Dev Breakfast · 2026-09-22

Linear's test suite has nearly quadrupled since the beginning of the year, but the wait time for PRs on CI has only dropped from over 6 minutes to just over 5 minutes, cutting the runner time for each test by about half. My judgment is that the real benefit of this work isn't in that one minute, but in the runner bill saved by steps like switching from GitHub Actions, using tsgo, and converting lint to pure AST.

Test volume quadrupled, PR wait only dropped from 6 minutes to 5 minutes

Linear's CTO assigned engineers an issue with the title 'CI costs too high', and also requested making CI faster. The reason is practical: Agents make coding an order of magnitude faster, but verification hasn't kept up. Each PR still has to queue for CI, so CI has become a bottleneck, infrastructure bills increase, and people and Agents wait together. They focused on two metrics: how long PRs wait on CI and how much runner time is consumed. The result is that the test suite has nearly quadrupled since the beginning of the year, wait time has dropped from over 6 minutes to just over 5 minutes, and runner time per test has been cut by about half. There are four types of changes: switching from GitHub Actions to third-party runners (34% faster on average with comparable metrics, tsc 52% faster), switching to tsgo to reduce type checking median time by 73%, rewriting dependency-type-dependent lint rules to pure AST static analysis (API lint time reduced by 68%, full-repo lint reduced by 55%), and reducing duplicate setups. If you're also held up by CI, first check which step is gating other tasks, then talk about optimization.

Sources:

• hackernews · julian_digital

Sandboxing isn't black magic: figure out who to lock up first

This article on software sandboxing basics opens by acknowledging that this area is largely unmapped: the components needed to implement sandboxes are scattered, and pioneers haven't yet assembled a 'world map' to guide newcomers. The author defines sandboxing based on a 2009 Hack In The Box presentation—limiting process privileges programmatically, not requiring admin privileges on the machine, and being able to drop privileges autonomously. He specifically notes that his 2025 views have diverged from that earlier framework, especially regarding 'whether to use superuser APIs'. The code examples in the article require the unreleased Emilua 0.11, needing the latest commit from the development branch. If you're working on plugins, script execution, or third-party code hosting, this is worth reading; but don't expect it to give you a ready-made security roadmap—it's more about saying: first, think clearly about what you want to restrict.

Sources:

• Software sandboxing: The basics (2025)

On the day MCP was questioned, five articles criticizing AI topped the charts simultaneously

On the same day's hot list, five not-so-flattering articles crowded in, with titles pointing to MCP, prompts, chatbots' 'mind-reading', and the AI bubble. One title asks: Was MCP always a bad idea? Another discusses the 'LLMentalist effect', likening chatbot persuasiveness to a fortune teller's cold reading—using statistically general statements to make you feel they're speaking specifically to you. Another report states that AI chatbots give wrong answers to financial queries 'most of the time'.

The common thread isn't the conclusions but the way they ask questions: they all ask the same thing—have we mistaken statistical fluency for understanding? For you who write code daily, only one point is practically actionable: wherever the model touches numbers, money, or state, the output must be externally verifiable, don't rely on it to self-assess correctness. I judge that the value of the MCP piece isn't in its conclusion, but in forcing you to think clearly about what this protocol layer actually saves you from doing.

Sources:

• The LLMentalist Effect (2023)
• MCP was always a bad idea?
• Prompts aren't Real
• AI chatbots give wrong answers to financial queries 'most of the time'
• I stopped drinking the AI Kool-Aid

macOS 27 can't turn off Siri: deleting model files brings it back after restart

Someone on macOS 27 found a workaround: first set Siri to 'Not Allowed' in 'Screen Time', then enter Recovery mode to mount the Data volume, delete the com_apple_MobileAsset_UAF_FM* model files under /System/Library/AssetsV2, and restart. The storage usage of Apple Intelligence indeed disappears. But the cost is real: mobileassetd also manages resources like voice, dictionary, and speech recognition; cutting it off en masse will have side effects; and the models will be re-downloaded after restart, so deleting files isn't a persistent solution. For machines with 8GB M1 and 256GB storage, several gigabytes isn't a small amount. Trying to turn off only Mac and not iPhone won't work either, as changing language syncs across all devices. My view is that system-level AI should have a real toggle, not require users to enter Recovery mode and delete files.

Sources:

• hackernews · ano-ther
• hackernews · semidror
• hackernews · alwillis

Python Workers GA: running FastAPI on the edge without JS glue

Cloudflare's Python Workers are now generally available, two years after the initial launch. The key change this time is that type conversion is built into the runtime and Python SDK: previously, to send a dict to a Queue, you had to write glue code like to_js({"key": "value"}, dict_converter=js.Object.fromEntries), but now you can directly use self.env.QUEUE.send({"key": "value"}). FastAPI, Django, and Flask can all run, underpinned by a Wasm-compiled Pyodide interpreter. The trade-off is clear: Pyodide means cold starts and package sizes need recalculation—don't equate it with native CPython. Look at the principles first; whether this work is worth it depends on how sensitive your interfaces are to latency.

Sources:

• Python Workers are now generally available

M5 Ultra Mac Studio review: 256 GB memory for local Agents

We've previously discussed benchmarks for running models locally on Mac; today brings generational progress: the M5 Ultra Mac Studio with 256 GB unified memory. The author tested it against the previous M3 Ultra (512 GB) and an RTX 5090 host for four days. The conclusion is that local Agents can finally be used as primary tools—he set Qwen3.8-Flash-Next as the default model for Open Minis and Hermes Agent, with long contexts and multi-turn loops no longer slowing down. The trade-offs are stated plainly: the 5090 still has advantages with higher memory bandwidth, but size, heat, and noise are factors. A reminder: the author identifies as not an AI developer, only doing inference without training or fine-tuning, so these numbers correspond to 'running ready-made models'. Also, the M6 Mac mini review mentions a $300 price hike, so the cost-effectiveness of this small machine line needs separate calculation. First see what it can replace in your workflow, then decide whether to pay for memory.

Sources:

• hackernews · piotrgrabowski
• hackernews · throw0101c

Sublime Text 4213: plugin host jumps from Python 3.8 to 3.14

Sublime Text has released Build 4213, upgrading the plugin host from Python 3.8 directly to 3.14, while the built-in Python 3.3 is disabled by default. Officials say future versions will completely remove it. If you have old plugins that won't run, you can set "disable_plugin_host_3.3": false to re-enable 3.3, but this is only a stopgap. Additionally, Windows minimum requirements are now Windows 10, macOS minimum is 10.13, C# syntax highlighting has been rewritten, and file icon themes and the open_terminal command have been added.

Sources:

• Sublime Text Build 4213

Google open-sources AX: claims running a billion Agent tasks per cluster

AX is Google's open-source Agent orchestration framework, describing tasks declaratively with task.yaml, and handling sandboxes, workspace mounting, and network isolation. Officials claim a single cluster can achieve up to a billion concurrent Agent sessions, with idle tasks checkpointed and restored within a second. It runs on the Agent Substrate runtime and includes 'generative workspace'—describing an environment in a sentence, it installs toolchains itself.

Don't rush to believe the numbers: the billion is the concurrent session limit 'per cluster', not measured throughput, and the original text provides no benchmarks or latency distributions. What's truly worth looking at is the fact that it treats Agents as stateful actors to schedule—the isolation, checkpoint-resuming, and active billing that annoy you when writing Agents, it aims to solve in one go.

Sources:

• AX – Google's Open Agentic Orchestrator

---

If CI is already gating your work, would you first modify the runner or first adjust type checking and lint? See you tomorrow morning at 8.

This issue selected 8 articles from 65 pieces of information from X / Hacker News / GitHub Trending over the past 24 hours (written hourly throughout the day, fact-checked, and selected for morning compilation). Content generated with LLM assistance, each with original source links; for important decisions, please cross-verify.