Prime Radiant

Superpowers 6 ships, claude-session-driver hits 4.0

A week of releases landing. Superpowers 6 shipped its full release train, claude-session-driver cut v4.0.0 — the release that ships the bash-to-TypeScript rewrite from the week before — and obol and slackline both stepped their versions. This is the last of the weeks we’re reaching back to fill in; the week of June 22 picks up right after it as our first originally-published recap.

The week in numbers

  • 15 repos active (5 featured, 10 also shipped)
  • 813 commits
  • 15 merged PRs
  • +162,740 / −67,055 lines
  • 4 commit identities
  • 10 tagged releases

A few of this week’s tagged releases are incremental point releases within active release trains (Superpowers 6’s .1/.2/.3 follow-ups, obol’s and slackline’s steady cadence), so the release count reflects a busy publishing week rather than that many distinct milestones.

Also shipped

obra/lace

A lightweight agentic coding environment (TypeScript).

83 commits 1 PR +13,717 −1,188

A “session-state architecture” pass: an O(tail) incremental turn-projection cache that avoids full-log rescans, a cross-process monotonic sequence authority (flock plus a head file, with self-healing detection of pid reuse on stale locks), and a non-lossy verbatim event_journal recall store. Plus prompt-cache byte-stability guardrails — golden-fixture determinism tests pinning exact request bytes for Anthropic, OpenAI, and Gemini, closing a parallel-tool-call cache-invalidation bug.

prime-radiant-inc/serf

A non-interactive coding agent that loops until the work is done, with native tool-calling across OpenAI, Anthropic, and Google models.

327 commits +69,401 −43,376

Still the week’s largest raw commit volume even without a release. A new serf-doctor diagnostic subsystem landed — tree/watches/transcript/forensics views over a session’s delegate and observer state, with a dedicated doctoring-serf skill — alongside a tool-fluency experimentation framework (live harness, prompt iteration, fixed-prompt reruns) and passive-observer sidecars that idle without generating tool-call churn. Install now defaults to a home-prefix layout.

prime-radiant-inc/superpowers-evals

The Quorum behavioral eval lab — drives real coding-agent CLIs through a QA agent and grades them on workflow compliance.

292 commits +65,752 −17,433

Design and early build of a “shared eval appliance”: a containerized, detached-worker job model with its own CLI, cancellation handling, git-ref preflight, and a trust-hardening pass (wrapper-environment trust, path validation, sanitation-bypass closure) — the eval harness’s own container story, distinct from stockyard. Plus Quorum dashboard decoupling (a standalone dashboard entrypoint, an OS-aware eligibility grid, 5-state cells) and a full-fidelity Pi normalizer.

prime-radiant-inc/llm-proxy

A transparent logging proxy for LLM API traffic — records every request and response for debugging, auditing, and analysis.

12 commits +2,372 −131

“Mantle” telemetry-contract work: run-scoped mantle routing required for cloud builds, telemetry-observation logging with pass-through provenance preserved on upstream failures, and safer cloud-build run-ID decoding.

obra/dotfiles

Jesse’s personal dotfiles.

6 commits +446 −86

Small quality-of-life fixes: Homebrew on PATH for non-interactive zsh, a guard against an inert stray root ~/.zshrc, and kimi-code added to the managed PATH.

obra/superpowers-marketplace

A curated Claude Code plugin marketplace.

4 commits +4 −4

Version-pin bumps only, tracking Superpowers v6.0.0/6.0.2/6.0.3 and claude-session-driver v4.0.0.

prime-radiant-inc/obol-go

The generated Go binding for obol.

1 commit

A single machine-generated release commit, keeping the Go binding in step with obol’s release train.

prime-radiant-inc/claude-plugin-stats

A daily scrape of Claude Code plugin install stats.

14 commits +4,606 −3,015

A quiet week — the daily scrape and chart rebuild running on schedule, all github-actions[bot], no human-authored change. The robot doing its job.