SlidingWindowTransformer

Transformer that maintains a sliding window of messages to manage context size.

Applies windowing in both transformBeforeLlmCall and transformAfterIteration to ensure bounded memory and context window usage.

Tool Call/Result Grouping

LLM APIs (OpenAI, Anthropic, Google, etc.) require that every ToolResultMessage must be preceded by an AssistantMessageWithToolCalls containing the corresponding tool call ID. This is a universal constraint across all major providers.

When truncation occurs, the sliding window might cut between an AssistantMessageWithToolCalls and its ToolResultMessages, leaving "orphaned" tool results that would cause API errors like:

"messages with role 'tool' must be a response to a preceding message with 'tool_calls'"

This transformer automatically drops any orphaned ToolResultMessages that appear at the start of the non-system message sequence after truncation.

Example

Before truncation (7 messages):

[System, User, Assistant+Calls(1,2), ToolResult(1), ToolResult(2), Assistant+Calls(3), ToolResult(3)]

After naive takeLast(4) with 1 system message (INVALID):

[System, ToolResult(2), Assistant+Calls(3), ToolResult(3)]
         ↑ ORPHAN - no parent Assistant+Calls for call_2

After fix (VALID):

[System, Assistant+Calls(3), ToolResult(3)]
         ↑ Valid start - parent exists for ToolResult(3)

Parameters