The ground floor of the dev stack got re-poured this week.

Coder built a GNU-screen-style terminal multiplexer in Zig on libghostty's VT core: sessions, a prefix key, and — their words — nothing else to learn. The interesting part is the scripting surface: boo send types into a session, peek --json returns the rendered screen as structured data, and wait blocks until specific text appears or output settles. No TTY required. Curl installer for Linux and macOS, MIT, v0.5.15.

wait until output settles is the tell: this multiplexer expects programs — agents included — to be driving the terminal, and it kills the sleep-and-poll loop every terminal-automation script is built on. Delete the tmux send-keys && sleep 2 && capture-pane incantation. Tradeoff: the minimalism is deliberate — screen's model, not tmux's, so power users will miss their layouts and plugins — and at 264 stars you're an early adopter of the thing holding your sessions.

The Ory people (Kratos, Hydra) open-sourced an API-key server in Go: issue, verify, and revoke keys at scale, plus the useful part — derive reduced-scope, short-lived JWTs and macaroons from a long-lived key offline, so agents, CI jobs, and services don't ping the auth server on every request. Single binary, Apache-2.0, Docker Compose for local dev.

This is the agent-credential problem getting named: today you hand a long-lived, full-power key to a process that will happily paste it into a log file. Minting a narrow token per task, without a network round trip, is the right shape for that. Delete the god-mode API key living in your agent's .env. Tradeoff: 97 stars and five releases is young even with Ory's pedigree, and cache-based verification means revocation propagates eventually, not instantly.

A single userspace process that turns any S3-compatible bucket into a real filesystem — NFS or 9P mounts, or a raw NBD block device you can put ext4 or a ZFS pool on. Data is encrypted (XChaCha20-Poly1305) and compressed (zstd/lz4) before it leaves your machine; a local cache keeps warm reads at microseconds. Curl or Docker install; AGPL-3.0 with a commercial option.

Object storage at fractions of a cent per gigabyte-month has been sitting there while we kept buying block storage; this makes it mountable without a NAS in between. Delete the aging Synology, or the EBS volume holding data you touch weekly. Tradeoff: cache misses cost 50–300ms of S3 round trip, the AGPL is a real conversation if you embed it, and your filesystem now has an availability dependency on your bucket.

A Rust OLTP database that does graph and vector (plus KV and documents) in one engine, now built on object storage — the v3 architecture that put it back on HN today. v3.0.5, Apache-2.0, 5k stars, curl installer, and helix chef walks setup interactively. The pitch is agent memory: the relationship-plus-embedding workloads agents generate, behind one query surface instead of three databases.

Agent memory keeps converging on the same architecture — Postgres for records, pgvector or Pinecone for similarity, a graph bolted on for relationships — and Helix is betting you'd rather run one engine than three with glue. Delete two of the three databases in your RAG stack, if the bet pays. Tradeoff: consolidation bets are lock-in bets, and the default dev instance is in-memory — persistence is a flag you have to remember to pass.

A local CLI that removes "um," "uh," and "er" from recordings: pip install erm, point it at a WAV, nothing uploads anywhere. Under the hood it's four stages — faster-whisper for word-level timestamps, audio analysis to catch the fillers Whisper silently drops from transcripts, cut points nudged to zero-crossings, then variable-length crossfades with room tone layered under the seams so the edits don't click.

It's also a model of scope discipline: it deliberately won't touch "like," "you know," repeated words, or long thinking pauses, on the theory that those are editorial calls, not noise. Delete the hour of razor-editing before every demo video or podcast cut. Tradeoff: Python plus ffmpeg setup is on you, and fillers are all it does — it's a scalpel, not an editor.