AgentCrew 0.21.4

This release improves reasoning continuity across turns, adds background summaries for long tool results, and refreshes provider and model support.

Highlights

Preserve OpenAI Responses reasoning across turns. Encrypted reasoning state is saved with assistant messages and replayed on compatible follow-up requests. Reasoning summaries also appear as thinking content, including when a stream is cancelled.
Summarize tool results during context shrinking. An optional background service creates compact summaries for eligible tool results. Summaries are bounded, cached, and persisted for chats; when one isn’t available, AgentCrew uses the existing truncated-result fallback. Enable or disable the feature in global settings with tool_result_summary_enabled.
Make agent streams request-local. A shared AgentResponseStream interface gives local and remote agents a consistent iteration, cancellation, and message-finalization lifecycle.

Added GPT-6 Astra, Sol, and Luna to OpenAI Codex, along with new and updated models across Command Code, DeepInfra, Fireworks, OpenCode Go, and Together AI.
Removed CrofAI support following the provider’s discontinuation.
Made conversation metadata writes atomic and added tool-summary inheritance when conversations are forked.
Redacted encrypted reasoning payloads from debug and error message dumps.
Updated configuration and getting-started documentation.
Added tests covering reasoning persistence and replay, stream cancellation, tool summaries, and model-selection behavior.