Prime Radiant

evener cuts three releases on an LLM-gateway and backlog-issue sweep, while superpowers retunes brainstorming for v7.0.1

Week of October 19, 2026 · Oct 19, 2026–Oct 25, 2026 · published Oct 23, 2026

prime-radiant-inc/evener tags three releases this week — v0.3.14, v0.3.15, and v0.3.16 — across a sustained sweep that spans LLM-gateway compatibility work, a new provider-connection UI, and a long tail of fixes and tests keyed to numbered backlog issues. obra/superpowers tags v7.0.1, retuning its brainstorming skill after community feedback on the v7.0.0 rebuild and restructuring three long skills so they survive Claude Code’s compaction cutoff intact. This week is still open: the data below runs through Friday afternoon, not the full calendar week.

The week in numbers

  • 10 repos active (2 featured, 8 also shipped)
  • 139 commits
  • 117 merged PRs
  • +134,525 / −15,514 lines
  • 6 commit identities
  • 4 tagged releases

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.

94 commits 94 PRs 🏷 3 releases +125,858 −15,075

Three tagged releases this week: v0.3.14, v0.3.15, v0.3.16 — all bare version tags with GitHub’s standard auto-generated changelog, no hand-written notes, consistent with every prior evener release. By far the most active repo again, with the work clustering into several threads: a cluster of LLM-gateway compatibility fixes (reading a gateway’s per-model reasoning-effort ladder and per-model dialect/token spelling from /models, honoring a gateway’s per-model “no-effort” verdict, retrying an in-band gateway error frame instead of treating it as permanent, recovering rejected JPEGs without losing the original image); a new provider-connection card unifying connection state into one UI, with a same-week fast-follow fixing retry-source handling and adding a race test; and a test-hardening sweep across a long tail of numbered backlog issues — branch names are almost all issue-NNNN-... — covering the transcript virtualizer’s prepend-measurement loop, memory search-result path scoping, duplicate overlay notices after Stop, notification body-entity escaping, mobile Back history-transition counting, shellguard witness-failure sharing, cascadeguard’s dotted ancestor spine, cross-tab display sync, composer async-projection timing (three separate PRs settling the same class of race before assertions run), activity-view geometry at phone widths, comment eviction beyond the reading cache limit, parked-search summaries, result-tool schema preservation across renames, detached-TMPDIR handling, and a turn-gate revalidation race against a replaced transcript writer. Alongside those: delegate/plan work (giving a delegate its whole plan at startup, then refactored to one prompt per current step plus one plan per root workflow); native-client work (iOS share-sheet session-attachment sharing, an Apple-convention iPhone design refinement, file links in assistant prose opening the Reader/document pane, queued-message edit/remove across browser and native, draft ownership preserved through an identity-changing resume); hub work (endpoint-configuration cleanup plus a same-week fast-follow for its own “Low” findings, a binary-install/restart split, connecting compatible hosts without forcing deployment); marketplace/plugin work (letting marketplace-add name a branch or tag, preserving published plugin revisions for existing sessions, keeping managed git metadata private); several perf trims (shared immutable catalog capability maps, shared floating-label pointer tracking, deduplicated failed project lookups, context-metrics sampled once per publication); and a docs audit publishing tested CLI/web tutorial tours, followed by a same-week repair pass for the audit’s own gaps. Sole human author: Jesse Vincent (staff, obra), plus two dependabot[bot] version-bump commits.

obra/superpowers

An agentic skills framework & software development methodology that works.

2 commits 12 PRs 🏷 v7.0.1

Tagged release v7.0.1, with a real hand-written release body (not an auto-generated changelog) — a brainstorming-flow retune after user feedback on the v7.0.0 rebuild: it now digs for existing project/spec context before asking anything, asks better-targeted follow-up questions instead of repeating a generic opener, only offers to skip its questions when it started unprompted, and lays out two or three approaches plus a short design doc before building, which the release’s own testing shows sharply cut how often it skipped straight to building without approval. Three long skills — subagent-driven-development, writing-skills, executing-plans — had supporting docs moved into references/ so they survive Claude Code’s compaction cutoff intact, after a reported session lost a skill’s round limit mid-compaction and ran far past it. Issue and PR templates now require a human-written summary, and the project will no longer accept agent-written PRs that fix someone else’s issue unless a maintainer asked for one. The release notes credit the original issue reporter and a list of community testers by GitHub handle — acknowledgments, not code contributors; none of them have commits in this window. Sole author: Jesse Vincent (staff, obra).

Also shipped

prime-radiant-inc/llm-proxy

Transparent logging proxy for LLM API traffic that auto-configures clients and records every request and response to Claude, OpenAI, and other providers for debugging, auditing, and analysis.

8 commits 3 PRs +4,703 −103

Log pruning is now fully off by default in the bare binary, with deployments that want it setting the limits explicitly; previously the off setting still left a background sweep running on a timer that did nothing. Optional model-call correlation (session/turn/logical-call IDs) is now carried through request-logging captures, stripped from upstream and Platform-signed headers but preserved in the local log. A real file-descriptor leak fixed: the logger never closed a per-session log file once opened, so descriptors grew unbounded and a pruned or deleted file’s disk space was never actually reclaimed while the descriptor stayed open; files now close after an idle timeout or when a session’s date directory rolls over, reopening cleanly on the next write. Contributors: Ada Sen and Drew Ritter (both staff, prime-radiant-inc).

prime-radiant-inc/github-triage

Claude Code plugin for triaging GitHub issues and pull requests, with a security-gated PR review workflow.

11 commits 2 PRs +1,129 −66

apply-triage-actions can now close an item as completed or duplicate instead of everything closing as “not planned,” losing the real reason and never linking a duplicate to its canonical. The session-id lookup now reads CLAUDE_CODE_SESSION_ID directly instead of a fragile ps/awk pipeline that both mis-parsed Claude Code’s argument substitution and only worked when a session was launched with an explicit flag. Also: a label-proposal checker that catches contradictions in cheap-model labelers’ output before applying them (in a real triage pass it caught contradictions that weren’t visible by eye), an exemption keeping quality labels off maintainers’ own items while still applying topical labels, and hardening against corrections or proposals that could vanish silently (typo’d override keys, type mismatches, last-wins collisions on items proposed twice). Contributors: Drew Ritter and Jesse Vincent (both staff).

prime-radiant-inc/prime-radiant-marketplace

Prime Radiant’s Claude Code plugin marketplace.

3 commits 1 PR +36 −8

Added the maintaining-documentation plugin to the marketplace catalog, and fixed two other entries that were added in parallel with an earlier HTTPS fix and had kept GitHub-SSH-style sources, which fail to clone for anyone without a configured SSH key. Sole author: Jesse Vincent (staff, obra).

obra/lace

Lightweight agentic coding environment.

6 commits 5 PRs +2,238 −74

Diagnostic model/worker-turn correlation IDs minted at the conversation boundary and carried through provider retries and compaction. Persona @includes can now fall back to embedder-supplied default files when a user persona’s include is missing locally, with a same-week follow-up fixing the fallback not reaching delegated subagents. Claude Haiku 5.5 added to the model catalog and made the new default small model. Trusted execution-context metadata is now forwarded to opted-in host MCP servers, fails closed when the trusted IDs are absent, and moved to a runtime-owned namespace rather than a shared one. Contributors: Drew Ritter and Jesse Vincent (both staff).

obra/superpowers-marketplace

Curated Claude Code plugin marketplace.

1 commit +1 −1

A single commit bumping the pinned superpowers entry to v7.0.1, matching that repo’s release this window. Sole author: Jesse Vincent (staff, obra).

Three more active repos this week had no human commits to attribute — prime-radiant-inc/claude-plugin-stats (the daily install-stats scrape and chart rebuild) and prime-radiant-inc/terminal-bench-analysis (a daily README regeneration with no new data) are bot-only, and obra/muse-skills’ commits look automated (a daily cadence, under a git identity with no linked GitHub account), though that isn’t confirmed.