mirror of
https://github.com/Significant-Gravitas/AutoGPT.git
synced 2026-02-04 03:45:12 -05:00
72ae2e9504
Consolidates context window management into a single reusable function in prompt.py that supports both truncation-only and summarization strategies. Changes: - Add CompressResult dataclass with compression metadata - Add compress_context() async function with configurable strategies: 1. Content truncation (always) 2. LLM summarization (if client provided) 3. Progressive message reduction 4. Middle-out deletion (fallback) 5. First/last trim (last resort) - Add _ensure_tool_pairs_intact() with full orphan tool_call_id handling - Add _summarize_messages_llm() for LLM-based summarization - Migrate llm_call() to use compress_context(client=None) for truncation-only - Remove compress_prompt() - replaced by compress_context() - Refactor chat service to use unified compress_context() - Fix: use async context manager for AsyncOpenAI client to prevent socket leaks - Fix: only create OpenAI client when api_key is provided - Fix: handle OpenAI-style tool messages (role=tool with string content) in truncation - Fix: use client.with_options(timeout=...) for per-request timeout (OpenAI SDK v1.x) - Add comprehensive tests for all new functions The unified function allows: - Chat service: client=openai_client for summarization - LLM blocks: client=None for truncation-only