Nothing in the API changes here. This release is about helping an AI client
pick the right tool — the complaint behind #2 — by spending the initialize
instructions on decisions rather than internals, and by making tools that get
confused for each other say what the other one is for.
Changed
- The
initializeinstructions now guide tool choice (#27, #2) — MCP clients inject this text straight into the model's context, and Nelson spent a third of it saying "All UNO operations run on the main thread for thread safety", which cannot change any decision a model makes. It now states the things that go wrong in practice: start fromdoc_list_openand never reach fordoc_createwhen the user names a document that already exists; callnav_outlinebefore editing a Writer document and address edits by bookmark rather than by counting paragraphs; the tool list depends on which document is active; work on a copy before a large edit - Tools on a decision boundary name the alternative (#27) —
doc_createnow says it will not open an existing document and points atdoc_list_open/doc_recent/doc_open; those three say what each other is for;text_insertdistinguishes itself fromtext_setandtext_apply_range. Descriptions are what a model reads when it chooses, so the ones that get confused for each other should say so - Tool groups are derived from the name (#27) — the hand-written
intentattribute was read by nothing and had drifted: 24 of 140 tools carried none, anddoc_open,doc_create,doc_closeanddoc_list_openwere all taggedmedia. Since every tool is nowdomain_verb, the group follows from the prefix (document,text,navigate,review,media,calc,draw,system), with an explicitgroupavailable when a prefix would mislead. A test asserts every declared tool resolves, so a new one cannot land ungrouped