Agent: optional cap on model requests per turn, then ask to continue

[ouster]

Problem

With local models on an OpenAI-compatible provider, an agent turn can keep going far past the point of being useful. #66686 stops the two clearest shapes, an identical tool call or a repeated reasoning passage. It does not catch a turn that makes progress-looking calls that are all different: rereading files with new ranges, or rerunning a command with small edits.

The current choice is between two settings. tool_permissions.default: "confirm" asks on every call, which is safe but slow. "allow" asks on none, so a long turn runs unattended.

Proposal

A setting such as agent.max_requests_per_turn, off by default. When a turn reaches N model requests, the agent pauses and shows "This turn has made N requests. Continue / Stop" in the thread. Continue allows another N.

• It counts model requests, not tool calls, so parallel tool calls in one response count once.
• It is separate from tool permissions, so it still applies under allow.
• VS Code has the same idea as chat.agent.maxRequests, which pauses and asks to confirm.

Questions

1. Would a setting like this be accepted? If so, should it be a number or off, and where should the prompt appear?
2. Should it share the stop notice and UI from agent: Stop a turn that repeats a tool call or a reasoning passage #66686, or stay separate?

I'm happy to open a PR once the shape is agreed.