superpowers v6.2.0, winpepper goes streaming, and serf's biggest week

Week of July 20, 2026 · Jul 20, 2026–Jul 26, 2026 · published Jul 26, 2026

Two releases anchor the week: superpowers cut v6.2.0, rolling up weeks of methodology work, and winpepper shipped a pair of alpha releases that quietly turned a local dictation tool into one that can also reach a cloud recognizer. Underneath the releases, serf logged its single heaviest push of the whole window — and several of the week’s merged fixes came from outside the core team.

The week in numbers

14 repos active (2 featured, 12 also shipped)
1,902 commits
19 merged PRs
+291,558 / −99,072 lines
9 commit identities
3 tagged releases

Featured

obra/superpowers

An agentic skills framework and software-development methodology that works.

4 commits 4 PRs 🏷 v6.2.0 +145 −13

v6.2.0 is the week’s headline release, gathering up work that had been merging over the prior weeks: the spec-driven-development plan-scoped workspace with its resume-based fix loop, a skills-compression sweep, and a Windows SessionStart fix that dispatches through Git-Bash. Two outside contributors landed fixes folded into the same release — a find-polluter.sh path-matching correction and a dangling-docs-anchor cleanup — a small sign of a framework picking up hands beyond the core team.

obra/winpepper

Windows-native local dictation: hold a hotkey, speak, release, and cleaned-up words appear in the focused app — Parakeet TDT v3 ASR with LlamaSharp cleanup, all local.

368 commits 5 PRs 🏷 2 releases +78,563 −5,912

The week read like a reliability pass — load-bearing-validation hardening for the mid-paste and settings-change races — but it culminated in two tagged alpha releases, and the release-grounded story is bigger than the diligence underneath it. AssemblyAI cloud ASR now sits alongside the local engine, sleep/resume incidents got a proper fix (an error taxonomy, recovery-driven clearing, hook reinstallation), the keyboard hook was hardened, and the app gained a per-user MSI installer plus a round of consumer polish (toast policy, pill fixes, ASR-config consolidation, live model swapping). As with superpowers, most of the commits came from outside — Dan Shapiro, with AI coding agents (Codex, Amplifier) committing alongside him.

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, using native tool-calling across OpenAI, Anthropic, and Google models.

1,467 commits 4 PRs +197,087 −91,164

serf’s busiest week of the window, almost all of it direct-push work on the session and pane architecture — a delegate-transcript UI audit, pane-routing repairs, nested-session owner promotion, and settings/session-routing placement contracts. Four concrete features merged on top of that rebuild: forking any message into the composer for editing, folding archived sessions behind a single disclosure grouped by project, prepopulating the session-path dropdown with recent projects, and per-turn job-notification parsing. Also in the mix: model-switch replay-parity work (asserting the same event kind in both live and replay paths) and serffuzz compile-gate fixes.

prime-radiant-inc/llm-proxy

A transparent logging proxy for LLM API traffic that auto-configures clients and records every request and response for debugging, auditing, and analysis.

4 commits 3 PRs +795 −31

An AWS SigV4 signing mode for the Anthropic passthrough, letting the proxy front Bedrock — hop-by-hop headers stripped before signing, with a clean mode-based rollback path. An Apache-2.0 license landed too.

obra/lace

A lightweight agentic coding environment.

5 commits 3 PRs +736 −421

The Bedrock provider was rewritten on AnthropicBedrockMantle for full Anthropic parity, SDK connection errors are now treated as retryable, and the integration suite got a test-hygiene fix. Thematically paired with llm-proxy’s SigV4 work this same week — both pushing Anthropic-via-AWS support.

prime-radiant-inc/stockyard

A coding-agent VM orchestrator: runs coding agents in isolated VMs — Firecracker micro-VMs on Linux with ZFS-based audit-trail snapshots, and Apple’s container tool on macOS.

16 commits +7,674 −731

Safe task-destruction work: requiring names to destroy named tasks, failing closed on identity errors, safely quoting destroy-confirmation names, plus hardened orphan-audit reconciliation and durable IP allocation.

prime-radiant-inc/superpowers-evals

A behavioral eval lab (Quorum) that drives real coding-agent CLIs through a QA agent and grades them on workflow compliance against scenario criteria and deterministic post-checks.

14 commits +3,343 −84

Two threads. First, Hermes-4 bring-up: a RED bootstrap verdict and mechanism autopsy, a clone-faithful plugin-staging fix, hardened staging (reject symlinks, expand ~), a provisioning adapter — and the honest negative finding that Hermes-4 via OpenRouter is structurally impossible to eval, because it exposes no tool-use endpoints. Second, a matched pair of “evidence” scenarios grading whether an agent tells the truth about what it built: a working feature with verified proof versus a broken feature with an honest report.

prime-radiant-inc/smevals

A framework for running evals against small (and large) models.

7 commits +1,628 −31

A per-task detail-view page, a run -n X flag to repeat runs and distinguish errors from failures, GitHub Actions CI, and a path-traversal fix in the serve command. Contributed by Simon Willison.

obra/blogosphere

A local-first, multi-platform blogging client for an 11ty blog that lives in a GitHub repo — the repo is the database.

4 commits +389 −24

A durability week: a “database is locked” publish failure fixed via single-statement rename pairs, and hardening against hard kills (Windows PAT persistence, bounded typing-buffer loss).