feat(hitl): wire up human-in-the-loop approval for system terminal co…

[Nithwin]

Description

This PR implements a critical safety feature: Human-in-the-Loop (HITL) Approval for system terminal commands. Previously, the agent could autonomously execute bash commands via the CommandTool in the background, which posed a security risk.

This update intercepts any run_command tool call, pauses the agent's background goroutine using Go channels, and triggers an interactive confirmation modal in the UI.

Key Changes

• system/command.go: Upgraded CommandTool to accept an ApprovalCallback func(cmd string) bool. If the user rejects, the tool returns a graceful rejection error to the AI so it can pivot its strategy.
• defaults/defaults.go: Updated RegisterAll to pass the approval callback to the system tools.
• chat/messages.go: Introduced ApprovalRequestMsg to bridge the background agent process with the main Bubble Tea event loop.
• chat/update.go & chat/model.go: Wired up a concurrency-safe channel m.approvalChan. It blocks standard inputs while waiting for the user to press y or n.
• chat/view.go: Replaces the standard chat input box with a yellow-bordered warning modal displaying the exact bash command waiting for approval.
• cmd/chat.go: Added headless fallback using fmt.Scanln for standard (y/N) terminal prompts outside the TUI.

Why this is necessary

Provides the "Aider/Claude Code" pair-programming experience. It ensures complete developer control over destructive shell commands (like rm, git push, or npm install) and prevents the LLM from getting stuck in an infinite command-execution loop.

How to Test

1. Run go run ./cmd/windmist
2. Ask the agent: "Create a folder called 'test-folder' using terminal commands"
3. Verify that the UI blocks execution and displays a yellow warning box.
4. Press n and verify the agent receives the rejection and apologizes.
5. Press y and verify the command executes successfully.