hearthstone builds a household RAG, mutter-butter stamps your windows
This was a from-scratch-products week. hearthstone went from its first commit to a working household knowledge assistant — a retrieval backend, a full iOS client, and an evaluation harness — in ninety commits. mutter-butter reimagined GNOME window management, turning minimized windows into draggable “stamps” on a shelf. And stockyard rebuilt its VM manager around a hypervisor-agnostic backend, then tuned a macOS microVM down to a 1.24-second cold boot to a live SSH session.
The week in numbers
- 10 repos active (3 featured, 7 also shipped)
- 241 commits
- 2 merged PRs
- +38,602 / −4,230 lines
- 4 commit identities
- 2 tagged releases
Featured
prime-radiant-inc/hearthstone
A household knowledge assistant — a retrieval service over your own documents, with an iOS client and a guest-sharing model.
90 commitsNinety commits, starting from the repo’s very first, built an entire product end to end. The backend came up as a provider-abstracted embed/chat layer, a section-aware Markdown chunker that preserves tables and prepends heading breadcrumbs, vector search via sqlite-vec scoped per household, and Google Drive ingestion that exports docs and converts them to Markdown through Pandoc. Authentication evolved mid-week from Google sign-in to passkeys plus email verification, with owner JWTs and short-lived guest sessions. The iOS app landed in parallel: Keychain-backed token storage, an SSE client streaming chat deltas, a numbered-citation system that surfaces only the sources the model actually cited, and a read-only in-app viewer that scrolls to and highlights the cited passage. The week closed on evaluation — a retrieval harness comparing four strategies (plain RAG, multi-query, multi-document, and full-context), an LLM-as-judge scoring answers against key facts, and a GEPA-inspired prompt optimizer running all its questions in parallel. Sole contributor: Matt Windbrook. (Line counts are omitted for this repo — the diff is dominated by generated and vendored files, so a raw total would mislead more than it informs.)
obra/mutter-butter
A GNOME Shell extension that turns minimized windows into draggable “stamps” on a shelf and snaps live windows to a grid.
53 commits +7,170 −804A new GNOME 50 extension that went through five rounds of design review before a line of implementation, then came together from focused subsystems: a grid engine with cell math and occupancy tracking, a focus-rank list, stamp-placement strategies, a scale manager handling clone lifecycle and auto-shrink, an input router with stable z-ordering and a reentrancy guard, and a drag handler that intercepts grab operations. The result is a window manager where minimized windows become “stamps” you drag around a shelf, live windows snap to a grid on any move, and dropping a window onto an occupied cell swaps the two. Sole contributor: Jesse Vincent.
Also shipped
prime-radiant-inc/stockyard
A VM manager for agent workloads.
34 commits +6,360 −527Stockyard grew a hypervisor-agnostic backend — a VMBackend interface with adapters for Firecracker and, new this week, macOS via vfkit, plus a rootfs-provisioner interface with ZFS, copy-based, and APFS-clonefile implementations. The macOS path got an Alpine arm64 rootfs tuned hard for boot speed (kernel-level DHCP, static-IP VirtioFS, SIGKILL-based instant destroy), landing a 1.24-second cold boot to a running SSH session. Sole contributor: Matt Windbrook.
obra/pepper-x
A local dictation app — the GUI sibling to ghost-pepper’s speech-to-text.
19 commits +17,714 −2,659Nineteen commits brought pepper-x to feature parity with ghost-pepper: streaming ASR, Qwen 3.5 cleanup, a full settings overhaul, an onboarding wizard, prerequisite-recovery flows, microphone controls with level feedback, and a GNOME Shell status-pill overlay. Sole contributor: Jesse Vincent.
prime-radiant-inc/iterative-development
A Claude Code plugin encoding an iterative, spec-driven development methodology.
21 commits +4,249 −9A new plugin scaffolded around a “walking skeleton” approach: format validators for requirements indexes, roadmaps, iteration logs, and skill files (each with fixtures), and a set of skill stubs — extracting requirements, scoping the simplest core, running an iteration, implementing tasks, and auditing progress. Sole contributor: Jesse Vincent.
prime-radiant-inc/gauntlet
An AI-powered QA testing framework.
11 commits 1 PR +1,427 −212Configuration moved into an API endpoint with a model dropdown in the UI, error handling gained a server-side error log with its own UI panel and a type-safe error-source union, and the interface picked up a persistent live-run banner and clearer “generate test cards” controls. Jesse Vincent did the bulk; Simon Willison contributed one commit.
obra/superpowers
An agentic skills framework and software-development methodology.
6 commits 1 PR 🏷 v5.0.7 +335 −18Shipped v5.0.7 — Copilot CLI support and OpenCode fixes — and added contributor guidelines aimed at reducing agentic-slop PRs, including agent-facing guardrails. The companion superpowers-marketplace tracked the release. Sole contributor: Jesse Vincent.
A couple of smaller notes round out the week: narcolepsyd tagged its first release, v0.1.0; and terminal-bench-analysis took in several thousand new benchmark trial results through automated commits, while Simon Willison wired a commit-message generator into its CI and added a link to the atom feed.