Droppy Code 1.6.0 folds a finished turn into its "Worked for …" line with the turn's steps behind the chevron, keeps a running turn's steps, and the agent's thinking, in the card below its bar, hands a chat stopped by its provider's usage limit a New thread that carries the work on, lets the heads leave the sidebar on their own, and puts a mark of your choosing in front of a project's name; Netlify joins the MCP catalog. Apple silicon, macOS 26 or later. Signed and notarized.
New features
- Netlify joins the MCP catalog: sign in with your Netlify account and a thread can create projects, deploy sites and manage environment variables.
- Projects wear a mark: each project can be given an emoji or an SF Symbol in Settings, General, Projects, and the activity list can draw that mark in front of a thread's name in place of the line that named the project, one line per thread, while what the thread is doing still reads at the end of the row. The choice of list sits on the General page and on the welcome screen, and a new install starts on the compact one.
- A chat that stops on its provider's usage limit now leaves a notice with two ways out beside it: Cancel stops the automatic continue and leaves the chat to you, and New thread stops the wait and opens a new chat of the same project holding a hand-off of the request and of what the stopped turn had already done, so another model can be picked there and sent it.
- Heads can leave the sidebar by themselves: with Auto-hide heads in the sidebar on (Settings, Hydra, Heads), a chat's head card shows while one of them runs and goes once every head of that chat has finished, coming back on its own when one starts again. Off by default, so heads stay where they are.
Bug fixes
- The GitLab MCP server no longer looks like a hiccup on GitLab's side: a top-level group that hasn't allowed MCP client access is answered by GitLab 19.4 and earlier with 404 instead of "MCP server disabled", so connecting ended in "gitlab.com didn't answer its MCP address (404), try again in a few minutes". That answer now opens the steps to turn MCP on for the group, and the app's own MCP proxy answers a path it has no route for with 502 rather than a 404 that reads as the server's reply.
- Scrolling through the chat and its panels no longer fights the reader: the timeline reads wheel ticks and a viewport that changes height as the reader's own movement rather than as content arriving, so rows landing or the composer growing never tugs anyone who is still scrolling, and the Settings pages settle a scroll burst without re-evaluating the page on every frame.
- The diff panel builds a file's rows once per change and drops a scroll request a newer one has replaced, the model list keeps its measured row heights without re-evaluating the page for every measurement, a pre-measured popover lays its content out once instead of on every parent update, the update notes on the About page scroll with the page instead of inside a scroller of their own, and the shortcuts page lays out whole so its scroller knows its content height.
Refinements
- A finished turn reads folded: it shows its final reply with the turn's plans and any warning, and the steps it took, its heads and its changed-files card wait behind the "Worked for …" line, which opens them in a card of their own.
- A running turn keeps its work on the working line: the progress card stays the card it is, and the steps the turn is taking, with the agent's thinking and what it says between them, open under it once its chevron is clicked; the reply that concludes the turn lands in the chat once the turn ends.
- The Chronological order switch is gone from Settings › Conversation: a turn reads folded until its line is clicked, which is what the app always does now, so the switch had nothing left to change.
Thanks
- René (@diazdesandi): confined
search_textto the project root with a--guard against flag smuggling, kept plan mode blocking edits after approve-for-turn, cleared the save timer handle beforesaveNowbails and stoppedWorkingTreeWatchtearing down the registered watch when a second caller lost the race (!307); one OpenAI-compatible session and one Keychain behind the DeepSeek, Meta and Z.ai providers (!308).
- Bryan Galdámez (@JosueGalRe): finished turns open chronologically and a head wears its profile (!283); native heads routed to head profiles on Claude, Codex and Copilot (!285); the pair editor keeps its effort row for a model whose catalog entry declares no efforts yet (!304); a model catalog probe that came back empty is tried again on the next ask instead of closing the provider's list for the rest of the run (!306).
- Nikita Tsyganov (@nik1tsyganov): named the parts of the Pi usage sum so it type-checks (!293).