mirror of
https://github.com/alexbelgium/hassio-addons.git
synced 2026-09-01 01:33:32 +02:00
a68ef0dde2
* fix(skill): stop the workflow scripts reporting verdicts they have not established
Every defect here is the same species: a check printed as passed that never ran, or
never proved what it claims. All six were reproduced before and after.
validate.sh
- `run`/`finish` were selected by filename and fed to bash -n and shellcheck. 25 of the
97 such files here are `#!/usr/bin/execlineb`, so 21 add-ons -- calibre_web and seerr
among them -- reported `local validation FAILED` and a wall of parse errors no matter
what the diff contained. One list, filtered by shebang, now feeds both checks; seerr
went from 15 shellcheck findings and two bash -n failures to the single finding the
test diff actually introduced.
- `grep -q "$ADDON/CHANGELOG.md"` was unanchored with `.` as a wildcard, so a diff
bumping zzz_archived_overseerr/CHANGELOG.md reported seerr's as updated -- a false
green on the one hard CI gate. Five such name collisions exist in this repo
(also birdnet-pi/battybirdnet-pi, mealie/social_to_mealie, plex/spotify_to_plex).
`grep -Fxq`.
- The --vs-master section skipped any file absent from origin/master, so a newly added
script's findings -- all of which are by definition added by the diff -- were never
reported, under a line reading "your diff introduced no new lint findings". An added
file now compares against an empty base. A deleted one is skipped: it was being linted
at a path that no longer exists, which turned every deletion into a fabricated
`openBinaryFile: does not exist` finding.
- That loop ran as the right-hand side of a pipe, so it could not have reached `fail`
even had it tried. It now runs in this shell and new findings fail the script; the
all-clear line is printed only when nothing was listed. Findings that merely moved
lines still cancel -- the comparison strips file:line:col before comm.
- `bash -n ok` stood for an add-on with no shell files at all, and hadolint could print
`clean` directly after printing findings (`A && {...} || C` with pipefail). Counted
and branched properly.
preflight.sh
- `MATCH -- this checkout corresponds to the running image` was concluded from
config.yaml's version equalling $BUILD_VERSION. Version is bumped once per PR, so any
later commit or a dirty tree matches while differing from what runs -- the one
conclusion the script exists to establish was the one it overstated.
pr_review.sh
- `watch` exhausting its minutes with checks still pending fell out of the loop and
exited 0, reporting success for checks that never settled. Unsettled is now exit 2.
Checks reported as `skipping` still count as passing, which is correct -- for this PR
itself, three jobs skipped because no */config.* changed, and that is the right
outcome, not a failure. But a skipped job tested nothing, so `watch` now says so.
Reviewed by Codex (gpt-5.6-sol), which corrected two claims in the audit behind this:
the .templates CHANGELOG assertion (CI skips the gate entirely for a template-only PR,
so that fix is not in this diff) and a tradeoff that did not exist. The deleted-file and
watch-timeout defects are its finds.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(skill): address review — cwd independence, no pass verdict for an empty check
---------
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>