First releases: books-for-bots, obol, and teststrip go public
Three repos cut public releases this week — two of them Rust tools aimed squarely at AI agents, one a macOS photo app making its first public appearance. Beneath the release headlines, the busiest work was quieter: serf ran a package-by-package fuzz-coverage campaign and the eval lab deepened its model-compatibility grid.
The week in numbers
- 12 repos active (3 featured, 9 also shipped)
- 1,764 commits
- 4 merged PRs
- +282,747 / −35,496 lines
- 7 commit identities
- 5 tagged releases
Featured
prime-radiant-inc/books-for-bots
A Rust CLI that converts EPUBs into a single YAML-headed Markdown file with per-chapter byte and line offsets — a navigation API that lets an LLM agent read a book token-efficiently.
8 commits 🏷 2 releases +1,348 −373books-for-bots shipped its first tagged releases this week. v0.1.0 brought a CI and tag-triggered release workflow that builds binaries for six platforms, plus an MIT license and a clippy/cargo fmt cleanup so CI can enforce lint and formatting going forward. The v0.1.1 follow-up added a --version flag so an agent can self-identify its build, fixed EPUB-internal path normalization at the load boundary, and retired a dead macOS CI runner label. Small, but it’s now a tool you can pin a version of.
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 and Node).
2 commits 🏷 v0.7.0 +10,873 −9,097obol’s v0.7.0 is a pure pricing-data release: it refreshes the bundled model-price snapshot and adds claude-sonnet-5 pricing, so cost estimates stay accurate as the model lineup moves.
prime-radiant-inc/teststrip
A macOS photo-culling app — catalog-first, non-destructive, AI-assisted. Early development; not ready for use.
675 commits 🏷 2 releases +89,735 −11,146teststrip went public this week and cut its first releases — a models-v1 model-asset release and a v0.1.0 app build. (The repo object was created this week, but the project already had a full week of history behind it; read this as “went public,” not “started.”) The week’s work hardened the import pipeline for trust: preflight free-space checks before a card import, heartbeated import/scan progress so the watchdog survives slow phases, a persisted default import destination, and a time-bounded preview scan so the photo count it shows you is accurate. Alongside that, security-scoped source bookmarks now persist, restore, and repair themselves across reconnects and relaunches, and the culling signals grew up — Apple feature-print extraction wired into similarity filters, signal-backed similarity stacks, and provider-driven Smart Collection and keyword suggestions.
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 — native tool-calling across OpenAI, Anthropic, and Google models.
971 commits +156,139 −13,163serf’s dominant thread this week was a fuzz-coverage completion campaign: closing fuzz-reachable coverage gaps package by package — agent, hub, doctor, contextmgr, execenv, TUI, appsource, jobstore — toward an exact-coverage target. The second real feature line was model-switching: validating a switch up front, gating it on turn state, broadcasting model and effort changes, persisting switch markers, and enforcing replay provenance when a session changes model mid-run. Some mobile-web polish for the spawn UI landed too (viewport handling, composer dock anchoring) — real, but a footnote next to the coverage push.
prime-radiant-inc/stockyard
A coding-agent VM orchestrator: runs agents in isolated VMs — Firecracker micro-VMs on Linux (with ZFS-based audit-trail snapshots) and Apple’s container tool on macOS.
2 commits 1 PR +1,344 −15One targeted reliability fix: stockyard now archives a VM’s console logs when it’s destroyed, so the audit trail survives teardown.
prime-radiant-inc/gauntlet
An AI-powered QA 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 1 PR +14 −3A single fix: give the Claude 5 named tier its full max-output-token cap.
prime-radiant-inc/superpowers-evals
A behavioral eval lab (Quorum) that drives real coding-agent CLIs — Claude, Codex, Gemini, Kimi, and more — through a QA agent and grades them on workflow compliance.
79 commits +20,154 −1,069The lab deepened its serf-as-harness coverage — OpenRouter preset-identity attestation and BYOK, a materialized campaign-level model-compatibility grid, and a pinned fix for serf’s streamed-response IDs. The launcher grew Mantle/Bedrock credential handling: default Claude calls to Bedrock with a direct opt-out, scrub host ANTHROPIC_API_KEY and AWS credentials so seeded credentials win, and reject Mantle credentials on Windows Claude. Grader tooling got a --grader-model flag (defaulting to claude-sonnet-5) and a Sonnet 4.6→5 drift screen that showed no behavioral drift.
prime-radiant-inc/smevals
A framework for running evals against small (and large) models.
1 commit +27 −0A small vocabulary addition, contributed by Simon Willison.
prime-radiant-inc/claude-plugin-stats
A daily scrape of Claude Code plugin install stats.
7 commits +55 −48Routine automated upkeep — the daily scrape and chart rebuild, github-actions[bot] only.
obra/lace
A lightweight agentic coding environment.
3 commits 1 PR +175 −14Reliability fixes to background-job notifications: a Slack follow-up hint added to every terminal job notification, and the runner’s last-seen event sequence used as the inject-drain watermark on the success path.
obra/dotfiles
Jesse’s personal dotfiles.
3 commits +268 −0A new /story-loop command — an agentic goal loop driven by end-to-end scenario cards — plus its design spec, and a small shell-env export for a keyring-backed CLI.
obra/superpowers
An agentic skills framework and software-development methodology that works.
12 commits 1 PR +2,615 −568Quiet week here: no direct commits, and the only merged change reverses course from the week before — a PR restoring Gemini CLI support after its removal was reverted.