feat(hook): include LLM response and stop reason in Stop hook payload

What feature would you like to see?

Summary

Currently, the Stop hook fires at the end of every agent turn but its payload is minimal:

{
  "hook_event_name": "Stop",
  "session_id": "abc123",
  "cwd": "/path/to/project",
  "stop_hook_active": false
}

There is no way for a hook to inspect what the agent actually said or why the turn ended. This limits use cases such as:

• Post-turn logging / auditing — recording the agent's final response to an external system
• Conditional approval — blocking the session if the agent's response contains certain patterns (e.g., "I am not sure", "I cannot help with that")
• Metrics / analytics — tracking turn outcomes (natural language vs. tool calls) without parsing wire logs

SubagentStop already provides a response field (the subagent's final summary). It would be consistent and useful for the root-agent Stop hook to expose the same class of information.

Proposed behavior

Enhance the Stop hook payload with two new fields:

Example payload after the change:

{
  "hook_event_name": "Stop",
  "session_id": "abc123",
  "cwd": "/path/to/project",
  "stop_hook_active": false,
  "stop_reason": "no_tool_calls",
  "response": "The fix is complete. I updated src/main.py and added a regression test."
}

Additional information

• Precedent: SubagentStop already includes "response" (truncated to 500 chars). This change brings parity to the root-agent lifecycle.
• Backwards compatibility: The new fields are optional and default to empty strings. Existing hooks that do not read them will continue to work unchanged.
• Implementation sketch: KimiSoul.run() already receives a TurnOutcome from _turn(). The final_message (a Message object) can have its text extracted via extract_text(" ").strip(), identical to how UserPromptSubmit extracts user text.

Start in KimiSoul.run(), following the TurnOutcome returned by _turn() and the existing Stop-hook payload construction; inspect final_message and the extract_text(" ").strip() precedent. Done means Stop payloads include response and stop_reason with the specified values and empty-string defaults, while existing fields remain unchanged.