LlmAgent.clone() keeps the original's sub-agent tools and finish_task #7894

🔴 Required Information

Describe the Bug:
LlmAgent.model_post_init adds a tool for each single_turn or task sub-agent (_SingleTurnAgentTool / _TaskAgentTool), and FinishTaskTool when the agent itself is in task mode. LlmAgent.clone() goes through BaseAgent.clone(), which uses model_copy. model_copy does not run model_post_init again, and tools is only shallow-copied, so the clone keeps the original's tool objects:

• The clone's sub-agent tools still wrap the original's sub-agents, not the clone's. A single_turn call from the clone runs the original sub-agent, so changes made to the clone's sub-agents are ignored.
• clone(update={"sub_agents": [...]}) does not add tools for the new sub-agents, and keeps the tools for the old ones. The model cannot call the new sub-agents. Calling an old one runs an agent that is no longer in the tree (single_turn), or fails with WorkflowConfigurationError: Task target agent '...' not found. (task).
• clone(update={"tools": [...]}) drops the sub-agent tools, so those sub-agents cannot be called at all.
• clone(update={"mode": ...}) or clone(update={"output_schema": ...}) leaves finish_task as it was on the original.

Steps to Reproduce:

1. Run the code below.

Expected Behavior:
A clone has the same tools it would get if it were constructed with its own fields: one tool per single_turn / task sub-agent of the clone, and finish_task matching the clone's mode and output schema.

True
False
['other'] ['other']
['helper']

Observed Behavior:

False
True
['other'] ['helper']
[]

Environment Details:

• ADK Library Version (pip show google-adk): 2.13.0 (also on main at e9e8f61)
• Desktop OS: macOS
• Python Version (python -V): 3.14.7

Model Information:

• Are you using LiteLLM: No
• Which model is being used: N/A (no model call is needed to reproduce)

🟡 Optional Information

Additional Context:
Running a clone through InMemoryRunner with a mocked model shows the same thing. When the clone's single_turn sub-agent has a changed instruction, the sub-agent's model still receives the original instruction.

task delegation itself looks up the target with root_agent.find_agent(fc.name) in _dispatch_task_fc, so a plain clone() of a task sub-agent still reaches the clone's sub-agent. The update cases above are where task breaks.

Minimal Reproduction Code:

from google.adk.agents import LlmAgent

helper = LlmAgent(name="helper", description="Helps", mode="single_turn")
root = LlmAgent(name="root", sub_agents=[helper])

clone = root.clone()
print(clone.tools[0].agent is clone.sub_agents[0])
print(clone.tools[0].agent is helper)

other = LlmAgent(name="other", description="Other", mode="single_turn")
clone = root.clone(update={"sub_agents": [other]})
print([a.name for a in clone.sub_agents], [t.name for t in clone.tools])

clone = root.clone(update={"tools": []})
print([t.name for t in clone.tools])

How often has this issue occurred?:

• Always (100%)