v0.32.0
Release date: 2026-09-12
Added
- Added an opt-in script to prepare existing session databases for local-shell provenance tracking, with a SQLite-safe backup before changes. Historical records remain unclassified; the script never runs automatically (457c9148).
Changed
- Renamed the running shell tool's “Stop command” button to “Interrupt” to clarify that it stops the command rather than the assistant turn (8a8e3e95).
Fixed
- Recover once from ChatGPT's code-less HTTP 400
Invalid previous_response_iderror by replaying the full current-turn context, including reasoning, tool calls, and completed results, without rerunning tools. Successful recovery resumes normal response-ID continuation (d8840418). - Preserve failed user turns and restore unfinished prompts, intermediate responses, and completed tool exchanges when reopening a saved session, even when completed-turn tool history is disabled. Replay respects checkpoint ordering, retries, forks, attachments, and compaction boundaries (9df5d6f5).
- Keep direct UI shell commands and their output out of model context, compaction, sub-agent snapshots, and prompt enhancement using explicit persisted provenance, while retaining visible history and ordinary bang-prefixed prompts (457c9148).