<?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: Michael Williams</title>
    <description>The latest articles on DEV Community by Michael Williams (@true_horison_37).</description>
    <link>https://dev.to/true_horison_37</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%2F4209581%2F1b3782d2-6f40-4686-8c58-4488635d32a7.png</url>
      <title>DEV Community: Michael Williams</title>
      <link>https://dev.to/true_horison_37</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/true_horison_37"/>
    <language>en</language>
    <item>
      <title>About 4% of October's commits in 30 busy repos carry an AI stamp</title>
      <dc:creator>Michael Williams</dc:creator>
      <pubDate>Tue, 17 Nov 2026 12:41:00 +0000</pubDate>
      <link>https://dev.to/true_horison_37/about-4-of-octobers-commits-in-30-busy-repos-carry-an-ai-stamp-237a</link>
      <guid>https://dev.to/true_horison_37/about-4-of-octobers-commits-in-30-busy-repos-carry-an-ai-stamp-237a</guid>
      <description>&lt;p&gt;Out of roughly 7,000 commits pushed to 30 busy open-source repos in October, about 4% carry a stamp saying an AI coding tool helped write them, and almost every one of those stamps comes from Claude Code.&lt;/p&gt;

&lt;p&gt;I picked the repos by hand from well-known, very active projects, and a third of them are AI or agent tools. On each default branch I went through every October commit apart from merges and counted the ones whose message has a trailer like &lt;code&gt;Co-Authored-By: Claude&lt;/code&gt; or &lt;code&gt;Assisted-by: ...&lt;/code&gt;. The counting was done by ai-code-sbom (&lt;a href="https://github.com/TrueHorison/ai-code-sbom" rel="noopener noreferrer"&gt;https://github.com/TrueHorison/ai-code-sbom&lt;/a&gt;), which I built after reading AT&amp;amp;T's supplier terms, which ask for an SBOM naming the AI tool behind each release. Per repo it was roughly:&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 &lt;span class="nt"&gt;--no-checkout&lt;/span&gt; &lt;span class="nt"&gt;--shallow-since&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;2026-09-25 &amp;lt;repo&amp;gt;
python ai_code_sbom.py &lt;span class="nt"&gt;--repo&lt;/span&gt; &amp;lt;repo&amp;gt; &lt;span class="nt"&gt;--since&lt;/span&gt; &amp;lt;last September commit&amp;gt; &lt;span class="nt"&gt;--until&lt;/span&gt; &amp;lt;last October commit&amp;gt; &lt;span class="nt"&gt;--format&lt;/span&gt; md
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;My first list had Textualize/rich and elysia on it, and neither had a single commit in October, so they got swapped out, along with kamal, which didn't have enough history to cover the month.&lt;/p&gt;

&lt;p&gt;Here's where the stamps turned up, plus a few of the zeros:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Repo&lt;/th&gt;
&lt;th&gt;Commits&lt;/th&gt;
&lt;th&gt;AI-stamped&lt;/th&gt;
&lt;th&gt;Share&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;sqlc-dev/sqlc&lt;/td&gt;
&lt;td&gt;12&lt;/td&gt;
&lt;td&gt;6&lt;/td&gt;
&lt;td&gt;50.0%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;crewAIInc/crewAI&lt;/td&gt;
&lt;td&gt;193&lt;/td&gt;
&lt;td&gt;76&lt;/td&gt;
&lt;td&gt;39.4%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;compiler-explorer/compiler-explorer&lt;/td&gt;
&lt;td&gt;174&lt;/td&gt;
&lt;td&gt;61&lt;/td&gt;
&lt;td&gt;35.1%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;modelcontextprotocol/typescript-sdk&lt;/td&gt;
&lt;td&gt;38&lt;/td&gt;
&lt;td&gt;6&lt;/td&gt;
&lt;td&gt;15.8%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;biomejs/biome&lt;/td&gt;
&lt;td&gt;407&lt;/td&gt;
&lt;td&gt;48&lt;/td&gt;
&lt;td&gt;11.8%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;conventional-changelog/commitlint&lt;/td&gt;
&lt;td&gt;54&lt;/td&gt;
&lt;td&gt;4&lt;/td&gt;
&lt;td&gt;7.4%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;pydantic/pydantic-ai&lt;/td&gt;
&lt;td&gt;1357&lt;/td&gt;
&lt;td&gt;48&lt;/td&gt;
&lt;td&gt;3.5%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;bytecodealliance/wasmtime&lt;/td&gt;
&lt;td&gt;375&lt;/td&gt;
&lt;td&gt;13&lt;/td&gt;
&lt;td&gt;3.5%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;openai/codex&lt;/td&gt;
&lt;td&gt;1282&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;0.0%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;anthropics/claude-code&lt;/td&gt;
&lt;td&gt;56&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;0.0%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;sveltejs/svelte&lt;/td&gt;
&lt;td&gt;40&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;0.0%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;django/django&lt;/td&gt;
&lt;td&gt;117&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;0.0%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;tokio-rs/tokio&lt;/td&gt;
&lt;td&gt;129&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;0.0%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;All 30&lt;/td&gt;
&lt;td&gt;7049&lt;/td&gt;
&lt;td&gt;275&lt;/td&gt;
&lt;td&gt;3.9%&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;crewAI was close to 40%, and sqlc only beats it because it had 12 commits all month. The top three repos between them hold about two-thirds of every stamp I found, and if you take them out, the other 27 repos come to well under 2%.&lt;/p&gt;

&lt;p&gt;Half the repos had no stamps at all. Svelte had none in 40 commits and django none in over a hundred, tokio was the same, and openai/codex pushed close to 1,300 commits in October without one. anthropics/claude-code itself had 56 commits and not one Claude Code stamp, which I'd guess (without checking) is squash merges.&lt;/p&gt;

&lt;p&gt;Nearly all of it is Claude Code, 264 of the 275 stamped commits, and the rest is a handful of Cursor, GitHub's Copilot coding agent, and &lt;code&gt;Assisted-by&lt;/code&gt; lines naming GPT or Gemini. A lot of that is defaults: Claude Code adds its co-author line unless you turn it off, so it was always going to win a count of stamps.&lt;/p&gt;

&lt;p&gt;So 4% is a floor. A stamp only exists if the tool writes one and the person leaves it in, and plenty of tools and people don't. Squash merges and rebases that rewrite the message drop the trailers as well, so a repo that squashes pull requests into one clean message can show zero even when the branch was full of AI commits. The sample isn't random either, and it probably leans towards projects that are relaxed about AI tools.&lt;/p&gt;

&lt;h2&gt;
  
  
  Projects should keep the stamps
&lt;/h2&gt;

&lt;p&gt;It's one line per commit, and it's the cheapest record of AI use there is that doesn't rely on someone squinting at the code and guessing. So I think projects should keep these lines, and the ones that squash-merge should carry the trailers into the squash message instead of letting them vanish.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Written with AI help.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>opensource</category>
      <category>git</category>
      <category>claudecode</category>
    </item>
    <item>
      <title>Every one of my 400k+ Claude Code calls came from a long main session</title>
      <dc:creator>Michael Williams</dc:creator>
      <pubDate>Wed, 11 Nov 2026 13:08:00 +0000</pubDate>
      <link>https://dev.to/true_horison_37/every-one-of-my-400k-claude-code-calls-came-from-a-long-main-session-3fje</link>
      <guid>https://dev.to/true_horison_37/every-one-of-my-400k-claude-code-calls-came-from-a-long-main-session-3fje</guid>
      <description>&lt;p&gt;Under my last post &lt;a class="mentioned-user" href="https://dev.to/junfengzhou01"&gt;@junfengzhou01&lt;/a&gt; asked whether my 400k+ context calls came from long sessions or from sub-agents reading big files, and I didn't know. It's the sessions. I split the same local logs into main session and sub-agent calls (7 Nov, 9,604 calls at list price) and got this:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;band       main calls  main cost  sub calls  sub cost
&amp;lt;100k             189     $15.72       3096   $121.48
100-200k          468     $39.50       3183   $175.73
200-400k          661     $64.08        764    $61.65
400k+            1243    $243.27          0     $0.00
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;No sub-agent ever got past 365k. Makes sense when you think about it: a sub-agent starts with an empty context and is done before it grows, while my main sessions stay open for days, which is also why your table will probably look nothing like mine if you clear often.&lt;/p&gt;

&lt;p&gt;Five sessions did all of the 400k+ spend, and three of them were 94% of it. Most of that money is cache reads, $160 or so of the $243, with output a small slice. Sub-agents are still half my total bill, but they get there by volume, 7,043 calls and most of them under 200k.&lt;/p&gt;

&lt;p&gt;&lt;a class="mentioned-user" href="https://dev.to/jgravelle"&gt;@jgravelle&lt;/a&gt; wondered in the replies if big file reads were what bloated the main sessions. I tried to measure what filled them and got a mess, because the browser tool's screenshots and page dumps swamp any character count. File reads came out at roughly a tenth. So probably not the main cause, though I wouldn't hang much on that number.&lt;/p&gt;

&lt;h2&gt;
  
  
  Coming back after a break is the expensive part
&lt;/h2&gt;

&lt;p&gt;On Sunday 18 October I opened a session I'd left running overnight. The first reply cost $3.70. Twelve hours later the cache was long gone, so it had to write 740k tokens back in before doing anything.&lt;/p&gt;

&lt;p&gt;21 of the 400k+ calls re-cached more than 100k tokens after a pause like that, and together they came to $56, close to a quarter of the 400k+ spend.&lt;/p&gt;

&lt;p&gt;Next time a session is past 400k and I'm about to walk away, I'll have Claude write a notes file and start fresh from that the next day. Caching a 20k-token note costs around 10 cents. I don't have real numbers on the saving yet.&lt;/p&gt;

&lt;p&gt;The numbers from the first post held up: 400k+ is now 13% of calls and 34% of cost.&lt;/p&gt;

&lt;p&gt;ctxtax shows the main vs sub-agent split per band now, if you want to run it on your own logs: &lt;a href="https://github.com/TrueHorison/ctxtax" rel="noopener noreferrer"&gt;https://github.com/TrueHorison/ctxtax&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Written with AI help.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>claudecode</category>
      <category>productivity</category>
      <category>discuss</category>
    </item>
    <item>
      <title>AT&amp;T wants your SBOM to name the AI that wrote the code, so I tried to produce one from git</title>
      <dc:creator>Michael Williams</dc:creator>
      <pubDate>Tue, 03 Nov 2026 13:37:00 +0000</pubDate>
      <link>https://dev.to/true_horison_37/att-wants-your-sbom-to-name-the-ai-that-wrote-the-code-i-tried-to-produce-one-55jg</link>
      <guid>https://dev.to/true_horison_37/att-wants-your-sbom-to-name-the-ai-that-wrote-the-code-i-tried-to-produce-one-55jg</guid>
      <description>&lt;p&gt;AT&amp;amp;T's supplier AI terms (v2.1, October 2025) have a clause aimed squarely at code. If a supplier's code was produced with an AI tool, the supplier has to submit a software bill of materials "including specification of the AI Solution" with the first delivery and every release after. AI-generated work has to be labelled "AI Generated". Each AI tool needs approval first, and if a supplier uses one that was not approved, AT&amp;amp;T has no obligation to pay for that work.&lt;/p&gt;

&lt;p&gt;AT&amp;amp;T is the only buyer I've found with wording this exact. A copy of Lockheed Martin's subcontract terms asks for "traceability of AI generated content" on request, which is looser. So I tried to answer the AT&amp;amp;T version for a real repo, using only what git already records.&lt;/p&gt;

&lt;p&gt;Most coding agents already stamp their commits. Claude Code adds a &lt;code&gt;Co-Authored-By: Claude&lt;/code&gt; line, and recent versions add a &lt;code&gt;Claude-Session&lt;/code&gt; link. GitHub's Copilot coding agent adds &lt;code&gt;Agent-Logs-Url&lt;/code&gt;, and Cursor and Aider have co-author lines of their own.&lt;/p&gt;

&lt;p&gt;I wrote a script that reads those stamps for every commit in a release and writes the result as a CycloneDX or SPDX file, plus a plain summary. It can also check the result against an allow-list of tools and providers and fail if something else shows up.&lt;/p&gt;

&lt;p&gt;I ran it on a public repo, pyrail, for one release (v0.4.1 to v0.5.0). 8 of 151 commits carried an AI stamp: 5 from Copilot's coding agent and 5 from Claude Code, with two squash merges carrying both. Those commits touched 20 of the 22 files changed across the release, and a "Claude Code only" policy failed and listed the Copilot ones.&lt;/p&gt;

&lt;p&gt;The AI commits added about 2,000 of the 8,350 lines, but that's an upper bound, because a trailer marks a whole commit and can't say which of its lines the agent wrote.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the standards don't have yet
&lt;/h2&gt;

&lt;p&gt;Neither CycloneDX nor SPDX has an agreed field for "this code was written by AI tool X with model Y". I put the tool under CycloneDX &lt;code&gt;formulation&lt;/code&gt;, which describes how the release was made, because top-level &lt;code&gt;services&lt;/code&gt; would claim the product calls the AI at runtime. In SPDX it's a &lt;code&gt;Tool&lt;/code&gt; the package &lt;code&gt;usesTool&lt;/code&gt;. Both validate against the official schemas. The mapping is my own. There is no field at all for the model, for the provider route (Anthropic direct, Bedrock, Vertex), or for how anyone found out the tool was used.&lt;/p&gt;

&lt;p&gt;It only sees AI use that left a trace. Inline editor suggestions leave none, and stamps can be switched off or lost in a squash, so if it finds nothing, that means nothing was stamped, and people could still have used AI. I left out "this looks AI-written" guessing on purpose, because those guesses can't name the tool or the provider, and that is what the clause asks for.&lt;/p&gt;

&lt;p&gt;The script is on GitHub as &lt;a href="https://github.com/TrueHorison/ai-code-sbom" rel="noopener noreferrer"&gt;ai-code-sbom&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;If a client has asked you for something like this, I'd really like to hear what they asked for and how you answered it. There's a &lt;a href="https://docs.google.com/forms/d/e/1FAIpQLSfHLjfFnQ249a2OB8p0qFRYL87-3nQegEcTcPnOnQ6dJpoPrA/viewform" rel="noopener noreferrer"&gt;one-minute form&lt;/a&gt;, or just say so in the comments.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Written with AI help.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>sbom</category>
      <category>git</category>
      <category>opensource</category>
    </item>
    <item>
      <title>I priced every call in my Claude Code logs. 12% of them cost 31%.</title>
      <dc:creator>Michael Williams</dc:creator>
      <pubDate>Wed, 28 Oct 2026 13:12:00 +0000</pubDate>
      <link>https://dev.to/true_horison_37/i-priced-every-call-in-my-claude-code-logs-12-of-them-cost-31-3gdg</link>
      <guid>https://dev.to/true_horison_37/i-priced-every-call-in-my-claude-code-logs-12-of-them-cost-31-3gdg</guid>
      <description>&lt;p&gt;I've been leaning on Claude Code hard for a few weeks, mostly long sessions with lots of sub-agents, and I priced every call in my logs at list price. 954 of my 7,802 calls sent more than 400k tokens of context.&lt;/p&gt;

&lt;p&gt;Claude Code keeps a &lt;code&gt;.jsonl&lt;/code&gt; file per session under &lt;code&gt;~/.claude/projects&lt;/code&gt;, and each API call is logged there with its token counts (input, cache writes, cache reads and output). The same message can show up on several lines, so I counted each message id once. I counted everything a call sent in (input, cache reads and cache writes) as its context, drew the big line at 400k, and priced the lot at Anthropic's list prices as of 26 Oct. Opus 5.5 is $4 in and $20 out per million tokens, with cache reads at $0.20. I'm on a plan, so these are what it would cost on the API.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;band        calls   % calls    cost    % cost
&amp;lt;100k        2787    35.7%   $114.83   20.3%
100-200k     2898    37.1%   $174.04   30.8%
200-400k     1163    14.9%   $101.71   18.0%
400k+         954    12.2%   $174.78   30.9%
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That's $565 in total, and sub-agents were $289 of it, a bit over half. Output tokens barely matter. Nearly all the cost is context being sent again and again, and while cache reads are cheap per token, a session that runs past a few hundred thousand tokens pays for all of them on every turn after.&lt;/p&gt;

&lt;p&gt;I also hit the 5-hour limit 14 times in about two weeks. In the 5 hours before each hit I'd used anywhere from about $19 to $43 of list-price usage, which is far too wide a spread to tell me what the limit actually is.&lt;/p&gt;

&lt;p&gt;What I'd try next, all of it untested so far, is to clear the session when a task is done instead of carrying on, start the next task from a short summary, and use a cheaper model for sub-agents that only read and search.&lt;/p&gt;

&lt;h2&gt;
  
  
  Run it on your own logs
&lt;/h2&gt;

&lt;p&gt;I put the script on GitHub as &lt;a href="https://github.com/TrueHorison/ctxtax" rel="noopener noreferrer"&gt;ctxtax&lt;/a&gt;, and the README says what it reads.&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;-O&lt;/span&gt; https://raw.githubusercontent.com/TrueHorison/ctxtax/main/ctxtax.py
python ctxtax.py
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The usage-limit threads are full of "my limit feels smaller" with no way to compare. &lt;code&gt;python ctxtax.py --meter&lt;/code&gt; prints a line for each limit hit with the day, how much you'd used in the 5 hours before and which models, and it never prints your prompts, so it's fine to paste. If a few people put theirs in the comments, we could finally see whether the limit really moved.&lt;/p&gt;

&lt;p&gt;What share of your cost comes from the 400k+ calls?&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Written with AI help.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>claudecode</category>
      <category>productivity</category>
      <category>discuss</category>
    </item>
  </channel>
</rss>
