trimReplayDanglingAssistantToolCalls only collected tool results that
immediately followed the assistant tool-call message. Models such as
gpt-5.3-codex-spark may emit the function_call item before the
explanation text within one response, so history replay order becomes
assistant[tool_call] -> assistant[text] -> tool[result]. The call was
misjudged as dangling and stripped while the tool result survived,
producing a function_call_output without a matching function_call that
the Responses API rejects with 400.
- widen the response collection window to skip interleaved plain
assistant text messages, and drop orphan tool results in the same pass
- synthesize a placeholder function_call (or drop the output when the
tool name is unknown) in normalizeOpenAIResponsesInput so conversations
already persisted with corrupted history can resume
- Added flushing of assistant text during provider completion to ensure no output is lost on transport failure.
- Updated checkpoint blob synchronization tests to validate behavior under various conditions, including terminal and non-terminal states.
- Introduced new functions for managing checkpoint terminal actions, improving clarity and maintainability of the code.
- Implemented additional tests for imported blob handling and conversation state restoration, ensuring robustness in data integrity across operations.
- Added support for checkpointing phases and blob management in the forwarder.
- Introduced new types and methods for handling checkpoint blobs, including queuing and publishing checkpoints.
- Enhanced the projector to build checkpoint projections with content-addressed blobs.
- Implemented tests to ensure proper checkpoint blob synchronization and handling of cancellation scenarios.
Store checkpoint turns as validated blobs and restore prefetched parent turns so forked conversations retain their replay history and rewind prefix.
Co-authored-by: Cursor <cursoragent@cursor.com>
上玄