serf grows a TUI, sprout begins, and the engineering notebook lands
This was an agent-infrastructure week end to end: two agent frameworks built in parallel, a tool to read back what they did, and a desktop app on the side.
The week in numbers
- 9 repos active (3 featured, 6 also shipped)
- 260 commits
- 8 merged PRs
- +108,686 / −3,958 lines
- 4 commit identities
- 4 tagged releases
Featured
prime-radiant-inc/serf
An agent framework in Go — sessions, tools, multi-provider model adapters, and a benchmark harness.
120 commits +74,852 −2,567serf’s largest week so far. It grew a real front end — a Bubble Tea terminal UI driven by an SSE server (ring buffer, reconnect catch-up, /status, /interrupt, /input), with a session picker, slash commands, theming, and rich tool renderers. Underneath, an append-only JSONL transcript system landed for durable session logging and transcript-based resume, including parent–child linkage for sub-agent transcripts. Live model listing arrived across the Anthropic, OpenAI, and Google adapters, the four-layer compaction collapsed to two, and a long run of prompt-and-steering work — TDD discipline, verification steps, content-filter recovery — came straight out of terminal-bench failure analysis.
prime-radiant-inc/sprout
A self-improving agent framework in TypeScript — agents that detect their own stumbles and rewrite a versioned “genome” to learn.
87 commits +19,583 −350A brand-new project, built up phase by phase in 87 commits. It starts with a unified LLM client routing across native Anthropic, OpenAI, and Gemini adapters, then adds a core agent loop with delegation, a git-versioned Genome storing agents/routing/memories, keyword and embedding recall wired into the system prompt, and — the interesting part — a Learn process: it watches for “stumbles,” emits learn signals, and applies mutations to its own genome while the agent runs, with programmatic rollback and pruning.
prime-radiant-inc/engineering-notebook
A tool that ingests agent session logs into a searchable, day-by-day journal.
24 commits +5,175 −215Also new this week: a Bun project that parses JSONL session logs, stores them in SQLite, summarizes them with an LLM, and serves the result as a Hono web app — a journal rendered as a day-by-day timeline, with conversations split on logical day boundaries rather than the wall clock.
Also shipped
obra/iMessagePrinter
A macOS app that reads iMessage history and exports conversations to PDF.
8 commits 🏷 3 releases +2,812 −147Zero to v1.0.0 in a week — a macOS iMessage viewer and PDF exporter, with a build-and-release workflow, an app icon, and fixes for message cross-contamination during export.
prime-radiant-inc/scenarios
A prompt-injection and evaluation test suite for LLM agents.
8 commits 1 PR +5,874 −308Simon Willison kept scenarios moving — a document-analysis scenario, automated tool-call extraction and evaluation scoring in run-scenario, and a demo walkthrough pitting gpt-4o-mini against gpt-5.2 across the set.
obra/superpowers
An agentic skills framework and software-development methodology.
6 commits 6 PRs +66 −38superpowers cut v4.3.1 — Cursor support and a Windows hook fix — with contributions from Drew Ritter and outside contributor abzhaw.
Rounding out the week: streamlinear tagged v1.1.0 with multi-workspace support (added, reverted, and readied for another pass), and there were small maintenance commits to cc-plugin-decision-log and the superpowers-marketplace registry.