<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Alex Ialanji</title>
    <description>The latest articles on DEV Community by Alex Ialanji (@alex_iolan).</description>
    <link>https://dev.to/alex_iolan</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4194717%2F1da96942-d73a-434a-a254-1a5f7faba2e5.jpg</url>
      <title>DEV Community: Alex Ialanji</title>
      <link>https://dev.to/alex_iolan</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/alex_iolan"/>
    <language>en</language>
    <item>
      <title>I made the model end every post with a question. The replies stayed the same</title>
      <dc:creator>Alex Ialanji</dc:creator>
      <pubDate>Thu, 19 Nov 2026 14:07:12 +0000</pubDate>
      <link>https://dev.to/alex_iolan/i-made-the-model-end-every-post-with-a-question-the-replies-stayed-the-same-6nb</link>
      <guid>https://dev.to/alex_iolan/i-made-the-model-end-every-post-with-a-question-the-replies-stayed-the-same-6nb</guid>
      <description>&lt;p&gt;&lt;em&gt;This article was created with the help of AI. The numbers come from my own Threads accounts.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;On October 10 I changed the prompt that writes the short posts for my two Russian-language AI news accounts on Threads, Claude News and Promtime. From that day every post ends with a question to the reader. Three weeks later Claude News was getting the same number of replies a day as before, a little under two, and each post was reaching fewer people.&lt;/p&gt;

&lt;p&gt;The idea was simple. A post that just retells a piece of news leaves the reader nothing to say. The old prompt already allowed a closing question, and the model never used it: 279 posts in a row went out without a single question mark.&lt;/p&gt;

&lt;p&gt;Before touching the live accounts I ran 24 real posts through five kinds of ending. "Who has already tried this?" spoke to too few readers. A forced choice between two options looked silly on business news. And an editorial stance made the model invent advice that wasn't in the story. The one that held up asked how the reader does things themselves: "Do you accept AI edits in a document one at a time or all at once?" It went into the prompt with one exception. Heavy news gets no question.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fakrqu2lcly30pw1vjwsf.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fakrqu2lcly30pw1vjwsf.png" alt="Replies per day on Threads before and after the closing question: Claude News 1.8 and 1.8, Promtime 2.0 and 1.0" width="800" height="500"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Claude News got about 1.8 replies a day before the change and about 1.8 after. On Promtime replies halved, though they had been sliding since early September.&lt;/p&gt;

&lt;p&gt;What did change was reach. The median Claude News post went from about 530 views to about 310, and the drop began right after the question appeared.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ffz50nnzjw7vn4gm7b4bp.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ffz50nnzjw7vn4gm7b4bp.png" alt="Median views per Threads post by week for both accounts: a drop in the week of October 12" width="800" height="500"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Whether the question caused it, this data can't say. Posts got a third longer, and other things changed in the accounts in the same weeks. But the people who did see a post answered twice as often.&lt;/p&gt;

&lt;p&gt;The most replied post of those three weeks ended with "Have you ever built yourself an app because the one you needed wasn't in the store?" It got seven replies from about 230 views. The runner-up asked "Have you had to stop working in the middle of the day because of Claude's five-hour limit?"&lt;/p&gt;

&lt;p&gt;Neither is the kind of question I picked in the trials. Both ask about something that has already happened to the reader, and questions like that drew more than twice as many replies per post as questions about habits or "what would you do". It is too early to make a rule of it. About half of those replies came under those two posts.&lt;/p&gt;

&lt;p&gt;The prompt still rotates four kinds of closing question. "Has this happened to you" gets one turn in four.&lt;/p&gt;

&lt;p&gt;The same newsroom publishes AI news in English as &lt;a href="https://promtime.net/en" rel="noopener noreferrer"&gt;Promtime&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>llm</category>
      <category>promptengineering</category>
      <category>socialmedia</category>
    </item>
    <item>
      <title>Next.js prefetches got me rate-limited on my own site. Googlebot had been fetching them by the thousand</title>
      <dc:creator>Alex Ialanji</dc:creator>
      <pubDate>Thu, 12 Nov 2026 14:07:18 +0000</pubDate>
      <link>https://dev.to/alex_iolan/nextjs-prefetches-got-me-rate-limited-on-my-own-site-googlebot-had-been-fetching-them-by-the-280</link>
      <guid>https://dev.to/alex_iolan/nextjs-prefetches-got-me-rate-limited-on-my-own-site-googlebot-had-been-fetching-them-by-the-280</guid>
      <description>&lt;p&gt;&lt;em&gt;This article was created with the help of AI. The numbers come from my own site's Search Console and server logs.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;On October 8 I opened my own news site, promtime.net, on my phone and got Cloudflare's error 1015 instead of a page. My own rate-limiting rule had blocked me, and all I had done was read posts on it.&lt;/p&gt;

&lt;p&gt;The cause was a Next.js feature that exists to make the site feel fast. Two days later Search Console showed me the same feature from Googlebot's side. Between August 10 and October 22 Googlebot made 25,943 requests to the site, and Search Console filed 16,284 of them under "Other file type": not pages, not images, not scripts.&lt;/p&gt;

&lt;h3&gt;
  
  
  What the router does
&lt;/h3&gt;

&lt;p&gt;When a &lt;code&gt;&amp;lt;Link&amp;gt;&lt;/code&gt; scrolls into view, the Next.js router prefetches the page behind it. It requests the same URL with an &lt;code&gt;_rsc&lt;/code&gt; parameter and an &lt;code&gt;RSC: 1&lt;/code&gt; header and gets back the React Server Components payload instead of HTML, so the click that follows feels instant. Googlebot renders pages with JavaScript, and the prefetches fire for it as well.&lt;/p&gt;

&lt;p&gt;All 990 example URLs Search Console lists for "Other file type" end in &lt;code&gt;?_rsc=&lt;/code&gt;. One page can have many of them, because Next.js derives the value from request headers. On September 28, within two minutes, Googlebot fetched one post with six different suffixes. Here are four:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;/c/ru/post/claude-mythos-solves-the-erdos-discrepancy-problem?_rsc=LIEXrFg38zfWwJsj
/c/ru/post/claude-mythos-solves-the-erdos-discrepancy-problem?_rsc=kFwwhQWGfJ28v95A
/c/ru/post/claude-mythos-solves-the-erdos-discrepancy-problem?_rsc=QQ8lwV8YUpwFDVCh
/c/ru/post/claude-mythos-solves-the-erdos-discrepancy-problem?_rsc=vcDMZA7KM_P7ZtS6
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  Six days in August
&lt;/h3&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fz5hj9s878q5wdil2fnfm.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fz5hj9s878q5wdil2fnfm.png" alt="Googlebot requests to promtime.net per day: a spike of &amp;quot;Other file type&amp;quot; requests on August 18-23, almost nothing after" width="800" height="500"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Almost all of it happened between August 18 and 23. On August 19 alone Googlebot fetched 8,984 of these URLs, while on an ordinary day it asked for fewer than a hundred URLs of every other kind. By September the row was down to about ten requests a day.&lt;/p&gt;

&lt;p&gt;The timing matches one change. On August 17-19 the site got newsroom pages, an archive with a month navigator and an infinite feed. Why the spike started, and why it then stopped on its own, these numbers can't show.&lt;/p&gt;

&lt;h3&gt;
  
  
  Who paid for it
&lt;/h3&gt;

&lt;p&gt;Right after that week the site's impressions in Google fell almost to zero, and the crawl is the obvious suspect. The data doesn't back it. Google's guide to crawl budget is written for sites with a million pages or more, or ten thousand if they change every day. Mine has a few thousand. The guide's other warning sign, a pile of URLs stuck in "Discovered - currently not indexed", did show up on my site, but in October, six weeks after the spike was over.&lt;/p&gt;

&lt;p&gt;The ones who paid were readers, me included.&lt;/p&gt;

&lt;p&gt;The rate-limiting rule on my Cloudflare plan sees only the path, not the query string, so to Cloudflare a prefetch looks exactly like someone opening a page. A post page carried more than forty tag and archive links, and the router prefetched each one as it came on screen. One post view could fire up to 58 requests, enough to trip the limit on its own.&lt;/p&gt;

&lt;h3&gt;
  
  
  What I changed
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Turned off prefetch on dense link lists:&lt;/strong&gt; tags, archive months, feed cards. In Next.js that's &lt;code&gt;prefetch={false}&lt;/code&gt; on &lt;code&gt;&amp;lt;Link&amp;gt;&lt;/code&gt;. Every other link still prefetches.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Doubled the Cloudflare rate limit.&lt;/strong&gt;&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Blocked the RSC URLs in robots.txt:&lt;/strong&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Disallow: /*?_rsc=
Disallow: /*&amp;amp;_rsc=
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The first one is now a standing rule in the project's notes: a dense list of links gets &lt;code&gt;prefetch={false}&lt;/code&gt; from the start.&lt;/p&gt;

&lt;h3&gt;
  
  
  How to check your own site
&lt;/h3&gt;

&lt;p&gt;In Search Console, open Settings, then Crawl stats, and click the "Other file type" row in the "By file type" block. The example URLs are at the bottom. If they end in &lt;code&gt;?_rsc=&lt;/code&gt;, you're looking at the same thing.&lt;/p&gt;

&lt;p&gt;In your server logs:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;grep Googlebot access.log | grep -c '_rsc='
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Since the robots.txt rule went in on October 10, my server logs haven't shown a single &lt;code&gt;_rsc&lt;/code&gt; request from Googlebot. Then again, it had almost stopped asking for them long before that. The spike ended on its own, back in August.&lt;/p&gt;

&lt;p&gt;The site from this story also publishes AI news in English, as &lt;a href="https://promtime.net/p/en" rel="noopener noreferrer"&gt;Promtime&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>nextjs</category>
      <category>seo</category>
      <category>webdev</category>
      <category>react</category>
    </item>
    <item>
      <title>The bigger Claude model cost 3x more per news post. A blind reviewer preferred the cheaper one</title>
      <dc:creator>Alex Ialanji</dc:creator>
      <pubDate>Thu, 05 Nov 2026 14:07:25 +0000</pubDate>
      <link>https://dev.to/alex_iolan/the-bigger-claude-model-cost-3x-more-per-news-post-a-blind-reviewer-preferred-the-cheaper-one-2dci</link>
      <guid>https://dev.to/alex_iolan/the-bigger-claude-model-cost-3x-more-per-news-post-a-blind-reviewer-preferred-the-cheaper-one-2dci</guid>
      <description>&lt;p&gt;&lt;em&gt;This article was created with the help of AI. The test, the posts and the numbers come from my own news pipeline.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;On October 8 I ran ten of my news posts through two Claude models. Opus 5.5 already wrote them for me every day. Fable 5.1 is the bigger model, and its price per token is 2.5 times higher. A blind reviewer preferred Opus in six of the ten posts, each time by a slight margin. Fable took twice as long and cost 3.2 times as much per post.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fzhg8qrlq124x2blcr1nn.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fzhg8qrlq124x2blcr1nn.png" alt="Ten news posts, two models: reviewer's pick, scores, invented facts, cost and time" width="800" height="500"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;I run two AI news feeds, Claude News and Promtime. Claude takes an English source and rewrites it into a post by each channel's rules. Claude News is for developers: Claude releases, the API, pricing. Promtime covers AI in general, in plainer language. Every story goes out in two languages and in several formats, from a long Telegram post to a short one for social networks, so the rewrites add up to a few dozen calls a day.&lt;/p&gt;

&lt;p&gt;Benchmarks mostly measure math and code. My task is narrower: retell a news story accurately, keep it lively and stay strictly inside the format. I wanted to see the difference on my own posts before paying for it.&lt;/p&gt;

&lt;p&gt;Claude Code built the test harness. I set the task and checked what came out.&lt;/p&gt;

&lt;h3&gt;
  
  
  The setup
&lt;/h3&gt;

&lt;p&gt;I took the five latest published posts from each feed, only those with a single source between 600 and 20,000 characters long.&lt;/p&gt;

&lt;p&gt;Both models got exactly what the live pipeline sends: the channel rules, a description of the style, examples of good posts and a hint for the ending. The same code assembled the request. Both were called directly through the API with the same settings, one run per post.&lt;/p&gt;

&lt;p&gt;I didn't judge them myself. A separate Claude session got the texts labelled A and B, shuffled for every post, so it never knew which model wrote what. It checked numbers and names against the source and scored each text from 1 to 5 on seven criteria: facts, headline, structure, language, ending, the short social post and the channel rules.&lt;/p&gt;

&lt;p&gt;The reviewer didn't get the last word either. Five of the distortions it reported were checked against the sources separately, and all five held up. About twenty more numbers and names from both models were spot-checked, with no mismatches.&lt;/p&gt;

&lt;h3&gt;
  
  
  Six to four, and every call a close one
&lt;/h3&gt;

&lt;p&gt;Opus won six posts, Fable four. The reviewer called all ten decisions a slight edge. Neither model invented a single fact. Opus made eight minor distortions, Fable seven.&lt;/p&gt;

&lt;p&gt;By criterion, Fable was a little more accurate on facts, 4.4 against 4.3. Opus was better at endings, 4.5 against 4.2, and at the short social post, 4.2 against 3.8.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fbrx8sy0njixqsmwh49rm.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fbrx8sy0njixqsmwh49rm.png" alt="Reviewer's mean scores for Opus 5.5 and Fable 5.1 on seven criteria" width="800" height="500"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The total hides a split. In the developer feed the reviewer picked Fable in four posts out of five. In the general feed it picked Opus in all five. Five posts per feed is too few to call it, but it points one way: dense, precise text wins when the reader is technical.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fgby88qogfuzdz9l6n5m0.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fgby88qogfuzdz9l6n5m0.png" alt="Which model the blind reviewer picked, post by post, in each feed" width="800" height="500"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Ten posts with one run each is enough to see that there's no big gap. It isn't enough to rank two models this close. The judge was a Claude session too, so its taste isn't neutral ground. And it's one task: retelling news from a given source under strict rules. Code, long reasoning and agent work are another matter.&lt;/p&gt;

&lt;h3&gt;
  
  
  What Fable did better, and where it slipped
&lt;/h3&gt;

&lt;p&gt;&lt;strong&gt;It digs deeper.&lt;/strong&gt; One source was a page of measurements for the new Haiku 5.5. Both models put the main number in the headline: the first token of an answer arrives after 415 seconds. Opus ended its post on the bright side: the wait at least comes cheap.&lt;/p&gt;

&lt;p&gt;Fable noticed something the price list doesn't show. The model is very verbose. Its run through the benchmark index took 440 million output tokens against a median of 100 million, so by the cost of one task it ranks only 46th of 182. That was the real story. It also took Fable 130 seconds and nearly five times the cost of the Opus post, which took 34.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;It is looser with limits.&lt;/strong&gt; In a story about Meta and Microsoft, its short social post came out at 248 characters against a hard ceiling of 240. Opus wrote 164. That was the only length violation in the whole run. It was also the first thing the reviewer named when it gave the post to Opus, even though it found Fable's text slightly more precise.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;It generalizes more boldly, and that's where it went wrong.&lt;/strong&gt; In a story about GPT-6, Fable wrote that Sol goes to paid plans and Luna to free ones. In the source, Luna goes to the Free and Go tiers, and Go is paid. The reviewer flagged it in the headline, the ending and the social post, and still rated Fable's headline as the stronger one. Opus took that post by a slight margin.&lt;/p&gt;

&lt;h3&gt;
  
  
  Why 2.5x on the price list became 3.7x on the bill
&lt;/h3&gt;

&lt;p&gt;By the price list, Fable costs 2.5 times as much as Opus for input and output tokens alike. Per post it came out 3.2 times as expensive by the median and 3.7 times on average.&lt;/p&gt;

&lt;p&gt;The difference is in the output. Both models read the same prompt, but Fable spends twice as many tokens on reasoning and the answer: 6.2 thousand on average against 3 thousand. Output tokens are the expensive ones, and that's also where the time goes: 63 seconds against 29 by the median, 130 at worst.&lt;/p&gt;

&lt;p&gt;And the prompt is big. Every rewrite carries 16-24 thousand input tokens: the channel rules, the style description, the examples of good posts. The source itself is the smaller part. On average, Fable's input alone cost more than a whole Opus call, answer included. Before writing a word, Fable had already cost more than a finished Opus post.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fdftp3g0pem1ji01iibg2.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fdftp3g0pem1ji01iibg2.png" alt="What one post costs in input and output, with an Opus post as the unit" width="800" height="500"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  What I kept
&lt;/h3&gt;

&lt;p&gt;Rewriting stayed on Opus 5.5, and it's still there a month later. Neither model made anything up, and both slipped on small details about equally often. For this task the rules and examples in the prompt seem to matter more than the class of the model.&lt;/p&gt;

&lt;p&gt;I'll come back to Fable if I start a feed where getting to the bottom of a story matters more than fitting the format. The Haiku 5.5 post shows it earns its price there.&lt;/p&gt;

&lt;p&gt;Both feeds from this test also publish in English, on &lt;a href="https://promtime.net/p/en" rel="noopener noreferrer"&gt;Promtime&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>claude</category>
      <category>ai</category>
      <category>llm</category>
      <category>machinelearning</category>
    </item>
    <item>
      <title>Six AI agents froze my production server. Each one did the right thing</title>
      <dc:creator>Alex Ialanji</dc:creator>
      <pubDate>Thu, 29 Oct 2026 14:07:34 +0000</pubDate>
      <link>https://dev.to/alex_iolan/six-ai-agents-froze-my-production-server-each-one-did-the-right-thing-4bo6</link>
      <guid>https://dev.to/alex_iolan/six-ai-agents-froze-my-production-server-each-one-did-the-right-thing-4bo6</guid>
      <description>&lt;p&gt;&lt;em&gt;This article was created with the help of AI. The incident, the logs and the numbers are from my own server.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;On September 25 at 19:01 UTC my server stopped responding. The site, the admin panel and the publishing queue all froze at once. Seventeen minutes later I hard-rebooted it.&lt;/p&gt;

&lt;p&gt;There was no attack and no traffic spike. I had given Claude Code a list of six tasks, and it started six agents in parallel. Each one dutifully installed dependencies, ran the type checker and the tests, and built the project. On the same machine that serves the site.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fcza92iiko16f4yjga6mu.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fcza92iiko16f4yjga6mu.png" alt="Memory asked for by one, two and six parallel builds against the 7 GB that were free" width="800" height="500"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  The arithmetic nobody did
&lt;/h3&gt;

&lt;p&gt;Take round numbers close to mine. The server has 10 GB of RAM, and the live services use about 3. That leaves 7.&lt;/p&gt;

&lt;p&gt;One typecheck or build of a project this size takes 1.5 to 3 GB. Six at once need 9 to 18. That doesn't fit into 7 however you slice it.&lt;/p&gt;

&lt;p&gt;Anyone could have added this up. Nobody did, not me and not the model.&lt;/p&gt;

&lt;h3&gt;
  
  
  Why the server didn't save itself
&lt;/h3&gt;

&lt;p&gt;I expected Linux to kill the hungriest process and let the rest carry on. It didn't. The journal has no out-of-memory kills at all.&lt;/p&gt;

&lt;p&gt;What it does have: system services reporting they were running "under memory pressure", and Docker saying it could not send a kill signal to a container process. Then nothing until the reboot.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fgf30c98v712na7usko7d.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fgf30c98v712na7usko7d.png" alt="The last journal lines before the freeze: memory pressure, a failed KILL, then 17 minutes of silence" width="800" height="500"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Swap did it. When RAM runs out, the kernel starts moving pages to disk and back, and that eats all the CPU time. On paper there is still memory, so the kernel sees no reason to kill anything. In practice the machine is busy only with itself. The server didn't crash, it hung, and that's worse. A crashed service restarts on its own. A hung one waits for a human.&lt;/p&gt;

&lt;h3&gt;
  
  
  Why the agent did it
&lt;/h3&gt;

&lt;p&gt;For agents, parallelism is a feature. Six independent tasks go to six workers, each with its own copy of the code in a separate git worktree, nobody in anyone's way. In the code, that held.&lt;/p&gt;

&lt;p&gt;Where they got in each other's way was memory, and nobody had told them about it. A model doesn't know how much RAM the machine has or that a live site runs next to it, unless it checks or someone tells it. And "finish the job and make sure it builds" is usually a good habit. Run six times over on a small server, that habit took the server down.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F54w7mlso9qeymkani5hh.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F54w7mlso9qeymkani5hh.png" alt="Six agents, each with dependencies, typecheck, tests and build, all landing on one server" width="800" height="500"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  What I changed
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;No more than two heavy jobs at once.&lt;/strong&gt; Heavy means a build, a typecheck of the whole project, or a Docker image build.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Each agent checks only its own part.&lt;/strong&gt; Types and tests for the package it changed, no full build.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Full builds run one at a time,&lt;/strong&gt; from the main session, not inside every agent.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The rule lives in CLAUDE.md,&lt;/strong&gt; the project notes Claude reads at the start of every session. Since then it brings the rule up on its own when the task list gets long.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  Doing the math for your machine
&lt;/h3&gt;

&lt;p&gt;Take total memory, subtract what's already in use, keep a gigabyte in reserve, and divide by the peak of one job. With my round numbers that's (10 - 3 - 1) / 3, so two jobs.&lt;/p&gt;

&lt;p&gt;You don't have to guess the peak. Measure it:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;free &lt;span class="nt"&gt;-h&lt;/span&gt;
/usr/bin/time &lt;span class="nt"&gt;-v&lt;/span&gt; npm run build 2&amp;gt;&amp;amp;1 | &lt;span class="nb"&gt;grep&lt;/span&gt; &lt;span class="s1"&gt;'Maximum resident'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The first command shows how much memory is free right now. The second runs the build and prints the peak of its largest process, in kilobytes. A build that forks parallel workers uses more than that in total, so round up.&lt;/p&gt;

&lt;p&gt;Two safeguards for when the math lets you down. First, don't build on the production server at all if you have a separate machine or CI. Second, install something like earlyoom, which kills the hungriest process early instead of waiting for the system to drown in swap.&lt;/p&gt;

&lt;p&gt;I put in the second one late, and I shouldn't have. Three weeks later the server hung again, this time with no agents involved: a nightly job that checks for dependency updates built a Docker image and started a second copy of a service next to the live ones. Same arithmetic, only nobody was around to do it.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fs3i9f75ln16x96oeswlw.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fs3i9f75ln16x96oeswlw.png" alt="What happens when memory runs out, without earlyoom and with it" width="800" height="500"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;This is what &lt;code&gt;/etc/default/earlyoom&lt;/code&gt; looks like on my box now:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nv"&gt;EARLYOOM_ARGS&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"-m 5 -s 100 --prefer ^(chrome|chromium|tsc|next-server)&lt;/span&gt;&lt;span class="nv"&gt;$ &lt;/span&gt;&lt;span class="s2"&gt;--avoid ^(postgres|dockerd|containerd|containerd-shim|systemd|sshd|journald)$"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;-s 100&lt;/code&gt; is the part that matters here. By default earlyoom acts only when both RAM and swap are nearly gone, and my server was already thrashing while swap was still free. With &lt;code&gt;-s 100&lt;/code&gt; it acts on low RAM alone. Builds and the browser go first, and the database and Docker are never touched.&lt;/p&gt;

&lt;p&gt;Each of the six agents did the right thing. The only thing wrong was how many there were.&lt;/p&gt;

&lt;p&gt;I collect findings like this about Claude and AI on &lt;a href="https://promtime.net/c/en" rel="noopener noreferrer"&gt;Promtime&lt;/a&gt;, the news site this server runs.&lt;/p&gt;

</description>
      <category>claude</category>
      <category>ai</category>
      <category>devops</category>
      <category>linux</category>
    </item>
    <item>
      <title>Claude Code quietly deletes your session history after 30 days. I noticed three and a half months in</title>
      <dc:creator>Alex Ialanji</dc:creator>
      <pubDate>Thu, 22 Oct 2026 14:07:28 +0000</pubDate>
      <link>https://dev.to/alex_iolan/claude-code-quietly-deletes-your-session-history-after-30-days-i-noticed-three-and-a-half-months-in-334j</link>
      <guid>https://dev.to/alex_iolan/claude-code-quietly-deletes-your-session-history-after-30-days-i-noticed-three-and-a-half-months-in-334j</guid>
      <description>&lt;p&gt;&lt;em&gt;This article was created with the help of AI. The numbers come from my own server.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;My first prompt to Claude Code on my work server is dated May 26. The oldest session transcript still on disk is from September 11. Everything in between is gone, and I never saw a warning.&lt;/p&gt;

&lt;p&gt;It isn't a bug. Claude Code keeps session transcripts for 30 days and then deletes them. The docs say so, on a page few people open. One line in the settings changes that, but only going forward.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F8uk3fqwjbe9t7xvtc62y.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F8uk3fqwjbe9t7xvtc62y.png" alt="Claude Code session transcripts by age" width="800" height="500"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  What disappears
&lt;/h3&gt;

&lt;p&gt;Every Claude Code session is written to a file under &lt;code&gt;~/.claude/projects/&lt;/code&gt;. It holds the whole conversation: your messages, the model's replies, every tool call and its result. &lt;code&gt;/resume&lt;/code&gt; reads these files to take you back into an old session.&lt;/p&gt;

&lt;p&gt;After 30 days the file is deleted. Nothing tells you. The session simply stops showing up in &lt;code&gt;/resume&lt;/code&gt;, and its subagent transcripts and any orphaned git worktrees from parallel tasks go with it.&lt;/p&gt;

&lt;p&gt;Some things survive:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Prompt history.&lt;/strong&gt; &lt;code&gt;history.jsonl&lt;/code&gt; keeps what you typed, unless HIPAA settings apply to your account. So my May questions are still there. The answers are not.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Auto memory.&lt;/strong&gt; The notes Claude keeps about a project are not part of the sweep.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Claude Desktop and Cowork sessions.&lt;/strong&gt; Since v2.1.248 they are kept at any age, unless you set &lt;code&gt;desktopSessionCleanupPeriodDays&lt;/code&gt;.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fzgiwbtiee4zmoc0nisvh.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fzgiwbtiee4zmoc0nisvh.png" alt="What the cleanup deletes and what survives" width="800" height="500"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  I checked my own server
&lt;/h3&gt;

&lt;p&gt;On October 11 I had 724 transcript files taking 861 MB. By age: 181 from the last week, 62 from the week before, 338 from the third week and 143 older than that. Not one older than 30 days.&lt;/p&gt;

&lt;p&gt;I can't check what was in the deleted ones. Probably bug hunts and notes on why something was built a certain way. Exactly what you open an old session for.&lt;/p&gt;

&lt;h3&gt;
  
  
  Has it already happened to you?
&lt;/h3&gt;

&lt;p&gt;Three commands:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;grep &lt;/span&gt;cleanupPeriodDays ~/.claude/settings.json
find ~/.claude/projects &lt;span class="nt"&gt;-name&lt;/span&gt; &lt;span class="s1"&gt;'*.jsonl'&lt;/span&gt; | &lt;span class="nb"&gt;wc&lt;/span&gt; &lt;span class="nt"&gt;-l&lt;/span&gt;
find ~/.claude/projects &lt;span class="nt"&gt;-name&lt;/span&gt; &lt;span class="s1"&gt;'*.jsonl'&lt;/span&gt; &lt;span class="nt"&gt;-mtime&lt;/span&gt; +30 | &lt;span class="nb"&gt;wc&lt;/span&gt; &lt;span class="nt"&gt;-l&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The first shows whether you set a retention period. If it prints nothing, you're on the 30-day default. The second counts all transcripts, the third only those older than 30 days. If the third prints zero and you've used Claude Code for more than a month, the sweep has already run.&lt;/p&gt;

&lt;h3&gt;
  
  
  One line in settings
&lt;/h3&gt;

&lt;p&gt;Add the key to &lt;code&gt;~/.claude/settings.json&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"cleanupPeriodDays"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="mi"&gt;3650&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That's about ten years. Three things to watch:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Zero doesn't mean never.&lt;/strong&gt; &lt;code&gt;0&lt;/code&gt; fails validation, and the minimum is 1. To keep things long, pick a big number.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;It only works forward.&lt;/strong&gt; Deleted files don't come back.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Disk adds up.&lt;/strong&gt; A month of my work took 861 MB, roughly 10 GB a year. I run a lot of agents, so yours may be far less.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  The catch: those files hold everything
&lt;/h3&gt;

&lt;p&gt;Transcripts are plain text, unencrypted. The docs are blunt about it: if a tool reads a &lt;code&gt;.env&lt;/code&gt; file or a command prints a credential, that value lands in the session file. That is why the default is short.&lt;/p&gt;

&lt;p&gt;The longer you keep history, the more secrets pile up on disk. If you raise the limit, deny reads of credential files with permission rules, and keep &lt;code&gt;~/.claude/projects/&lt;/code&gt; out of unencrypted backups.&lt;/p&gt;

&lt;p&gt;And if you'd rather keep nothing, &lt;code&gt;CLAUDE_CODE_SKIP_PROMPT_HISTORY&lt;/code&gt; stops writing transcripts and prompt history, &lt;code&gt;--no-session-persistence&lt;/code&gt; does the same for a single &lt;code&gt;-p&lt;/code&gt; run, and &lt;code&gt;claude purge&lt;/code&gt; wipes what is stored for one project.&lt;/p&gt;

&lt;h3&gt;
  
  
  What I picked
&lt;/h3&gt;

&lt;p&gt;365 days. A year covers nearly every time I go back to an old decision, and the disk use and the secrets stay manageable.&lt;/p&gt;

&lt;p&gt;Funny that a tool which holds your whole project in its head forgets its own history after a month.&lt;/p&gt;

&lt;p&gt;The details are in the &lt;a href="https://code.claude.com/docs/en/claude-directory#cleaned-up-automatically" rel="noopener noreferrer"&gt;Claude Code docs&lt;/a&gt;. I collect findings like this on &lt;a href="https://promtime.net/c/en" rel="noopener noreferrer"&gt;Promtime&lt;/a&gt;, the news site I edit.&lt;/p&gt;

</description>
      <category>claude</category>
      <category>ai</category>
      <category>devtools</category>
      <category>productivity</category>
    </item>
  </channel>
</rss>
