v2.20.0
What's changed
- Automatic provider quota-exhaustion wait and resume: When an OAuth provider (such as Codex, GitHub Copilot, or Grok) or any configured model hits rate-limiting or quota exhaustion with a known reset window, the agent enters an interruptible quota-waiting state instead of aborting. Displays an in-transcript countdown timer and model banner, and automatically resumes work once the window resets.
- Multimodal image tool returns for MCP: MCP tools returning image MIME types (
ImageContentand image embedded resources) deliver structuredImageDataBlockparts directly to vision models while preserving concise textual summaries in transcripts and persistent history. - Default new user model updated to
opencode:big-pickle: New installations and unconfigured profiles default toopencode:big-picklefor out-of-the-box free inference without token indirection. - Consolidated subagent orchestration: Lead agent subagent orchestration is consolidated onto the unified
delegatetool, removing deprecated legacyteam_*tools across the backend and prompts. - Reliable plan mode switching on approval: Proposed plan approval reliably switches the session to Code mode before dispatching implementation prompts, showing an in-button loading spinner during transition and preventing mode mismatch rejections.
- Pinned markdown table copy button: Markdown table copy buttons remain pinned in the top-right corner of table containers during horizontal scroll instead of drifting across cells.
Full changelog: v2.19.0...v2.20.0