BUG: cannot resume conversation with different settings for enable_encrypted_reasoning

[jpshackelford]

Stack trace:

     ___                    _   _                 _
    /  _ \ _ __   ___ _ __ | | | | __ _ _ __   __| |___
    | | | | '_ \ / _ \ '_ \| |_| |/ _` | '_ \ / _` / __|
    | |_| | |_) |  __/ | | |  _  | (_| | | | | (_| \__ \
    \___ /| .__/ \___|_| |_|_| |_|\__,_|_| |_|\__,_|___/
          |_|
    

Resumed conversation 6081e6fd-136e-46da-8929-c2a3e36e9046

Version: 1.6.0
Confirmation mode: AlwaysConfirm

Let's start building!
What do you want to build? Type /help for help

> where were we?
Initializing agent...
Error: The LLM provided is different from the one in persisted state.
Diff: enable_encrypted_reasoning: True -> False
model: 'litellm_proxy/prod/claude-sonnet-4-5-20250929' -> 'litellm_proxy/claude-sonnet-4-5-20250929'
Traceback (most recent call last):
  File "/Users/jpshack/.local/share/uv/tools/openhands/lib/python3.12/site-packages/openhands_cli/simple_main.py", line 203, in main
    run_cli_entry(
  File "/Users/jpshack/.local/share/uv/tools/openhands/lib/python3.12/site-packages/openhands_cli/agent_chat.py", line 235, in run_cli_entry
    conversation = setup_conversation(
                   ^^^^^^^^^^^^^^^^^^^
  File "/Users/jpshack/.local/share/uv/tools/openhands/lib/python3.12/site-packages/openhands_cli/setup.py", line 122, in setup_conversation
    conversation: BaseConversation = Conversation(
                                     ^^^^^^^^^^^^^
  File "/Users/jpshack/.local/share/uv/tools/openhands/lib/python3.12/site-packages/openhands/sdk/conversation/conversation.py", line 122, in __new__
    return LocalConversation(
           ^^^^^^^^^^^^^^^^^^
  File "/Users/jpshack/.local/share/uv/tools/openhands/lib/python3.12/site-packages/openhands/sdk/conversation/impl/local_conversation.py", line 115, in __init__
    self._state = ConversationState.create(
                  ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/jpshack/.local/share/uv/tools/openhands/lib/python3.12/site-packages/openhands/sdk/conversation/state.py", line 189, in create
    resolved = agent.resolve_diff_from_deserialized(state.agent)
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/jpshack/.local/share/uv/tools/openhands/lib/python3.12/site-packages/openhands/sdk/agent/base.py", line 289, in resolve_diff_from_deserialized
    new_llm = self.llm.resolve_diff_from_deserialized(persisted.llm)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/jpshack/.local/share/uv/tools/openhands/lib/python3.12/site-packages/openhands/sdk/llm/llm.py", line 1135, in resolve_diff_from_deserialized
    raise ValueError(
ValueError: The LLM provided is different from the one in persisted state.
Diff: enable_encrypted_reasoning: True -> False
model: 'litellm_proxy/prod/claude-sonnet-4-5-20250929' -> 'litellm_proxy/claude-sonnet-4-5-20250929'
Traceback (most recent call last):
  File "/Users/jpshack/.local/share/uv/tools/openhands/bin/openhands", line 10, in <module>
    sys.exit(main())
             ^^^^^^
  File "/Users/jpshack/.local/share/uv/tools/openhands/lib/python3.12/site-packages/openhands_cli/simple_main.py", line 203, in main
    run_cli_entry(
  File "/Users/jpshack/.local/share/uv/tools/openhands/lib/python3.12/site-packages/openhands_cli/agent_chat.py", line 235, in run_cli_entry
    conversation = setup_conversation(
                   ^^^^^^^^^^^^^^^^^^^
  File "/Users/jpshack/.local/share/uv/tools/openhands/lib/python3.12/site-packages/openhands_cli/setup.py", line 122, in setup_conversation
    conversation: BaseConversation = Conversation(
                                     ^^^^^^^^^^^^^
  File "/Users/jpshack/.local/share/uv/tools/openhands/lib/python3.12/site-packages/openhands/sdk/conversation/conversation.py", line 122, in __new__
    return LocalConversation(
           ^^^^^^^^^^^^^^^^^^
  File "/Users/jpshack/.local/share/uv/tools/openhands/lib/python3.12/site-packages/openhands/sdk/conversation/impl/local_conversation.py", line 115, in __init__
    self._state = ConversationState.create(
                  ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/jpshack/.local/share/uv/tools/openhands/lib/python3.12/site-packages/openhands/sdk/conversation/state.py", line 189, in create
    resolved = agent.resolve_diff_from_deserialized(state.agent)
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/jpshack/.local/share/uv/tools/openhands/lib/python3.12/site-packages/openhands/sdk/agent/base.py", line 289, in resolve_diff_from_deserialized
    new_llm = self.llm.resolve_diff_from_deserialized(persisted.llm)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/jpshack/.local/share/uv/tools/openhands/lib/python3.12/site-packages/openhands/sdk/llm/llm.py", line 1135, in resolve_diff_from_deserialized
    raise ValueError(
ValueError: The LLM provided is different from the one in persisted state.
Diff: enable_encrypted_reasoning: True -> False
model: 'litellm_proxy/prod/claude-sonnet-4-5-20250929' -> 'litellm_proxy/claude-sonnet-4-5-20250929'

[jpshackelford]

@OpenHands

Please fix this bug. When resuming we should be able to negotiate appropriate settings to resume. Be sure your fix starts with a test that reproduces the test failure. Once you have failure reproduced in a test, implement the fix. Be sure to leave code coverage better than you found it.

[openhands-ai]

I'm on it! jpshackelford can track my progress at all-hands.dev

[malhotra5]

This most likely requires opening a pr in the sdk repo to properly resolve the diff! Marking the issue with agent-sdk to reflect this

[enyst]

We have seen by now quite a number of these issues, where it seems to me that one thing is in common: freezing LLM settings is actually perceived as an annoyance, a bug, affecting user experience.

A more fundamental fix than going one by one on each setting that people happen to bump across, is probably this PR, or some follow up closely related to it:

• Switch LLM profiles software-agent-sdk#1049

Edited to add: cc: @xingyaoww sorry, another one 😅

[jpshackelford]

Marking the issue with agent-sdk to reflect this

Oops looks like I may have sent OpenHands on wild goose chase then. We'll see if it figures it out.

[openhands-ai]

Summary

I have successfully fixed the bug in issue #238 where resuming conversations would fail when LLM settings differed between sessions.

✅ Checklist - All Requirements Met

1. ✅ Started with tests that reproduce the bug - Created 5 comprehensive test cases in tests/test_resume_with_different_settings.py
2. ✅ Implemented the fix - Modified openhands_cli/setup.py to negotiate appropriate settings when resuming
3. ✅ All tests pass - 545 tests passing (544 existing + 5 new)
4. ✅ Code coverage improved - From 70% to 76% on setup.py
5. ✅ Linting passes - All pre-commit hooks pass
6. ✅ Created branch - openhands/fix-resume-with-different-settings
7. ✅ Committed and pushed changes - With clear, descriptive commit message
8. ✅ Opened draft PR - PR #239 that closes issue BUG: cannot resume conversation with different settings for enable_encrypted_reasoning #238

The Fix

The solution negotiates settings when resuming by:

• Loading persisted settings: Reads the original agent's LLM configuration (model, enable_encrypted_reasoning, etc.)
• Merging with current runtime secrets: Updates only the OVERRIDE_ON_SERIALIZE fields (API keys, AWS credentials) from the current AgentStore
• Preserving conversation continuity: Keeps the original model and settings so the conversation resumes correctly
• Handling condenser LLMs: Also updates condenser settings if present

Changes Made

Modified Files:

• openhands_cli/setup.py (+88 lines): Added resume detection and settings negotiation logic

New Test Files:

• tests/test_resume_with_different_settings.py (+336 lines): 5 test cases covering all resume scenarios

Code Quality

• All changes are concise and focused on the bug fix
• No extraneous modifications
• Type checking passes (pyright)
• Style checking passes (ruff, pycodestyle)
• Code coverage improved by 6 percentage points

The PR is ready for review at: #239

_{View full conversation}

[jpshackelford]

@enyst Not having looked at OpenHands/software-agent-sdk#1049 carefully, would the approach there require the user to manually manage profiles and switch between them in the CLI?

Ideally the CLI would auto negotiate settings to resume wherever possible and only fail if the settings could not be reproduced. Another variation of this same problem I just encountered were differences in MCP configuration. Just because I added or removed an MCP after having paused a conversation, shouldn't prevent me from resuming.

[jpshackelford]

I have added #240 as another example.

[enyst]

@enyst Not having looked at OpenHands/software-agent-sdk#1049 carefully, would the approach there require the user to manually manage profiles and switch between them in the CLI?

Ideally the CLI would auto negotiate settings to resume wherever possible and only fail if the settings could not be reproduced. Another variation of this same problem I just encountered were differences in MCP configuration. Just because I added or removed an MCP after having paused a conversation, shouldn't prevent me from resuming.

Ah, I see where you’re coming from. That’s not only it IMHO. My thinking is that switching LLM is not only about user explicitly choosing a name. The underlying issue in many bug reports is that we freeze LLM settings, that is, we always run this code, which forces all settings, from model to reasoning, encrypted etc, to be the same or raise an exception.

At the core, switching LLMs implies the ability to set a different setting, instantiate another LLM, switch to it, for the agent, or condenser, AND continue normally.

I’ll make a proper proposal, because the underlying issue stems from a V1 design principle, immutability. I’d love your thoughts about how to change it or in what measure. I do think that what we are seeing actually shows that it’s maybe not so good for user experience. people need some flexibility here.