Add e2e suite 27: retried LLM turn must keep conversation history

[ling-senpeng13]

Adds an e2e regression suite for the bug — "LLM_CHAT_COMPLETE retry loses conversation history, causing duplicate tool calls".

The tests follows the steps described below to reproduce the issue. There are 2 test cases: test_retried_llm_turn_keeps_conversation_history is expected to fail and test_retried_tool_task_is_unaffected  is expected to pass.

Steps To Reproduce

Minimal repro using a native agent, two sequential tools:

import time
from conductor.ai.agents import Agent, AgentRuntime, tool

@tool
def get_customer_info(customer_id: str) -> dict:
    """Look up a customer's account information, including their home city, by customer ID."""
    return {"customer_id": customer_id, "name": "Jane Doe", "city": "Seattle"}

@tool
def get_weather(city: str) -> str:
    """Get the current temperature, wind speed, and humidity for a city."""
    return f"{city}: 17.7C, wind 6.4 km/h, humidity 73%"

agent = Agent(
    name="retry_context_repro",
    model="OpenAI/gpt-5-mini",
    tools=[get_customer_info, get_weather],
    instructions=(
        "First call get_customer_info to find the customer's city, then "
        "call get_weather for that city, then summarize both."
    ),
)

with AgentRuntime() as runtime:
    handle = runtime.start(agent, "What's the weather like where customer 4471 lives?")
    print(handle.execution_id)

Start the agent. Let it complete its first tool call (get_customer_info) so it's genuinely in-progress on a second LLM turn.

In the Conductor UI, terminate the execution while the second (or any subsequent) LLM_CHAT_COMPLETE task is IN_PROGRESS, not while a tool task is running.

Use the UI's Retry action to resume from the failed/canceled task.

Inspect the retried LLM_CHAT_COMPLETE task's inputData.messages, it contains only [system, user], with none of the already-completed tool_call/tool message pairs from prior iterations, even though those tool tasks show COMPLETED in the same workflow's task list.

Observe the model, having no memory of the already-successful get_customer_info call, calls it a second time. This is a duplicate execution, not just a display quirk: a second get_customer_info task instance actually runs.

The following LLM turn (the next fresh, non-retried dispatch) does correctly receive the full history, including both get_customer_info calls, and produces a coherent final answer, the model itself narrates the duplication (e.g., prefacing its answer with "Continuing from before...").

For contrast, terminating instead while a tool task (e.g., get_weather) is IN_PROGRESS and retrying it behaves correctly: the retried tool task just re-runs with the same arguments, and the next LLM turn correctly sees both the failed and successful attempts in its history (e.g., "Tool execution failed (status: CANCELED)" followed by the successful retry's result), producing a clean, non-duplicated final answer.

Expected behavior

Retrying a canceled/failed LLM_CHAT_COMPLETE task should receive the same fully-reconstructed conversation history (all prior completed tool calls and results) that a freshly-scheduled LLM_CHAT_COMPLETE task for the next iteration would receive, not the bare static template. The retry mechanism shouldn't silently drop context that's demonstrably still present in the workflow's own task execution history.