<?xml version="1.0" encoding="UTF-8"?>
<feed xmlns="http://www.w3.org/2005/Atom">
  <title>Notes — Wisp</title>
  <subtitle>Write-ups of the traps hit while building Wisp, a desktop AI agent with a VRM body. Measurements and mistakes, not announcements.</subtitle>
  <link href="https://wisp.rengaworks.com/blog/feed.xml" rel="self"/>
  <link href="https://wisp.rengaworks.com/blog/"/>
  <id>https://wisp.rengaworks.com/blog/</id>
  <updated>2026-11-04T00:00:00Z</updated>
  <author><name>Renga</name></author>
  <entry>
    <title>A finished revision can still leave a stale review copy</title>
    <link href="https://wisp.rengaworks.com/blog/review-copy-must-match-revision/"/>
    <id>https://wisp.rengaworks.com/blog/review-copy-must-match-revision/</id>
    <updated>2026-11-04T00:00:00Z</updated>
    <summary>A local article revision and its cloud review copy disagreed. Separate the working original, the intended snapshot, and the text read back from the review location.</summary>
  </entry>
  <entry>
    <title>An unknown note category should stop the write</title>
    <link href="https://wisp.rengaworks.com/blog/unknown-note-category-stops-write/"/>
    <id>https://wisp.rengaworks.com/blog/unknown-note-category-stops-write/</id>
    <updated>2026-10-25T00:00:00Z</updated>
    <summary>A note-saving tool now rejects missing, unknown, or conflicting destinations before creating a directory. What the saved fixture tests establish, and what remains unverified.</summary>
  </entry>
  <entry>
    <title>A generated video still needs a receipt and a stopped runtime</title>
    <link href="https://wisp.rengaworks.com/blog/verify-video-receipt-and-release/"/>
    <id>https://wisp.rengaworks.com/blog/verify-video-receipt-and-release/</id>
    <updated>2026-10-18T00:00:00Z</updated>
    <summary>A stale widget exposed separate completion checks for generation, media validity, byte receipt, engine shutdown, and runtime release.</summary>
  </entry>
  <entry>
    <title>A CSV comparison result expires when its inputs change</title>
    <link href="https://wisp.rengaworks.com/blog/csv-results-expire-with-inputs/"/>
    <id>https://wisp.rengaworks.com/blog/csv-results-expire-with-inputs/</id>
    <updated>2026-10-15T00:00:00Z</updated>
    <summary>Use an input epoch and a per-job token to keep old comparisons and reports out of the current UI.</summary>
  </entry>
  <entry>
    <title>A no-network test needs a request that makes it fail</title>
    <link href="https://wisp.rengaworks.com/blog/no-network-test-needs-negative-control/"/>
    <id>https://wisp.rengaworks.com/blog/no-network-test-needs-negative-control/</id>
    <updated>2026-10-05T00:00:00Z</updated>
    <summary>Separate local Blob loads from HTTP traffic, observe blocked attempts, and check that an intentional fetch fails the test.</summary>
  </entry>
  <entry>
    <title>\"Zero matches\" and \"couldn't read the file\" look identical</title>
    <link href="https://wisp.rengaworks.com/blog/checks-that-cannot-be-loosened/"/>
    <id>https://wisp.rengaworks.com/blog/checks-that-cannot-be-loosened/</id>
    <updated>2026-09-02T00:00:00Z</updated>
    <summary>I wrote the checks before fixing anything, then spent the project finding defects in my own checks. Every one of them looked correct when written.</summary>
  </entry>
  <entry>
    <title>Three independent proposals won't find the error. Adding a refuter killed one of them.</title>
    <link href="https://wisp.rengaworks.com/blog/refuter-agent/"/>
    <id>https://wisp.rengaworks.com/blog/refuter-agent/</id>
    <updated>2026-08-28T00:00:00Z</updated>
    <summary>Five agents: three perspectives, one refuter, one synthesizer. The refuter killed a proposal, found the assumption all three shared, and caught two errors in the data I handed them.</summary>
  </entry>
  <entry>
    <title>I published three posts in one day and two of them were never seen by anyone</title>
    <link href="https://wisp.rengaworks.com/blog/one-post-per-day/"/>
    <id>https://wisp.rengaworks.com/blog/one-post-per-day/</id>
    <updated>2026-08-23T00:00:00Z</updated>
    <summary>Not low views. Zero. A quarter of everything I had written had never been shown to a single reader, and neither tags nor timing explained it.</summary>
  </entry>
  <entry>
    <title>The jitter wasn't in the interpolation. It was in the schedule.</title>
    <link href="https://wisp.rengaworks.com/blog/jitter-is-a-schedule-bug/"/>
    <id>https://wisp.rengaworks.com/blog/jitter-is-a-schedule-bug/</id>
    <updated>2026-08-23T00:00:00Z</updated>
    <summary>One symptom, three unrelated causes, and the order to check them in. Including a threshold-free test for telling a real kink from a smooth curve.</summary>
  </entry>
  <entry>
    <title>My VRM was rejected at upload and every viewer I own said it was fine</title>
    <link href="https://wisp.rengaworks.com/blog/accessor-bounds-drift/"/>
    <id>https://wisp.rengaworks.com/blog/accessor-bounds-drift/</id>
    <updated>2026-08-22T00:00:00Z</updated>
    <summary>The mesh data was correct. The declaration about the mesh data was wrong — and rendering never reads the declaration, so nothing local could see it.</summary>
  </entry>
  <entry>
    <title>I set the font to the largest size and found the same bug eleven times</title>
    <link href="https://wisp.rengaworks.com/blog/xxxl-broke-japanese-worse/"/>
    <id>https://wisp.rengaworks.com/blog/xxxl-broke-japanese-worse/</id>
    <updated>2026-08-16T00:00:00Z</updated>
    <summary>Reading the code found none of them. Screenshots found all eleven — and Japanese broke worse than English, because Japanese can wrap anywhere.</summary>
  </entry>
  <entry>
    <title>My dev environment was hiding eight production bugs</title>
    <link href="https://wisp.rengaworks.com/blog/dev-env-hid-eight-bugs/"/>
    <id>https://wisp.rengaworks.com/blog/dev-env-hid-eight-bugs/</id>
    <updated>2026-08-15T00:00:00Z</updated>
    <summary>Every convenience in my development setup was concealing a defect on the other side of it. Counting the conveniences found the bugs faster than looking for the bugs.</summary>
  </entry>
  <entry>
    <title>My evidence pipeline was saving Cloudflare block pages as evidence</title>
    <link href="https://wisp.rengaworks.com/blog/block-pages-as-evidence/"/>
    <id>https://wisp.rengaworks.com/blog/block-pages-as-evidence/</id>
    <updated>2026-08-15T00:00:00Z</updated>
    <summary>A third-party wrapper died, and the failure came back as a plausible-looking image with an HTTP 200 attached. What broke, and what I changed.</summary>
  </entry>
  <entry>
    <title>I redacted my username from a screen recording three times and missed it every time</title>
    <link href="https://wisp.rengaworks.com/blog/mask-screen-recording/"/>
    <id>https://wisp.rengaworks.com/blog/mask-screen-recording/</id>
    <updated>2026-08-14T00:00:00Z</updated>
    <summary>Fixed rectangles cannot hide a dialog that animates in. I tracked the text with an FFT correlation, missed it twice more, and finally scanned the finished file with a machine — which found one more frame nobody was looking at.</summary>
  </entry>
  <entry>
    <title>My approval dialog grew until the buttons went off screen</title>
    <link href="https://wisp.rengaworks.com/blog/approval-dialog/"/>
    <id>https://wisp.rengaworks.com/blog/approval-dialog/</id>
    <updated>2026-08-14T00:00:00Z</updated>
    <summary>Showing the user everything the agent is about to do, and letting them actually choose, are two different properties. I only had the first one, and it took a demo recording to find out.</summary>
  </entry>
  <entry>
    <title>Where to split a sentence for streaming TTS is decided by one number</title>
    <link href="https://wisp.rengaworks.com/blog/streaming-tts-split-point/"/>
    <id>https://wisp.rengaworks.com/blog/streaming-tts-split-point/</id>
    <updated>2026-08-11T00:00:00Z</updated>
    <summary>The split point for streaming text-to-speech follows from the real-time factor. I measured that factor on a machine with a full disk, got it 4x wrong, and shipped a constant that was never going to fail loudly enough to notice.</summary>
  </entry>
</feed>
