Document the macOS build and the chat export macOS is on the Releases page now rather than build-it-yourself, so Installation names the two DMGs and explains the quarantine flag Gatekeeper trips over. Tools and Capabilities gains the export: where the button is, what is written, and what is deliberately left out. Test count corrected to around 990. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Add machine specifications, and what shutdown stops Installation gains minimum and recommended sets, naming real parts rather than bare numbers, taken from the VRAM ladder the app uses to pick a model. Troubleshooting gains an entry for something still running after a close: what Draggy stops, and why an Ollama it did not start is left alone. Test count corrected to around 950. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Lead with what Draggy does, and drop the em dashes Home now opens on the same sentence as the repository description, that Draggy talks, browses and works with your files, rather than on the privacy pitch. Every em dash across the twelve pages is rewritten as a colon, a comma, parentheses or a full stop. Nav lists, status labels and module lists take a colon; the rest are sentences. Test count corrected to around 940. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Document PDFs, extensions, compaction and hybrid search Adds an Extensions page for the MCP catalogue, and brings the rest of the wiki in line with what the app now does: PDF reading and writing, the folded-conversation behaviour in long chats, keyword-and-vector search in the library, and the allowlist on what the model may fetch. Corrects two claims that had gone stale: the test count, and the note saying nothing runs the suite for you. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Document the thinking modes properly, and what a per-message model load means Settings described thinking mode by its internal names rather than the labels on screen, and did not say what Fast now does: it turns reasoning off outright instead of asking for less of it. Troubleshooting gains the symptom that led to the 1.1.5 fix — a model load in the timings under every reply, which means the model is being unloaded between messages rather than held.
Say what CI actually does: CodeQL scans, nothing runs the test suite
Correct both model ladders, and note the licence and how models are judged The chat table was missing every Gemma 4 rung — 2, 6, 10, 16 and 24 GB — and four of the six rows in the talk table named the wrong model: 4 GB is Llama 3.2 3B rather than Qwen 3 1.7B, 8 GB is Gemma 3 4B, 16 GB is Llama 3.1 8B, and there is no 6 GB rung at all. Both now match the tables in the source. Small models are no longer hidden from the chat picker, and what is hidden is decided by asking Ollama what a model can do rather than by matching its name, so Choosing a Model and Settings both say so. Architecture was missing appData.cjs and logger.cjs from the main-process listing. Building from Source now mentions that the translation tests run in both directions, that nothing runs npm run check on a pull request, and links CONTRIBUTING.md. Home carries the GPL notice.
Document the in-app browser, the ad blocker and its switch, and the session split
Reflect automatic embedding sizing, the moved voice settings, and the current search providers
Remove the fast helper model, now that Draggy doesn't have one The question-routing helper model was dropped in v1.1.0: the classification it did was already running on the main model, not the helper it downloaded, so it added latency and disk space for nothing. Also moves the talk-model picker's documented location to Settings > Personalization, where it now lives, and refreshes the test count.
Add the Draggy wiki