[Bug] Cloud routines fail silently when GitHub MCP removed but system prompt still requires it

Bug Description
Title: Cloud routines that use GitHub silently break — the env stopped shipping the GitHub MCP but the system prompt still forces it

My scheduled cloud routine that bumps PR reviewers has been dead for days. Only noticed because it started DMing me an error twice a day.

The sandbox system prompt says: "You do NOT have access to the gh CLI, hub CLI, or direct GitHub API access. Instead, use the GitHub MCP server tools ... for ALL GitHub interactions."

Back on 2026-08-07 that was fine — the env auto-provided the mcp__github__* tools and the routine used them (session cse_017D4K537t3oBr9E29WGi8Ay). Around 2026-08-17 the env stopped providing the GitHub MCP, but the prompt still tells the model to use it. So now there's no GitHub tool at all. The model falls back to running gh, which isn't installed → gh: command not found → the routine fails every single run (cse_01WYwdF6yMguxmtLEgZn2acU, 2026-08-18).

And there's no way around it inside the sandbox. I tested installing gh at runtime and the model refused — correctly, because that same system prompt forbids it (cse_01UDdF18U2BvdcADjNyGW5EC).

So the bug is the contradiction: the prompt requires the GitHub MCP, the env no longer ships it, and the model is right to refuse every fallback. Anything touching GitHub just dies quietly.

Fix should be one of: keep bundling the GitHub MCP in the default cloud env, or stop injecting the "use the GitHub MCP for all GitHub" instruction when no GitHub MCP is attached. Either way, surface a real error instead of failing silent.

Env: env_015fSP9vhUALLyPZWkjaSi4D

Environment Info

• Platform: darwin
• Terminal: ghostty
• Version: 2.1.234
• Feedback ID: b69ef7ec-88ee-4d75-bd25-8a68af4cdc90

Errors

[]