Skip to content

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.

You can set them as follows:

agent:
  history_processors:
    - type: last_n_observations
      n: 5

sweagent.agent.history_processors.DefaultHistoryProcessor pydantic-model

Bases: BaseModel

Config:

  • extra: forbid

Fields:

  • type (Literal['default'])

type pydantic-field 

type: Literal['default'] = 'default'

Do not change. Used for (de)serialization.

sweagent.agent.history_processors.LastNObservations pydantic-model

Bases: BaseModel

Elide all but the last n observations or remove tagged observations.

This is our most classic history processor, used in the original paper to elide but the last 5 observations. Elided observations are replaced by "Old environment output: (n lines omitted)".

Typical configuration:

agent:
  history_processors:
    - type: last_n_observations
      n: 5

as for example in use in the SWE-agent 0.7 config at https://github.com/SWE-agent/SWE-agent/blob/main/config/sweagent_0_7/07.yaml

For most use cases, you only need to set n.

Note that using this history processor will break prompt caching (as the history of every query will change every time due to the elided observations). There are some workarounds possible with the polling parameter.

However, most SotA models can now fit a lot of context, so generally this history processor is not always needed anymore.

Config:

  • extra: forbid

Fields:

Validators:

n pydantic-field 

n: int

Number of observations to keep.

polling pydantic-field 

polling: int = 1

How many steps to keep between updating the number of observations to keep. This is useful for caching, as we want to remove more and more messages, but every time we change the history, we need to cache everything again. Effectively, we will now keep between n and n+polling observations.

always_remove_output_for_tags pydantic-field 

always_remove_output_for_tags: set[str] = {'remove_output'}

Any observation with a tags field containing one of these strings will be elided, even if it is one of the last n observations.

always_keep_output_for_tags pydantic-field 

always_keep_output_for_tags: set[str] = {'keep_output'}

Any observation with a tags field containing one of these strings will be kept, even if it is not one of the last n observations.

type pydantic-field 

type: Literal['last_n_observations'] = 'last_n_observations'

Do not change. Used for (de)serialization.

validate_n pydantic-validator 

validate_n(n: int) -> int
Source code in sweagent/agent/history_processors.py 
132
133
134
135
136
137
@field_validator("n")
def validate_n(cls, n: int) -> int:
    if n <= 0:
        msg = "n must be a positive integer"
        raise ValueError(msg)
    return n

sweagent.agent.history_processors.TagToolCallObservations pydantic-model

Bases: BaseModel

Adds tags to history items for specific tool calls.

Config:

  • extra: forbid

Fields:

type pydantic-field 

type: Literal['tag_tool_call_observations'] = 'tag_tool_call_observations'

Do not change. Used for (de)serialization.

tags pydantic-field 

tags: set[str] = {'keep_output'}

Add the following tag to all observations matching the search criteria.

function_names pydantic-field 

function_names: set[str]

Only consider observations made by tools with these names.

sweagent.agent.history_processors.CacheControlHistoryProcessor pydantic-model

Bases: BaseModel

This history processor adds manual cache control marks to the history. Use this when running with anthropic claude.

Config:

  • extra: forbid

Fields:

type pydantic-field 

type: Literal['cache_control'] = 'cache_control'

Do not change. Used for (de)serialization.

last_n_messages pydantic-field 

last_n_messages: int = 2

Add cache control to the last n user messages (and clear it for anything else). In most cases this should be set to 2 (caching for multi-turn conversations). When resampling and running concurrent instances, you want to set it to 1. If set to <= 0, any set cache control will be removed from all messages.

last_n_messages_offset pydantic-field 

last_n_messages_offset: int = 0

E.g., set to 1 to start cache control after the second to last user message. This can be useful in rare cases, when you want to modify the last message after we've got the completion and you want to avoid cache mismatch.

tagged_roles pydantic-field 

tagged_roles: list[str] = ['user', 'tool']

Only add cache control to messages with these roles.

sweagent.agent.history_processors.RemoveRegex pydantic-model

Bases: BaseModel

This history processor can remove arbitrary content from history items

Config:

  • extra: forbid

Fields:

remove pydantic-field 

remove: list[str] = ['<diff>.*</diff>']

Regex patterns to remove from history items

keep_last pydantic-field 

keep_last: int = 0

Keep the last n history items unchanged

type pydantic-field 

type: Literal['remove_regex'] = 'remove_regex'

Do not change. Used for (de)serialization.

Our projects
SWE-rex SWE-ReX SWE-smith SWE-smith SWE-bench SWE-bench sb-cli sb-cli