Fix agent reconciliation to allow agent_context updates

[jpshackelford]

Problem

When resuming a conversation from a different directory or switching between ACP and regular CLI mode, the agent state validation fails with a ValueError showing differences in:

• Skill paths (absolute paths change with working directory)
• Skills list (ACP mode adds user_provided_resources skill)
• System message suffix (contains working directory)

This prevents users from resuming conversations across different environments.

Solution

Modified resolve_diff_from_deserialized() in agent/base.py to reconcile agent_context during conversation resume, similar to how LLM configuration is already handled. The reconciliation now:

1. Updates agent_context to use the current environment's values
2. Allows skill changes between sessions
3. Handles working directory changes gracefully

Testing

Added comprehensive test test_conversation_restart_with_different_agent_context() that simulates:

• Creating a conversation in ACP mode with specific skills
• Resuming the conversation in regular CLI mode with different skills
• Verifying the agent_context is properly reconciled

Test validates the fix

Without the fix (code commented out), the test fails with the exact error from the bug report:

FAILED - ValueError: The Agent provided is different from the one in persisted state.
Diff: agent_context:
  skills:
    0:
      content: 'Information about the current project' -> 'You may encounter...'
      name: 'project_info' -> 'user_provided_resources'
  system_message_suffix: 'You current working directory is: /Users/jpshack' -> '.../all-hands'

With the fix, the test passes along with all other reconciliation tests.

All 44 tests in the cross directory pass ✓

Impact

• ✅ Users can now resume conversations from any directory
• ✅ Switching between ACP and regular CLI mode works seamlessly
• ✅ Backward compatible - existing conversations continue to work
• ✅ No breaking changes to the API

[github-actions]

Coverage Report •

File	Stmts	Miss	Cover	Missing
openhands-sdk/openhands/sdk/agent
base.py	159	18	88%	154, 160, 204–205, 211–213, 226, 234–235, 269, 316, 323, 336, 373–374, 384–385
TOTAL	12427	5612	54%

[xingyaoww]

Thank you!