Changelog¶

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[0.0.5] - 2026-03-21¶

LimitWarnerProcessor — standalone history processor that injects warning SystemPromptParts as request, context-window, or total-token limits approach (#10, by @Gby56)
create_limit_warner_processor() factory function
WarningOn type alias for selecting warning categories

on_before_compress callback — called before compression with messages and cutoff index
on_after_compress callback — called after compression; return a string to re-inject into context as SystemPromptPart
Continuous message persistence via messages_path — saves all messages to JSON on every call
Guided compaction — focus parameter on compact() and request_compact() for topic-focused summaries
request_compact(focus) method — deferred manual compaction
compact(messages, focus) method — direct LLM-based compaction
max_tokens auto-detection from genai-prices via resolve_max_tokens(model_name)
model_name parameter for auto-detection
Async token counting — TokenCounter now accepts sync or async callables (#6)
async_count_tokens() helper function
BeforeCompressCallback, AfterCompressCallback type aliases
6 runnable examples in examples/

max_tokens default: 200_000 → None (auto-detect, fallback 200K)
keep default: ("messages", 20) → ("messages", 0) (only summary survives)
Validation: 0 now allowed for messages/tokens (negative still rejected)

ContextManagerMiddleware — dual-protocol middleware (history processor + AgentMiddleware)
Shared cutoff algorithms — _cutoff.py module
ModelType type alias