>_ Eng Coffee Sips

All editions · EP.002 · 2026-08-24

Eng Coffee Sips EP.002

What are we doing about git?

Eng Coffee Sips is a short factual based newsletter for articles that might interest engineers or engineering leaders, not just AI news. Short and to the point to skim over your morning coffee.

What I’m Building This Week goat run simulator (goatrun.bltech.app), bug fixes on plinth (centralized auth), unifi network optimization, nfl know your team (know-your-team.bltech.app), software factory review (private), paywall remover (ethically and legal).

By Pete Shima · August 24, 2026

Today’s edition is 981 words, a 4-minute read.

  1. GitHub has another long outage, shares incident post as well as follow up. GitHub had a 7hr 47 min outage that impacted most of GitHub services on August 17. Increased load caused a load balancer snowball effect across Istio and HAProxy. This combined with an autoscaling policy that watched host service but not sidecar limits, and a VS Code retry storm, caused an extensive outage. The follow up post has a graph that shows the sheer scale of increases in merges, commits and new repos that is growing exponentially.

    Why should you care?

    GitHub has come under a lot of scrutiny for its continued outages. This re-sparked more conversations on who is building a GitHub alternative. The follow up post leaves a lot to be desired, but at least shows the sheer scale they are trying to deal with, which is no small feat.

    ( GitHub Blog / GitHub Status / @kdaigle / @Vlad_GitHub / @kovyrin / @vasilios_s )

  2. Git at any scale This post covers the technical limitations of git at scale. If you were ever curious how this works, this post is for you. It then covers Spokes and Continuity technologies that have been used to help scale git. The Shopify CEO re-implemented it over the weekend: https://x.com/tobi/status/2091678506992222258

    ( Cursor / @cursor_ai / @tobi )

  3. Durable Git on Cloudflare dgit is a git server for Cloudflare Workers and for your own machines with celld. Each repository is a Durable Object: a small server with a name and a private SQLite database that holds the repository’s objects and refs, speaks the git smart HTTP protocol to a stock git client, and renders a cgit-style web interface. There is no origin server, no filesystem, and no GitHub in the critical path.

    ( GitHub / @undefined_void )

  4. “Black box tests are your most valuable asset” — Paul Stack. AI models are generally good at pumping out tests. This article talks specifically about the benefit of black box acceptance tests, which exercise the compiled binary the way a user would rather than the implementation. Stack’s argument is that when an agent writes both the code and its unit tests, the same misunderstanding ends up in both, so they pass while the implementation is still wrong.

    ( stack72.dev )

  5. Slack launches Slack Code. Salesforce launched Slack Code on August 20, which lets a team tag a coding agent into a conversation and have it open a project channel where everyone can watch the work, audit diffs as they are proposed, and approve before anything ships.

    ( Salesforce / @Benioff )

  6. Warp announced Warp Factories, their version of software factories. Factories are “cloud software factories where specialized agents work engineering requests from intake to a mergeable pull request” — a coordinator Warp calls the foreman takes a work item in from Slack, GitHub, Linear or Jira and routes it through separate triage, specification, implementation and review agents, with the whole factory defined in version-controlled files. Completed runs are graded by scorers you write, and a self-improvement loop groups the failures they flag into follow-up runs that propose fixes to the application code or to the factory’s own definition.

    ( Warp / @zachlloydtweets / @zachlloydtweets )

  7. Block open sourced Berd, the desktop app its own teams use to work with agents Block’s problem was fragmentation rather than capability — it had goose, Claude Code and Codex internally, but “working with them meant navigating different interfaces, configuration systems, and ways of managing context.” Berd consolidates conversations, files, instructions, agents and skills around persistent projects, and gives each agent a visual identity on the reasoning that “different agents can look distinct because they are distinct.” It completes a three-layer open source stack: Berd for private work, Buzz for multiplayer, and goose as the shared runtime underneath — the same goose Block donated to the Linux Foundation’s Agentic AI Foundation, alongside Anthropic’s MCP and OpenAI’s AGENTS.md.

    ( Block / @blocks / Linux Foundation / @BlockWhaleXYZ )

  8. PlanetScale explains data topology, the config that tells a sharded database where rows live PlanetScale published a primer on data topology in Neki, its Postgres sharding system — a JSON config mapping logical tables to physical shards so routers know where a query goes, rather than every application tracking that itself.

    ( PlanetScale / @sarmag77 / Stripe )

  9. Databricks walks through serving ML features with sub-second freshness The problem is that batch feature pipelines add “minutes to hours of lag,” which is useless when a fraud check “must decide in milliseconds within a user pressing purchase.” Databricks laid out the path it uses to get there: Kafka into Spark Real-Time Mode running rolling-window aggregations against local RocksDB state, written into Lakebase, then joined onto inference requests at serving time.

    ( Databricks / @ramxcodes )

  10. OpenAI says Codex cleared a five year migration in two weeks OpenAI published a case study claiming Asana migrated its frontend tests from Enzyme to React Testing Library in two calendar weeks, a project previously estimated at five years and roughly $6 million in staffing. It ran on a five sentence prompt with up to four agents working in parallel, cost about $12,000 in model and infrastructure, and a human reviewed every change. The claim drew immediate pushback that the five year baseline was overblown and that the comparison sets calendar time against engineer time — the actual work was about 1.5 weeks of engineering effort.

    ( OpenAI / @GergelyOrosz / @WesRoth )

  11. The What is Software Engineering Now? Conversation continues on X. There are software engineers building software without caring about the outcome, must be someone else’s job. Will the move to more outcome based engineering break these engineers?

    ( @samlambert / @getifyX )

  12. You can now change Claude’s default code output style natively. There have been a lot of skills or modifications to Claude to provide more concise output. Caveman (https://github.com/juliusbrussee/caveman) hit 100,135 stars as well as the more advanced adhd skill (https://github.com/uditakhourii/adhd).

    ( @ClaudeDevs )

  13. The game industry is being restructured by content abundance, not layoffs Jacob Navok argues the real disruption in games is algorithmic discovery meeting an unlimited supply of content — Steam releases went from 8,000 a year to 20,000 over seven years, PlayStation from 600 to 2,400, and Roblox’s new Build app points at hundreds of thousands annually. He expects Codex-class tools to make building a game trivial, local LLMs on next-generation consoles to enable AI game masters, and the surviving publishers to compete on design, IP, and marketability rather than on production budgets.

    ( @JNavok )