clearance ships a v1 editor, sprout restructures its genome, and agent-eval tooling takes off
Two threads ran through the week. One was a new native app — clearance, a local-first macOS markdown editor — going from an empty repo to a tagged v1.1.0 in seven days. The other was agent evaluation: serf poured most of its week into an eval dashboard and a trajectory-export format, a new harness called gauntlet appeared, and the dashboard itself was spun out into its own repo. Underneath it all, sprout kept 165 commits of self-improvement work moving.
The week in numbers
- 11 repos active (2 featured, 9 also shipped)
- 460 commits
- 4 merged PRs
- +101,376 / −16,574 lines
- 7 commit identities
- 7 tagged releases
Featured
prime-radiant-inc/sprout
A self-improving agent framework that detects its own failures and rewrites its genome to avoid them.
165 commits, centered on a genome restructure: a rootDir + overlay resolution model for merged agent lookup, per-project genome data directories, and export-learnings simplified to overlay-only resolution — all pushed through several rounds of code review. Alongside the plumbing came safety work: guardrails injected dynamically into system prompts at the tool boundary, a fix preventing agent hallucination when tool resolution returns zero tools, support for workspace-tool-only agents, and a project-memory agent added to the root agent tree.
prime-radiant-inc/clearance
A native macOS markdown editor — local-first, with a network-blocked rendering view.
A new app that shipped fast: 93 commits from “bootstrap clearance macos app” through a string of releases ending at v1.1.0. Along the way it grew a recent-files model, markdown frontmatter parsing and a rendered view with a frontmatter table, a CodeMirror editor with syntax highlighting (assets vendored for a fully local editor), a solarized light/dark theme, drag-to-popout windows with disk-change reload, and a security posture that blocks network resources in both the rendered view and the runtime webview. Outside contributors acme (2 commits) and rodbegbie (1) pitched in.
93 commits 2 PRs 🏷 7 releases +18,285 −1,600prime-radiant-inc/serf
An agent framework in Go — this week, mostly its evaluation dashboard and trajectory export.
85 commits, with staff engineer Drew Ritter contributing 9. The bulk went into the eval dashboard: proper task-lifecycle states (queued/running/verifying/pass/fail), run metadata and task deduplication, tool-call duration instrumentation surfaced in trajectories, and detecting task status from agent output rather than result-file timestamps. The other strand was ATIF — a native trajectory-export format — with v1.6 types, a converter handling thinking blocks and redacted content, and an --export-atif flag that writes a trajectory at session close for the harbor viewer.
Also shipped
prime-radiant-inc/gauntlet
An LLM-agent evaluation harness (“vet”) — runs scenarios against agents through CLI, web, and TUI adapters.
A new project, 63 commits from design doc to a working harness: a story-card parser, an evidence logger, an LLM provider abstraction (Anthropic + OpenAI), and three adapters — a subprocess CLI adapter, a web adapter wrapping a forked superpowers-chrome CDP library, and a TUI adapter driving tmux sessions. It also gained an API server with scenario CRUD, containerized execution via Docker, a “fanout” generator for producing scenario variations, and the beginnings of a React UI.
63 commits +21,062 −772prime-radiant-inc/harbor-eval-analysis-dashboard
A web dashboard for analyzing agent evaluation runs.
13 commits, starting by extracting the eval dashboard out of serf into a standalone repo. It added per-trial result dots, filtering/sorting/grouping in the task-history view, trajectory-step timestamps, an /api/containers endpoint that parses docker ps, and an active-containers view with auto-refresh and run linking.
obra/superpowers
An agentic skills framework and software-development methodology.
22 commits, including work from outside contributor karuturi (2 commits).
22 commits +4,467 −113Elsewhere: lace took 8 commits; scribble (5) gained a slack_reply tool for threaded replies to arbitrary messages, documented in its constitution; packnplay (2) took contributions from outside contributors clkao and schuyler; and beeper-message-sync and smallest-agent each saw a couple of commits.