Dev Breakfast · 2026-10-19

Codex Desktop's Crashpad pending dumps have no cap—at least 5GB written per day, most likely because some process keeps crashing and restarting. If you keep the desktop version open regularly, go check your disk usage first—this kind of problem doesn't touch code logic, but it will really eat your machine alive.

Codex Desktop writes 5GB of crash dumps a day

A GitHub issue points out that Codex Desktop keeps generating Crashpad pending dumps with no upper limit, writing at least 5GB in a single day. Crashpad is a crash-collection component common to Chromium-based apps; normally it only writes one file when a crash occurs, but now it has become a continuous write, most likely because some process keeps crashing and restarting. If you keep Codex Desktop open all the time, it's worth taking a look at disk usage before your system drive fills up and you notice. This kind of problem doesn't involve code logic, but it will genuinely eat your machine.

Sources:

• Codex Desktop continuously generates Crashpad pending dumps

DeepSeek opens search & fetch APIs: $0.75 per 1,000 successful requests

DeepSeek has turned its search engine and web-scraping capabilities into two JSON APIs: /v1/tools/search returns structured results by query, and /v1/tools/fetch converts a single web page into Markdown. Authentication reuses the existing API key, and pricing is $0.75 per 1,000 successful requests. search supports query operators like site:, -site:, and exact phrases; count allows up to 50 results; and turning on need_content brings back the page body directly. For people building Agents and RAG, this saves you from maintaining your own scraping pipeline and anti-bot setup. But how a "successful request" is counted, and whether pages that fail to fetch are charged, aren't stated in the original—you should load-test it yourself before integrating. Map out the failure paths first, then talk about saving money.

Sources:

• Search & Fetch | DeepSeek API Docs

Websites start closing their doors to AI agents: scraping is no longer allowed by default

The Verge reports that more and more websites are blocking AI agents, no longer treating scraping as something allowed by default. The original piece only has a headline and a lede; the specific blocking methods and which sites are involved aren't elaborated. For people who write code, this means the crawling, aggregation, and Agent data-fetching flows you build—paths that used to get through with just robots.txt and a UA—are being shut off by more proactive mechanisms. Don't rush to complain; think about it the other way around: sites are closing up mostly because their bandwidth and content are being taken for free. My judgment is that licensed data interfaces will become increasingly valuable, and switching your data-fetching path from "scrape" to "integrate" early is less work than patching afterwards.

Sources:

• Websites are closing their doors to AI agents

Sphinx: answer a question before you push, and it blocks your AI code if you get it wrong

Sphinx is an open-source Git hook. Once installed (npm install -g sphinx-understand, requires Node 22.19+, runs on macOS, Linux, or WSL), when you or your agent push to a protected branch, it reads the diff and generates a question about that code—such as "if the event ID in a webhook is stored before charge(), what happens on a retry?". Get it wrong and it follows up with the next scenario; only a correct answer lets you through. It supports Codex, Claude Code, Cursor, OpenCode, Pi, and plain Git. The answers aren't sent back to the coding agent, and it doesn't read .env or secrets. If you're in a hurry you can skip it with sphinx yolo, but every skip is logged.

The value of this thing isn't "quizzing you"—it's that it forces you to actually read the agent-generated diff. The only caveat I'd raise: the questions are generated by a model, testing what it thinks matters, not the real risk points. Don't treat it as a quality gate.

Sources:

• hackernews · critiquebrowser

Anthropic's proposal: regulation only triggers at 50 deaths or $1B in damages

When we covered Anthropic's employment impact a few issues ago, we saw a fairly gradualist path—job structures shifting slowly, with no hard thresholds. Today's item is the same company from another angle: it has put forward an AI safety bill proposal with very specific trigger lines—50 deaths or $1 billion in damages—regulation kicks in only above that line. The report also notes that this design would effectively make the Department of Commerce the only agency capable of stopping a model.

What does this have to do with writing code? In the short term, you don't need to change anything, but it's worth knowing what the regulatory posture looks like: the threshold is set on large-scale harm after the fact, rather than pre-release access review. In other words, the gate before a model goes live—this proposal basically has no intention of adding one. My view: quantifiable trigger lines are at least more workable than vague phrasing like "potentially harmful," but consolidating the stop-power into a single department is itself a path that should be questioned.

Sources:

• Anthropic's proposed AI safety bill only kicks in if 50 deaths or $1B in damages

---

This issue selected 5 items from 50 pieces of information across X / Hacker News / GitHub Trending over the past 24 hours (collected hourly throughout the day, fact-checked, and compiled in the morning). Content was generated with LLM assistance; each item is linked to its original source—please cross-verify for important decisions.