Feature: Real-time Steering (Mid-generation prompting / Coaching)

[shdomi8599]

Describe the feature you'd like

I would like the ability to provide "steering" advice or corrections to the Continue agent while it is in the middle of generating a response or executing an autonomous loop, without having to stop/cancel the current generation.

How it would work:

1. While the AI is streaming a response, the chat input box remains active (or a "Steer" button appears).
2. The user types a correction (e.g., "Actually, use the V2 API for this, not V1").
3. This message is queued.
4. Continue finishes its current stream/tool call smoothly.
5. Before Continue makes its next LLM request in the loop, it injects the queued user message into the context.
6. The AI acknowledges the new instruction and course-corrects.

Why is this needed?

When watching the AI write code, developers often spot an incorrect assumption early on. Stopping the generation abruptly loses the current context and forces a restart. A "steering queue" allows true pair programming, where the human can gently nudge the AI back on track without breaking its momentum.

Note: I am the author of @steer-agent/core and am currently building this natively into Roo Code. I think Continue's architecture (especially with its powerful autocomplete and chat features) would be a perfect fit for this paradigm.

Proposed Solution

Implement a simple SteeringQueue that intercepts the message flow between the UI and the Agent Loop, injecting pending advice just before the next execute or LLM prompt generation step.

[shdomi8599]

I'd like to claim this issue. I have experience building this architecture (@steer-agent/core) and would love to contribute a native integration to Continue. Please let me know if I can go ahead and open a draft PR.