<?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: Shanmukh</title>
    <description>The latest articles on DEV Community by Shanmukh (@shanmukhdheekonda).</description>
    <link>https://dev.to/shanmukhdheekonda</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%2F4246023%2F98b6e516-6bc9-40e4-927c-e8c76510ba97.png</url>
      <title>DEV Community: Shanmukh</title>
      <link>https://dev.to/shanmukhdheekonda</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/shanmukhdheekonda"/>
    <language>en</language>
    <item>
      <title>Your AI agent skills are an unversioned dependency</title>
      <dc:creator>Shanmukh</dc:creator>
      <pubDate>Sun, 15 Nov 2026 13:45:58 +0000</pubDate>
      <link>https://dev.to/shanmukhdheekonda/your-ai-agent-skills-are-an-unversioned-dependency-i8g</link>
      <guid>https://dev.to/shanmukhdheekonda/your-ai-agent-skills-are-an-unversioned-dependency-i8g</guid>
      <description>&lt;h1&gt;
  
  
  Your AI agent skills are an unversioned dependency
&lt;/h1&gt;

&lt;p&gt;A skill is a prompt that lands in a model's system context. Most teams install&lt;br&gt;
them the way we installed libraries in 2009: clone a repo, copy a folder, hope&lt;br&gt;
nobody touches it. Then the folder drifts and the agent quietly loads something&lt;br&gt;
nobody reviewed.&lt;/p&gt;

&lt;p&gt;I maintain Skillporter, a package manager for AI agent skills. This is the design&lt;br&gt;
problem behind it, not a feature tour.&lt;/p&gt;

&lt;h2&gt;
  
  
  The real problem is paths, not prompts
&lt;/h2&gt;

&lt;p&gt;Every harness reads skills from a different directory:&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Claude Code        .claude/skills
OpenAI Codex       .agents/skills
Cursor             .agents/skills
GitHub Copilot     .github/skills
Windsurf           .windsurf/skills
Cline              .cline/skills
OpenCode           .opencode/skills
Hermes Agent       $HERMES_HOME/skills
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;

&lt;p&gt;Eight harnesses share .agents/skills as their primary project path, and six more&lt;br&gt;
read it as a documented compatibility path. That directory is the trap. Remove a&lt;br&gt;
skill to clean up Cursor and you have just broken Codex, Gemini CLI, Zed, OpenHands&lt;br&gt;
and Goose for the rest of the team.&lt;/p&gt;

&lt;p&gt;Two other approaches fail in predictable ways. Symlinks break inside Docker&lt;br&gt;
builds, git worktrees and Windows checkouts. Copy scripts drift, because nothing&lt;br&gt;
records what the copy was supposed to be.&lt;/p&gt;

&lt;h2&gt;
  
  
  Treat the skill as a dependency
&lt;/h2&gt;

&lt;p&gt;The package manager model maps cleanly:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;a manifest you commit, listing skills and target harnesses&lt;/li&gt;
&lt;li&gt;a resolution step that fetches sources and validates structure&lt;/li&gt;
&lt;li&gt;a static scan before anything is written&lt;/li&gt;
&lt;li&gt;one install per harness path, with shared paths written once&lt;/li&gt;
&lt;li&gt;a lockfile with a content hash per skill&lt;/li&gt;
&lt;li&gt;
&lt;p&gt;snapshots, so rollback is possible after a bad upstream change&lt;/p&gt;

&lt;p&gt;uvx skillporter add anthropics/skills --skill pdf --harness all&lt;/p&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Real output from a clean workspace:&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;+ Installed pdf across 19 harness(es) (+12 created, ~0 updated, =7 deduplicated/unchanged)
  + Claude Code          created    (copy via .claude/skills)
  + GitHub Copilot       created    (copy via .github/skills)
  + Hermes Agent         created    (copy via .hermes/skills)
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;

&lt;p&gt;Twelve directories, seven shared paths deduplicated. That tally is the whole&lt;br&gt;
argument for a resolver: a copy script would have written nineteen copies and&lt;br&gt;
called it success.&lt;/p&gt;

&lt;h2&gt;
  
  
  Drift is the thing worth detecting
&lt;/h2&gt;

&lt;p&gt;Once a skill is pinned, the interesting question stops being what is installed and&lt;br&gt;
becomes what changed.&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;skillporter sync          # reconcile workspace with manifest, write the lock
skillporter lock --check  # exit 1 on drift, exit 2 if the check could not run
skillporter outdated      # compare version, commit and tree hash upstream
skillporter rollback pdf  # restore the last snapshot
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;

&lt;p&gt;The exit code split matters more than it looks. A pipeline needs to distinguish a&lt;br&gt;
failed check from a check that never ran, otherwise a broken tool reads as a pass.&lt;/p&gt;

&lt;h2&gt;
  
  
  Skills are a supply chain input, so be honest about scanning
&lt;/h2&gt;

&lt;p&gt;A static scanner finds known dangerous patterns: curl piped to shell, base64&lt;br&gt;
payloads, credential reads, risky permission requests. It reports finding codes&lt;br&gt;
and a risk score, and it never executes skill code during install.&lt;/p&gt;

&lt;p&gt;What it cannot do is catch novel prompt injection written in ordinary prose. A URL&lt;br&gt;
can serve harmless content when inspected and something else when the agent fetches&lt;br&gt;
it later. If a vendor tells you their scanner proves a skill is safe at runtime,&lt;br&gt;
they are selling you certainty they do not have. The docs for this one say so&lt;br&gt;
explicitly, which is the main reason I trust the rest of them.&lt;/p&gt;

&lt;h2&gt;
  
  
  Read-only by default when a model is the caller
&lt;/h2&gt;

&lt;p&gt;The MCP mode is worth a paragraph. An MCP server is designed to be called by a&lt;br&gt;
model, which means its arguments may be hostile. So the server exposes detect,&lt;br&gt;
diagnose, scan, inspect and matrix, and denies add, install, sync and remove unless&lt;br&gt;
started with --allow-writes. The workspace root comes from a flag or an environment&lt;br&gt;
variable, not from tool arguments, and traversal, escaping symlinks and UNC paths&lt;br&gt;
are rejected.&lt;/p&gt;

&lt;h2&gt;
  
  
  What this is not
&lt;/h2&gt;

&lt;p&gt;Not a marketplace. Not a sandbox. It does not run skills, and it cannot tell you&lt;br&gt;
what a skill will do when a model follows it. What it does is make the set of&lt;br&gt;
skills deterministic, visible and reversible.&lt;/p&gt;

&lt;h2&gt;
  
  
  Status, honestly
&lt;/h2&gt;

&lt;p&gt;v3.2.0, MIT, Python 3.9+, zero runtime dependencies, single maintainer. Nineteen&lt;br&gt;
adapters, of which 1 is runtime verified, 3 are integration tested and 15 are&lt;br&gt;
fixture tested. The docs state the tier per harness because pretending otherwise&lt;br&gt;
would be the fastest way to lose the trust of the people who hit the bugs.&lt;/p&gt;

&lt;p&gt;Install, if you want to try it in a throwaway repo:&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;uvx skillporter demo
uvx skillporter diagnose .
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;

&lt;p&gt;Site: &lt;a href="https://skillporter.dev" rel="noopener noreferrer"&gt;https://skillporter.dev&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Repo: &lt;a href="https://github.com/Shanmukh-l33t/skillporter" rel="noopener noreferrer"&gt;https://github.com/Shanmukh-l33t/skillporter&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;If a harness path is wrong on your machine, that is the bug report I want most.&lt;/p&gt;

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