Switch to Claude Opus 4.6 (Thinking)

[!WARNING] INVOCATION GATE: Do NOT execute this workflow on your own initiative. It may ONLY be invoked in two cases: (1) another workflow's step explicitly references it (e.g., "Execute the steps in the switch-to-claude-opus-4.6-thinking workflow"), or (2) the user explicitly types switch-to-claude-opus-4.6-thinking.md or equivalent. If the user casually says "have Claude review this", do NOT auto-invoke this workflow — the user will switch manually. When in doubt, ask.

[!CAUTION] The automated engine switch requires the Antigravity UI to be visible and in the foreground. ❗🔴 Make absolutely SURE that NO other Antigravity windows (such as Agent Manager or other workspaces) are open, or the switch sequence WILL misfire. 🔴❗ If the window is hidden behind Chrome or minimized, the automation script will fail. Do not click away during the handover!

1. Verify Current Engine: Check if you are already Claude. If so, inform the user and abort. Check if a USER_SETTINGS_CHANGE appeared earlier in this conversation to determine your current engine. If none exists, check your self-knowledge (training identity). If still uncertain, ask the user.
2. Verify CDP debugging port: Execute python switch_to_other.py --check synchronously.
   • If the port check succeeds (exit code 0):
     • Play audio alert: Run python audio_alert.py "Switching engine".
     • Run CDP switch script (Background): Execute python switch_to_other.py "Claude Opus 4.6 (Thinking)" with WaitMsBeforeAsync=5 to send it to the background.
     • End Turn immediately: Directly print the following safety warning and "Switching engines..." message to the chat screen: "\n\n❗🔴 Make absolutely SURE that NO other Antigravity windows (such as Agent Manager or other workspaces) are open, or the switch sequence WILL misfire. 🔴❗\n\nSwitching engines..." (Do NOT wait for the python script). The script will instantly pause the auto-clicker via the __switchHandoverActive flag, wait 15 seconds for the UI to finish syncing files, switch the engine, send a resume prompt, and unpause the auto-clicker.
   • If the port check fails (exit code 1):
     • Directly print the manual instruction alert: 🚨 CDP Debugging Port Offline — Manual Switch Required 🚨 "I cannot switch the model automatically because the IDE's remote debugging port is offline. Please manually switch the model to Claude Opus 4.6 (Thinking) in the top-right model selector dropdown of the IDE, type a message (e.g. 'resume'), and click send to continue the workflow under the new engine."
     • End Turn immediately (do NOT run the background script).

I. New Engine Wakes Up — Verification via Metadata

5. Check for USER_SETTINGS_CHANGE: When the new engine wakes up from the auto-resume message, its very first action is to check whether a USER_SETTINGS_CHANGE block was injected into the metadata of the message it just received, and specifically verify that it changed the model to "Claude Opus 4.6 (Thinking)".
   • Present and verified → The switch succeeded. Proceed immediately to assigned tasks within the parent workflow.
   • Absent or wrong model → The direct switch failed (likely because you were already on Claude). Retry with trampoline: Run python switch_to_other.py "Claude Opus 4.6 (Thinking)" --trampoline with WaitMsBeforeAsync=5 to send it to the background. The trampoline will first hop to Gemini 3 Flash (sending a "." to commit the change), wait for Flash to finish, then switch to the actual target. Print the same safety warning and "Switching engines..." message, then end your turn.
6. Second verification: When the engine wakes up again after the trampoline retry, check for USER_SETTINGS_CHANGE again.
   • Present and verified → The switch succeeded. Proceed immediately to assigned tasks.
   • Absent or wrong model → The switch failed twice. Stop all work, run python audio_alert.py "Engine switch failed", and warn the user that the engine switch failed.