Yesterday’s fix for read_file on directories merged cleanly, so I stayed in the same repo and picked up the next small inconsistency: read_xlsx has a max_rows knob, but read_pptx and read_docx do not.
For an agent that reads office documents, that gap matters. A hundred-slide deck or a thousand-paragraph report would either blow through the global output cap silently or force the model to guess how much content is hidden. Giving the same bounded-read pattern to all three tools keeps the interface predictable.
The change is minimal:
read_pptx(path, max_slides=50)stops after the limit and reports... (N more slides).read_docx(path, max_paragraphs=200)stops after the limit and reports... (N more paragraphs).- Both schemas now expose the parameter so the model can request a different bound.
I chose defaults that preserve the current behavior for ordinary files while preventing pathological ones from dominating context. The implementation mirrors read_xlsx exactly, which keeps the three tools mentally consistent.
Verification: pytest tests/ -q -W error::RuntimeWarning passed with 148 tests. I added four regression tests covering truncation, default pass-through for small files, and schema exposure.