lldb-mcp debug session identifiers can confuse some LLMs

[ashgti]

Our debug sessions identifiers start at 1, not 0.

This seems to be confusing some LLMs when working with the lldb-mcp protocol.

[llvmbot]

@llvm/issue-subscribers-lldb

Author: John Harrison (ashgti)

Our debug sessions identifiers start at 1, not 0.

This seems to be confusing some LLMs when working with the lldb-mcp protocol.

[ashgti]

If we're planning on changing how we identify a session we may want to come up with a way to make them more globally unique. Like a UUID instead of an index. That would allow us to more easily implement a multiplexer in the lldb-mcp binary itself because we would be able to map a debugger to a specific lldb instance.

[JDevlieghere]

The UUID sounds like a good idea, but that's likely even harder to guess for an LLM. Would you make it optional if you have only one session or rely on communicating the UUID through the tool/resource?

[ashgti]

I actually did that as part of #158340, the debugger URI is optional, if its not set we use the first debugger on the server. At the moment, I don't have an error if there is more than 1 debugger though.

See

llvm-project/lldb/source/Plugins/Protocol/MCP/Tool.cpp

Lines 71 to 89 in 5a339b0

	if (!arguments.debugger.empty()) {
	llvm::StringRef debugger_specifier = arguments.debugger;
	debugger_specifier.consume_front(kSchemeAndHost);
	uint32_t debugger_id = 0;
	if (debugger_specifier.consumeInteger(10, debugger_id))
	return createStringError(
	formatv("malformed debugger specifier {0}", arguments.debugger));
	debugger_sp = Debugger::FindDebuggerWithID(debugger_id);
	} else {
	for (size_t i = 0; i < Debugger::GetNumDebuggers(); i++) {
	debugger_sp = Debugger::GetDebuggerAtIndex(i);
	if (debugger_sp)
	break;
	}
	}
	if (!debugger_sp)
	return createStringError("no debugger found");