Bug: chat continuation history is rebuilt twice when server injects CONTINUATION header

Summary

chat continuations can accidentally duplicate conversation history and grow context much faster than expected.

The server-level continuation reconstruction emits a header like:

=== CONVERSATION HISTORY (CONTINUATION) ===
...
=== END CONVERSATION HISTORY ===

But SimpleTool currently checks for the exact substring:

"=== CONVERSATION HISTORY ==="

That check does not match the actual (CONTINUATION) header. As a result, when a request reaches SimpleTool with already-embedded history, the tool treats it as missing history and reconstructs the thread again.

Impact

This causes duplicated history in the prompt and can also store an already-expanded prompt back into the conversation thread as a new user turn. Multi-turn chat continuations then inflate quickly and may hit downstream context limits earlier than expected.

Symptoms may include:

• repeated or unexpectedly large === CONVERSATION HISTORY blocks
• continuation turn counts increasing faster than expected
• downstream provider errors such as context length exceeded / request too large
• poor behavior for long-running chat continuation workflows

Suspected Root Cause

server.py reconstructs continuation history before tool execution.

tools/simple/base.py then tries to detect whether history is already embedded:

if "=== CONVERSATION HISTORY ===" in field_value:
    prompt = field_value
else:
    # reconstruct again

The actual header contains a suffix:

=== CONVERSATION HISTORY (CONTINUATION) ===

So the detection misses.

Suggested Fix

Use a broader sentinel check, for example:

def _has_embedded_conversation_history(prompt: str) -> bool:
    return (
        "=== CONVERSATION HISTORY" in prompt
        and "=== END CONVERSATION HISTORY ===" in prompt
    )

Then use that helper in SimpleTool instead of checking for the exact old header.

Suggested Test

Add a test covering the server-emitted header:

prompt = """=== CONVERSATION HISTORY (CONTINUATION) ===
Thread: 12345678-1234-1234-1234-123456789012
Previous conversation turns:

--- Turn 1 (Agent) ---
hello

=== END CONVERSATION HISTORY ===

=== NEW USER INPUT ===
continue
"""

assert ChatTool._has_embedded_conversation_history(prompt) is True

Notes

This appears independent of any particular downstream provider. Providers with stricter context limits simply expose the issue sooner.