A light week: evener's busiest repo again, superpowers-evals' campaign push
A light week. No repo was created in-window and none cut a tagged release, so nothing gets the featured treatment this time — everything below is also-shipped. prime-radiant-inc/evener is the busiest repo again, cutting over to a new provider registry and working through a long tail of hardening across its hub, agent, and web layers. prime-radiant-inc/superpowers-evals runs one continuous campaign-infrastructure arc. The rest of the week is brief. As with the last couple of posts, this week is still open: the data below runs through Friday evening, not the full calendar week.
The week in numbers
- 7 repos active (0 featured, 7 also shipped)
- 793 commits
- 117 merged PRs
- +262,351 / −161,676 lines
- 5 commit identities
- 0 tagged releases
Also shipped
prime-radiant-inc/evener
A 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.
695 commits 113 PRs +231,444 −148,915By far the most active repo this week, with broad simultaneous work across the hub, agent, and web layers. The headline is a models.dev-backed provider registry cut over as a flag-day migration, landing a new providers.toml schema with an auto-conversion path from the old one. Alongside it: per-connection serial worker dispatch for the app server, unified tool-call disclosure and verbosity controls in the web UI, job_watch timers, and delegate task-list handling (presence-based add/update, seeded task lists). Underneath, a long tail of hardening: drain/completion race fixes, scratch-disposal cleanup, credential-leak guards (never harvest a credential carried in a URL path; stop echoing pasted secrets from malformed refs), mobile web fixes (composer above the keyboard, larger tap targets), and test-suite robustness work (real-completion waits, ETXTBSY guards, deterministic cancel classification). Contributor: Jesse Vincent (staff), with one dependabot[bot] dependency-bump commit.
prime-radiant-inc/superpowers-evals
Behavioral eval lab (Quorum) for the superpowers project that drives real coding-agent CLIs through a QA agent and grades them on workflow compliance.
69 commits +17,706 −1,806The second-most active repo this week, all direct pushes rather than merged PRs. One continuous campaign-infrastructure arc: sealing a multiharness signature campaign (frontier and open models, later adjusting the model roster), and a frontier and open models signature campaign run through two launches — one cancelled on platform findings, one sealed. Underneath both, a run of fixes to the campaign “seal act” (digest-verified regeneration, atomic markdown-then-JSON publication, sealed-state preservation across a post-append cancel race, orphan-kill and verified-death reaching the tmux subject host), credential routing for Bedrock-backed graders, and a new “descriptive fold” report schema (counts, medians, contention counters, provenance findings). Sole author: Drew Ritter (staff).
prime-radiant-inc/alltheagents.org
Comprehensive census of coding agent harnesses.
16 commits 3 PRs +10,295 −8,644A polish-and-launch-prep week on a repo that only turned up in last week’s post as brand-new: renamed to “All the Agents,” with an About page and new contribution entry points (issue forms, a PR template, site links), “other”-category entries hidden, client-side pagination with a size control added, and the site served from the alltheagents.org apex domain. The home page was reworked too — simplified layout, a Prime Radiant footer credit, sticky footer. Separately, the underlying census data got two content passes: entries added from the 2389-research GitHub org, then a false-negative audit and a broader recategorization under a tightened loop-ownership definition. Contributors: Kattni (staff) on the site and UX work, Jesse Vincent (staff) on the census-data reclassification.
prime-radiant-inc/gauntlet
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 +36 −5One commit: routing Bedrock/Mantle-prefixed Anthropic model ids to the Anthropic provider. Author: Drew Ritter (staff).
obra/lace
Lightweight agentic coding environment.
1 commit 1 PR +159 −1One commit, one merged PR: passing tool-result image blocks through to the model. Author/merger: Jesse Vincent (staff).
The remaining two active repos were bot-only this week: claude-plugin-stats ran its usual daily install-stats scrape and chart rebuild, and terminal-bench-analysis saw one automated README regeneration with no new data — no human commits on either.