[agent] Steering commands not working -- should be editable and injected at next agent breakpoint

[acertainKnight]

Feature / Bug

The UI should support "steering commands" -- user inputs that get queued and injected at the next natural breakpoint in the agent's processing stream. These commands should be:

1. Editable by the user after typing them, up until the moment they are actually sent/injected
2. Injected at the next breakpoint in the agent stream (e.g., between tool calls, between reasoning steps), not requiring the agent to fully finish its current turn

Current behavior

This functionality is not working. It's unclear whether the feature is partially implemented but broken, or not yet implemented.

Expected behavior

1. While the agent is mid-response (streaming, executing tool calls, etc.), the user can type a steering command in the input field
2. The command sits in an editable queue -- the user can modify or cancel it before it fires
3. At the next natural breakpoint in the agent's execution loop, the steering command gets injected as input
4. The agent processes the steering command and adjusts its behavior accordingly

Use case

This is critical for interactive development and debugging workflows. When the agent is going down the wrong path mid-response (e.g., making unnecessary tool calls, generating a long unwanted output, or misunderstanding the task), the user needs a way to redirect without:

• Waiting for the full response to complete
• Using the stop button (which is broken per [agent] Stop agent stream command does not fully halt agent execution -- agent freezes then errors #52)
• Losing the agent's current execution context

Context

• Reported by user during onboarding session
• Closely related to [agent] Stop agent stream command does not fully halt agent execution -- agent freezes then errors #52 (stop stream not working) -- both are about mid-stream user control over agent execution
• The editability requirement distinguishes this from a simple "inject message" -- the user needs to be able to draft and revise the steering command while the agent is still running

[acertainKnight]

Clarification from user: The implementation can be simpler than originally described. It's essentially a message queue where:

• User can type/queue messages while the agent is mid-stream
• Only the last element in the queue needs to be editable (i.e., the one that hasn't been injected yet)
• Once a message is injected at a breakpoint, it's locked -- only the pending/unsent one stays editable

This is closer to a standard message queue with an editable draft at the tail, not a fully revisable multi-message buffer.