How does the model target an edit? That's the part I'd want to poke at first.

I run a multi-agent pipeline where models write structured content through MCP, and the challenge was addressing, not generation. "Replace the paragraph about pricing" looks great in a demo and then hits the wrong span the moment the doc gets long or the same tool call fires twice. I fixed it by giving every editable node a stable id and keying writes on a content hash, so a re-run changes nothing by accident.

So on revise: when the model edits, is it pointing at an id/anchor, or matching on the text it remembers seeing? And if the edit tool fires twice with the same args, do you get one change or two?

The other thing I'd want to know is DOCX fidelity. Do you hand the model a structured view of the doc or flattened text, and does formatting survive once an edit round-trips? That's usually where these start leaking.

Nice work shipping this solo. Let's connect on HF: https://huggingface.co/Awaliuddin