<?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: voidcraft.world</title>
    <description>The latest articles on DEV Community by voidcraft.world (@voidcraft_world).</description>
    <link>https://dev.to/voidcraft_world</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%2F4090413%2F3d5fc64a-8e78-4ad5-97b9-d20b6852ad59.png</url>
      <title>DEV Community: voidcraft.world</title>
      <link>https://dev.to/voidcraft_world</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/voidcraft_world"/>
    <language>en</language>
    <item>
      <title>Run your coding agents in tmux and let a local model tell you which replies to read</title>
      <dc:creator>voidcraft.world</dc:creator>
      <pubDate>Fri, 20 Nov 2026 19:07:45 +0000</pubDate>
      <link>https://dev.to/voidcraft_world/run-your-coding-agents-in-tmux-and-let-a-local-model-tell-you-which-replies-to-read-2k30</link>
      <guid>https://dev.to/voidcraft_world/run-your-coding-agents-in-tmux-and-let-a-local-model-tell-you-which-replies-to-read-2k30</guid>
      <description>&lt;p&gt;In two weeks I sent my coding agents 1,523 prompts and read every closing message. I was fed up, so I put a small local model in front of them. It left me 696 to read.&lt;/p&gt;

&lt;p&gt;The tool is TLDRack. It is open source (MIT) and runs on macOS with Claude Code or Codex. Here is how to get it running and what you get.&lt;/p&gt;

&lt;h2&gt;
  
  
  What it does
&lt;/h2&gt;

&lt;p&gt;TLDRack runs your coding agents side by side in tmux and shows every one of them on one local page. Sessions name themselves from your prompt, so you can find a task at a glance. When a turn finishes, a local model reads the closing message and marks it done or needs you.&lt;/p&gt;

&lt;h2&gt;
  
  
  Install
&lt;/h2&gt;

&lt;p&gt;You need git, tmux, ttyd, Node 22.18 or newer, uv, and one agent CLI you have logged in to once.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;brew install tmux ttyd node uv

git clone https://github.com/voidcraft-world/tldrack
&lt;span class="nb"&gt;cd&lt;/span&gt; tldrack
uv tool install --editable ./bridge   &lt;span class="c"&gt;# puts `tldrack` on your PATH&lt;/span&gt;
tldrack doctor                        &lt;span class="c"&gt;# every missing part prints its fix&lt;/span&gt;
tldrack hooks --install               &lt;span class="c"&gt;# your agent CLIs report their state to the rack&lt;/span&gt;
&lt;span class="nb"&gt;cd&lt;/span&gt; /path/to/your/project              &lt;span class="c"&gt;# new agents start where you run `tldrack up`&lt;/span&gt;
tldrack up                            &lt;span class="c"&gt;# builds the page, then opens http://127.0.0.1:7682/&lt;/span&gt;
tldrack laya install                  &lt;span class="c"&gt;# the model's runtime and base weights, about 800 MB&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;It takes about ten minutes, and most of that is downloads.&lt;/p&gt;

&lt;h2&gt;
  
  
  What you see
&lt;/h2&gt;

&lt;p&gt;One row per agent session, each with a name it gave itself and a grade once its turn ends. Next to that: the selected agent's terminal, its worktree, a live view of what every agent is doing, and any files or drawings it made for you to review.&lt;/p&gt;

&lt;p&gt;The grade is the part that saves reading. The model answers one question about the closing message: does this say work is still open? It works with no training at all, and that is your first run. After that you fix its misses with one tap per turn, and those grades train a version fitted to your agents.&lt;/p&gt;

&lt;h2&gt;
  
  
  Three ways to run a big task
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Individual sessions that you manage yourself. The least money, and the slowest.&lt;/li&gt;
&lt;li&gt;One iterative agent with sub-agents (/goal, /loop). The middle.&lt;/li&gt;
&lt;li&gt;A LEAD with workers, for work that splits cleanly. The fastest route to a good result, and the most expensive.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;That ranking is my judgment from running all three. I have not measured it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Limits
&lt;/h2&gt;

&lt;p&gt;The model judges the message, not the work. A turn that claims it is done and is not still reads as done. Treat the grade as where to look first.&lt;/p&gt;

&lt;p&gt;Video (one minute): &lt;a href="https://youtu.be/dnTCRYkfKiY" rel="noopener noreferrer"&gt;https://youtu.be/dnTCRYkfKiY&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Repo: &lt;a href="https://github.com/voidcraft-world/tldrack" rel="noopener noreferrer"&gt;https://github.com/voidcraft-world/tldrack&lt;/a&gt;&lt;/p&gt;

</description>
      <category>agents</category>
      <category>ai</category>
      <category>automation</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Reject hair, become Saitama: the habit bar that rises as you do</title>
      <dc:creator>voidcraft.world</dc:creator>
      <pubDate>Tue, 25 Aug 2026 06:26:52 +0000</pubDate>
      <link>https://dev.to/voidcraft_world/reject-hair-become-saitama-the-habit-bar-that-rises-as-you-do-3j87</link>
      <guid>https://dev.to/voidcraft_world/reject-hair-become-saitama-the-habit-bar-that-rises-as-you-do-3j87</guid>
      <description>&lt;p&gt;British Cycling hired Dave Brailsford in 2003. His theory was that if you find one percent in the saddle, one percent in the pillow a rider sleeps on and one percent in how the team washes its hands, the small stuff compounds into something no single fix can buy. Britain won eight cycling golds in Beijing and eight more in London.&lt;/p&gt;

&lt;p&gt;I wanted that as a number on my own habits, and the naive version does not survive contact with arithmetic.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why literal compounding fails
&lt;/h2&gt;

&lt;p&gt;One percent a day for a year is &lt;code&gt;1.01^365&lt;/code&gt;, about 37x. If your habit score is bounded at 100 and you are sitting at 70, the whole runway from 70 to 100 is only 1.43x. You run out of room in roughly five weeks, and every day after that is a failure by construction.&lt;/p&gt;

&lt;p&gt;So the target is not compounded from a fixed origin. It is:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="n"&gt;target&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;min&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;100&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nc"&gt;EWMA_14&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;score&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mf"&gt;1.01&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A 14 day exponentially weighted baseline, plus one percent, capped. The half life is 14 days on purpose, so one heroic Sunday does not set an unreachable bar for Monday. On a bad week the baseline sags and the target comes back within reach, which is the behaviour you want from something meant to restart a habit rather than grade it.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the score is made of
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="n"&gt;score&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;100&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mf"&gt;0.70&lt;/span&gt;&lt;span class="o"&gt;*&lt;/span&gt;&lt;span class="n"&gt;cadence&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="mf"&gt;0.15&lt;/span&gt;&lt;span class="o"&gt;*&lt;/span&gt;&lt;span class="n"&gt;recovery&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="mf"&gt;0.15&lt;/span&gt;&lt;span class="o"&gt;*&lt;/span&gt;&lt;span class="n"&gt;breadth&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Cadence is earned over owed, per habit, weighted. Recovery asks how often one miss turns into two, over a trailing 28 days. Breadth is how many habits are still alive in the last 14 days, measured against your own peak rather than a constant I invented.&lt;/p&gt;

&lt;p&gt;Nothing is stored. The whole series is a pure function of which days you checked off, replayed on the client each time the tab opens, which means changing a habit's weight retroactively moves the entire history instead of leaving a stale number on a row.&lt;/p&gt;

&lt;h2&gt;
  
  
  Habits are not one shape
&lt;/h2&gt;

&lt;p&gt;A habit is every day, or N days a week where the rest days are genuinely free, or bonus only, where logging adds credit and skipping is never a miss. Each carries a weight between half and triple, applied symmetrically: doubling what a habit earns also doubles what missing it costs. An asymmetric bonus would let one habit paper over another's miss, which defeats the point of a single number.&lt;/p&gt;

&lt;p&gt;A month of clean cadence banks a spare, spent automatically on the first missed period. A streak that resets on one miss mostly punishes the week your mother visits and cooks, which is not a lapse in the habit.&lt;/p&gt;

&lt;p&gt;Logging happens by tapping an NFC chip stuck where the habit actually happens. Full write-up and a clip of it running: &lt;a href="https://voidcraft.world/blog/one-percent-better.html" rel="noopener noreferrer"&gt;https://voidcraft.world/blog/one-percent-better.html&lt;/a&gt;&lt;/p&gt;

</description>
    </item>
  </channel>
</rss>
