v2026.7.22

Breaking Changes

Added

• Added configurable Claude text tool-call recovery across both ModelRuntime streaming entry points, models.json definitions/overrides, persisted sessions, and isolated Anthropic/OpenAI mock-loop QA.
• Added the look_at vision-delegation tool with model capability gating and configurable settings/context support.
• Added hot reload for configuration resources, including a documented extension protocol, change events, hash-based self-write suppression, and safe lifecycle/mode handling.
• Added /btw for parallel side questions, progressive web-fetch feedback, and incremental terminal rendering for tool activity and results.
• Added Alibaba Cloud Model Studio Token Plan provider setup and model support.

Changed

Fixed

• Fixed the interactive render hot path re-materializing the entire session every frame: SessionManager.getEntries(), no-arg getBranch(), and getSessionName() are now memoized behind a monotonic mutation counter, eliminating repeated deep copies of large sessions at frame rate.
• Fixed OpenAI Responses replay to omit invalid custom tool-call item IDs.
• Fixed compaction budgeting and diagnostics, including bounded summary streams and preservation of queued input after failures.
• Fixed MCP lazy-start races so registered tools and instructions reliably rehydrate.
• Fixed queued steering/follow-up delivery and tool-result reveal behavior across session rebinds.

Removed