From a849556a33105f0e2d8d491d38c04a5db45c3f8d Mon Sep 17 00:00:00 2001
From: Jean-Pierre Martineau <johnputer@meshterm.net>
Date: Wed, 9 Sep 2026 02:43:07 -0400
Subject: [PATCH] A release with notes should not publish 'No changelog entry'
 for losing a race

---
 .claude/skills/release/SKILL.md      | 16 +++++++++++++---
 .github/workflows/github-release.yml | 27 ++++++++++++++++-----------
 2 files changed, 29 insertions(+), 14 deletions(-)

diff --git a/.claude/skills/release/SKILL.md b/.claude/skills/release/SKILL.md
index 032d7f2..74733e1 100644
--- a/.claude/skills/release/SKILL.md
+++ b/.claude/skills/release/SKILL.md
@@ -121,9 +121,19 @@ gh run watch --exit-status
 gh release view vX.Y.Z
 ```
 
-Five platform builds plus the release job, so it takes a few minutes. Report the release
-URL and whether the notes came out of the changelog or fell back to "No changelog entry".
-If a build fails, the tag is already public: fix forward with a new patch version rather
+Five platform builds plus the release job, so it takes a few minutes. **Check the notes
+actually came out of the changelog** — a green run is not proof, because the fallback is a
+successful step:
+
+```
+gh release view vX.Y.Z --json body --jq .body | head -3
+```
+
+If that reads "No changelog entry for X.Y.Z.", the extraction failed rather than the
+changelog being empty. Fix the workflow, then repair the published notes in place with
+`gh release edit vX.Y.Z --notes-file` — the binaries are fine and the tag does not move.
+
+If a *build* fails, the tag is already public: fix forward with a new patch version rather
 than deleting and re-pushing a tag people may have fetched.
 
 ## At 0.9.0 the changelog resets
diff --git a/.github/workflows/github-release.yml b/.github/workflows/github-release.yml
index 37f9c48..6c68ab6 100644
--- a/.github/workflows/github-release.yml
+++ b/.github/workflows/github-release.yml
@@ -74,17 +74,22 @@ jobs:
         run: |
           # The section for this version, between its own heading and the next one. If the
           # changelog has nothing for it, say so plainly rather than inventing notes.
-          if awk -v v="$VERSION" '
-                $0 ~ "^## \\[" v "\\]" {found=1; next}
-                found && /^## \[/ {exit}
-                found {print}
-              ' CHANGELOG.md | sed '/^$/d' | head -1 | grep -q .; then
-            awk -v v="$VERSION" '
-                $0 ~ "^## \\[" v "\\]" {found=1; next}
-                found && /^## \[/ {exit}
-                found {print}
-              ' CHANGELOG.md > notes.md
-          else
+          #
+          # Extracted once, into the file, and the *file* is what gets tested. It used to
+          # run twice — once down a `| sed | head -1 | grep -q .` pipeline to decide, then
+          # again for real — and this step runs under `bash -e -o pipefail`, where that
+          # pipeline is a race it can lose. `head -1` exits the moment it has its line and
+          # closes the pipe; if `sed` still has bytes to write it dies on SIGPIPE, pipefail
+          # promotes that to a failed condition, and a version with a perfectly good
+          # changelog section publishes "No changelog entry" instead. It only bites once a
+          # section outgrows one buffered write, which is why every entry under 4 KB was
+          # fine and 0.3.1 was not — while 0.3.0, the same size, happened to win.
+          awk -v v="$VERSION" '
+              $0 ~ "^## \\[" v "\\]" {found=1; next}
+              found && /^## \[/ {exit}
+              found {print}
+            ' CHANGELOG.md > notes.md
+          if ! grep -q '[^[:space:]]' notes.md; then
             echo "No changelog entry for $VERSION." > notes.md
           fi
           echo "--- notes ---"; cat notes.md