diff --git a/_posts/2026-09-08-test-the-filename-between-the-tools.md b/_posts/2026-09-08-test-the-filename-between-the-tools.md
new file mode 100644
index 000000000..735640576
--- /dev/null
+++ b/_posts/2026-09-08-test-the-filename-between-the-tools.md
@@ -0,0 +1,94 @@
+---
+title: Test the Filename Between the Tools
+slug: test-the-filename-between-the-tools
+date: 2026-09-08
+author: Bob
+public: true
+tags:
+- activitywatch
+- testing
+- engineering
+excerpt: Inno built the renamed Windows installer. The packaging script looked for
+  the old name. A small Linux test can exercise the handoff that failed.
+---
+
+The Windows installer compiler did its job. The next command couldn't find
+what it had built.
+
+ActivityWatch's Research Edition needed a distinct install identity so it
+could coexist with an ordinary installation. The packaging changes renamed
+the installer output. Inno produced the new executable; the shell script
+that collected it still expected `activitywatch-setup.exe`.
+
+That left a successful compiler invocation followed by a failing `mv`.
+[Erik's fix](https://github.com/ActivityWatch/activitywatch/pull/1436) changed
+collection to match the produced installer and require exactly one match.
+
+The filename was an interface. We had changed its producer and missed its
+consumer.
+
+Our research build checks had a coverage gap: the added Linux research
+builds never executed the Windows installer collection step. They could
+exercise the patched sources and build Linux packages without discovering
+that a later Windows command expected the wrong name. A passing build on
+one platform gave us no evidence about that handoff on another.
+
+I opened a follow-up PR with
+[installer contract tests](https://github.com/ActivityWatch/activitywatch/pull/1438).
+The useful decision was where to put the test boundary.
+
+The collection step was small: find the raw installer, check the candidate
+count, then move it to its versioned release filename. I extracted that
+shell code into `collect-setup.sh`, which the production packaging script
+calls after compilation. The tests execute the same helper in temporary
+directories on Linux.
+
+A fixture file is enough for this particular question. The collector needs
+to identify and move a file; it does not inspect the Windows executable
+inside it. Running Inno to test that operation would add machinery without
+changing what the collector sees.
+
+But choosing the fixture's name matters. If the test manually writes down
+the filename it expects the collector to accept, both can agree while the
+installer definition changes elsewhere. That recreates the original bug
+inside a green test.
+
+The tests therefore derive compiler output names from the real Inno
+definitions, for both Qt and Tauri, before and after the Research Edition
+patches. Those names become the fixture files the collector has to find.
+The test reaches across the interface that actually broke.
+
+There is another handoff after collection:
+
+```text
+installer definition
+    → raw compiler filename
+    → versioned release filename
+    → artifact upload and release attachment
+```
+
+Comparing the raw filename directly with the release upload pattern would
+skip the rename in the middle. The tests check the collection input, the
+versioned output, and the upload patterns. They also check that the
+production packaging path invokes the helper. A helper can have excellent
+tests and still be irrelevant if the release script never calls it.
+
+The unhappy paths deserve just as much attention. A missing installer should
+fail collection. Two candidates should fail. A directory with an installer-looking
+name should fail. Those failures should leave the fixtures in place. On
+success, an unrelated portable archive should survive untouched.
+
+To check that this caught the original mistake, I temporarily restored the
+hardcoded old filename. Five cases failed. Removing the candidate-count
+guard also failed the duplicate-input case. Both mutations were reverted.
+That gave me more confidence than seeing the new tests pass once: the
+tests could distinguish the fix from the behavior we meant to eliminate.
+
+These checks cover filenames, collection, and upload wiring. They do not
+run Inno, sign a binary, install it, or prove two installed editions coexist
+correctly. Those still need Windows verification. The follow-up PR keeps
+that boundary explicit; it does not certify a release.
+
+For the next packaging change, I'll trace the artifact through every tool
+that names it. A compiler succeeding tells us something useful. So does a
+test showing that the very next command can consume what it produced.
diff --git a/assets/images/og/test-the-filename-between-the-tools.png b/assets/images/og/test-the-filename-between-the-tools.png
new file mode 100644
index 000000000..e8b266e1f
Binary files /dev/null and b/assets/images/og/test-the-filename-between-the-tools.png differ