Payload Structure for Agent Hooks | Generated by AI

Home 2026.05 English 中文

---

Here’s the complete picture. The payload is constructed in agent/shell_hooks.py:465-481 (_serialize_payload). For your post_llm_call hook, this is what gets piped to stdin:

{
  "hook_event_name": "post_llm_call",
  "tool_name": null,
  "tool_input": null,
  "session_id": "sess_abc123",
  "cwd": "/Users/lzwjava/projects/jekyll-ai-blog",
  "extra": {
    "user_message": "what the user asked",
    "assistant_response": "the agent's final response text",
    "conversation_history": [
      {"role": "system", "content": "..."},
      {"role": "user", "content": "..."},
      {"role": "assistant", "content": "..."}
    ],
    "model": "deepseek/deepseek-v4-pro",
    "platform": "cli"
  }
}

What goes where:

Field	Source
hook_event_name	Always "post_llm_call"
tool_name	Always null for non-tool events
tool_input	Always null for non-tool events
session_id	kwargs.get("session_id")
cwd	Path.cwd() at render time
extra	Everything else not in _TOP_LEVEL_PAYLOAD_KEYS ({"tool_name", "args", "session_id", "parent_session_id"})

The extra keys come from the invocation in run_agent.py:15804-15812:

_invoke_hook("post_llm_call",
    session_id=self.session_id,
    user_message=original_user_message,       # → extra
    assistant_response=final_response,        # → extra
    conversation_history=list(messages),      # → extra
    model=self.model,                          # → extra
    platform=getattr(self, "platform", None) or "",  # → extra
)

Other hooks have different payloads. Quick reference:

Hook	tool_name	tool_input	Notable extra keys
pre_tool_call	"terminal" etc.	{"command": "..."}	task_id, tool_call_id
post_tool_call	"terminal" etc.	{"command": "..."}	task_id, result, duration_ms, tool_call_id
pre_llm_call	null	null	user_message, platform, conversation_history
post_llm_call	null	null	user_message, assistant_response, conversation_history, model, platform
on_session_start	null	null	platform, session_id
on_session_end	null	null	platform, session_id
subagent_stop	null	null	child_role, duration_ms, result_summary

So in your script, you can access extra.assistant_response to show a snippet of the answer, extra.model to know which model was used, or extra.platform to know where it ran. Want me to update the notification to include any of those?

---

Back Donate