[Bug]: Resuming a conversation with restored stats can desync LLM.metrics and Telemetry.metrics, breaking token/cost accounting

[ShreySatapara]

Is there an existing issue for the same bug?

• I have searched existing issues and this is not a duplicate.

Bug Description

I ran into an issue while resuming a conversation from persisted state.

1. Run the conversation using conversation.run()
2. Whenever it encounter's an error due to API rate limit or some other system issues. It stores the cost tracking at that point in base_state.json in persistnat dir.
3. When I resume the conversation, the costs in telemetry logs dosent get updated, neither it is added to base_state.json.

Suspected root cause

ConversationStats.register_llm() restores metrics into the LLM via llm.restore_metrics(...).
However, restore_metrics() only updates self._metrics and does not update the Telemetry object if it has already been created.
So after restore, these two can diverge:

• llm.metrics → restored metrics object
• llm.telemetry.metrics → stale old metrics object
  Since telemetry writes token/cost/latency into self.metrics, resumed accounting may go into the wrong object.`

Expected Behavior

No response

Actual Behavior

No response

Steps to Reproduce

No response

Installation Method

uv pip install openhands-sdk

If you selected "Other", please specify

No response

SDK Version

1.11.4, main branch

Version Confirmation

• I have confirmed this bug exists on the LATEST version of OpenHands SDK

Python Version

3.12.7

Model Name (if applicable)

gpt-4o, gpt-5.1-codex-mini, GPT-5.2-chat

Operating System

Linux

Logs and Error Messages

No response

Minimal Code Sample

from openhands.sdk.llm.llm import LLM
from openhands.sdk.llm.utils.metrics import Metrics

llm = LLM(model="gpt-4o-mini", api_key="dummy")

# Force telemetry creation
_ = llm.telemetry

old_metrics = llm.metrics
restored = Metrics(model_name=llm.model)

llm.restore_metrics(restored)

print("llm.metrics is restored:", llm.metrics is restored)
print("llm.telemetry.metrics is restored:", llm.telemetry.metrics is restored)
print("llm.telemetry.metrics is llm.metrics:", llm.telemetry.metrics is llm.metrics)

# OBSERVED BEHAVIOR
# llm.metrics is restored: True
# llm.telemetry.metrics is restored: False
# llm.telemetry.metrics is llm.metrics: False

Screenshots and Additional Context

No response