Prime Radiant

winpepper is born, claude-session-driver hits v2

Week of May 11, 2026 · May 11, 2026–May 17, 2026 · published May 17, 2026

A dense week. winpepper — Windows-native local dictation — was created and built out through numbered milestones in seven days. claude-session-driver cut a real v2.0.0, dropping its “fake tool-approval gate,” and episodic-memory landed end-to-end Codex support across four point releases. The eval harness lifted into superpowers the week before split out into its own repo, superpowers-evals, and superpowers-chrome made browser-dialog handling a first-class concern.

The week in numbers

  • 19 repos active (6 featured, 13 also shipped)
  • 729 commits
  • 18 merged PRs
  • +1,743,243 / −41,596 lines
  • 9 commit identities
  • 9 tagged releases

A note on the line count: this week’s raw added-lines figure is dominated by a single automated commit — a bot regenerating a large leaderboard data file in an analysis repo (~1.5M generated lines). That’s data, not hand-written code; read the headline total with that in mind and look to the per-repo work below for the real shape of the week.

obra/claude-session-driver

Launch, control, and monitor other Claude Code sessions as workers via tmux.

13 commits 2 PRs 🏷 2 releases +2,229 −798

v2.0.0 dropped the fake tool-approval gate; v2.0.1 disallowed AskUserQuestion inside workers. In between came three rounds of user-feedback fixes: a universal worker handle, spec-aligned paths, status and handoff commands, and a default worker listing. Sole contributor: Jesse Vincent.

obra/episodic-memory

A memory and recall layer for coding-agent sessions.

24 commits 1 PR 🏷 4 releases +5,751 −740

Codex support landed end-to-end across four releases (v1.3.0 → v1.4.1): transcript parsing, session summarization, plugin packaging, hook synchronicity, and production hardening (raw-HTML escaping, local-shell-output handling). On the reliability side, v1.4.1 drained a zero-exchange sync backlog that had been silently stalling summarization, recall-trigger reliability improved, and the /search-conversations slash command was dropped. Jesse Vincent drove it; one commit came from outside contributor minyek.

obra/superpowers-chrome

A Chrome-control plugin via the DevTools Protocol.

37 commits 🏷 v2.1.0 +6,168 −132

v2.1.0 made dialog handling first-class: a DialogRefusedError thrown rather than silently swallowed on browser dialogs, a permission shim registered at attach, dialog-awareness threaded through the mouse / keyboard / capture / CDP-connection attachers, and Tier-C smoke tests against real Chrome. Sole contributor: Jesse Vincent.

obra/winpepper

Windows-native local dictation — hold a hotkey, speak, release. Parakeet TDT v3 ASR plus LlamaSharp cleanup, entirely local.

156 commits 🏷 v0.6.0-alpha +42,706 −427

This is winpepper’s creation week — and unlike every later week already covered on this blog, all of it is Jesse Vincent, before the outside contributors arrive. The README is loud about its own provenance (“entirely agentic, never human-tested”), and the tool was built out in plan-numbered milestones in a single week: Plan 4 (history / lab / models tab plus a downloader), Plan 5 (post-paste learning, a Diagnostics page, crash handling via MiniDumpWriter), and Plan 6 (WiX MSI packaging, autostart, a code-signing wrapper, nightly CI smoke), cutting a v0.6.0-alpha. Sole contributor: Jesse Vincent.

prime-radiant-inc/superpowers-evals

A behavioral eval lab (Quorum) for the superpowers project.

15 commits 3 PRs +17,786 −201

Split out this week from the drill/evals/ harness that had been lifted into superpowers the week before. It shipped a Codex-native-hook eval backend and Pi backend support in the same week, plus initial security hardening — one of three merged PRs explicitly targeting Codex security controls ahead of wider use. Contributors: Drew Ritter and Matt Windbrook.

Also shipped

obra/lace

A lightweight agentic coding environment.

16 commits +3,259 −435

Persona-bundle support took shape: Ent initialization and a session/create that accepts persona config, ordered user search paths, and frontmatter-plus-body parsing — plus a provider tool-name sanitizer. Sole contributor: Jesse Vincent.

prime-radiant-inc/gauntlet

An AI-powered QA testing framework.

153 commits +31,113 −2,402

A bash tool for agents (allowlisted env, output caps, process-tree reaping) and a “shell-as-session” model for the TUI and CLI adapters (spawn bash in a tmux pane or scratch cwd, single-SIGKILL close escalation), plus a fetch_credential tool threaded through the orchestrator and an eight-story-card TodoMVC example fixture. Sole contributor: Matt Windbrook.

prime-radiant-inc/serf

A non-interactive coding agent.

251 commits +117,593 −9,818

A JSON-naming convention pass (camelCase JSON, kebab-case TOML/CLI, enforced by a new serf-namingcheck lint), OpenAI device-code login for headless sessions, and OAuth-state-over-env-var precedence fixes. Sole contributor: Jesse Vincent.

prime-radiant-inc/scribble

A self-hosted Slack knowledge bot.

17 commits 1 PR +1,693 −357

A public-repo hardening pass merged — pinned Actions, Dependabot, CODEOWNERS, trimmed internal references — alongside disabling the Claude Agent SDK’s built-in auto-memory (Scribble owns its own wiki-backed memory) and an OpenTelemetry advisory patch. Sole contributor: Drew Ritter.

obra/superpowers

An agentic skills framework and software-development methodology.

12 commits 11 PRs +629 −12,578

Eleven PRs merged (credited by PR author — the in-window commit rollup reads empty, same squash/backdate pattern as prior weeks): cross-platform skill-compatibility prose with source-verified per-runtime tool references, a Pi extension plus eval backend, and several Codex-track fixes. Jesse Vincent merged much of it; two outside contributors landed one PR each — fuleinist (a stale .cursor-plugin reference cleanup) and stablegenius49 (a writing-skills @-reference fix).

prime-radiant-inc/greenfield

A Claude Code plugin that reverse-engineers behavioral specs from a codebase.

1 commit +22 −22

One commit, from outside contributor pumarogie: restructured the skills into per-skill directories, each with its own SKILL.md.

Smaller changes filled out the week: slackline took a Node-24 CI bump and an app-provisioning fix; ts-libghostty landed two more “bobbihack” fixes (both Matt Windbrook); bot-toolkit added an autoMemory opt-out (Drew Ritter); ghost-pepper saw a usage report and Granola auto-import from outside contributor Matt Hartman (matthartman); and superpowers-marketplace bumped its version pins. A github-actions[bot] commit in terminal-bench-analysis regenerated a leaderboard data file — the automated line-count spike noted above.