Two new repos go public, and superpowers turns to Codex
Two brand-new repositories opened to the public this week — a Linear work-tracking plugin and a native-Linux terminal workspace manager — while superpowers cut two releases that pivot it hard toward Codex. Underneath the headlines, serf spent the week hardening itself with a fuzzing-and-coverage campaign, and teststrip raced its photo-culling workflow forward. Activity spread across both prime-radiant-inc and obra.
The week in numbers
- 10 repos active (3 featured, 7 also shipped)
- 1,687 commits
- 10 merged PRs
- +303,075 / −37,482 lines
- 3 commit identities
- 2 tagged releases
A note on the line count: added lines this week run high because two of the busiest repos are early-stage projects standing themselves up (a lot of scaffolding), and serf’s coverage campaign harvests real-traffic test corpus — generated and vendored material rather than all hand-written diff. We report the raw figure and flag it.
Featured
obra/superpowers
An agentic skills framework and software-development methodology that works.
36 commits 7 PRs 🏷 2 releases +1,475 −1,223Two releases this week — v6.1.0 and v6.1.1 — and the throughline is Codex. v6.1.0 shipped Codex marketplace install support, a leaner per-session bootstrap, and removed the Gemini CLI path; v6.1.1 followed with a run of Codex-hook correctness fixes: repairing SessionStart hook re-registration, adding Codex portal packaging, fixing the plugin manifest category, and suppressing hook auto-discovery when the hooks object is empty. There’s a smaller editorial thread in here too — trimming per-harness tool-mapping boilerplate and compressing the bootstrap prose so the skills cost less context to load — but the week’s real story is making superpowers a first-class citizen on Codex.
obra/insanitty
Native Linux (GTK4 / libadwaita) terminal workspace manager — a port of Fantastty, built on embedded Ghostty with tmux-backed workspaces and a QUIC remote engine.
90 commits +13,786 −1,164New this week, and it arrived at a sprint. In a single week the project went from an empty repo to a working native-Linux terminal workspace manager: a Swift/GTK4 app skeleton around an embedded Ghostty terminal, with a tmux control-mode protocol as the foundation. The user-facing features landed in the same window — splits, tabs, browser tabs, layout and session persistence across restart, workspace rename/archive/trash, drag-to-reorder — alongside a native Swift QUIC client and a ported remote-grid protocol that renders a remote workspace right in the GUI, interop-tested against the real helper. Packaging and CI stood up in parallel: Swift-toolchain CI, .tar.gz and .rpm targets, an MIT license matched to Fantastty, and end-to-end regression screenshots refreshed as each piece landed.
prime-radiant-inc/agent-plugin-linear-use
Keep AI-driven work tracked in Linear: a Claude Code / Codex skill + hook that finds or creates a ticket and moves it through your workflow.
3 commits +483 −70New and public on purpose. The initial release is the linear-use plugin itself — a skill plus a hook that automatically finds or creates a Linear ticket for the work an agent is doing and moves it through its workflow states, so AI-driven work stays tracked without a human minding the board. A same-week adversarial-review pass tightened the first cut, and a small fix corrected the commit author email — attribution matters on a repo we publish as a worked example of how we run Prime Radiant.
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.
938 commits 1 PR +217,393 −27,272The busiest repo of the week by commit volume, and the work was mostly serf turning its tools on itself. The throughline — captured in the one merged PR, “corpus refresh + coverage-floor ratchet” — is a multi-phase hardening campaign: schema-aware fuzzing of tool arguments, a failure-to-regression promoter that turns each new crash into a permanent test, real-traffic corpus harvesting, and coverage-floor ratchets applied module by module across the llm, auth, agent, hub, and TUI packages. A second, real thread is TUI and web polish — a provider-grouped, prettified, badged model picker, a colorblind-safe status vocabulary that leans on distinct shapes rather than color alone, an amber→blue recolor for “needs you” states, and the Lucide icon set vendored for one status language across surfaces — plus a settings-UI refactor split into per-concern files and some early self-healing tool-call repair. But the headline is coverage: this was the week serf got serious about not breaking.
prime-radiant-inc/teststrip
macOS photo-culling app — catalog-first, non-destructive, AI-assisted. Early development; not ready for use.
589 commits +65,430 −5,451A project in the thick of early development, and it covered a lot of ground in a week — from foundation to features. The early days laid the plumbing: a Swift package bootstrap, the asset/catalog domain model, a SQLite catalog repository, a decode-and-preview pipeline, and folder-ingest planning. By week’s end it was shipping the culling workflow itself: top-ranked stack culling, signal-backed compare recommendations, a compare “keep all,” and a compact rapid-cull rationale. A people/face-review workflow arrived too — manual review actions, naming selected photos, persisting confirmed people groups — and Smart Collections became contextual and actionable with a suggestion engine, preset composition, and typed rules. Throughout, the “alpha plan” doc got refreshed after nearly every slice. Not ready for use yet, but moving fast.
prime-radiant-inc/llm-proxy
Transparent logging proxy for LLM API traffic that auto-configures clients and records every request and response for debugging, auditing, and analysis.
8 commits +402 −26One coherent theme: provenance. The proxy gained a config-driven upstream allowlist (default-open) wired in at construction, then run-envelope-attributed logging routed through the synthetic-session path so every request carries attribution back to the run that made it. A “no-pollution” test for that attribution was made load-bearing to keep it honest, with a review-driven follow-up fix.
prime-radiant-inc/superpowers-evals
Behavioral eval lab (Quorum) that drives real coding-agent CLIs through a QA agent and grades them on workflow compliance against scenario criteria and deterministic post-checks.
6 commits 2 PRs +417 −41A quiet week here, and honestly reported as one: two small merged PRs adding pinned Claude credentials — claude-sonnet-5 and claude-sonnet-4-6 — so the eval harness can target specific model versions reliably. That’s the week.
obra/superpowers-marketplace
Curated Claude Code plugin marketplace.
2 commits +2 −2Two version-bump commits tracking the superpowers releases — pointing the marketplace listing first at v6.1.0, then at v6.1.1.
prime-radiant-inc/claude-plugin-stats
A daily scrape of Claude Code plugin install stats.
14 commits +3,668 −2,213All github-actions[bot] — the daily stats scrape and chart-data rebuild doing its job. Included for completeness; it’s the robot, not a person.
obra/dotfiles
Jesse’s personal dotfiles.
1 commit +19 −20A single snapshot commit — housekeeping, but real activity in the window, so it’s on the list.