<?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: RobertJJJ</title>
    <description>The latest articles on DEV Community by RobertJJJ (@jjjjjjjjnnjnn).</description>
    <link>https://dev.to/jjjjjjjjnnjnn</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%2F4223026%2Fb46a5501-34f1-42fe-909e-5663f0c820c5.png</url>
      <title>DEV Community: RobertJJJ</title>
      <link>https://dev.to/jjjjjjjjnnjnn</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/jjjjjjjjnnjnn"/>
    <language>en</language>
    <item>
      <title>What a repair layer can and cannot do on DSH's llm/stream seam</title>
      <dc:creator>RobertJJJ</dc:creator>
      <pubDate>Thu, 19 Nov 2026 08:23:51 +0000</pubDate>
      <link>https://dev.to/jjjjjjjjnnjnn/what-a-repair-layer-can-and-cannot-do-on-dshs-llmstream-seam-2ejm</link>
      <guid>https://dev.to/jjjjjjjjnnjnn/what-a-repair-layer-can-and-cannot-do-on-dshs-llmstream-seam-2ejm</guid>
      <description>&lt;p&gt;A fused tool call once cost a whole turn: the provider streamed two JSON objects into one call index, the adapter threw MALFORMED_RESPONSE, and the turn ended.&lt;/p&gt;

&lt;p&gt;dsh-remedy is a patch set I built for that and the neighbouring failures on DeepSeek Harness's seams. It runs as a resident plugin on llm/stream. It splits a fused call only when every fragment can be proven to parse on its own; otherwise it rethrows the provider's failure. A chunk the wrapper cannot read is passed through rather than turned into a new exception.&lt;/p&gt;

&lt;p&gt;There is also an offline half. scan reads this host's session logs and reports fused calls, truncations and turns that ended without an answer. explain names which patch covers a failure class, without pretending an unread profile is "off".&lt;/p&gt;

&lt;p&gt;Measured on this tree: 54 registered patches, 451 tests, six gates, and a cold install from the registry printing OK 8 fixture(s). npm latest is 0.4.20 on both packages. Those are build and host readings, not an audience estimate.&lt;/p&gt;

&lt;p&gt;Three boundaries: it does not patch app.asar, it is not a plugin market, and macOS is only a CI job here. External bug reports stay leads until the shape is reproduced locally.&lt;/p&gt;

&lt;p&gt;Repo: &lt;a href="https://github.com/jjjjjjjjnnjnn/dsh-remedy" rel="noopener noreferrer"&gt;https://github.com/jjjjjjjjnnjnn/dsh-remedy&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;What do you think about "rethrow rather than fake a fix" as the default for a repair layer?&lt;/p&gt;

</description>
      <category>deepseek</category>
      <category>devtools</category>
      <category>opensource</category>
      <category>debugging</category>
    </item>
    <item>
      <title>19 visitors, 1 star, 1,416 clones: what a launch actually looked like</title>
      <dc:creator>RobertJJJ</dc:creator>
      <pubDate>Tue, 17 Nov 2026 21:38:33 +0000</pubDate>
      <link>https://dev.to/jjjjjjjjnnjnn/19-visitors-1-star-1416-clones-what-a-launch-actually-looked-like-43pi</link>
      <guid>https://dev.to/jjjjjjjjnnjnn/19-visitors-1-star-1416-clones-what-a-launch-actually-looked-like-43pi</guid>
      <description>&lt;p&gt;We shipped a developer tool for a plugin ecosystem, published four packages to npm, and wrote a launch post. Two weeks later the repository had &lt;strong&gt;47 views from 19 unique visitors&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;That is the number I want to start with, because everything else in this post is only interesting next to it.&lt;/p&gt;

&lt;h3&gt;
  
  
  The numbers, in the order they mislead you
&lt;/h3&gt;

&lt;p&gt;&lt;strong&gt;Clones: 1,416 in 14 days.&lt;/strong&gt; Sounds like traction. It is 30× the view count, which is the tell: browsers do not clone a repository 30 times per visit. CI runners, mirrors and scanners do. Clone count measures automation, not attention.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;npm downloads: 2,996 in 31 days.&lt;/strong&gt; Also sounds like traction. Look at the shape instead of the total:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;10-29   634   &amp;lt;- release day
10-30   192
10-31    17
11-01   192
11-02   357
11-08    15
11-09   183   &amp;lt;- next release
11-10    30
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Release days spike, the days between sit at 30–57. That is not 30–57 people a day; that is mirrors and caches refreshing. Our own measurement rule says a single-day burst does not count as demand, and this is what that rule looks like from the inside.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Stars: 1. Forks: 0.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The number that decided things for us:&lt;/strong&gt; we had pre-registered a test before launching — reactions ≥10 on the launch thread, at least one third-party reference, and ≥20 weekly downloads, judged over a two-week window. Final reading: &lt;strong&gt;0 reactions, 0 third-party references, 1,212 weekly downloads&lt;/strong&gt;. One of three, and the two zeros were never going to move, because you cannot manufacture a third party citing you.&lt;/p&gt;

&lt;p&gt;So the honest summary is: the delivery worked, the distribution did not.&lt;/p&gt;

&lt;h3&gt;
  
  
  What "the delivery worked" means
&lt;/h3&gt;

&lt;p&gt;This part is checkable, so I will be specific. Four packages published; registry bytes compared against the local tarballs; a clean-room install with an empty npm config exiting 0; the CLI running; the bundle importing. 18 checks, all green in the release evidence file.&lt;/p&gt;

&lt;p&gt;None of that got us a reader. It only ensured that a reader who arrived would not bounce off a broken install.&lt;/p&gt;

&lt;h3&gt;
  
  
  What we did, and what it bought
&lt;/h3&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Action&lt;/th&gt;
&lt;th&gt;Cost&lt;/th&gt;
&lt;th&gt;What changed&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Rewrote the README around one problem&lt;/td&gt;
&lt;td&gt;a day&lt;/td&gt;
&lt;td&gt;nothing measurable in 14 days&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Added a GitHub Release&lt;/td&gt;
&lt;td&gt;an hour&lt;/td&gt;
&lt;td&gt;nothing measurable&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Set the repository homepage&lt;/td&gt;
&lt;td&gt;a minute&lt;/td&gt;
&lt;td&gt;nothing measurable&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Icon on the marketplace tile&lt;/td&gt;
&lt;td&gt;an hour&lt;/td&gt;
&lt;td&gt;card no longer falls back to an avatar — no data yet&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Launch post in the ecosystem's showcase category&lt;/td&gt;
&lt;td&gt;a day&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;0 reactions&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The marketplace listing was already there and stayed there. The listing gets downloads (2,953 as of today), but downloads without stars is the same shape as clones without views.&lt;/p&gt;

&lt;h3&gt;
  
  
  The diagnosis that changed our plan
&lt;/h3&gt;

&lt;p&gt;We had been assuming the bottleneck was the pitch: maybe the front page was unclear, maybe the description was too long. So we measured who was actually visiting, by path.&lt;/p&gt;

&lt;p&gt;The top paths were the repository home, &lt;code&gt;/issues&lt;/code&gt;, and a handful of &lt;code&gt;data/results/*.json&lt;/code&gt; blobs. All of it traced back to &lt;strong&gt;one person&lt;/strong&gt;: a user who had filed the single external issue, and who was reading the evidence files we had pasted into it.&lt;/p&gt;

&lt;p&gt;Nineteen unique visitors, and nearly all of the human traffic was one person checking our work.&lt;/p&gt;

&lt;p&gt;That moved the bottleneck from "what do they read when they arrive" to "who arrives at all". Those need different work, and we had been doing the first one.&lt;/p&gt;

&lt;h3&gt;
  
  
  What we are doing instead
&lt;/h3&gt;

&lt;p&gt;The ecosystem has one place where people are already in the failure state we solve for: the upstream discussions, where plugin users report what broke. In a one-week window, &lt;strong&gt;345 of 845 recent threads matched a failure cluster, and 242 of those had no reply at all&lt;/strong&gt; — 70%.&lt;/p&gt;

&lt;p&gt;So the plan changed from "publish another announcement" to "answer the threads that have no answer". Not a link drop: a reply that opens with a read-only command the reporter can run, explains what the output means, and states what it does not prove.&lt;/p&gt;

&lt;p&gt;Two things about that, honestly:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;It does not scale. Each answer takes real reading.&lt;/li&gt;
&lt;li&gt;It may produce nothing measurable either. A reply in someone else's thread is contact, not a signal, and we deliberately do not count it as one.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;But it is the only channel where the audience is already assembled and already in pain, and it is the only one our own data supports.&lt;/p&gt;

&lt;h3&gt;
  
  
  Three things I would tell someone shipping a niche developer tool
&lt;/h3&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Pre-register what "worked" means, then publish the result even when it fails.&lt;/strong&gt; We wrote the three thresholds before the launch post. That is the only reason this article can say "0 reactions" instead of quietly redefining success.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Separate installation health from attention.&lt;/strong&gt; They fail for completely different reasons, and a green install suite feels like progress while nothing arrives.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Check whether your only visitors are the one person who filed an issue.&lt;/strong&gt; Path-level data is small and unglamorous, and it told us more than the totals did.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;If you are building something for a small ecosystem, the ceiling may be lower than you think — not because the tool is bad, but because the number of people who have the problem is small. Measuring that early is cheaper than rewriting your landing page twice.&lt;/p&gt;

&lt;p&gt;If you want to check any of it, the repository and the evidence files this post cites are public: &lt;a href="https://github.com/jjjjjjjjnnjnn/dsh-triage" rel="noopener noreferrer"&gt;https://github.com/jjjjjjjjnnjnn/dsh-triage&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Boundary: these numbers describe published artifacts and public traffic, not your machine — and none of them is a user count.&lt;/p&gt;

</description>
      <category>opensource</category>
      <category>showdev</category>
      <category>discuss</category>
      <category>productivity</category>
    </item>
  </channel>
</rss>
