Notes

Cloudflare published a best practices guide for Durable Objects — especially useful for feeding to agents during early design stages and code review. Covers how to design around single-threaded, globally-unique instances with persistent storage.

@Jnatanh's pi-autoresearch-studio adds granular experiment-to-PR selection with auto-resolved dependencies to the Pi Coding Agent workflow. Built as a Pi extension, it lets you cherry-pick which experimental runs become pull requests instead of committing everything automatically.

@bcherny's tip for power users: Claude Code has deep git worktree support for running multiple agent sessions in parallel on the same repo. Use claude -w to start in a new worktree, or check the worktree option in Claude Desktop. Essential when you have dozens of agents working simultaneously.

@marcelpociot's Polyscope uses copy-on-write clones instead of complicated worktrees for managing multiple project checkouts. Way faster, uses less disk space, and includes all dependencies without complex setup scripts.

@_chenglou's new TypeScript library brings fast, accurate text measurement to the browser without touching the DOM. The pure userland algorithm can lay out entire web pages bypassing CSS and reflows — foundational work for the next generation of high-performance UI tooling.

@penberg from Turso clarifies their SQLite compatibility promise: databases are fully compatible with SQLite's file and WAL formats, you can always migrate back, and any incompatible features (like concurrent writes via BEGIN CONCURRENT) require explicit opt-in. When Turso falls short of compatibility, they treat it as a bug to fix rather than a feature difference to document.

@micLivs makes the case against vector search for most memory use cases. Their pi napkin tool uses BM25 and TF-IDF on plain Obsidian markdown files—no embeddings, no vector DB. On HotpotQA multi-hop questions it hit 99.8% recall and 79.6% F1 versus 90%/41% for the original and 29%/19% for Mem0. Sometimes a good map beats a complex pipeline.

Cloudflare Workers now lets you declare required secrets in wrangler.jsonc or TOML — Wrangler validates them locally, generates TypeScript types, and blocks deploys if any are missing. Small DX win that catches config drift before it hits production.

@M1Astra archived Anthropic's Claude Mythos blog post before it was taken down — a rare look at internal documentation about Claude's character, values, and the philosophical foundations behind how it was trained to engage with the world.

@appfactory's Agent-CI brings GitHub Actions workflows to local execution with a twist: when a step fails, it pauses and lets your AI agent fix the issue before retrying just that step. Uses the standard GH Actions Docker image but with near-zero caching latency, designed for the tight feedback loops agents need without burning through CI minutes.

Claude Code now supports .worktreeinclude files — drop one in your project root to automatically copy gitignored files like .env, .env.local, or config/secrets.json when Claude creates a worktree. Handy for keeping your agent's isolated branches functional without manual setup.

pi-lab lets you A/B test tools in the pi coding agent by intercepting calls, forking into isolated lanes, comparing outputs, and merging the winner back. Still experimental, but a promising approach for iterating on agent tooling without flying blind.

@jerryjliu0 on LiteParse — a fast document parser that lets agents read any PDF on the internet. No VLM under the hood, so it runs anywhere. Latest update adds URL parsing and buffer/stream support. Designed to plug into 40+ different agents.

Gemini in Google Calendar now handles meeting scheduling by analyzing availability and time zones across busy calendars. Small quality-of-life improvement for anyone who's played scheduling Tetris across multiple participants.

@FabianHiller makes the case for Valibot over Zod: 88% smaller bundles (1.91 kB vs 16.57 kB), 16× faster initialization (54 μs vs 735 μs), plus cleaner pipelines and precise TypeScript inference. If you ship to web, edge, or serverless, the numbers are hard to ignore.

@dinokupinic previews Drizzle ORM v1.0 — migration folders reworked (no more journal.json), validator packages (drizzle-zod etc.) moved into drizzle-orm itself, and Relational Queries v2. The 1.0 stabilization everyone's been waiting for.

@adibhanna ships tsm v0.6 — a native terminal multiplexer that doesn't wrap your terminal in a server. Instead of re-emulating VT output, it delegates splits to your terminal's own API. Every pane is a real native surface with GPU rendering, ligatures, and scrollback preserved. Supports cmux, kitty, Ghostty, WezTerm, plus workspace manifests and agent sidebar sync.

@Lovable ships AI-powered penetration testing for vibe-coded apps — a swarm of agents runs OWASP Top 10 checks, privilege escalation tests, and data exposure scans. Validated findings sync back as actionable issues with formal pentest reports for SOC 2 and ISO 27001. What used to take weeks and cost $5k-$50k, now built into the platform.

@claudeai ships auto mode in Claude Code — instead of approving every file write and bash command (or skipping permissions entirely), auto mode lets Claude make permission decisions on your behalf with safeguards checking each action before it runs. The middle ground between trust-nothing and YOLO.