Prime Radiant

Two tools from scratch: blogosphere and smevals

Week of July 13, 2026 · Jul 13, 2026–Jul 19, 2026 · published Jul 19, 2026

The story this week is two tools that didn’t exist on Monday and shipped by Sunday: blogosphere, a local-first blogging client built from an empty repo to mobile builds in seven days, and smevals, a small-model eval framework Simon Willison stood up from scratch and released twice off the same week’s work. Around them, teststrip and obol cut releases, and commits landed from a noticeably wider circle of contributors across both orgs.

The week in numbers

  • 14 repos active (4 featured, 10 also shipped)
  • 1,035 commits
  • 23 merged PRs
  • +234,606 / −53,681 lines
  • 7 commit identities
  • 4 tagged releases

Two of the active repositories this week — terminal-bench-analysis and claude-plugin-stats — are automated daily-scrape jobs. Their churn is bot-generated, not human work, so we’ve kept it out of the framing below.

obra/blogosphere

A local-first, multi-platform blogging client for an 11ty blog that lives in a GitHub repo — the repo is the database.

35 commits +49,070 −2,223

New this week — and not a soft launch. The corrected commit log shows a full application built from zero: a Tauri v2 + Vite + React scaffold under TS-strict and Biome, a from-scratch sync engine (serialized pushes, a stale-head guard, an activity log, partial-validation pushes, discard-changes), a rewritten document-style editor surface, image fetch-on-miss resolution, and Android and iOS mobile builds. A late “delight pass” layered on publish-URL previews, ⌘K, a deploy watch, resume, versions, and live view. An adversarial review swept a batch of confirmed findings before ship — rename collisions, push/pull races, quit-flush, SSRF hardening. The public-launch essentials — MIT license, README, a boot-failure screen, platform-honest connect copy — are real too; they’re the last layer on top of a much bigger build.

prime-radiant-inc/smevals

A framework for running evals against small (and large) models.

30 commits 🏷 2 releases +2,251 −333

Also new this week: Simon Willison built smevals from scratch and cut two releases off that same-week work. It’s a full CLI-plus-reporting tool, not a cleanup pass — initial run and grade commands, a Checklist→Grader rename, an HTML report app with multiple graders and clickable tags, a terminal Markdown report command, multi-grader support, and example evals (haiku, Pelican, markdown-tables). A README-generating docs command, Python 3.10+ support, and the SMEVAL_*→SMEVALS_* environment rename round it out.

prime-radiant-inc/teststrip

A macOS photo-culling app — catalog-first, non-destructive, AI-assisted. Early development; not ready for use.

246 commits 🏷 v0.2.0 +41,651 −5,711

Cut v0.2.0 on the back of a culling-workflow redesign: a dedicated culling-workstation shell, simplified verdict copy (“Keep” / “Toss” instead of percentages), reads-card signal ordering, and burst-rail thumbnail and caption polish. It also persisted a burst-stack capture-interval preference, added a run tracker with a full completion summary, fixed Library-grid activation to open the Library loupe rather than the culling loupe, and repositioned and muted the RAW+JPEG badge.

prime-radiant-inc/obol

Read an AI-agent transcript and estimate what it cost — a Rust core with C-ABI bindings for Python, Go, and TypeScript (Bun + Node).

4 commits 🏷 v0.8.0 +9,974 −9,672

Cut v0.8.0, driven by a pricing-data refresh: the bundled snapshot now includes the gpt-5.6 model family, alongside CI and lint repairs (cargo-fmt drift, and a version test whose expectation is now derived from the workspace manifest rather than hard-coded).

Also shipped

prime-radiant-inc/serf

A non-interactive coding agent: give it a prompt and it reads, writes, runs commands, and searches code in a loop until the work is done, using native tool-calling across OpenAI, Anthropic, and Google models.

589 commits 8 PRs +119,015 −34,747

This week’s merged work centered on queue and steering UX: edit or cancel an unconsumed queued message, promote a queued follow-up to active steering per-message, render user-sent steering as user messages rather than system dividers, and keep a sticky per-session draft of the unsent composer. On mobile web, the sidebar ”⋯” menus now open as a full-width modal sheet and the phone sidebar behaves as a drawer instead of a squeezed rail. Agent-side, delegate-turn limits became configurable with honest occupancy and drive-budget reporting, and subagent inline activity now renders as the tool call’s stated purpose. A CI naming-gate fix tidied the snake_case REST contract.

prime-radiant-inc/llm-proxy

A transparent logging proxy for LLM API traffic that auto-configures clients and records every request and response for debugging, auditing, and analysis.

11 commits +1,369 −120

Response-capture and attribution work: stamp the path and termination reason on every response-record exit path (including abort, dial, and body-read failures), stamp metering_provider / upstream / capture_version on request lines, and derive the provider from the host. The Bedrock and Mantle relay legs got hardening — paired request-creation and dial-failure handling, termination-EOF stamping, and new test coverage for the relay legs.

prime-radiant-inc/superpowers-evals

A behavioral eval lab (Quorum) for the superpowers project that drives real coding-agent CLIs through a QA agent and grades them on workflow compliance.

89 commits 3 PRs +6,625 −494

A heavy experiment-documentation week, built around verifying the SDD fix-loop redesign that landed in obra/superpowers this same week: audits, hardening, codex-precheck routes, hostile-probe verdicts, contingency reruns, and final verdicts with economics. Infrastructure fixes — CI hermeticity, container-build repairs, and a container agent-CLI refresh — kept CI green.

obra/superpowers

An agentic skills framework and software-development methodology that works.

11 commits 9 PRs +2,349 −226

All merged-PR work, no direct commits. The SDD plan-scoped workspace ledger landed — per-plan artifact directories, a self-identifying ledger, end-of-plan cleanup. Alongside it, a skills editorial pass stripped social-proof and self-selling detritus from a dozen skills (eval-gated), reframed TDD’s testing anti-patterns, and modernized the branch-finishing skill’s rationalization-table pattern.

obra/winpepper

Windows-native local dictation. Hold a hotkey, speak, release — cleaned-up words appear in the focused app. Parakeet TDT v3 ASR + LlamaSharp cleanup, all local.

3 commits 2 PRs +1,433 −120

Two merged PRs, both from an outside contributor: keep the app responsive during model downloads, and fix modifier-only hotkey capture and key-swallowing.

prime-radiant-inc/greenfield

A Claude Code plugin that reverse-engineers clean behavioral specs, test vectors, and acceptance criteria from any codebase, with a provenance trail so a fresh team can reimplement without inheriting the original’s structure.

1 commit 1 PR +22 −22

An outside contributor restructured the skills into per-skill directories, each with its own SKILL.md — a layout fix ahead of wider sharing.

prime-radiant-inc/gauntlet

An AI-powered QA testing framework that uses LLMs to test web apps, CLI tools, and TUI programs from markdown story cards, returning structured pass/fail verdicts with evidence.

1 commit +201 −0

A LICENSE file added by Eden — housekeeping ahead of sharing the repo more widely.

prime-radiant-inc/terminal-bench-analysis

Fetches detailed JSON results from the Terminal Bench 2 leaderboard, loads them into SQLite, and publishes queryable analysis via Datasette.

1 commit +1 −2

An automated README regeneration with no new underlying data — a bot job, not human work this week.