fix(ci): auto-release never detects feat/fix once the log exceeds the pipe buffer

[Weegy]

Problem

The release pipeline has not cut a release since v0.56.0 (2026-06-30), despite 19 feat:/fix: commits landing on main. Every auto-release run since has succeeded while logging No feat/fix commits since v0.56.0 — skipping release, which also skipped publish-images and desktop-apps (no images, no desktop installers for a month).

Root cause

The bump detection piped the entire git log into grep -q:

elif printf '%s\n' "$msgs" | grep -qE '^feat(\(.+\))?:'; then

grep -q exits at the first match. If printf still has data buffered at that point, it gets EPIPE. Under set -o pipefail the pipeline's exit status is that failed writer, not grep's 0 — so the condition evaluates as no match precisely because there was a match.

Below ~64 KiB (the pipe buffer) printf completes before grep exits, so the pipeline is correct and the workflow worked for months. Past that threshold it inverts. And it wedges permanently rather than intermittently: no release → no new tag → a longer commit range on the next run.

The tell is visible in the log of run 30552280319 — printf: write error: Broken pipe on script lines 13 and 15, which map to exactly the feat and fix grep lines:

latest tag: v0.56.0 (range: v0.56.0..HEAD)
...: line 13: printf: write error: Broken pipe
...: line 15: printf: write error: Broken pipe
No feat/fix commits since v0.56.0 — skipping release.

Reproduced deterministically against the real range (655 KiB of log): bump=[] with 19 matching commits present.

Fix

• Write the log to a file and let grep read the file. A file read cannot EPIPE, so the pipefail interaction disappears at the source rather than being papered over (e.g. by dropping pipefail or adding || true, both of which would mask genuine failures).
• Report the range's commit count on the skip path. No feat/fix among 0 commits is a normal skip; among 19 commits is visibly wrong. This class of bug was silent for a month because the message was plausible either way.

.github/workflows/auto-release.yml was the only occurrence of the pattern in CI. The | grep -q uses in desktop-apps.yml read tiny codesign/file output and are unaffected.

Verification

Replayed the real v0.56.0..main range with the fixed script, mutations (git tag/git push/gh release create) stubbed:

latest tag: v0.56.0 (range: v0.56.0..origin/main)
Releasing v0.57.0 (bump=minor)
STUB git tag v0.57.0
STUB gh release create v0.57.0
GITHUB_OUTPUT -> new_tag=v0.57.0

• generate-changelog.mjs notes v0.56.0 main exits 0 and produces 39 lines across Added / Changed / Fixed.
• new_tag is set, so publish-images and desktop-apps are no longer skipped.

Note on merging

Merging this PR is itself a push to main, so it will trigger auto-release and cut v0.57.0 covering the full month of accumulated commits — one release with the complete backlog rather than the 19 individual ones that were missed. Images and desktop installers will build for it.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}