v0.1.9
PuPu 0.1.9 — Early Preview Release 🚀
A desktop AI client with a file-system chat explorer, memory, workspace context, and runtime management for both local and cloud models.
What's New in 0.1.9
This release is about one idea: everything you can give the assistant now lives in one place. Toolkits, MCP servers, and skills used to be separate concepts with separate flows — 0.1.9 folds them into a single Plugin Store, and adds skills as a first-class way to work.
- 🧩 One Plugin Store — toolkits, MCP servers, and skills together — the old toolkit modal is gone, replaced by a proper in-app store. A Discover page opens on a featured pick, curated essentials, and collections that expand inline to show what's inside. A unified sidebar sorts the whole catalog by type — Toolkits / MCP / Skills — and every entry gets a real detail page with its README, permissions, and where it came from. Curated picks install in one click, with no credentials to set up first.
- ⌨️ Skills, as slash commands — this is the big one. Skills declared by any installed plugin appear directly in the
/menu in the composer. Pick one and it expands inline into a ready-to-edit template, and using it selects that plugin for just that run — nothing to configure, nothing that sticks around afterward. - 📚 Skill packs — install a whole workflow at once — browse and install curated skill packs (Prompt Master, Frontend Design, and more) in one click from the Skills section, or point PuPu at a local folder of
SKILL.mdfiles to turn your own into a pack. - 🛟 Long runs that survive a crash — a long agent run is no longer tied to the life of the app. Interactions are now durable: each execution attempt is recorded and bound to its session, so if the app or the runtime dies mid-run, the work reconnects to the right conversation on restart instead of being lost or duplicated. Pending tool confirmations survive with it.
- 🔁 Steadier long-running execution — extensive hardening of the run loop itself: execution control, recovery, and the durable job worker were rebuilt so that long, many-tool runs hold up. This release was soak-tested with multi-agent runs driving hundreds of tool calls, and they now complete without a single retried attempt.
- ✨ A real startup experience — the blank loading window is gone. PuPu now boots through a single themed overlay with one continuous progress bar from the very first frame until the app is ready, over an ambient blurred field derived from your own theme colors.
⚠️ Install failures explain themselves — when an MCP install fails, the actual underlying cause now surfaces on the card and detail panel instead of leaving you at a silent dead end.
Fixes & Polish
- 🧠 Memory contract hardening — memory calls now go through a transparent contract proxy, ending a whole class of silent breakage where unrecognized arguments were quietly dropped instead of raising.
- 💾 Sturdier settings storage — settings moved further onto a transactional local database: resetting is a single atomic operation, pending writes are drained on quit so nothing is lost on exit, and the storage page groups everything by category.
- 🖌️ Consistent surfaces everywhere — dropdowns, segmented controls, switches, menus, chips, and code/diff blocks were swept onto shared sizing, radius, and border rules, so the same control looks the same in every modal — and light/dark parity is noticeably tighter.
- 💬 Composer fixes — the empty-state greeting no longer overlaps the floating attach panel, dropdown options stop compressing when a list scrolls, and dropdown scrolling no longer chains into the panel behind it.
- 🌍 Translation sweep — all 11 shipped languages were filled in for the new plugins and skills surfaces.
- 🧪 Release quality — fixes across the release QA pipeline, end-to-end credential handling, and web build warnings so the release gate runs clean.
⚠️ Known Limitations
This is still an early preview. Some features are incomplete and edge-case bugs may still exist. If you hit one, please open an issue.
On the Horizon
Following the v0.1.5 convention, a short forward-looking note — roadmap previews, not release commitments. Scope and timing may change.
- Custom model providers — bring your own OpenAI-compatible endpoint, with its own base URL and key, and have its models flow through the catalog and model picker like any built-in provider. Landing behind the scenes now.
- Computer use — letting PuPu see the screen and drive the mouse and keyboard, gated behind explicit consent and per-model support. In active development; not yet enabled.
- Agent Builder — a visual way to define a reusable agent's prompt, model defaults, and tool access, then call that agent directly from chat for repeatable workflows.
