweagent.agent.history_processors.DefaultHistoryProcessor
::: sweagent.agent.history_processors.LastNObservations
::: sweagent.agent.history_processors.TagToolCallObservations
+::: sweagent.agent.history_processors.CacheControlHistoryProcessor
+
+::: sweagent.agent.history_processors.RemoveRegex
\ No newline at end of file
diff --git a/sweagent/agent/history_processors.py b/sweagent/agent/history_processors.py
index 35fe9025..84d9b017 100644
--- a/sweagent/agent/history_processors.py
+++ b/sweagent/agent/history_processors.py
@@ -223,6 +223,10 @@ class ClosedWindowHistoryProcessor(BaseModel):
class CacheControlHistoryProcessor(BaseModel):
+ """This history processor adds manual cache control marks to the history.
+ Use this when running with anthropic claude.
+ """
+
type: Literal["cache_control"] = "cache_control"
"""Do not change. Used for (de)serialization."""
@@ -263,8 +267,13 @@ class CacheControlHistoryProcessor(BaseModel):
class RemoveRegex(BaseModel):
+ """This history processor can remove arbitrary content from history items"""
+
remove: list[str] = ["<diff>.*</diff>"]
+ """Regex patterns to remove from history items"""
+
keep_last: int = 0
+ """Keep the last n history items unchanged"""
type: Literal["remove_regex"] = "remove_regex"
"""Do not change. Used for (de)serialization."""