ValueError: Invalid AML XML: not well-formed (invalid token): line 1, column 0

[liu1018-0326]

Every time I run python -m applications.vibegraph_demo.main directly, the following error occurs:
raise ValueError(f"Invalid AML XML: {exc}") from exc
ValueError: Invalid AML XML: not well-formed (invalid token): line 1, column 0
The same thing happens with different models.

[AckerlyLau]

Thanks for reporting this.

Could you please share:

1. The exact model name and provider/base URL you are using
2. Whether you are using the Responses API or Chat Completions API
3. Your MASFactory version or commit hash

VibeGraphing requires the build model to generate a relatively large, structurally valid AML document, so we recommend using a capable reasoning/coding model for the build stage. I tested a fresh build locally with GPT-5.5, and the AML generation, compilation, and workflow execution completed successfully.

Please also note that the demo reuses the generated AML cache. Changing the model does not regenerate the workflow while the cache exists. Before testing another model, remove it:

To capture the model output for debugging, you can temporarily add the following immediately after build_workflow.build() in masfactory/components/vibe/vibe_graph.py:

from pathlib import Path
from masfactory import Agent

def dump_model_output(agent, response, _messages):
    content = response.get("content", "")
    with Path("vibegraph-model-output.log").open("a", encoding="utf-8") as log:
        log.write(f"\n===== {agent.name} =====\n")
        log.write(str(content))
        log.write("\n")

build_workflow.hook_register(
    Agent.Hook.THINK_COMPLETED,
    dump_model_output,
    recursion=True,
    target_type=Agent,
)

Then delete the AML cache and run the demo again. Please share the generated vibegraph-model-output.log and the first few lines of applications/vibegraph_demo/assets/cache/workflow.aml, after removing any sensitive information.