0.170.0

• Fix compaction reusing an old summary when the model returns text without calling compact_chat; report failure instead. #629
• Preserve hosted web search and image generation in Codex parallel requests without changing Lite routing. #628
• Fix auto-approved tool calls waiting on an earlier call's manual approval; they now start right away while approvals are still asked one at a time.
• Fix calls to the same MCP server running one at a time across all chats; only stdio writes are serialized now.
• Support MCP progress notifications: each progress report resets the tool-call timeout, and timed-out calls are cancelled on the server.
• Forward MCP tool progress to clients: toolCallRunning is re-sent with a progress field (progress, total, message), at most once per second per call.