Feature Request

Problem:
Currently, steer mode only injects messages at tool boundaries. If the agent is mid-generation or waiting on a long tool call, the user's message queues until the next boundary. This defeats the purpose of real-time steering.

Desired behavior:
True mid-stream injection — when a user sends a message while the agent is actively generating:

1. Pause/abort the current generation cleanly
2. Inject the user's message into context
3. Resume generation with awareness of the new input

Prior art:
Manus AI has this feature and it's reportedly very useful for redirecting agents mid-task without losing context or having to wait.

Use case:
User starts a complex task, realizes they want to add constraints or redirect, sends a follow-up message. Agent should see it immediately and adjust, not finish the entire original task first.

Current workaround:

• Abort triggers (stop, abort) kill the run entirely
• steer mode only works at tool boundaries
• No way to inject mid-thinking/streaming

Would love to see this as an enhancement to queue modes or as a new soft-steer option.

Filed by: Marcus Wylde (via OpenClaw agent)
cc: @marcuswylde