v2026.7.4

Added

• Added ctx.updateToolHookStatus() so tool_call/tool_result extension handlers can report what they are doing in the live "Running PreToolUse/PostToolUse hook" TUI status row
• Added inherited generated model catalog refreshes from models.dev, including newly listed models such as Kimi K2.7 Code for GitHub Copilot and Fable 5 providers (#6256).

Changed

Fixed

• Fixed TUI screen corruption from external stdout writes: console.log from libraries or extensions while a TUI owns the terminal (interactive mode, startup dialogs, config selector) is now hidden from the screen and appended, redacted, to the debug log — matching the existing stderr guard.
• Fixed the live tool hook status row showing a generic running builtin:hooks label instead of the running command hook's configured statusMessage (falling back to its sanitized command text)
• Fixed startup model resolution to await available-model lookups before selecting defaults.
• Fixed pnpm self-update failures to show a prune hint when package-manager metadata blocks a self-update (#6279 by @rajp152k).
• Fixed the edit tool schema to allow model-invented extra replacement fields instead of rejecting otherwise valid edits (#6278).
• Fixed inherited OAuth device-code polling to honor the server-provided slow_down interval so GitHub Copilot login recovers instead of appearing to hang when polls arrive early (#6187).
• Fixed inherited OpenAI Codex WebSocket sessions to rotate cached connections before the backend's 60-minute limit, avoiding connection-limit failures on long sessions (#6268).
• Fixed inherited retry classification for Cloudflare 524 timeout responses (#6239).

Removed

• Removed default attribution headers from Vercel AI Gateway requests.