paul/audi-app
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Signierskript richtet den Klon selbst aus - kein Handgriff mehr noetig

Browse Source 
Der Waechter brach bisher ab und legte dem Menschen einen Befehl vor. Jetzt
bringt er den Klon selbst auf origin/main, solange das nachweislich verlustfrei
ist: reiner Rueckstand wird vorgespult; Commits, die nur lokal liegen, deren
Tree aber dem des gemeinsamen Vorfahren gleicht, tragen inhaltlich nichts bei
und werden verworfen (Reflog haelt sie 90 Tage). Nur Commits, die auch den
Inhalt aendern, fuehren weiterhin zum Abbruch mit beiden Wegen zur Wahl.

Damit entfaellt die Vorbereitung vor einem Signierlauf vollstaendig - wichtig,
weil die bisher dokumentierte Vorbereitung ausgerechnet `git pull` war, also
genau der Befehl, der auf einem Klon von vor dem Rewrite die entfernten Commits
zurueckgeholt haette.

Fuenf Lagen gegen Wegwerf-Repositories geprueft: gleich, zurueck, inhaltslos
voraus, echte ungepushte Arbeit, echte Divergenz.

APPLE_DEV_BACKLOG.md, OFFEN.md und AGENTS.md (Abschnitt AJ) nachgezogen.

Co-Authored-By: Claude Opus 5 <[email protected]>
This commit is contained in:
tobias
2026-09-07 22:01:32 +02:00
co-authored by Claude Opus 5
parent 185a044c5e
commit 5b840ea510
4 changed files with 78 additions and 57 deletions
Download Patch File Download Diff File Expand all files Collapse all files
AGENTS.md
+23 -11
View File
@@ -4916,18 +4916,30 @@ staleness bug this project keeps fixing elsewhere (cache-busting, OTA-bundle-ver
tree is clean and `HEAD` is exactly `origin/main` (fetches first, aborts with a clear message
otherwise) — added right after `cd "$APP"`, before the web build.
**Nachtrag 2026-09-07: the guard now distinguishes two cases, and the reason is a real
trap.** On that day `main`'s history was rewritten once - three commits at the tip, all of
which contained nothing but a document that has since moved to its own repository, were
**Nachtrag 2026-09-07: the guard repairs the clone itself where that is provably
lossless.** On that day `main`'s history was rewritten once - three commits at the tip, all
of which contained nothing but a document that has since moved to its own repository, were
removed and the branch force-pushed back to `0e84248`. **Every clone made before that date
is now *ahead* of `origin/main`, not behind it**, so `git pull` there is exactly wrong: the
merge pulls the removed commits back into the history, and the next push restores them on
the server - without even needing `--force`. Git cannot tell that apart from genuine
unpushed work, so the script does not decide: it lists the commits that exist only locally,
names both possibilities (push them, or `git fetch origin && git reset --hard origin/main`)
and aborts. Only the plain behind-case still recommends `git pull`. Both branches verified
against throwaway repositories built to each shape, not just parsed - the first version of
this fix classified the situation as "diverged" and would have told the Mac to push.
holds commits `origin/main` does not**, so `git pull` there is exactly wrong: the merge
pulls the removed commits back into the history, and the next push restores them on the
server - without even needing `--force`.
Git cannot tell that apart from genuine unpushed work by the commit graph alone, but it can
by the **trees**: if `HEAD`'s tree equals the merge-base's tree, the local-only commits
contribute no content at all, and discarding them cannot lose a single file. That is the
test the script now runs (`git diff --quiet $(git merge-base …) HEAD`). In that case it
resets to `origin/main` itself, prints what it discarded, and carries on - the clean-tree
check one step earlier guarantees no uncommitted work is in the way, and the reflog keeps
the commits for 90 days. The same reasoning covers a clone that is merely **behind**: the
script fast-forwards it rather than demanding a `git pull`, so the documented preparation
for a signing run is now *nothing at all* - which matters, because telling a human to run
`git pull` is precisely the instruction that would have re-created the removed commits on
the wrong clone. Only when the local commits *do* change content does it abort and leave
the decision to the human, naming both routes and warning off `git pull`.
All five shapes (equal, behind, content-free ahead, real unpushed work, genuine divergence)
verified against throwaway repositories built to each shape, not just parsed - an earlier version of this fix
classified the situation as "diverged" and would have told the Mac to push.
---