fix(sdk): disable switch_llm by default (#16442)

[KHARSHAVARDHAN-eng]

HUMAN:

• A human has tested these changes.

AGENT:

---

Why

Fixes #16442 by disabling switch_llm by default in the SDK agent settings and profiles, and annotating SwitchLLMTool with destructiveHint=True to require user confirmation when enabled.

Summary

• Changed enable_switch_llm_tool default value from True to False in OpenHandsAgentSettings and AgentProfile.
• Annotated SwitchLLMTool with destructiveHint=True so model switching operations trigger confirmation flows under policies requiring user approval.
• Updated SDK unit tests in test_switch_llm.py, test_settings.py, test_resolver.py, and test_agent_profile.py.

Issue Number

Fixes #16442

How to Test

Run SDK unit tests:

poetry run pytest software-agent-sdk/tests/sdk/tool/test_switch_llm.py software-agent-sdk/tests/sdk/test_settings.py software-agent-sdk/tests/sdk/profiles/test_resolver.py software-agent-sdk/tests/sdk/profiles/test_agent_profile.py

[enyst]

All the git changes don't seem related to the PR, was this unintentional?

[enyst]

Hey, thank you for the proposal! I'm not sure that this is the root cause, or the right fix.
Also please see inline a little comment on the extraneous changes.

Could you show logs where the agent switches LLM and that caused a problem?

[KHARSHAVARDHAN-eng]

Hi @enyst, thank you for the review!

Here is a summary of the updates made to address your feedback:

1. Removed Extraneous Git Changes:

   • Rebased fix/disable-switch-llm-tool-default-16442 onto latest main and dropped the unrelated commit (84df1d5), removing all /api/git/diff files from this PR.
   • Pushed the updated branch so the PR now strictly contains only the 7 files relevant to enable_switch_llm_tool.

2. LLM Switching Rationale (#16442):

   • When enable_switch_llm_tool defaulted to True, any environment with saved LLM profiles automatically exposed SwitchLLMTool to the agent.
   • During task execution, models occasionally invoke switch_llm unprompted when encountering difficult steps, changing the active model/credentials mid-conversation without user intent.
   • Setting enable_switch_llm_tool default to False ensures model-switching remains an explicit user opt-in, while destructiveHint=True guarantees user confirmation is prompted when enabled under active verification policies.

All 207 unit tests (test_switch_llm.py, test_settings.py, test_resolver.py, test_agent_profile.py) and ruff linter/formatter checks pass cleanly.

Let me know if any other required changes has to be made!! Thanks!!!