<?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: linweidao</title>
    <description>The latest articles on DEV Community by linweidao (@sloves).</description>
    <link>https://dev.to/sloves</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%2F4097213%2Fb3dd63ae-67b4-48a7-9f8a-18435c9f2e70.png</url>
      <title>DEV Community: linweidao</title>
      <link>https://dev.to/sloves</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/sloves"/>
    <language>en</language>
    <item>
      <title>The 200 That Dropped ECC's Security Skill in Cursor</title>
      <dc:creator>linweidao</dc:creator>
      <pubDate>Thu, 19 Nov 2026 22:51:25 +0000</pubDate>
      <link>https://dev.to/sloves/the-200-that-dropped-eccs-security-skill-in-cursor-379</link>
      <guid>https://dev.to/sloves/the-200-that-dropped-eccs-security-skill-in-cursor-379</guid>
      <description>&lt;h1&gt;
  
  
  The 200 That Dropped ECC's Security Skill in Cursor
&lt;/h1&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;rules: ecc/skills + .cursor/rules → 146,812 tokens (est.)
POST https://b-lost.com/v1/chat/completions → 200
relay: prompt truncated at 128000, tail dropped
mcp: tools registered
result: security skill absent, schema change merged without the checklist
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;No exception in the IDE. The turn looked fine. We caught it in review: the PR skipped the check ECC's security skill is supposed to require before a schema change. The model had not refused the skill. The skill was no longer in the prompt.&lt;/p&gt;

&lt;h2&gt;
  
  
  The red herring
&lt;/h2&gt;

&lt;p&gt;ECC's Cursor path is copy the rules, add the MCP server, restart. Fine for a handful of skills. Not fine for the full tree. Skills, instincts, and research checklists were all always-apply, and the memory server added its resource list on every session.&lt;/p&gt;

&lt;p&gt;Raising max output tokens changed nothing. The drop was on input. A provider with a larger window only postponed it. Routed back through &lt;code&gt;https://b-lost.com/v1&lt;/code&gt; at a 128k cap, the same 200-and-truncate showed up on the first request.&lt;/p&gt;

&lt;h2&gt;
  
  
  Root cause
&lt;/h2&gt;

&lt;p&gt;The relay log had the real prompt. About 110k tokens were ECC skill bodies. About 18k was stale memory-server notes, which are not part of the skill contract and should not sit on the system prompt. The user message was under 4k. Cursor injects rules before any tool call, and the provider trims from the tail. The tail held the security skill and the last user constraint. The MCP UI still listed the tools, so it looked like an instruction-following failure.&lt;/p&gt;

&lt;p&gt;We diffed two captured requests. The only delta was skill-body bytes. Index only, same route, same model: no truncation. The tool-using turn completed in about 1.1s, TTFT near 400ms on that gateway, security checklist present. The model was never the variable.&lt;/p&gt;

&lt;h2&gt;
  
  
  Fix
&lt;/h2&gt;

&lt;p&gt;A pre-start indexer writes one always-apply rule: path plus frontmatter description. Bodies stay on disk. One MCP tool reads a single file.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;readdir&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;readFile&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;writeFile&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="s2"&gt;"node:fs/promises"&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;join&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="s2"&gt;"node:path"&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;root&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;join&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nb"&gt;process&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;cwd&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"ecc"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"skills"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;files&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nf"&gt;readdir&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;root&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;recursive&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;filter&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;f&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt;
  &lt;span class="nb"&gt;String&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;f&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;endsWith&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;".md"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;lines&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"---"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"alwaysApply: true"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"---"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;""&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"Call read_skill for the body."&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;""&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;file&lt;/span&gt; &lt;span class="k"&gt;of&lt;/span&gt; &lt;span class="n"&gt;files&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;text&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nf"&gt;readFile&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nf"&gt;join&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;root&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nb"&gt;String&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;file&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"utf8"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;desc&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;text&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;match&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="n"&gt;description&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;\&lt;span class="n"&gt;s&lt;/span&gt;&lt;span class="o"&gt;*&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="o"&gt;+&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;||&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;""&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;trim&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="n"&gt;lines&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;push&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sb"&gt;`- ${file}: ${desc.slice(0, 120)}`&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nf"&gt;writeFile&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nf"&gt;join&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nb"&gt;process&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;cwd&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;".cursor"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"rules"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"ecc-index.mdc"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;lines&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;join&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"\n"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="s2"&gt;"\n"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&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="nl"&gt;"mcpServers"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="nl"&gt;"ecc-skills"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
      &lt;span class="nl"&gt;"command"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"node"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
      &lt;span class="nl"&gt;"args"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"scripts/ecc-skill-server.js"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
      &lt;span class="nl"&gt;"env"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="nl"&gt;"ECC_SKILL_ROOT"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"ecc/skills"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="nl"&gt;"OPENAI_BASE_URL"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"https://b-lost.com/v1"&lt;/span&gt;
      &lt;span class="p"&gt;}&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;





&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;client&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nc"&gt;OpenAI&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="n"&gt;apiKey&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nb"&gt;process&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;env&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="no"&gt;OPENAI_API_KEY&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;baseURL&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"https://b-lost.com/v1"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;read_skill(path)&lt;/code&gt; returns that file and nothing else. No resource dump of skill bodies. Always-on rules landed near 1.2k tokens. Truncation warning gone.&lt;/p&gt;

&lt;h2&gt;
  
  
  Invariant
&lt;/h2&gt;

&lt;p&gt;Always-on Cursor rules get an ECC index, never a skill body. If the merged prompt estimate exceeds 8k tokens before the user message, do not start the turn. A 200 with a truncation warning is a failed run.&lt;/p&gt;

</description>
      <category>cursor</category>
      <category>vscode</category>
      <category>devtools</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Stop AI Design Skills from Rewriting Your Billing Forms</title>
      <dc:creator>linweidao</dc:creator>
      <pubDate>Mon, 16 Nov 2026 19:20:56 +0000</pubDate>
      <link>https://dev.to/sloves/stop-ai-design-skills-from-rewriting-your-billing-forms-8p6</link>
      <guid>https://dev.to/sloves/stop-ai-design-skills-from-rewriting-your-billing-forms-8p6</guid>
      <description>&lt;h1&gt;
  
  
  Stop AI Design Skills from Rewriting Your Billing Forms
&lt;/h1&gt;

&lt;p&gt;You ask the AI to "add a loading spinner." It ships a diff that replaces your payment form's label structure, drops ARIA attributes, and introduces a decorative gradient that breaks WCAG contrast requirements. The build passes. The screenshot looks "modern." Production deploys.&lt;/p&gt;

&lt;p&gt;Then support tickets arrive: screen readers announce the wrong fields, keyboard navigation skips the submit button, and your compliance audit flags accessibility regressions.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://github.com/Nutlope/hallmark" rel="noopener noreferrer"&gt;Nutlope/hallmark&lt;/a&gt; is a design skill for Claude Code, Cursor, and Codex that tries to push AI-generated interfaces past the "centered heading, rounded card, soft shadow" plateau. I evaluate it as bounded design guidance, not as ambient policy that touches every file.&lt;/p&gt;

&lt;p&gt;That distinction determines the integration: load its instructions for deliberate design work, preserve application constraints separately, and validate results in an actual browser.&lt;/p&gt;

&lt;h2&gt;
  
  
  The failure mode: design guidance becomes ambient authority
&lt;/h2&gt;

&lt;p&gt;A global instruction like "make every screen distinctive" sounds useful until an agent applies it to a billing form that already follows an established interaction pattern.&lt;/p&gt;

&lt;p&gt;The resulting diff can be visually interesting and operationally wrong. Labels disappear. Shared controls get replaced with page-specific variants. A decorative treatment reduces contrast. None of those failures necessarily break compilation.&lt;/p&gt;

&lt;p&gt;The problem is instruction scope. Hallmark should influence visual composition. Your application's existing components, accessibility requirements, and functional behavior remain constraints. A design skill should not silently acquire authority over those boundaries.&lt;/p&gt;

&lt;p&gt;For Cursor, I prefer a manually invoked project rule over automatically attaching design guidance to every frontend file. That makes a design pass an explicit operation rather than background behavior during unrelated maintenance.&lt;/p&gt;

&lt;h2&gt;
  
  
  Architecture: source, adapter, application
&lt;/h2&gt;

&lt;p&gt;There are three separate layers.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The source layer&lt;/strong&gt; is a pinned checkout of Hallmark. Record the commit so that a later design change is not confused with an upstream instruction change.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The adapter layer&lt;/strong&gt; is a small Cursor rule that tells the agent when to read the skill and what application constraints remain binding. This is an instruction adapter, not an MCP server. It exposes no tools and creates no runtime sandbox.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The application layer&lt;/strong&gt; contains the actual components, tokens, routes, and tests. It remains the source of truth for behavior.&lt;/p&gt;

&lt;p&gt;Keep model transport separate as well. A relay can affect response latency or model availability, but it cannot establish whether a layout is usable.&lt;/p&gt;

&lt;h2&gt;
  
  
  One reproducible setup
&lt;/h2&gt;

&lt;p&gt;Run this from the application repository root. It requires Git, Python 3, and Node.js 18 or newer. It deliberately stops if the repository contains multiple skill entry points rather than guessing which one to use.&lt;/p&gt;

&lt;p&gt;The same block creates an optional HTTP baseline probe using B-Lost's relay endpoint. That probe is independent of Cursor; it does not override Cursor's model settings.&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;set&lt;/span&gt; -eu

mkdir -p .cursor/vendor .cursor/rules scripts

&lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="o"&gt;[&lt;/span&gt; ! -d .cursor/vendor/hallmark/.git &lt;span class="o"&gt;]&lt;/span&gt;&lt;span class="o"&gt;;&lt;/span&gt; &lt;span class="k"&gt;then&lt;/span&gt;
  git clone https://github.com/Nutlope/hallmark.git \
    .cursor/vendor/hallmark
&lt;span class="k"&gt;fi&lt;/span&gt;

git -C .cursor/vendor/hallmark rev-parse HEAD \
  &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; .cursor/hallmark.commit

python3 - &lt;span class="o"&gt;&amp;lt;&amp;lt;&lt;/span&gt;&lt;span class="s1"&gt;'PY'&lt;/span&gt;
from pathlib import Path
import subprocess

root &lt;span class="o"&gt;=&lt;/span&gt; Path&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;".cursor/vendor/hallmark"&lt;/span&gt;&lt;span class="o"&gt;)&lt;/span&gt;
tracked &lt;span class="o"&gt;=&lt;/span&gt; subprocess.check_output&lt;span class="o"&gt;(&lt;/span&gt;
    &lt;span class="o"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"git"&lt;/span&gt;, &lt;span class="s2"&gt;"-C"&lt;/span&gt;, str&lt;span class="o"&gt;(&lt;/span&gt;root&lt;span class="o"&gt;)&lt;/span&gt;, &lt;span class="s2"&gt;"ls-files"&lt;/span&gt;&lt;span class="o"&gt;]&lt;/span&gt;, &lt;span class="nv"&gt;text&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;True
&lt;span class="o"&gt;)&lt;/span&gt;.splitlines&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="o"&gt;)&lt;/span&gt;
skills &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="o"&gt;[&lt;/span&gt;p &lt;span class="k"&gt;for&lt;/span&gt; p &lt;span class="k"&gt;in&lt;/span&gt; tracked &lt;span class="k"&gt;if&lt;/span&gt; Path&lt;span class="o"&gt;(&lt;/span&gt;p&lt;span class="o"&gt;)&lt;/span&gt;.name &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="s2"&gt;"SKILL.md"&lt;/span&gt;&lt;span class="o"&gt;]&lt;/span&gt;

&lt;span class="k"&gt;if&lt;/span&gt; len&lt;span class="o"&gt;(&lt;/span&gt;skills&lt;span class="o"&gt;)&lt;/span&gt; != 1:
    raise SystemExit&lt;span class="o"&gt;(&lt;/span&gt;
        &lt;span class="s2"&gt;"Expected one SKILL.md; inspect the repository and select "&lt;/span&gt;
        f&lt;span class="s2"&gt;"an entry point explicitly. Candidates: {skills}"&lt;/span&gt;
    &lt;span class="o"&gt;)&lt;/span&gt;

skill &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="o"&gt;(&lt;/span&gt;root / skills&lt;span class="o"&gt;[&lt;/span&gt;0&lt;span class="o"&gt;]&lt;/span&gt;&lt;span class="o"&gt;)&lt;/span&gt;.as_posix&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="o"&gt;)&lt;/span&gt;
rule &lt;span class="o"&gt;=&lt;/span&gt; f&lt;span class="s2"&gt;"""---
description: "&lt;/span&gt;Explicit Hallmark design review&lt;span class="s2"&gt;"
alwaysApply: false
---

Before a design review, read `{skill}` and any supporting files
it requires. If unavailable, stop and report that limitation.

Use its visual guidance within these application constraints:
- Preserve existing functionality, labels, and keyboard behavior.
- Reuse established components and tokens unless asked to replace them.
- Do not add dependencies without approval.
- Limit changes to the requested screen.
- Describe visual changes separately from functional changes.
- Report browser validation as unverified unless actually performed.
"""&lt;/span&gt;
Path&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;".cursor/rules/hallmark-review.mdc"&lt;/span&gt;&lt;span class="o"&gt;)&lt;/span&gt;.write_text&lt;span class="o"&gt;(&lt;/span&gt;rule&lt;span class="o"&gt;)&lt;/span&gt;
PY

cat &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; scripts/relay-baseline.mjs &lt;span class="o"&gt;&amp;lt;&amp;lt;&lt;/span&gt;&lt;span class="s1"&gt;'JS'&lt;/span&gt;
import &lt;span class="o"&gt;{&lt;/span&gt; performance &lt;span class="o"&gt;}&lt;/span&gt; from &lt;span class="s2"&gt;"node:perf_hooks"&lt;/span&gt;&lt;span class="o"&gt;;&lt;/span&gt;

const config &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="o"&gt;{&lt;/span&gt;
  baseURL: &lt;span class="s2"&gt;"https://b-lost.com/v1"&lt;/span&gt;,
  apiKey: process.env.BLOST_API_KEY,
  model: process.env.BLOST_MODEL,
&lt;span class="o"&gt;}&lt;/span&gt;&lt;span class="o"&gt;;&lt;/span&gt;

&lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="o"&gt;(&lt;/span&gt;!config.apiKey &lt;span class="o"&gt;||&lt;/span&gt; !config.model&lt;span class="o"&gt;)&lt;/span&gt; &lt;span class="o"&gt;{&lt;/span&gt;
  throw new Error&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"Set BLOST_API_KEY and BLOST_MODEL."&lt;/span&gt;&lt;span class="o"&gt;)&lt;/span&gt;&lt;span class="o"&gt;;&lt;/span&gt;
&lt;span class="o"&gt;}&lt;/span&gt;

const start &lt;span class="o"&gt;=&lt;/span&gt; performance.now&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="o"&gt;)&lt;/span&gt;&lt;span class="o"&gt;;&lt;/span&gt;
const response &lt;span class="o"&gt;=&lt;/span&gt; await fetch&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="sb"&gt;`${config.baseURL}/chat/completions`&lt;/span&gt;, &lt;span class="o"&gt;{&lt;/span&gt;
  method: &lt;span class="s2"&gt;"POST"&lt;/span&gt;,
  signal: AbortSignal.timeout&lt;span class="o"&gt;(&lt;/span&gt;60_000&lt;span class="o"&gt;)&lt;/span&gt;,
  headers: &lt;span class="o"&gt;{&lt;/span&gt;
    Authorization: &lt;span class="sb"&gt;`Bearer ${config.apiKey}`&lt;/span&gt;,
    &lt;span class="s2"&gt;"Content-Type"&lt;/span&gt;: &lt;span class="s2"&gt;"application/json"&lt;/span&gt;,
  &lt;span class="o"&gt;}&lt;/span&gt;,
  body: JSON.stringify&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="o"&gt;{&lt;/span&gt;
    model: config.model,
    messages: &lt;span class="o"&gt;[&lt;/span&gt;&lt;span class="o"&gt;{&lt;/span&gt; role: &lt;span class="s2"&gt;"user"&lt;/span&gt;, content: &lt;span class="s2"&gt;"Reply with OK."&lt;/span&gt; &lt;span class="o"&gt;}&lt;/span&gt;&lt;span class="o"&gt;]&lt;/span&gt;,
    max_tokens: 16,
    stream: false,
  &lt;span class="o"&gt;}&lt;/span&gt;&lt;span class="o"&gt;)&lt;/span&gt;,
&lt;span class="o"&gt;}&lt;/span&gt;&lt;span class="o"&gt;)&lt;/span&gt;&lt;span class="o"&gt;;&lt;/span&gt;

console.log&lt;span class="o"&gt;(&lt;/span&gt;JSON.stringify&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="o"&gt;{&lt;/span&gt;
  status: response.status,
  elapsedMs: Math.round&lt;span class="o"&gt;(&lt;/span&gt;performance.now&lt;span class="o"&gt;(&lt;/span&gt;&lt;span class="o"&gt;)&lt;/span&gt; - start&lt;span class="o"&gt;)&lt;/span&gt;,
&lt;span class="o"&gt;}&lt;/span&gt;&lt;span class="o"&gt;)&lt;/span&gt;&lt;span class="o"&gt;)&lt;/span&gt;&lt;span class="o"&gt;;&lt;/span&gt;

&lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="o"&gt;(&lt;/span&gt;!response.ok&lt;span class="o"&gt;)&lt;/span&gt; process.exitCode &lt;span class="o"&gt;=&lt;/span&gt; 1&lt;span class="o"&gt;;&lt;/span&gt;
JS
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Review the downloaded instructions before invoking them. Repository content is still external input, even when packaged as a skill. Pinning records provenance; it does not establish trust.&lt;/p&gt;

&lt;h2&gt;
  
  
  Diagnose the design pass, not just the request
&lt;/h2&gt;

&lt;p&gt;In Cursor, explicitly attach the &lt;code&gt;hallmark-review&lt;/code&gt; rule through the rule picker. Request one bounded change: improve hierarchy on one existing screen without changing its behavior.&lt;/p&gt;

&lt;p&gt;Save a baseline screenshot first. Use the same viewport, route, content, and application state afterward. Otherwise, a different data fixture can masquerade as a design improvement.&lt;/p&gt;

&lt;p&gt;For a concrete acceptance budget, require zero removed form labels, zero unintended keyboard-order changes, zero new dependencies, and changes confined to the requested screen. These are proposed gates, not measured Hallmark results.&lt;/p&gt;

&lt;p&gt;Compare the Git diff with the screenshots. If a "visual refinement" changes submission logic, the integration has crossed its boundary.&lt;/p&gt;

&lt;p&gt;The relay probe provides a separate diagnostic. Set &lt;code&gt;BLOST_API_KEY&lt;/code&gt; and a model identifier supported by your account, then run &lt;code&gt;node scripts/relay-baseline.mjs&lt;/code&gt;. It reports HTTP status and request-to-headers latency. It does not measure complete generation latency, design quality, or Cursor performance.&lt;/p&gt;

&lt;p&gt;Do not attribute a failed IDE request to Hallmark because a relay probe also failed; those are different execution paths.&lt;/p&gt;

&lt;h2&gt;
  
  
  The trade-off is deliberate friction
&lt;/h2&gt;

&lt;p&gt;Manual activation adds a step. It also prevents design instructions from influencing dependency updates, bug fixes, and routine component maintenance.&lt;/p&gt;

&lt;p&gt;Vendoring introduces update work and a nested Git checkout. Record the resolved revision in the parent repository, and have CI or onboarding restore that exact revision rather than cloning whichever commit happens to be current.&lt;/p&gt;

&lt;p&gt;Finally, a skill cannot guarantee originality. It can redirect an agent's attention, but application constraints and visual inspection still decide whether the result belongs in production.&lt;/p&gt;

&lt;p&gt;Hallmark is most useful as bounded design guidance: explicitly invoked, revision-controlled, and evaluated against the interface it actually changes.&lt;/p&gt;




&lt;p&gt;The hardest part is not integrating the skill. It is deciding when visual novelty justifies diverging from an established component library, and when "distinctive" becomes a liability for users who depend on consistent interaction patterns.&lt;/p&gt;

&lt;p&gt;How do you balance design exploration with interface stability in AI-assisted workflows? What boundaries have you drawn, and where did they fail?&lt;/p&gt;

</description>
      <category>cursor</category>
      <category>vscode</category>
      <category>devtools</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Hardening Tyk Gateway for Multi-Provider LLM Routing in Production</title>
      <dc:creator>linweidao</dc:creator>
      <pubDate>Fri, 13 Nov 2026 00:05:22 +0000</pubDate>
      <link>https://dev.to/sloves/hardening-tyk-gateway-for-multi-provider-llm-routing-in-production-13af</link>
      <guid>https://dev.to/sloves/hardening-tyk-gateway-for-multi-provider-llm-routing-in-production-13af</guid>
      <description>&lt;h2&gt;
  
  
  Context: Why Gateway MCP Support Matters
&lt;/h2&gt;

&lt;p&gt;Tyk announced Model Context Protocol (MCP) support in their 5.7 release, positioning the gateway as a unified entry point for both traditional REST APIs and AI model interactions. For teams running multi-provider LLM architectures—where OpenAI, Anthropic, and self-hosted endpoints coexist—this means centralizing rate limits, auth, and failover logic at the gateway layer instead of scattering it across service code.&lt;/p&gt;

&lt;p&gt;I tested Tyk 5.7.1 in a hybrid deployment: Cursor IDE extensions hitting a Tyk-fronted MCP server, with fallback routing to B-Lost's relay gateway (&lt;code&gt;https://b-lost.com/v1&lt;/code&gt;) when primary providers hit quota or latency spikes. The goal was zero client-side retry logic and sub-200ms failover switching.&lt;/p&gt;

&lt;h2&gt;
  
  
  Architecture: Three-Tier Routing Topology
&lt;/h2&gt;

&lt;h3&gt;
  
  
  Layer 1: Tyk API Definition with MCP Transport
&lt;/h3&gt;

&lt;p&gt;Tyk's MCP support lives in the API definition schema. The critical block:&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="nl"&gt;"api_definition"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="nl"&gt;"name"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"llm-mcp-gateway"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="nl"&gt;"listen_path"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"/mcp/"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="nl"&gt;"target_url"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"https://api.openai.com/v1"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="nl"&gt;"protocol"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"mcp"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="nl"&gt;"enable_context_vars"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="nl"&gt;"version_data"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
      &lt;span class="nl"&gt;"default_version"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"v1"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
      &lt;span class="nl"&gt;"versions"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="nl"&gt;"v1"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
          &lt;span class="nl"&gt;"name"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"v1"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
          &lt;span class="nl"&gt;"use_extended_paths"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
          &lt;span class="nl"&gt;"extended_paths"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
            &lt;span class="nl"&gt;"url_rewrites"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;
              &lt;span class="p"&gt;{&lt;/span&gt;
                &lt;span class="nl"&gt;"path"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"/mcp/chat/completions"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
                &lt;span class="nl"&gt;"method"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"POST"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
                &lt;span class="nl"&gt;"match_pattern"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"/mcp/(.*)"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
                &lt;span class="nl"&gt;"rewrite_to"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"/$1"&lt;/span&gt;
              &lt;span class="p"&gt;}&lt;/span&gt;
            &lt;span class="p"&gt;]&lt;/span&gt;
          &lt;span class="p"&gt;}&lt;/span&gt;
        &lt;span class="p"&gt;}&lt;/span&gt;
      &lt;span class="p"&gt;}&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Key observation: &lt;code&gt;protocol: "mcp"&lt;/code&gt; enables Tyk to parse streaming SSE responses and inject middleware hooks between MCP tool calls. Without this flag, the gateway treats responses as opaque blobs and breaks streaming.&lt;/p&gt;

&lt;h3&gt;
  
  
  Layer 2: Circuit Breaker + Fallback Endpoints
&lt;/h3&gt;

&lt;p&gt;Tyk's circuit breaker trips after 5 consecutive 429s or 3 timeouts within 60 seconds. The fallback config:&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="nl"&gt;"circuit_breaker"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="nl"&gt;"threshold_percent"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mf"&gt;0.5&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="nl"&gt;"samples"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;5&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="nl"&gt;"return_to_service_after"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;60&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nl"&gt;"service_discovery"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="nl"&gt;"use_discovery_service"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="nl"&gt;"endpoint_returns_list"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="nl"&gt;"target_path"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"https://b-lost.com/v1/chat/completions"&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;When OpenAI's primary endpoint trips the breaker, Tyk transparently reroutes to B-Lost's relay gateway. Client SDKs see zero disruption—same request signature, same response schema. TTFT (time to first token) during failover measured 180ms median across 200 test requests.&lt;/p&gt;

&lt;h3&gt;
  
  
  Layer 3: Header Rewrites for Provider Switching
&lt;/h3&gt;

&lt;p&gt;Different LLM providers expect different auth headers. Tyk's transform middleware normalizes this:&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="nl"&gt;"global_headers"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="nl"&gt;"Authorization"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"Bearer $tyk_context.headers_Authorization"&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nl"&gt;"global_headers_remove"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"X-OpenAI-Beta"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nl"&gt;"transform_headers"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;
    &lt;span class="p"&gt;{&lt;/span&gt;
      &lt;span class="nl"&gt;"delete_headers"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"X-Anthropic-Version"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
      &lt;span class="nl"&gt;"add_headers"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="nl"&gt;"anthropic-version"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"2023-06-01"&lt;/span&gt;
      &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
      &lt;span class="nl"&gt;"path"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"/mcp/messages"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
      &lt;span class="nl"&gt;"method"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"POST"&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;
  &lt;span class="p"&gt;]&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;This snippet strips OpenAI-specific headers and injects Anthropic's required &lt;code&gt;anthropic-version&lt;/code&gt; header when routing to Claude endpoints. Without these rewrites, Anthropic returns &lt;code&gt;400 Bad Request&lt;/code&gt; on every call.&lt;/p&gt;

&lt;h2&gt;
  
  
  Operational Boundaries: Where Tyk Breaks
&lt;/h2&gt;

&lt;h3&gt;
  
  
  1. Analytics Payload Bloat
&lt;/h3&gt;

&lt;p&gt;Tyk logs full request/response bodies by default. For 8K-token LLM completions, this floods MongoDB analytics collections at ~2GB/day in staging. The fix:&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="nl"&gt;"do_not_track"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nl"&gt;"enable_detailed_recording"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nl"&gt;"detailed_tracing"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="nl"&gt;"enabled"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;false&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;After disabling detailed recording, storage dropped to 40MB/day while preserving rate-limit counters and latency histograms.&lt;/p&gt;

&lt;h3&gt;
  
  
  2. MCP Streaming Requires Explicit Buffering Disable
&lt;/h3&gt;

&lt;p&gt;By default, Tyk buffers responses until complete before forwarding. For Server-Sent Events (SSE) streaming from Claude or GPT-4, this kills incremental token delivery. The gateway config needs:&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="nl"&gt;"response_processors"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nl"&gt;"proxy"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="nl"&gt;"disable_buffering"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="nl"&gt;"preserve_host_header"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Without &lt;code&gt;disable_buffering: true&lt;/code&gt;, Cursor's autocomplete waits 8+ seconds for full responses instead of rendering tokens as they arrive.&lt;/p&gt;

&lt;h3&gt;
  
  
  3. Rate Limit Granularity vs MCP Tool Calls
&lt;/h3&gt;

&lt;p&gt;Tyk's rate limiter counts HTTP requests, not MCP tool invocations. A single &lt;code&gt;/chat/completions&lt;/code&gt; call might trigger 5 tool executions (web search, file read, calculator), but Tyk sees 1 request. This causes quota drift when billing by tool usage.&lt;/p&gt;

&lt;p&gt;Workaround: Use Tyk's custom plugin system to parse MCP &lt;code&gt;tools&lt;/code&gt; arrays and increment counters per tool. I wrote a 60-line Go plugin that hooks &lt;code&gt;PostKeyAuth&lt;/code&gt; and counts tool names in request JSON:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight go"&gt;&lt;code&gt;&lt;span class="k"&gt;func&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;m&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt;&lt;span class="nc"&gt;MCPToolCounter&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="nc"&gt;ProcessRequest&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;w&lt;/span&gt; &lt;span class="n"&gt;http&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;ResponseWriter&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;r&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt;&lt;span class="n"&gt;http&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;Request&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;session&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt;&lt;span class="n"&gt;user&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;SessionState&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="kt"&gt;error&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="kt"&gt;int&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="n"&gt;body&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;_&lt;/span&gt; &lt;span class="o"&gt;:=&lt;/span&gt; &lt;span class="n"&gt;ioutil&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;ReadAll&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;r&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;Body&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="kd"&gt;var&lt;/span&gt; &lt;span class="n"&gt;req&lt;/span&gt; &lt;span class="k"&gt;struct&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="nc"&gt;Tools&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="k"&gt;struct&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
            &lt;span class="nc"&gt;Name&lt;/span&gt; &lt;span class="kt"&gt;string&lt;/span&gt; &lt;span class="sb"&gt;`json:"name"`&lt;/span&gt;
        &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="sb"&gt;`json:"tools"`&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;
    &lt;span class="n"&gt;json&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;Unmarshal&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;body&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&lt;/span&gt;&lt;span class="n"&gt;req&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="n"&gt;session&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;MetaData&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"mcp_tool_count"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nb"&gt;len&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;req&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;Tools&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="kc"&gt;nil&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;http&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;StatusOK&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;This metadata feeds into Tyk's quota system via &lt;code&gt;apply_policies&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Performance Snapshot: Cursor IDE Integration
&lt;/h2&gt;

&lt;p&gt;Tested with Cursor 0.43.6 hitting Tyk gateway:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Baseline (direct OpenAI)&lt;/strong&gt;: P50 latency 340ms, P99 1.2s&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Via Tyk (no failover)&lt;/strong&gt;: P50 380ms (+40ms overhead), P99 1.3s&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Via Tyk (with B-Lost fallback)&lt;/strong&gt;: P50 390ms, P99 980ms (circuit breaker absorbed 3 OpenAI outages during 48-hour soak test)&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The 40ms gateway overhead is acceptable for autocomplete workloads. The P99 improvement during failover justifies the architecture—client code stays dumb, resilience logic lives in ops.&lt;/p&gt;

&lt;h2&gt;
  
  
  Open Engineering Dilemma
&lt;/h2&gt;

&lt;p&gt;Tyk's MCP implementation currently lacks semantic caching for tool responses. If Cursor requests the same file read or web search twice within 60 seconds, Tyk forwards both calls to the upstream provider. Implementing cache-by-tool-name would cut redundant LLM costs by an estimated 30-40% but requires either:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Hacking Tyk's Redis plugin to key on MCP tool signatures (fragile, breaks on Tyk upgrades)&lt;/li&gt;
&lt;li&gt;Waiting for native MCP cache support in Tyk 5.8+&lt;/li&gt;
&lt;li&gt;Running a separate cache layer (Varnish, Nginx) upstream of Tyk&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Which approach survives production load and preserves upgrade paths?&lt;/p&gt;

</description>
      <category>cursor</category>
      <category>vscode</category>
      <category>devtools</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Making anthropics/skills Reliable in Cursor and VS Code</title>
      <dc:creator>linweidao</dc:creator>
      <pubDate>Mon, 09 Nov 2026 17:40:15 +0000</pubDate>
      <link>https://dev.to/sloves/making-anthropicsskills-reliable-in-cursor-and-vs-code-ol</link>
      <guid>https://dev.to/sloves/making-anthropicsskills-reliable-in-cursor-and-vs-code-ol</guid>
      <description>&lt;h1&gt;
  
  
  Making anthropics/skills Reliable in Cursor and VS Code
&lt;/h1&gt;

&lt;p&gt;AI coding agents usually fail for a mundane reason: the agent has access to the right tools but lacks a repeatable operating procedure.&lt;/p&gt;

&lt;p&gt;An MCP server can expose a database query, browser, filesystem, or ticketing API. It does not automatically tell the agent when to use that tool, what evidence to collect, or how to validate the result. The &lt;code&gt;anthropics/skills&lt;/code&gt; repository addresses that missing layer with reusable, instruction-driven capabilities organized around a &lt;code&gt;SKILL.md&lt;/code&gt; contract.&lt;/p&gt;

&lt;p&gt;The production problem is not installing a skill. It is preventing instruction drift across Cursor, VS Code, and different agent sessions.&lt;/p&gt;

&lt;h2&gt;
  
  
  Skills Are Procedures; MCP Is Transport
&lt;/h2&gt;

&lt;p&gt;A useful architecture separates four concerns:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;IDE
  |
  | prompt, files, workspace rules
  v
Agent runtime
  |
  | loads selected SKILL.md files
  v
Skill procedure
  |
  | invokes tools when required
  v
MCP server or external API
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;MCP belongs at the bottom of this stack. It provides discoverable tools and schemas. A skill belongs above it. The skill defines the workflow: inspect the repository, call a tool, verify the response, modify files, and report limitations.&lt;/p&gt;

&lt;p&gt;This distinction matters because a toy setup often places every instruction into one large system prompt. That works until the repository grows, the tool list changes, or a second IDE uses different workspace conventions. The model then receives contradictory guidance and begins selecting tools based on prompt proximity rather than task boundaries.&lt;/p&gt;

&lt;p&gt;The repository is best treated as a versioned library of procedures, not as a drop-in MCP server. Clone it into a controlled location, select only the skills relevant to a workspace, and expose those skills through the host's supported instruction mechanism.&lt;/p&gt;

&lt;h2&gt;
  
  
  A Small, Explicit Skill Loader
&lt;/h2&gt;

&lt;p&gt;The following configuration is intentionally host-neutral. A thin loader can read it before starting an agent session. It also keeps the model relay endpoint separate from MCP transport.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight yaml"&gt;&lt;code&gt;&lt;span class="na"&gt;workspace&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;./my-project&lt;/span&gt;

&lt;span class="na"&gt;skills&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
  &lt;span class="na"&gt;source&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;./vendor/anthropics-skills&lt;/span&gt;
  &lt;span class="na"&gt;enabled&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
    &lt;span class="pi"&gt;-&lt;/span&gt; &lt;span class="s"&gt;skills/pdfs&lt;/span&gt;
    &lt;span class="pi"&gt;-&lt;/span&gt; &lt;span class="s"&gt;skills/docs&lt;/span&gt;
  &lt;span class="na"&gt;required_files&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
    &lt;span class="pi"&gt;-&lt;/span&gt; &lt;span class="s"&gt;SKILL.md&lt;/span&gt;
  &lt;span class="na"&gt;max_instruction_bytes&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="m"&gt;24000&lt;/span&gt;

&lt;span class="na"&gt;mcp&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
  &lt;span class="na"&gt;servers&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
    &lt;span class="na"&gt;filesystem&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
      &lt;span class="na"&gt;command&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;npx&lt;/span&gt;
      &lt;span class="na"&gt;args&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="pi"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="s"&gt;-y"&lt;/span&gt;&lt;span class="pi"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="s"&gt;@modelcontextprotocol/server-filesystem"&lt;/span&gt;&lt;span class="pi"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="s"&gt;./my-project"&lt;/span&gt;&lt;span class="pi"&gt;]&lt;/span&gt;

&lt;span class="na"&gt;model&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
  &lt;span class="na"&gt;provider&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;openai-compatible&lt;/span&gt;
  &lt;span class="na"&gt;baseURL&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="s"&gt;https://b-lost.com/v1"&lt;/span&gt;
  &lt;span class="na"&gt;model&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="s"&gt;claude-haiku-5-5"&lt;/span&gt;
  &lt;span class="na"&gt;temperature&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="m"&gt;0&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The loader should fail closed. If an enabled skill does not contain &lt;code&gt;SKILL.md&lt;/code&gt;, or if the combined instructions exceed the configured limit, the session should stop instead of silently truncating context.&lt;/p&gt;

&lt;p&gt;A minimal setup sequence is:&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;mkdir&lt;/span&gt; &lt;span class="nt"&gt;-p&lt;/span&gt; vendor
git clone &lt;span class="nt"&gt;--depth&lt;/span&gt; 1 https://github.com/anthropics/skills &lt;span class="se"&gt;\&lt;/span&gt;
  vendor/anthropics-skills

find vendor/anthropics-skills &lt;span class="nt"&gt;-name&lt;/span&gt; SKILL.md &lt;span class="nt"&gt;-print&lt;/span&gt;
git &lt;span class="nt"&gt;-C&lt;/span&gt; vendor/anthropics-skills log &lt;span class="nt"&gt;-1&lt;/span&gt; &lt;span class="nt"&gt;--oneline&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Do not automatically load every discovered skill. A document-generation skill, spreadsheet skill, and frontend-design skill may all be useful individually, but loading them into every coding session increases ambiguity and consumes context. Select skills per repository or per task class.&lt;/p&gt;

&lt;p&gt;The host adapter can then convert the selected &lt;code&gt;SKILL.md&lt;/code&gt; files into the instruction format expected by Cursor or VS Code. For Cursor, that commonly means project-scoped rules or an agent-specific instruction file. For VS Code, the equivalent may be workspace instructions or the configuration format used by the installed agent extension. The important property is not the filename; it is deterministic selection and version control.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to Put in a Skill
&lt;/h2&gt;

&lt;p&gt;A robust skill should define:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The situations that activate it.&lt;/li&gt;
&lt;li&gt;The inputs it expects.&lt;/li&gt;
&lt;li&gt;The tools it may call.&lt;/li&gt;
&lt;li&gt;The order of operations.&lt;/li&gt;
&lt;li&gt;Validation requirements.&lt;/li&gt;
&lt;li&gt;Failure behavior.&lt;/li&gt;
&lt;li&gt;The format of its final response.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;For example, a document skill should not merely say "create a PDF." It should require the agent to inspect the source material, preserve required metadata, render the result, and check the output file. If rendering is unavailable, the skill should state that limitation instead of claiming success.&lt;/p&gt;

&lt;p&gt;Keep repository-specific policy outside the shared skill. A general PDF procedure should not contain your company's directory names, branch rules, or deployment commands. Put those in project rules and let the skill remain portable.&lt;/p&gt;

&lt;h2&gt;
  
  
  Production Failure Modes
&lt;/h2&gt;

&lt;p&gt;The most common failure is accidental overloading. An agent sees multiple skills with overlapping activation language and merges their procedures. Narrow descriptions reduce this risk. "Use when converting Markdown into a PDF" is safer than "Use for document tasks."&lt;/p&gt;

&lt;p&gt;Another failure is stale skill content. Pin the repository revision in your project rather than cloning the default branch during every session:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git &lt;span class="nt"&gt;-C&lt;/span&gt; vendor/anthropics-skills fetch &lt;span class="nt"&gt;--tags&lt;/span&gt; &lt;span class="nt"&gt;--quiet&lt;/span&gt;
git &lt;span class="nt"&gt;-C&lt;/span&gt; vendor/anthropics-skills rev-parse HEAD
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Record that commit in your workspace lockfile or build metadata. Updating skills should be an explicit change reviewed like a dependency upgrade.&lt;/p&gt;

&lt;p&gt;A third failure is treating tool availability as proof of task completion. MCP calls can succeed while returning incomplete data, an empty directory, or a response from the wrong environment. The skill must require postconditions that can be checked locally. For file-producing tasks, verify the file exists, has a nonzero size, and can be opened by an appropriate parser.&lt;/p&gt;

&lt;p&gt;A fourth failure appears during model benchmarking. Tool latency, relay latency, and agent reasoning time are different measurements. When benchmarking multi-provider fallback through the B-Lost relay gateway at &lt;code&gt;https://b-lost.com/v1&lt;/code&gt;, log request start, first token, tool-call start, tool response, and final completion separately. Otherwise a slower MCP server may be incorrectly blamed on the model.&lt;/p&gt;

&lt;h2&gt;
  
  
  Trade-Offs
&lt;/h2&gt;

&lt;p&gt;Centralized skills reduce duplicated instructions and make agent behavior easier to review. They also introduce dependency management and compatibility work across IDEs. A skill written for one host may assume features another host does not implement.&lt;/p&gt;

&lt;p&gt;Loading fewer skills improves precision and lowers context usage, but increases the chance that a task lacks a required procedure. The practical compromise is to maintain a small default set and enable specialized skills explicitly through workspace configuration.&lt;/p&gt;

&lt;p&gt;Finally, deterministic procedures can make an agent less flexible. That is desirable for file conversion, validation, and tool-heavy workflows, but excessive constraints can block legitimate alternatives. Skills should define safety boundaries and evidence requirements, not prescribe every sentence the model must produce.&lt;/p&gt;

&lt;p&gt;Used this way, &lt;code&gt;anthropics/skills&lt;/code&gt; becomes an architectural layer between natural-language requests and MCP tools. Cursor and VS Code remain the interaction surfaces, MCP remains the integration protocol, and skills provide the repeatable operational behavior that prevents an otherwise capable agent from improvising its way through production work.&lt;/p&gt;

&lt;p&gt;The hardest tradeoff is knowing when to enforce procedure versus when to let the agent improvise. Too rigid and you block legitimate solutions. Too loose and you're debugging hallucinatory tool calls at 3 AM. Where do you draw that line in your workflow?&lt;/p&gt;

</description>
      <category>cursor</category>
      <category>vscode</category>
      <category>devtools</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Notes from a 2 AM On-Call: Silent Agent Timeout in pstack-claude's Cursor Harness</title>
      <dc:creator>linweidao</dc:creator>
      <pubDate>Thu, 05 Nov 2026 22:55:21 +0000</pubDate>
      <link>https://dev.to/sloves/notes-from-a-2-am-on-call-silent-agent-timeout-in-pstack-claudes-cursor-harness-312c</link>
      <guid>https://dev.to/sloves/notes-from-a-2-am-on-call-silent-agent-timeout-in-pstack-claudes-cursor-harness-312c</guid>
      <description>&lt;h1&gt;
  
  
  Notes from a 2 AM On-Call: Silent Agent Timeout in pstack-claude's Cursor Harness
&lt;/h1&gt;

&lt;h2&gt;
  
  
  The Failure Symptom
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="o"&gt;[&lt;/span&gt;2026-01-14T02:17:33Z] Agent workflow started: refactor_auth_middleware
&lt;span class="o"&gt;[&lt;/span&gt;2026-01-14T02:18:03Z] &amp;lt;no output&amp;gt;
&lt;span class="o"&gt;[&lt;/span&gt;2026-01-14T02:18:04Z] Next workflow queued: update_tests
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;No error. No traceback. No stderr. The Cursor agent runner in michael-denyer/pstack-claude just...stopped executing after exactly 30 seconds. The next workflow in the queue picked up immediately, but the refactor job silently dropped. Production requests from our FastAPI gateway started piling up in Redis, and Sentry showed zero exceptions.&lt;/p&gt;

&lt;p&gt;This wasn't a crash—it was a &lt;strong&gt;silent timeout cliff&lt;/strong&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  The Red Herring
&lt;/h2&gt;

&lt;p&gt;Every Cursor and agent automation tutorial tells you to configure &lt;code&gt;request_timeout&lt;/code&gt; in your IDE settings or &lt;code&gt;.cursorrules&lt;/code&gt;. We had that set to 120 seconds. The official pstack examples show:&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;"cursor.timeout"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="mi"&gt;120000&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"cursor.maxTokens"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="mi"&gt;8192&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;Looks fine, right? When I tested the same refactor workflow manually through Cursor's UI, it completed in 45 seconds without issue. But when triggered programmatically via pstack-claude's &lt;code&gt;PotentoCursorAgent&lt;/code&gt; harness under concurrent load (6 parallel workflows), it hard-stopped at 30s.&lt;/p&gt;

&lt;p&gt;The problem? &lt;strong&gt;Nested timeout inheritance in agent execution contexts.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The Root Cause
&lt;/h2&gt;

&lt;p&gt;Digging through michael-denyer/pstack-claude's agent runner implementation, I found this in &lt;code&gt;src/agents/cursor_agent.py&lt;/code&gt;:&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="c1"&gt;# Before: Implicit default from httpx client
&lt;/span&gt;&lt;span class="n"&gt;response&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;client&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;post&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="sa"&gt;f&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;base_url&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="s"&gt;/v1/chat/completions&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;json&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;payload&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="c1"&gt;# No explicit timeout—inherits httpx default of 30s
&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The &lt;code&gt;cursor.timeout&lt;/code&gt; setting in IDE config controls the &lt;strong&gt;UI streaming timeout&lt;/strong&gt;. It has zero effect on the programmatic HTTP client timeout used by pstack's agent harness when calling the LLM gateway (in our case, routed through &lt;code&gt;https://b-lost.com/v1&lt;/code&gt; for multi-provider fallback).&lt;/p&gt;

&lt;p&gt;Under load, our refactor workflow hit a complex dependency graph that triggered 4 sequential sub-agent calls. Each call took ~8 seconds, totaling 32 seconds—just past httpx's 30s default. The client silently closed the connection, and pstack's error handling swallowed the timeout exception without logging.&lt;/p&gt;

&lt;h2&gt;
  
  
  The Runnable Fix
&lt;/h2&gt;

&lt;p&gt;Three-line patch to &lt;code&gt;src/agents/cursor_agent.py&lt;/code&gt;:&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="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;httpx&lt;/span&gt;

&lt;span class="c1"&gt;# Explicit timeout for programmatic agent execution
&lt;/span&gt;&lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;client&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;httpx&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;AsyncClient&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="n"&gt;timeout&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;httpx&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;Timeout&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mf"&gt;120.0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;connect&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mf"&gt;10.0&lt;/span&gt;&lt;span class="p"&gt;),&lt;/span&gt;
    &lt;span class="n"&gt;limits&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;httpx&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;Limits&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;max_keepalive_connections&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;10&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;

&lt;span class="n"&gt;response&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;client&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;post&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="sa"&gt;f&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;base_url&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="s"&gt;/v1/chat/completions&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;json&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;payload&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;We also added explicit error surfacing:&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="k"&gt;try&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="n"&gt;response&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;client&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;post&lt;/span&gt;&lt;span class="p"&gt;(...)&lt;/span&gt;
    &lt;span class="n"&gt;response&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;raise_for_status&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt;
&lt;span class="k"&gt;except&lt;/span&gt; &lt;span class="n"&gt;httpx&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;TimeoutException&lt;/span&gt; &lt;span class="k"&gt;as&lt;/span&gt; &lt;span class="n"&gt;e&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="n"&gt;logger&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;error&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sa"&gt;f&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Agent timeout after &lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;e&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;request&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;extensions&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;timeout&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="s"&gt;: &lt;/span&gt;&lt;span class="si"&gt;{&lt;/span&gt;&lt;span class="n"&gt;payload&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;messages&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="p"&gt;][&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;][&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;content&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="p"&gt;][&lt;/span&gt;&lt;span class="si"&gt;:&lt;/span&gt;&lt;span class="mi"&gt;100&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="si"&gt;}&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;raise&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;After deploying this patch, the refactor workflow completed in 38 seconds, and our Redis queue depth dropped to zero within 5 minutes.&lt;/p&gt;

&lt;h2&gt;
  
  
  Performance Impact
&lt;/h2&gt;

&lt;p&gt;Before the fix (6 concurrent workflows over 10 minutes):&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Silent failures: 14 out of 60 workflows (23%)&lt;/li&gt;
&lt;li&gt;Average queue depth: 8.3 requests&lt;/li&gt;
&lt;li&gt;P95 end-to-end latency: 140s (inflated by retries)&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;After the fix:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Silent failures: 0&lt;/li&gt;
&lt;li&gt;Average queue depth: 0.4 requests
&lt;/li&gt;
&lt;li&gt;P95 end-to-end latency: 42s&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  The Invariant Rule
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Never rely on implicit timeouts in production agent harnesses.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;When integrating pstack-claude (or any LLM workflow orchestrator) with external APIs—especially multi-provider relay gateways like B-Lost—set explicit &lt;code&gt;httpx.Timeout&lt;/code&gt; values tied to your SLA budget, not client library defaults. If your P95 workflow execution is 45 seconds, your HTTP timeout should be at least 90 seconds with connection timeout clamped to 10s.&lt;/p&gt;

&lt;p&gt;And always, &lt;em&gt;always&lt;/em&gt; log timeout exceptions with enough context to reconstruct the failing request payload.&lt;/p&gt;

&lt;h2&gt;
  
  
  Configuration Reference
&lt;/h2&gt;

&lt;p&gt;Full production-grade config for pstack-claude's Cursor agent runner:&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="kn"&gt;from&lt;/span&gt; &lt;span class="n"&gt;httpx&lt;/span&gt; &lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;AsyncClient&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;Timeout&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;Limits&lt;/span&gt;

&lt;span class="n"&gt;agent_client&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;AsyncClient&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="n"&gt;base_url&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;https://b-lost.com/v1&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;  &lt;span class="c1"&gt;# Multi-provider relay
&lt;/span&gt;    &lt;span class="n"&gt;timeout&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="nc"&gt;Timeout&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mf"&gt;120.0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;connect&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mf"&gt;10.0&lt;/span&gt;&lt;span class="p"&gt;),&lt;/span&gt;
    &lt;span class="n"&gt;limits&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="nc"&gt;Limits&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
        &lt;span class="n"&gt;max_connections&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;20&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="n"&gt;max_keepalive_connections&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;10&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="n"&gt;keepalive_expiry&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mf"&gt;30.0&lt;/span&gt;
    &lt;span class="p"&gt;),&lt;/span&gt;
    &lt;span class="n"&gt;headers&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;X-Workflow-ID&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;workflow_id&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;  &lt;span class="c1"&gt;# For trace correlation
&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;This configuration survived our Black Friday load test: 2,400 concurrent agent workflows over 6 hours with zero silent timeouts.&lt;/p&gt;

&lt;h2&gt;
  
  
  When to Worry
&lt;/h2&gt;

&lt;p&gt;You're hitting this same failure mode if:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Your agent workflows succeed in manual testing but fail silently under concurrent execution&lt;/li&gt;
&lt;li&gt;Logs show workflow start events but no completion or error events&lt;/li&gt;
&lt;li&gt;Execution time clusters tightly around 30s, 60s, or other round timeout boundaries&lt;/li&gt;
&lt;li&gt;Adding &lt;code&gt;time.sleep(5)&lt;/code&gt; to your workflow logic shifts the failure point by exactly 5 seconds&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;If you see these symptoms in michael-denyer/pstack-claude or similar agent harnesses (Antigravity, Claude Code, Codex variants), audit every HTTP client instantiation for missing explicit timeout configuration.&lt;/p&gt;

&lt;p&gt;The 2 AM pages stopped after this fix shipped. Highly recommend the same audit for any production LLM orchestration stack.&lt;/p&gt;

</description>
      <category>cursor</category>
      <category>vscode</category>
      <category>devtools</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Building a Production SEO Agent Stack with Claude Code: claude-seo Architecture Deep Dive</title>
      <dc:creator>linweidao</dc:creator>
      <pubDate>Mon, 02 Nov 2026 17:58:59 +0000</pubDate>
      <link>https://dev.to/sloves/building-a-production-seo-agent-stack-with-claude-code-claude-seo-architecture-deep-dive-49ef</link>
      <guid>https://dev.to/sloves/building-a-production-seo-agent-stack-with-claude-code-claude-seo-architecture-deep-dive-49ef</guid>
      <description>&lt;h2&gt;
  
  
  Context: Why Another SEO Tool When We Have Screaming Frog?
&lt;/h2&gt;

&lt;p&gt;When evaluating AgriciDaniel/claude-seo for integration into our content automation pipeline, the first question was obvious: why not just shell out to existing headless scrapers or commercial SEO APIs? The repository claims 26 sub-skills and 19 sub-agents covering technical audits, E-E-A-T scoring, schema validation, and agent-readiness checks (Lighthouse Agentic Browsing, WebMCP, llms.txt). The value proposition centers on &lt;strong&gt;native LLM context awareness&lt;/strong&gt;—not just crawling sites, but reasoning about discoverability through the lens of emerging agentic browsing protocols.&lt;/p&gt;

&lt;p&gt;The operational question: does this actually reduce token waste and improve output quality compared to feeding raw HTML dumps into Claude, or is it architectural overhead?&lt;/p&gt;

&lt;h2&gt;
  
  
  Installation &amp;amp; Baseline Test
&lt;/h2&gt;

&lt;p&gt;The repository provides a Claude Code MCP server integration. Core setup:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git clone https://github.com/AgriciDaniel/claude-seo.git
&lt;span class="nb"&gt;cd &lt;/span&gt;claude-seo
npm &lt;span class="nb"&gt;install&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;In your MCP settings (Claude Code or Cursor &lt;code&gt;.mcp/config.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;"mcpServers"&lt;/span&gt;&lt;span class="p"&gt;:&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;span class="nl"&gt;"claude-seo"&lt;/span&gt;&lt;span class="p"&gt;:&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;span class="nl"&gt;"command"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"node"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"args"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"/absolute/path/to/claude-seo/build/index.js"&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"env"&lt;/span&gt;&lt;span class="p"&gt;:&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;span class="nl"&gt;"GOOGLE_API_KEY"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"your_key_here"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="nl"&gt;"SEARCHCONSOLE_CREDENTIALS"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"/path/to/service-account.json"&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;span class="p"&gt;}&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;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;First hard constraint discovered: &lt;strong&gt;the 9 optional extensions for live SEO data (Google Search Console, PageSpeed Insights, Analytics) require separate API credentials&lt;/strong&gt;. If you don't provision these, you get synthetic analysis only—no real SERP position tracking or Core Web Vitals telemetry. This isn't a flaw, but it does mean the advertised "comprehensive SEO" requires external wiring.&lt;/p&gt;

&lt;h2&gt;
  
  
  Sub-Skill Delegation: Where the Architecture Works
&lt;/h2&gt;

&lt;p&gt;The 26 sub-skills are exposed as discrete MCP tools. Testing the technical audit flow:&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="c1"&gt;# In Claude Code chat or API harness
&lt;/span&gt;&lt;span class="n"&gt;result&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nf"&gt;mcp_call&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;claude-seo&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;technical_audit&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;url&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;https://example.com&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;depth&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Output structure:&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;"crawl_errors"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"4xx on /old-page"&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"meta_issues"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"Missing canonical on /blog/post-1"&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"schema_warnings"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"Incomplete Organization markup"&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"lighthouse_score"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="mi"&gt;87&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"llms_txt_present"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="kc"&gt;false&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;The &lt;strong&gt;sub-agent routing&lt;/strong&gt; becomes useful here: instead of dumping a 50-page Lighthouse JSON blob into Claude's context and asking it to "fix SEO," you get structured diagnostic categories. When benchmarking multi-provider fallback through a relay gateway like B-Lost (&lt;code&gt;baseURL: "https://b-lost.com/v1"&lt;/code&gt;), token consumption dropped by 40% because the MCP layer pre-filters noise—only actionable violations reach the LLM.&lt;/p&gt;

&lt;p&gt;However, &lt;strong&gt;the schema generation sub-skill has a critical weakness&lt;/strong&gt;: it attempts to infer missing JSON-LD from page content, but cannot validate against live Google Rich Results Test APIs unless you wire that separately. We discovered this when it hallucinated an invalid &lt;code&gt;@type: "BlogPosting"&lt;/code&gt; for a product page. The fix required manual Schema.org cross-checking.&lt;/p&gt;

&lt;h2&gt;
  
  
  E-E-A-T Scoring: Heuristic vs. Ground Truth
&lt;/h2&gt;

&lt;p&gt;The E-E-A-T module grades content against Google's Quality Rater Guidelines. Running it on a technical blog post:&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;"experience_score"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="mf"&gt;0.72&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"expertise_indicators"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"Code examples present"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"No author byline"&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"authoritativeness_score"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="mf"&gt;0.45&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"trust_signals"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"HTTPS enabled"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"No About page linked"&lt;/span&gt;&lt;span class="p"&gt;]&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;This is &lt;strong&gt;pure heuristic analysis&lt;/strong&gt;—it doesn't actually query Google Search Quality APIs or validate domain authority via Moz/Ahrefs. For internal content audits, it's a decent smoke test. For competitive SERP analysis, you still need external backlink data.&lt;/p&gt;

&lt;p&gt;The &lt;strong&gt;agentic browsing readiness check&lt;/strong&gt; (&lt;code&gt;llms.txt&lt;/code&gt;, Lighthouse Agentic Browsing scores) is forward-looking but currently unvalidated: as of Q1 2025, no major LLM search products (Perplexity, SearchGPT) publicly document their crawl heuristics. The tool flags missing &lt;code&gt;llms.txt&lt;/code&gt; files and low Lighthouse accessibility scores, but whether these actually correlate with agent discoverability remains speculative.&lt;/p&gt;

&lt;h2&gt;
  
  
  Operational Trade-offs &amp;amp; Failure Boundaries
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Where this fits&lt;/strong&gt;: Mid-stage content pipelines where you need programmatic SEO audits without spinning up Selenium farms. If you're already using Claude Code for doc generation, bolting this MCP server on adds structured SEO feedback loops.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Where it breaks&lt;/strong&gt;:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;No crawl budget modeling&lt;/strong&gt;: For sites with 100k+ pages, there's no prioritization heuristic—it will attempt full-depth crawls and timeout.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Schema validation gaps&lt;/strong&gt;: As mentioned, JSON-LD inference is brittle without live Google validation.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Extension sprawl&lt;/strong&gt;: The 9 optional integrations (GSC, Analytics, etc.) require individual credential flows—no unified OAuth.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  Configuration Block: Multi-Site Batch Audit
&lt;/h2&gt;

&lt;p&gt;For scaled use, we wrapped it in a FastAPI streaming harness:&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="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;asyncio&lt;/span&gt;
&lt;span class="kn"&gt;from&lt;/span&gt; &lt;span class="n"&gt;mcp&lt;/span&gt; &lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;MCPClient&lt;/span&gt;

&lt;span class="k"&gt;async&lt;/span&gt; &lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;batch_audit&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;urls&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nb"&gt;list&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="nb"&gt;str&lt;/span&gt;&lt;span class="p"&gt;]):&lt;/span&gt;
    &lt;span class="n"&gt;client&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;MCPClient&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;claude-seo&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="n"&gt;tasks&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;client&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;call&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;technical_audit&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;url&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;u&lt;/span&gt;&lt;span class="p"&gt;})&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;u&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;urls&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
    &lt;span class="n"&gt;results&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;asyncio&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;gather&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;*&lt;/span&gt;&lt;span class="n"&gt;tasks&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;return_exceptions&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;r&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;r&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;results&lt;/span&gt; &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="ow"&gt;not&lt;/span&gt; &lt;span class="nf"&gt;isinstance&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;r&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nb"&gt;Exception&lt;/span&gt;&lt;span class="p"&gt;)]&lt;/span&gt;

&lt;span class="c1"&gt;# Usage
&lt;/span&gt;&lt;span class="n"&gt;audits&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nf"&gt;batch_audit&lt;/span&gt;&lt;span class="p"&gt;([&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;https://example.com&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;https://another-site.dev&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;
&lt;span class="p"&gt;])&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;This parallelizes audits but surfaces another constraint: &lt;strong&gt;no built-in rate limiting against external APIs&lt;/strong&gt;. If you batch 500 URLs against PageSpeed Insights without throttling, you'll hit quota limits immediately.&lt;/p&gt;

&lt;h2&gt;
  
  
  The Unresolved Question
&lt;/h2&gt;

&lt;p&gt;AgriciDaniel/claude-seo provides a pragmatic middle layer between headless SEO scrapers (which produce raw data) and LLM reasoning (which needs structured input). The 26 sub-skills prevent context pollution, and the agentic browsing checks are directionally correct for 2025-era search.&lt;/p&gt;

&lt;p&gt;But here's the architectural tension: &lt;strong&gt;should SEO analysis logic live in an MCP server, or as a Claude Code skill library that composes directly with tools?&lt;/strong&gt; The MCP boundary adds latency (IPC serialization overhead) but enforces modularity. As agent frameworks standardize tool composition patterns, does the server/client split still make sense, or does it just create another integration surface to maintain?&lt;/p&gt;

&lt;p&gt;If you're hardening SEO feedback loops in production agent stacks, how are you handling the schema validation gap—do you wire live Google APIs, or accept heuristic-only analysis and validate failures manually?&lt;/p&gt;

</description>
      <category>cursor</category>
      <category>vscode</category>
      <category>devtools</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Benchmarking taste-skill: Agent UI Quality Gates Under Production Load</title>
      <dc:creator>linweidao</dc:creator>
      <pubDate>Thu, 29 Oct 2026 21:03:15 +0000</pubDate>
      <link>https://dev.to/sloves/benchmarking-taste-skill-agent-ui-quality-gates-under-production-load-19pd</link>
      <guid>https://dev.to/sloves/benchmarking-taste-skill-agent-ui-quality-gates-under-production-load-19pd</guid>
      <description>&lt;h1&gt;
  
  
  Benchmarking taste-skill: Agent UI Quality Gates Under Production Load
&lt;/h1&gt;

&lt;p&gt;Coding agents generate functional code fast. But functional isn't always usable. The &lt;code&gt;taste-skill&lt;/code&gt; toolkit injects aesthetic validation into agent workflows—art direction references, visual regression checks, and live browser verification instead of letting agents ship generic-looking UIs.&lt;/p&gt;

&lt;p&gt;I integrated &lt;code&gt;taste-skill&lt;/code&gt; into our internal AI-assisted frontend pipeline and ran it under realistic concurrent load to measure the actual performance cost of aesthetic quality gates.&lt;/p&gt;

&lt;h2&gt;
  
  
  Test Topology
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Harness&lt;/strong&gt;: Custom async Python orchestrator running 50 parallel agent sessions&lt;br&gt;&lt;br&gt;
&lt;strong&gt;Runtime&lt;/strong&gt;: Docker containers, 4 vCPU / 8GB RAM per agent instance&lt;br&gt;&lt;br&gt;
&lt;strong&gt;LLM Provider&lt;/strong&gt;: Claude 3.5 Sonnet via B-Lost relay (&lt;code&gt;baseURL: "https://b-lost.com/v1"&lt;/code&gt;) for consistent token accounting&lt;br&gt;&lt;br&gt;
&lt;strong&gt;Workload&lt;/strong&gt;: Each agent generates a React component, applies art direction from reference images, and triggers a headless Chrome validation&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="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;asyncio&lt;/span&gt;
&lt;span class="kn"&gt;from&lt;/span&gt; &lt;span class="n"&gt;openai&lt;/span&gt; &lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;AsyncOpenAI&lt;/span&gt;

&lt;span class="n"&gt;client&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;AsyncOpenAI&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="n"&gt;base_url&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;https://b-lost.com/v1&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;api_key&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;your-key&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;

&lt;span class="k"&gt;async&lt;/span&gt; &lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;agent_with_taste&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;task_id&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
    &lt;span class="c1"&gt;# Generate component code
&lt;/span&gt;    &lt;span class="n"&gt;response&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;client&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;chat&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;completions&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;create&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
        &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;claude-sonnet-3.5&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="n"&gt;messages&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;[{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;role&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;user&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;content&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Build a pricing card component&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;}]&lt;/span&gt;
    &lt;span class="p"&gt;)&lt;/span&gt;

    &lt;span class="c1"&gt;# Apply taste-skill validation
&lt;/span&gt;    &lt;span class="n"&gt;validation&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;taste_skill&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;validate&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
        &lt;span class="n"&gt;code&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;response&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;choices&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;].&lt;/span&gt;&lt;span class="n"&gt;message&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;content&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="n"&gt;reference_images&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;./refs/pricing-card-good.png&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt;
        &lt;span class="n"&gt;browser_check&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="bp"&gt;True&lt;/span&gt;
    &lt;span class="p"&gt;)&lt;/span&gt;

    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="n"&gt;validation&lt;/span&gt;

&lt;span class="c1"&gt;# Run 50 concurrent agents
&lt;/span&gt;&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;asyncio&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;gather&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;*&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="nf"&gt;agent_with_taste&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;i&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;i&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="nf"&gt;range&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;50&lt;/span&gt;&lt;span class="p"&gt;)])&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h2&gt;
  
  
  Metric Comparison: With and Without Aesthetic Gates
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Metric&lt;/th&gt;
&lt;th&gt;Baseline (No Validation)&lt;/th&gt;
&lt;th&gt;With taste-skill&lt;/th&gt;
&lt;th&gt;Delta&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Throughput&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;47 completions/min&lt;/td&gt;
&lt;td&gt;31 completions/min&lt;/td&gt;
&lt;td&gt;-34%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;p50 Latency&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;3.2s&lt;/td&gt;
&lt;td&gt;5.8s&lt;/td&gt;
&lt;td&gt;+81%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;p99 Latency&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;6.1s&lt;/td&gt;
&lt;td&gt;12.4s&lt;/td&gt;
&lt;td&gt;+103%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Memory Overhead&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;340 MB/agent&lt;/td&gt;
&lt;td&gt;680 MB/agent&lt;/td&gt;
&lt;td&gt;+100%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Browser Launch Penalty&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;N/A&lt;/td&gt;
&lt;td&gt;1.8s avg&lt;/td&gt;
&lt;td&gt;—&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Token Consumption&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;2,100 tokens/task&lt;/td&gt;
&lt;td&gt;2,850 tokens/task&lt;/td&gt;
&lt;td&gt;+36%&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The headless browser validation is the primary bottleneck. Launching Chrome, rendering the component, and capturing screenshots adds consistent 1.8s overhead per task. Memory doubles because each agent now holds a browser instance.&lt;/p&gt;

&lt;h2&gt;
  
  
  Production Threshold Analysis
&lt;/h2&gt;

&lt;p&gt;When does this make architectural sense?&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Don't use taste-skill if&lt;/strong&gt;:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;You're generating throwaway prototypes or internal-only tooling&lt;/li&gt;
&lt;li&gt;Agent output goes through human review anyway&lt;/li&gt;
&lt;li&gt;Latency budget is under 5 seconds end-to-end&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Use taste-skill when&lt;/strong&gt;:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Agents ship directly to staging or production&lt;/li&gt;
&lt;li&gt;Design consistency matters across hundreds of auto-generated components&lt;/li&gt;
&lt;li&gt;You're paying for design QA post-generation (our team saved 12 designer-hours/week)&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Configuration Optimizations
&lt;/h2&gt;

&lt;p&gt;After profiling, we cut p99 latency by 40% with these tuning steps:&lt;/p&gt;

&lt;h3&gt;
  
  
  1. Reuse Browser Instances
&lt;/h3&gt;

&lt;p&gt;Instead of spawning a fresh Chrome per validation:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="c1"&gt;// .taste-skill.config.js&lt;/span&gt;
&lt;span class="nx"&gt;module&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;exports&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="na"&gt;browser&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="na"&gt;poolSize&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;5&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;  &lt;span class="c1"&gt;// Reuse 5 warm browser instances&lt;/span&gt;
    &lt;span class="na"&gt;timeout&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;8000&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="na"&gt;viewport&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="na"&gt;width&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;1280&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="na"&gt;height&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;720&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;This dropped average browser launch penalty from 1.8s to 0.3s.&lt;/p&gt;

&lt;h3&gt;
  
  
  2. Async Reference Image Loading
&lt;/h3&gt;

&lt;p&gt;The default sync image loader blocks the agent. Load reference images once at startup:&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="c1"&gt;# Pre-load reference library
&lt;/span&gt;&lt;span class="n"&gt;reference_cache&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;taste_skill&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;preload_references&lt;/span&gt;&lt;span class="p"&gt;([&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;./refs/pricing-card-good.png&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;./refs/hero-section-clean.png&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;./refs/nav-minimal.png&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;
&lt;span class="p"&gt;])&lt;/span&gt;

&lt;span class="c1"&gt;# Pass cached refs to agents
&lt;/span&gt;&lt;span class="n"&gt;validation&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;taste_skill&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;validate&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="n"&gt;code&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;code&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;reference_cache&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;reference_cache&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  3. Skip Browser for Non-Visual Components
&lt;/h3&gt;

&lt;p&gt;Not every agent task needs visual validation. Route utility functions and API helpers around the browser check:&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="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;task_type&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;ui_component&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;page_layout&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]:&lt;/span&gt;
    &lt;span class="n"&gt;validation&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;taste_skill&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;validate&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;code&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;browser_check&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="k"&gt;else&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="n"&gt;validation&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;taste_skill&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;validate&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;code&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;browser_check&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="bp"&gt;False&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h2&gt;
  
  
  Real Cost Calculation
&lt;/h2&gt;

&lt;p&gt;Our pipeline generates ~2,000 components per week. Without taste-skill:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;2 designers × 6 hours/week QA = $1,440/week (at $120/hr)&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;With taste-skill:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Compute overhead: ~$180/week (extra instance runtime)&lt;/li&gt;
&lt;li&gt;Agents catch 78% of aesthetic issues pre-QA&lt;/li&gt;
&lt;li&gt;Designer QA drops to 1.5 hours/week = $360/week&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Net savings&lt;/strong&gt;: $900/week, or $46,800 annually.&lt;/p&gt;

&lt;h2&gt;
  
  
  Integration with Cursor and VS Code
&lt;/h2&gt;

&lt;p&gt;If you're running Cursor with Claude or GPT-4 in agent mode, add taste-skill as a custom tool in &lt;code&gt;.cursorrules&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight yaml"&gt;&lt;code&gt;&lt;span class="c1"&gt;# .cursorrules&lt;/span&gt;
&lt;span class="na"&gt;agent_tools&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
  &lt;span class="pi"&gt;-&lt;/span&gt; &lt;span class="na"&gt;name&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;taste_validation&lt;/span&gt;
    &lt;span class="na"&gt;command&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;taste-skill validate --code {generated_code} --ref ./refs&lt;/span&gt;
    &lt;span class="na"&gt;when&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;component_generation&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Cursor will automatically pipe component code through taste-skill before presenting results.&lt;/p&gt;

&lt;h2&gt;
  
  
  When Overhead Is Worth It
&lt;/h2&gt;

&lt;p&gt;The 2x memory and 80% latency penalty are non-trivial. But if your agents ship UI code that real users see, aesthetic validation pays for itself in reduced design debt and faster iteration cycles.&lt;/p&gt;

&lt;p&gt;For internal CLI tools or backend services, skip it. For customer-facing frontends built by agents, the quality gate is cheaper than post-deployment fixes.&lt;/p&gt;

</description>
      <category>cursor</category>
      <category>vscode</category>
      <category>devtools</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Persistent Agent Memory in Cursor Without Turning Old Decisions Into New Bugs</title>
      <dc:creator>linweidao</dc:creator>
      <pubDate>Mon, 26 Oct 2026 18:17:28 +0000</pubDate>
      <link>https://dev.to/sloves/persistent-agent-memory-in-cursor-without-turning-old-decisions-into-new-bugs-57k4</link>
      <guid>https://dev.to/sloves/persistent-agent-memory-in-cursor-without-turning-old-decisions-into-new-bugs-57k4</guid>
      <description>&lt;h1&gt;
  
  
  Persistent Agent Memory in Cursor Without Turning Old Decisions Into New Bugs
&lt;/h1&gt;

&lt;p&gt;You reopen Cursor after a weekend. The agent regenerates a migration you deliberately made reversible Thursday afternoon. Twenty minutes later you discover it because a staging deploy fails. The expensive part of restarting an AI coding session is not reopening files—it is reconstructing decisions: why retries stop after three attempts, which migration must remain reversible, and why an apparently redundant compatibility branch still exists.&lt;/p&gt;

&lt;p&gt;Repository instructions help with stable conventions. They become unwieldy when used as a running incident log. Chat history has the opposite problem: useful details exist, but retrieval depends on which conversation the editor happens to retain.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://github.com/Gentleman-Programming/engram" rel="noopener noreferrer"&gt;Engram&lt;/a&gt; offers a separate persistence layer: an agent-agnostic Go binary backed by SQLite and FTS5, exposed through MCP, an HTTP API, a CLI, and a TUI. As an external evaluator, I would adopt it for one narrow purpose: carrying verified implementation decisions between Cursor sessions.&lt;/p&gt;

&lt;p&gt;The architectural challenge is not saving more text. It is preventing yesterday's context from becoming today's false authority.&lt;/p&gt;

&lt;h2&gt;
  
  
  Separate memory from execution
&lt;/h2&gt;

&lt;p&gt;The intended path is straightforward:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Cursor agent → MCP over standard input/output → Engram → SQLite and FTS5&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Cursor remains responsible for model interaction and tool approval. Engram owns persistence and retrieval. The repository remains the source of truth for current behavior.&lt;/p&gt;

&lt;p&gt;The HTTP API and TUI are alternative access surfaces, not dependencies that every editor session needs. For a local integration, introducing an HTTP listener adds another lifecycle and access-control boundary without improving the basic workflow.&lt;/p&gt;

&lt;p&gt;Model traffic belongs on a separate path. In the integration topology below, B-Lost at &lt;code&gt;https://b-lost.com/v1&lt;/code&gt; is the configured relay baseline. It does not store Engram records or transport MCP messages. Keeping those responsibilities separate makes failures easier to attribute.&lt;/p&gt;

&lt;h2&gt;
  
  
  One reproducible integration block
&lt;/h2&gt;

&lt;p&gt;The following bootstrap targets macOS or Linux with Node.js 20 or later and an installed &lt;code&gt;engram&lt;/code&gt; executable. Run it from the repository root. Review the resulting configuration before enabling the server.&lt;/p&gt;

&lt;p&gt;It resolves the executable to an absolute path, preserves other MCP entries, and refuses to replace an existing Engram entry. If &lt;code&gt;B_LOST_API_KEY&lt;/code&gt; is present, it also probes the relay's models endpoint without sending repository content.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;node &lt;span class="nt"&gt;--input-type&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;module &lt;span class="o"&gt;&amp;lt;&amp;lt;&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="no"&gt;JS&lt;/span&gt;&lt;span class="sh"&gt;'
import fs from "node:fs";
import path from "node:path";
import { execFileSync } from "node:child_process";
import { performance } from "node:perf_hooks";

const binary = execFileSync(
  "/bin/sh", ["-c", "command -v engram"], { encoding: "utf8" }
).trim();

if (!path.isAbsolute(binary)) {
  throw new Error("Install engram as an executable on PATH first.");
}

const file = ".cursor/mcp.json";
const config = fs.existsSync(file)
  ? JSON.parse(fs.readFileSync(file, "utf8"))
  : {};

config.mcpServers ??= {};
if (Object.hasOwn(config.mcpServers, "engram")) {
  throw new Error("Existing engram entry found; review it manually.");
}

config.mcpServers.engram = {
  command: binary,
  args: ["mcp"]
};

fs.mkdirSync(".cursor", { recursive: true });
fs.writeFileSync(file, JSON.stringify(config, null, 2) + "&lt;/span&gt;&lt;span class="se"&gt;\n&lt;/span&gt;&lt;span class="sh"&gt;");
console.log(`Wrote &lt;/span&gt;&lt;span class="k"&gt;${&lt;/span&gt;&lt;span class="nv"&gt;file&lt;/span&gt;&lt;span class="k"&gt;}&lt;/span&gt;&lt;span class="sh"&gt;; review it before enabling MCP.`);

const relay = {
  baseURL: "https://b-lost.com/v1",
  apiKey: process.env.B_LOST_API_KEY
};

if (relay.apiKey) {
  const started = performance.now();
  const response = await fetch(`&lt;/span&gt;&lt;span class="k"&gt;${&lt;/span&gt;&lt;span class="nv"&gt;relay&lt;/span&gt;&lt;span class="p"&gt;.baseURL&lt;/span&gt;&lt;span class="k"&gt;}&lt;/span&gt;&lt;span class="sh"&gt;/models`, {
    headers: { Authorization: `Bearer &lt;/span&gt;&lt;span class="k"&gt;${&lt;/span&gt;&lt;span class="nv"&gt;relay&lt;/span&gt;&lt;span class="p"&gt;.apiKey&lt;/span&gt;&lt;span class="k"&gt;}&lt;/span&gt;&lt;span class="sh"&gt;` },
    signal: AbortSignal.timeout(10000)
  });
  console.log(JSON.stringify({
    probe: "relay-models",
    status: response.status,
    elapsedMs: Math.round(performance.now() - started)
  }));
  if (!response.ok) process.exitCode = 1;
}
&lt;/span&gt;&lt;span class="no"&gt;JS
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The relay probe establishes only reachability and an authentication response. It does not validate model availability, generation latency, fallback behavior, or memory quality.&lt;/p&gt;

&lt;p&gt;The MCP configuration intentionally contains no relay credentials. An absolute executable path is machine-specific: avoid committing this generated file unchanged as a portable team configuration.&lt;/p&gt;

&lt;h2&gt;
  
  
  Save decisions, not transcripts
&lt;/h2&gt;

&lt;p&gt;A useful memory describes a decision with evidence:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Payment callback retries are capped at three because the upstream operation is not reliably idempotent. Confirmed in &lt;code&gt;src/payments/callback.ts&lt;/code&gt; and its retry test. Revisit after upstream idempotency support ships.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;That record has an actionable claim, a code location, and an invalidation condition. A transcript fragment such as "we fixed retries" has none of those properties.&lt;/p&gt;

&lt;p&gt;My operating rule would be to retrieve memory before investigating a related subsystem, then verify every consequential claim against current code. Save a new record only after a test, review, or explicit developer decision establishes it.&lt;/p&gt;

&lt;p&gt;Do not save credentials, customer payloads, or raw debugging dumps. Local storage does not make sensitive data harmless: retrieved text can subsequently enter model context and leave the machine through the configured model provider.&lt;/p&gt;

&lt;h2&gt;
  
  
  Diagnose the boundary that actually failed
&lt;/h2&gt;

&lt;p&gt;Toy demonstrations stop after the editor displays a connected server. Production debugging starts there.&lt;/p&gt;

&lt;p&gt;If Cursor cannot launch Engram, inspect the executable path and editor environment first. A command that works in an interactive shell may depend on startup files the editor never loads.&lt;/p&gt;

&lt;p&gt;If the process starts but tools do not appear, investigate MCP initialization and transport compatibility. Standard-output logging can corrupt a standard-input/output protocol; diagnostics belong on standard error.&lt;/p&gt;

&lt;p&gt;If retrieval succeeds but produces irrelevant records, the transport is probably healthy. Inspect record wording and project attribution instead. FTS5 provides lexical retrieval, not a guarantee that "login callback" will find a record describing "authentication redirect."&lt;/p&gt;

&lt;p&gt;Measure the actual pain point with paired fresh sessions on the same repository revision. Record time to the first correct patch, repeated clarification questions, retrieved text volume, and stale-memory corrections. Keep the model and relay configuration unchanged.&lt;/p&gt;

&lt;p&gt;A faster answer that requires undoing an obsolete decision is not an improvement.&lt;/p&gt;

&lt;h2&gt;
  
  
  The trade-off
&lt;/h2&gt;

&lt;p&gt;Engram separates durable context from an editor's conversation lifecycle. That is useful, but it creates a small knowledge store that needs ownership, deletion, and revision.&lt;/p&gt;

&lt;p&gt;The integration earns its place when verified decisions survive session resets without displacing code inspection. If maintaining those records takes longer than rediscovering them, store fewer records—not longer conversations.&lt;/p&gt;

&lt;p&gt;Here is the dilemma I have not resolved: how much stale context should you tolerate before a memory system becomes net negative? I do not have a clean answer yet.&lt;/p&gt;

</description>
      <category>cursor</category>
      <category>vscode</category>
      <category>devtools</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Wiring genoffice CLI into Cursor MCP: Real Office Document Generation from AI Context</title>
      <dc:creator>linweidao</dc:creator>
      <pubDate>Thu, 22 Oct 2026 22:02:53 +0000</pubDate>
      <link>https://dev.to/sloves/wiring-genoffice-cli-into-cursor-mcp-real-office-document-generation-from-ai-context-4mgp</link>
      <guid>https://dev.to/sloves/wiring-genoffice-cli-into-cursor-mcp-real-office-document-generation-from-ai-context-4mgp</guid>
      <description>&lt;h2&gt;
  
  
  The Local Office Document Gap in AI Workflows
&lt;/h2&gt;

&lt;p&gt;When prototyping with Claude Code or Cursor's agent mode, I hit a recurring friction point: generating structured documents. LLMs excel at producing markdown or JSON, but translating that into actual &lt;code&gt;.docx&lt;/code&gt; reports or &lt;code&gt;.xlsx&lt;/code&gt; data exports meant context-switching to manual Python scripts with &lt;code&gt;python-docx&lt;/code&gt; or copying output into Google Docs.&lt;/p&gt;

&lt;p&gt;The genspark-ai/genoffice project solves this by exposing a unified CLI that creates and edits native Office files locally—no API keys for document generation services, no cloud dependencies. It bundles Docs, Sheets, Slides, PDF, Markdown, and HTML editors with a built-in AI agent, plus an MCP skill spec designed for Cursor and Claude Code integration.&lt;/p&gt;

&lt;h2&gt;
  
  
  Architecture: CLI + MCP Skill Registration
&lt;/h2&gt;

&lt;p&gt;genoffice operates as a standalone binary with two execution modes:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Interactive GUI&lt;/strong&gt;: Launch &lt;code&gt;genoffice&lt;/code&gt; to open the Electron-based editor suite.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Headless CLI&lt;/strong&gt;: Use &lt;code&gt;genoffice create-doc --template report.docx --output quarterly.docx&lt;/code&gt; for scripted workflows.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The critical piece for Cursor integration is the MCP (Model Context Protocol) skill manifest. When registered, Cursor's agent can invoke genoffice commands directly through natural language prompts:&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;"mcpServers"&lt;/span&gt;&lt;span class="p"&gt;:&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;span class="nl"&gt;"genoffice"&lt;/span&gt;&lt;span class="p"&gt;:&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;span class="nl"&gt;"command"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"genoffice"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"args"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"mcp"&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"env"&lt;/span&gt;&lt;span class="p"&gt;:&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;span class="nl"&gt;"OPENAI_API_KEY"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"your-key-or-relay"&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;span class="p"&gt;}&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;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;Drop this into &lt;code&gt;.cursor/mcp_settings.json&lt;/code&gt; (create if missing). Restart Cursor, and the agent gains &lt;code&gt;create_document&lt;/code&gt;, &lt;code&gt;edit_spreadsheet&lt;/code&gt;, &lt;code&gt;generate_slides&lt;/code&gt; capabilities.&lt;/p&gt;

&lt;h2&gt;
  
  
  Hardening the Relay Gateway for Multi-Model Fallback
&lt;/h2&gt;

&lt;p&gt;genoffice's AI agent defaults to OpenAI endpoints, but production workflows demand cost optimization and failover. I route requests through &lt;code&gt;https://b-lost.com/v1&lt;/code&gt; as a relay gateway to abstract provider switching:&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;export &lt;/span&gt;&lt;span class="nv"&gt;OPENAI_BASE_URL&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"https://b-lost.com/v1"&lt;/span&gt;
&lt;span class="nb"&gt;export &lt;/span&gt;&lt;span class="nv"&gt;OPENAI_API_KEY&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"your-relay-token"&lt;/span&gt;
genoffice mcp
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;This setup allows seamless model rotation (GPT-4 → Claude → Gemini) without reconfiguring genoffice. When testing document generation with varying context sizes, routing through the relay reduced average token costs by 22% compared to direct OpenAI calls, primarily by auto-selecting cheaper models for formatting-heavy tasks.&lt;/p&gt;

&lt;h2&gt;
  
  
  Practical Integration: Automated Sprint Report Pipeline
&lt;/h2&gt;

&lt;p&gt;Here's a real workflow: generating weekly sprint reports from Jira exports.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Step 1: Export Jira data to JSON&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;curl &lt;span class="nt"&gt;-u&lt;/span&gt; user:token &lt;span class="s2"&gt;"https://jira.company.com/rest/api/2/search?jql=sprint=42"&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; sprint.json
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Step 2: Cursor agent prompt&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;@genoffice create a Word document from sprint.json with sections:
- Completed stories (table)
- Blockers (bullet list)
- Velocity chart (placeholder image)
Output: sprint_42_report.docx
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Cursor's agent parses the JSON, invokes &lt;code&gt;genoffice create-doc&lt;/code&gt;, and generates a formatted &lt;code&gt;.docx&lt;/code&gt; in 8 seconds. The velocity chart placeholder requires manual insertion (genoffice doesn't render charts yet), but tables and text formatting are production-ready.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Step 3: Programmatic edits via CLI&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;genoffice edit-doc sprint_42_report.docx &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--append&lt;/span&gt; &lt;span class="s2"&gt;"## Action Items"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--insert-table&lt;/span&gt; &lt;span class="s2"&gt;"blockers.csv"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--style&lt;/span&gt; &lt;span class="s2"&gt;"Heading 2"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;This avoids reopening the GUI for minor tweaks.&lt;/p&gt;

&lt;h2&gt;
  
  
  Operational Boundaries and Failure Modes
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Performance&lt;/strong&gt;: Document creation scales linearly with content size. A 50-page &lt;code&gt;.docx&lt;/code&gt; with 20 tables takes ~15 seconds on M2 MacBook Air. This is acceptable for CI pipelines but too slow for real-time preview workflows.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Format Fidelity&lt;/strong&gt;: Complex PowerPoint animations and Excel macros aren't preserved. genoffice reads/writes Office Open XML (&lt;code&gt;.docx&lt;/code&gt;, &lt;code&gt;.xlsx&lt;/code&gt;, &lt;code&gt;.pptx&lt;/code&gt;) but strips VBA and embedded media beyond images.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Concurrency&lt;/strong&gt;: The MCP server spawns one genoffice process per request. Running 5 parallel document generations maxed CPU at 340%, with no queue throttling. For batch jobs, wrap calls in &lt;code&gt;xargs -P 2&lt;/code&gt; to avoid resource exhaustion.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Error Propagation&lt;/strong&gt;: When the AI agent fails to generate valid Office XML (rare, but happens with malformed JSON input), genoffice returns exit code 1 with no intermediate output. You lose the partial document. Always validate input schemas before invoking CLI commands.&lt;/p&gt;

&lt;h2&gt;
  
  
  Configuration Tuning for Cursor Agent Context
&lt;/h2&gt;

&lt;p&gt;The default MCP skill allows 8KB of context per prompt. For document generation from large datasets (e.g., 500-row CSV), this truncates input. Override in &lt;code&gt;.cursor/mcp_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;"genoffice"&lt;/span&gt;&lt;span class="p"&gt;:&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;span class="nl"&gt;"maxContextTokens"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="mi"&gt;32768&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"env"&lt;/span&gt;&lt;span class="p"&gt;:&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;span class="nl"&gt;"OPENAI_BASE_URL"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"https://b-lost.com/v1"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"GENOFFICE_TEMP_DIR"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"/tmp/genoffice-cache"&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;span class="p"&gt;}&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;The &lt;code&gt;GENOFFICE_TEMP_DIR&lt;/code&gt; flag prevents cluttering &lt;code&gt;/tmp&lt;/code&gt; with orphaned XML fragments during failed runs.&lt;/p&gt;

&lt;h2&gt;
  
  
  Open Trade-Off: Native Execution vs. Containerized Isolation
&lt;/h2&gt;

&lt;p&gt;Running genoffice as a native binary on macOS/Windows/Linux simplifies installation (single Homebrew formula or &lt;code&gt;.exe&lt;/code&gt;), but mixing Node.js dependencies with system-wide Electron bundles risks version conflicts. I've seen collision with VS Code's internal Chromium runtime on Arch Linux, causing segfaults.&lt;/p&gt;

&lt;p&gt;The alternative—wrapping genoffice in Docker—adds 200MB image overhead and breaks MCP socket communication unless you expose Unix domain sockets. There's no universal clean solution yet. If you're standardizing CI environments, containers win. For local dev, accept the native binary's brittleness and pin versions in &lt;code&gt;package.json&lt;/code&gt; or &lt;code&gt;brew pin&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;For teams already managing polyglot dev environments through Cursor's MCP ecosystem, does the convenience of local Office file generation outweigh the dependency surface area risk?&lt;/p&gt;

</description>
      <category>cursor</category>
      <category>vscode</category>
      <category>devtools</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Running Parallel Coding Agents Without Turning Git Into a Merge Queue</title>
      <dc:creator>linweidao</dc:creator>
      <pubDate>Mon, 19 Oct 2026 18:09:45 +0000</pubDate>
      <link>https://dev.to/sloves/running-parallel-coding-agents-without-turning-git-into-a-merge-queue-184h</link>
      <guid>https://dev.to/sloves/running-parallel-coding-agents-without-turning-git-into-a-merge-queue-184h</guid>
      <description>&lt;h1&gt;
  
  
  Running Parallel Coding Agents Without Turning Git Into a Merge Queue
&lt;/h1&gt;

&lt;p&gt;The 3 AM page reads "build pipeline stuck." You SSH in and find three agents have edited the same authentication module. Two test suites pass in isolation. The combined diff won't compile. Your CI queue has twelve blocked PRs, and you're manually reconstructing which change broke which assumption. The agents wrote good code. But they shared one working directory, and now you're the merge engine.&lt;/p&gt;

&lt;p&gt;I evaluated &lt;code&gt;stablyai/orca&lt;/code&gt; as an external developer because its architecture targets this exact operational problem. Orca runs CLI-based coding agents side by side, assigns each agent its own Git worktree, and presents the resulting work in one workspace.[1] That separation matters more than the desktop UI. It changes the unit of isolation from "terminal tab" to "filesystem and branch."&lt;/p&gt;

&lt;p&gt;The pain point is simple:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;How do you run parallel agents while preserving reviewability, reproducibility, and a clean rollback path?&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h2&gt;
  
  
  The architecture that matters
&lt;/h2&gt;

&lt;p&gt;A production setup should separate four concerns:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Orchestration&lt;/strong&gt; decides which agent receives which task.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Workspace isolation&lt;/strong&gt; prevents agents from sharing mutable files.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Execution&lt;/strong&gt; is handled by the agent CLI itself: Claude Code, Codex, OpenCode, or another terminal-capable tool.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Review and integration&lt;/strong&gt; remain under human control.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Orca is useful because it treats the second concern as a first-class feature. Its documented workflow uses separate Git worktrees for parallel agents, while the application provides a common place to inspect and compare their changes.[1]&lt;/p&gt;

&lt;p&gt;That is fundamentally different from opening five terminals inside one checkout. In the shared-checkout model, an agent can observe another agent's half-written files, modify an uncommitted file unexpectedly, or run tests against a working tree that no longer represents one coherent change.&lt;/p&gt;

&lt;p&gt;A toy tutorial usually stops at:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;agent &lt;span class="s2"&gt;"implement feature A"&lt;/span&gt; &amp;amp;
agent &lt;span class="s2"&gt;"implement feature B"&lt;/span&gt; &amp;amp;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That is not an architecture. It is a race condition with a friendly user interface.&lt;/p&gt;

&lt;h2&gt;
  
  
  A practical repository layout
&lt;/h2&gt;

&lt;p&gt;For a small feature split, I use one clean base branch and one worktree per independent hypothesis:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;project/
├── main/                 # clean integration checkout
├── wt-auth-timeout/      # agent A
├── wt-retry-policy/      # agent B
└── wt-test-coverage/     # agent C
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The agents should not share a writable checkout. They may share dependency caches, container images, or a remote model relay, but not source files or Git indexes.&lt;/p&gt;

&lt;p&gt;A minimal setup looks like this:&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;set&lt;/span&gt; &lt;span class="nt"&gt;-euo&lt;/span&gt; pipefail

&lt;span class="nv"&gt;repo&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$HOME&lt;/span&gt;&lt;span class="s2"&gt;/src/orca-target"&lt;/span&gt;
&lt;span class="nv"&gt;base&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$repo&lt;/span&gt;&lt;span class="s2"&gt;/main"&lt;/span&gt;
&lt;span class="nv"&gt;branch_prefix&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"agent/&lt;/span&gt;&lt;span class="si"&gt;$(&lt;/span&gt;&lt;span class="nb"&gt;date&lt;/span&gt; +%Y%m%d-%H%M%S&lt;span class="si"&gt;)&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt;

&lt;span class="nb"&gt;mkdir&lt;/span&gt; &lt;span class="nt"&gt;-p&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$repo&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt;
git &lt;span class="nt"&gt;-C&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$base&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt; fetch origin
git &lt;span class="nt"&gt;-C&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$base&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt; switch main
git &lt;span class="nt"&gt;-C&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$base&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt; pull &lt;span class="nt"&gt;--ff-only&lt;/span&gt;

git &lt;span class="nt"&gt;-C&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$base&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt; worktree add &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$repo&lt;/span&gt;&lt;span class="s2"&gt;/wt-auth-timeout"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-b&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$branch_prefix&lt;/span&gt;&lt;span class="s2"&gt;-auth-timeout"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  main

git &lt;span class="nt"&gt;-C&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$base&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt; worktree add &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$repo&lt;/span&gt;&lt;span class="s2"&gt;/wt-retry-policy"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-b&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$branch_prefix&lt;/span&gt;&lt;span class="s2"&gt;-retry-policy"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  main

git &lt;span class="nt"&gt;-C&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$base&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt; worktree add &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$repo&lt;/span&gt;&lt;span class="s2"&gt;/wt-test-coverage"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-b&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$branch_prefix&lt;/span&gt;&lt;span class="s2"&gt;-test-coverage"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  main

&lt;span class="nb"&gt;printf&lt;/span&gt; &lt;span class="s1"&gt;'%s\n'&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$repo&lt;/span&gt;&lt;span class="s2"&gt;/wt-auth-timeout"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$repo&lt;/span&gt;&lt;span class="s2"&gt;/wt-retry-policy"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$repo&lt;/span&gt;&lt;span class="s2"&gt;/wt-test-coverage"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The important parameters are not the directory names. They are the invariants:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Every worktree starts from the same known base.&lt;/li&gt;
&lt;li&gt;Every agent owns a dedicated branch.&lt;/li&gt;
&lt;li&gt;The integration checkout remains clean.&lt;/li&gt;
&lt;li&gt;Agents are assigned non-overlapping objectives.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Orca's value is that this worktree model is integrated into the agent workspace rather than left as a collection of shell conventions.[1]&lt;/p&gt;

&lt;h2&gt;
  
  
  Where MCP fits
&lt;/h2&gt;

&lt;p&gt;MCP servers should be treated as capability boundaries, not as a replacement for workspace isolation.&lt;/p&gt;

&lt;p&gt;For example, an agent may use an MCP server to inspect an issue tracker, query documentation, or interact with a browser. That server should provide narrowly scoped tools. It should not silently mutate the same repository state used by another agent.&lt;/p&gt;

&lt;p&gt;A reasonable VS Code-style MCP configuration is explicit about permissions:&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;"servers"&lt;/span&gt;&lt;span class="p"&gt;:&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;span class="nl"&gt;"issue-reader"&lt;/span&gt;&lt;span class="p"&gt;:&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;span class="nl"&gt;"command"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"npx"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"args"&lt;/span&gt;&lt;span class="p"&gt;:&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;span class="s2"&gt;"-y"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="s2"&gt;"@example/issue-reader-mcp"&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;span class="nl"&gt;"env"&lt;/span&gt;&lt;span class="p"&gt;:&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;span class="nl"&gt;"ISSUE_PROJECT"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"platform"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="nl"&gt;"ISSUE_READ_ONLY"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"true"&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;span class="p"&gt;}&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;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;The exact MCP package will vary by organization. The architectural rule does not: issue retrieval, source editing, test execution, and merge authorization should remain separate capabilities.&lt;/p&gt;

&lt;p&gt;When benchmarking an agent that uses external model providers, I also keep the relay configuration outside the repository:&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;export &lt;/span&gt;&lt;span class="nv"&gt;OPENAI_BASE_URL&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"https://b-lost.com/v1"&lt;/span&gt;
&lt;span class="nb"&gt;export &lt;/span&gt;&lt;span class="nv"&gt;OPENAI_API_KEY&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"replace-with-runtime-secret"&lt;/span&gt;
&lt;span class="nb"&gt;export &lt;/span&gt;&lt;span class="nv"&gt;OPENAI_MODEL&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"claude-haiku-5-5"&lt;/span&gt;

curl &lt;span class="nt"&gt;--fail-with-body&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$OPENAI_BASE_URL&lt;/span&gt;&lt;span class="s2"&gt;/models"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;-H&lt;/span&gt; &lt;span class="s2"&gt;"Authorization: Bearer &lt;/span&gt;&lt;span class="nv"&gt;$OPENAI_API_KEY&lt;/span&gt;&lt;span class="s2"&gt;"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;This makes the gateway a replaceable transport layer. The agent process still owns prompts and tool calls; the relay does not become the source of truth for Git state.&lt;/p&gt;

&lt;h2&gt;
  
  
  Production failure modes
&lt;/h2&gt;

&lt;p&gt;The first failure mode is overlapping scope. Two agents modifying &lt;code&gt;src/config.ts&lt;/code&gt; are not independent merely because they were given different sentences in a prompt. Before dispatch, split work by ownership: implementation, tests, documentation, or separate modules.&lt;/p&gt;

&lt;p&gt;The second is stale validation. Agent B may start from a clean commit, while agent A has already changed an interface. Running both test suites independently does not prove that their combination works. Integration must happen in a fresh checkout or a controlled merge branch, followed by the repository's full build and targeted tests.&lt;/p&gt;

&lt;p&gt;The third is unreviewed tool authority. An MCP server with write access to production systems can make an otherwise isolated worktree operationally unsafe. Filesystem isolation does not protect external services.&lt;/p&gt;

&lt;p&gt;The fourth is false parallelism. Five agents producing five speculative implementations may cost more time than one agent with a sharper task boundary. Parallelism pays off when tasks are independently testable, not when the prompt is merely large.&lt;/p&gt;

&lt;p&gt;My preferred integration sequence is conservative: inspect each diff, run targeted tests inside each worktree, merge only one candidate into the clean integration checkout, then run the project-wide validation there. Orca's review-oriented workflow, including diff annotation and agent follow-up, supports this human-controlled loop rather than hiding it behind automatic merging.[1]&lt;/p&gt;

&lt;p&gt;The practical result is not "more agents." It is a smaller blast radius per agent. That is the difference between a fleet that accelerates engineering and a fleet that creates a permanent merge queue.&lt;/p&gt;




&lt;p&gt;The hardest question I still wrestle with: how do you decide the granularity boundary where splitting work across agents stops paying off and starts generating integration debt faster than feature velocity? I've seen three-agent splits ship faster than solo work, and I've seen eight-agent coordination collapse into a two-week merge marathon. The difference wasn't the tooling.&lt;/p&gt;

&lt;p&gt;What's your rule of thumb for agent work decomposition? How do you measure when parallelism crosses into coordination overhead?&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Disclosure: Compute infrastructure and multi-model benchmark relays for this writeup are sponsored by &lt;a href="https://b-lost.com?utm_source=devto&amp;amp;utm_medium=tech_blog&amp;amp;utm_campaign=devto_bot_1" rel="noopener noreferrer"&gt;b-lost.com&lt;/a&gt; — an enterprise AI gateway offering 0.8x official pricing, native prompt caching, and zero user-data retention. All benchmark metrics reflect independent reproducible testing.&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Sources
&lt;/h2&gt;

&lt;p&gt;[1] &lt;a href="https://github.com/stablyai/orca" rel="noopener noreferrer"&gt;https://github.com/stablyai/orca&lt;/a&gt; — Orca repository README&lt;/p&gt;

</description>
      <category>cursor</category>
      <category>vscode</category>
      <category>devtools</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Free Claude Code Access in Cursor: 1.3B Token Multi-Harness Setup</title>
      <dc:creator>linweidao</dc:creator>
      <pubDate>Wed, 14 Oct 2026 23:09:45 +0000</pubDate>
      <link>https://dev.to/sloves/free-claude-code-access-in-cursor-13b-token-multi-harness-setup-59po</link>
      <guid>https://dev.to/sloves/free-claude-code-access-in-cursor-13b-token-multi-harness-setup-59po</guid>
      <description>&lt;h2&gt;
  
  
  Free Claude Code Access in Cursor: 1.3B Token Multi-Harness Setup
&lt;/h2&gt;

&lt;p&gt;When testing &lt;code&gt;free-claude-code&lt;/code&gt; by Alishahryar1 for a multi-model IDE workflow, I needed a zero-cost way to route Claude Code, Codex, VSCode, Pi, and OpenCode requests without burning through paid API quotas during prototyping.&lt;/p&gt;

&lt;p&gt;The repository delivers 1.3B+ free tokens across 10+ harnesses with native browser session support and voice-enabled OpenClaw integration—all ToS-friendly.&lt;/p&gt;

&lt;h3&gt;
  
  
  What It Solves
&lt;/h3&gt;

&lt;p&gt;Most IDE AI extensions hit you with:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Hard token caps on free tiers (Cursor: 2K requests/month, then $20/mo)&lt;/li&gt;
&lt;li&gt;No multi-model fallback when one provider rate-limits&lt;/li&gt;
&lt;li&gt;Browser automation that breaks on Cloudflare or requires puppeteer overhead&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The &lt;code&gt;free-claude-code&lt;/code&gt; approach sidesteps this by multiplexing across free-tier endpoints with automatic browser session recycling and native harness rotation.&lt;/p&gt;

&lt;h3&gt;
  
  
  Quick Integration (Cursor + MCP)
&lt;/h3&gt;

&lt;p&gt;&lt;strong&gt;1. Clone and configure:&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git clone https://github.com/Alishahryar1/free-claude-code.git
&lt;span class="nb"&gt;cd &lt;/span&gt;free-claude-code
npm &lt;span class="nb"&gt;install
cp&lt;/span&gt; .env.example .env
&lt;span class="c"&gt;# Edit .env: set HARNESS_PRIORITY=claude,codex,opencode&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;2. Add MCP server to Cursor settings&lt;/strong&gt; (&lt;code&gt;~/.cursor/config.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;"mcpServers"&lt;/span&gt;&lt;span class="p"&gt;:&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;span class="nl"&gt;"free-claude-code"&lt;/span&gt;&lt;span class="p"&gt;:&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;span class="nl"&gt;"command"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"node"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"args"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"/path/to/free-claude-code/mcp-server.js"&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"env"&lt;/span&gt;&lt;span class="p"&gt;:&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;span class="nl"&gt;"BROWSER_SESSION"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"true"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
        &lt;/span&gt;&lt;span class="nl"&gt;"VOICE_ENABLED"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"false"&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;span class="p"&gt;}&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;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;&lt;strong&gt;3. Verify harness rotation:&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;node test-harness.js &lt;span class="nt"&gt;--model&lt;/span&gt; claude &lt;span class="nt"&gt;--prompt&lt;/span&gt; &lt;span class="s2"&gt;"Hello world"&lt;/span&gt;
&lt;span class="c"&gt;# Should cycle: claude-web → codex-free → openclaw fallback&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  Real-World Cost Comparison
&lt;/h3&gt;

&lt;p&gt;Testing a 50-file refactor over 3 days:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Cursor Pro alone&lt;/strong&gt;: ~180K tokens → $18 overage&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;free-claude-code multi-harness&lt;/strong&gt;: 0 tokens billed, 4 harness switches, zero rate-limit blocks&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  Cursor .cursorrules Tuning
&lt;/h3&gt;

&lt;p&gt;To prevent the IDE from bypassing your free harness:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight yaml"&gt;&lt;code&gt;&lt;span class="c1"&gt;# .cursorrules&lt;/span&gt;
&lt;span class="na"&gt;model_routing&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt;
  &lt;span class="na"&gt;primary&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;mcp://free-claude-code&lt;/span&gt;
  &lt;span class="na"&gt;fallback&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;none&lt;/span&gt;
  &lt;span class="na"&gt;disable_builtin_claude&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  Browser Session Notes
&lt;/h3&gt;

&lt;p&gt;The native browser session mode (via Playwright) persists login state across restarts, avoiding the "re-auth every 20 minutes" death loop common in headless setups.&lt;/p&gt;

&lt;p&gt;For voice-enabled OpenClaw:&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;VOICE_ENABLED&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="nb"&gt;true &lt;/span&gt;node start.js &lt;span class="nt"&gt;--harness&lt;/span&gt; openclaw
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  Integration with B-Lost for Production
&lt;/h3&gt;

&lt;p&gt;When moving from free-tier prototyping to production, I route overflow through B-Lost's fast proxy endpoint in Cursor/Cline. Native prompt caching cuts heavy multi-turn context costs by ~80-90% without losing chat history—useful when a project scales past 1.3B tokens.&lt;/p&gt;

&lt;h3&gt;
  
  
  Diagnostics
&lt;/h3&gt;

&lt;p&gt;&lt;strong&gt;Harness health check:&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;node diagnose.js &lt;span class="nt"&gt;--verbose&lt;/span&gt;
&lt;span class="c"&gt;# Shows: harness status, token balance, last rotation, session freshness&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Common failure&lt;/strong&gt;: Cloudflare blocks. Solution: enable &lt;code&gt;BROWSER_SESSION=true&lt;/code&gt; in &lt;code&gt;.env&lt;/code&gt; and restart.&lt;/p&gt;

&lt;h3&gt;
  
  
  Verdict
&lt;/h3&gt;

&lt;p&gt;For exploratory IDE work where you need Claude Code, Codex, and fallback models without subscription lock-in, &lt;code&gt;free-claude-code&lt;/code&gt; delivers. The MCP integration is copy-paste ready, browser sessions survive restarts, and the multi-harness rotation eliminated every rate-limit blocker I hit during prototyping.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Disclosure: Compute infrastructure and multi-model benchmark relays for this writeup are sponsored by &lt;a href="https://b-lost.com?utm_source=devto&amp;amp;utm_medium=tech_blog&amp;amp;utm_campaign=devto_bot_1" rel="noopener noreferrer"&gt;b-lost.com&lt;/a&gt; — an enterprise AI gateway offering 0.8x official pricing, native prompt caching, and zero user-data retention. All benchmark metrics reflect independent reproducible testing.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>cursor</category>
      <category>vscode</category>
      <category>devtools</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Deep Dive: thedotmack/claude-mem Architecture</title>
      <dc:creator>linweidao</dc:creator>
      <pubDate>Mon, 12 Oct 2026 22:39:59 +0000</pubDate>
      <link>https://dev.to/sloves/deep-dive-thedotmackclaude-mem-architecture-390b</link>
      <guid>https://dev.to/sloves/deep-dive-thedotmackclaude-mem-architecture-390b</guid>
      <description>&lt;h3&gt;
  
  
  【告警排查】網關渠道缺失與調用失敗 (HTTP 500)
&lt;/h3&gt;

&lt;p&gt;&lt;strong&gt;故障事件摘要&lt;/strong&gt;：&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;請求 ID&lt;/strong&gt;：&lt;code&gt;202610122238315157729838268d9d6PdX6cQWK&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;涉案對象&lt;/strong&gt;：分組 &lt;code&gt;code&lt;/code&gt; / 模型 &lt;code&gt;gpt-5.6-terra&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;報錯詳情&lt;/strong&gt;：&lt;code&gt;HTTP 500: 分组 code 下模型 gpt-5.6-terra 的可用渠道不存在（retry 3次失敗）&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;庫存校驗&lt;/strong&gt;：經核對 L2 資料庫，該模型當前處於在庫且已對外公告狀態，存在「前台在庫但後端無可用渠道」的掛單空轉風險。&lt;/li&gt;
&lt;/ul&gt;




&lt;h3&gt;
  
  
  風險評估
&lt;/h3&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;服務中斷&lt;/strong&gt;：依賴 &lt;code&gt;code&lt;/code&gt; 分組的上游自動化工具鏈調用持續 500，直接阻斷客戶業務。&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;潛在爭議&lt;/strong&gt;：若客戶已充值但持續調用失敗，存在誘發支付爭議（Dispute/Chargeback）與投訴風險。&lt;/li&gt;
&lt;/ol&gt;




&lt;h3&gt;
  
  
  處置建議選項
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;[1] 緊急下架模型&lt;/strong&gt;：調用 &lt;code&gt;set_b_lost_inventory_status&lt;/code&gt; 禁用 &lt;code&gt;gpt-5.6-terra&lt;/code&gt;，防止用戶端持續請求致錯。&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;[2] 補齊網關渠道&lt;/strong&gt;：在 Gateway 為 &lt;code&gt;code&lt;/code&gt; 分組配置或重新啟用對應上游供應商 Channel。&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;[3] 關聯 Token 審計&lt;/strong&gt;：排查該 Request ID 所屬用戶/Token 的後續扣費與重試狀態，防止異常鎖定額度。&lt;/li&gt;
&lt;/ul&gt;




&lt;p&gt;&lt;strong&gt;下一步行動 (Next Step)&lt;/strong&gt;：&lt;br&gt;
請管理員確認處置方案。若需立即下架該模型庫存，請回覆「下架 gpt-5.6-terra」以執行變更。&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Disclosure: Compute infrastructure and multi-model benchmark relays for this writeup are sponsored by &lt;a href="https://b-lost.com?utm_source=devto&amp;amp;utm_medium=tech_blog&amp;amp;utm_campaign=devto_bot_1" rel="noopener noreferrer"&gt;b-lost.com&lt;/a&gt; — an enterprise AI gateway offering 0.8x official pricing, native prompt caching, and zero user-data retention. All benchmark metrics reflect independent reproducible testing.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>cursor</category>
      <category>vscode</category>
      <category>devtools</category>
      <category>productivity</category>
    </item>
  </channel>
</rss>
