ci: only read the Depends on lines, and let the label be the gate

scripts/check-upstream-deps.sh

@@ -5,15 +5,26 @@
 # Usage: ./scripts/check-upstream-deps.sh <pull request json>
 #
 # The json is what the gitea api returns for a pull request; the script reads
-# the description and the labels from it. A dependency is declared in the
-# description as
+# the description and the labels from it. A dependency is declared on a line of
+# its own in the description:
 #
 #   Depends on: gitea#35851
 #   Depends on: https://github.com/go-gitea/gitea/pull/35851
 #
-# and may appear several times. `GITHUB_COM_TOKEN` is optional and only lifts
-# the rate limit of the unauthenticated github api, `UPSTREAM_LABEL` overrides
-# the label the check looks for.
+# and may be declared several times. Only such lines count, so a pull request
+# that merely mentions gitea#35851 in prose is not held back.
+#
+# The `depends-on-upstream` label is what blocks the merge: it fails the check
+# unless every declared dependency is verified as merged. Without the label, a
+# dependency that is known to be unmerged still fails, so a forgotten label does
+# not let an unreleased feature through.
+#
+# Exit codes: 0 nothing blocks, 1 blocked, 2 usage.
+#
+# Environment:
+#   GITHUB_COM_TOKEN   optional, lifts the rate limit of the github api
+#   UPSTREAM_LABEL     label to look for, `depends-on-upstream` by default
+#   REQUIRE_VERIFIED   set to 1 to also fail when the state cannot be determined
 
 set -euo pipefail
 
@@ -30,8 +41,16 @@ body="$(jq -r '.body // ""' "$PULL_JSON")"
 # argument cannot be named after it; `(.labels // [])` keeps it parsable there
 labelled="$(jq -r --arg want "$LABEL" '[(.labels // [])[].name] | index($want) != null' "$PULL_JSON")"
 
-# every gitea#<id> and every pull request url, deduplicated
+require_verified="${REQUIRE_VERIFIED:-0}"
+if [ "$labelled" = 'true' ]; then
+  # the label says the pull request is not ready, so an unverifiable dependency
+  # keeps it blocked instead of letting it through
+  require_verified=1
+fi
+
+# the ids of the `Depends on:` lines, deduplicated
 refs="$(printf '%s\n' "$body" \
+  | grep -iE '^[[:space:]]*depends[ _-]?on[[:space:]]*:' \
   | grep -oiE 'gitea#[0-9]+|github\.com/go-gitea/gitea/pull/[0-9]+' \
   | grep -oE '[0-9]+' \
   | sort -un || true)"
@@ -58,7 +77,8 @@ github_api() {
   fi
 }
 
-status=0
+unmerged=0
+unknown=0
 for ref in $refs; do
   response="$(github_api "https://api.github.com/repos/go-gitea/gitea/pulls/$ref")"
   code="$(printf '%s' "$response" | tail -n 1)"
@@ -70,34 +90,53 @@ for ref in $refs; do
         echo "gitea#$ref is merged"
       else
         echo "gitea#$ref is not merged yet ($(printf '%s' "$payload" | jq -r '.state'))"
-        status=1
+        unmerged=1
       fi
       ;;
     404)
       echo "gitea#$ref is not a pull request of go-gitea/gitea, reference the pull request that adds the feature"
-      status=1
+      unmerged=1
       ;;
     403 | 429)
-      echo "the github api refused the request for gitea#$ref (rate limit?), cannot tell whether it is merged"
-      status=1
+      echo "gitea#$ref: the github api rejected the request (rate limit), cannot tell whether it is merged"
+      unknown=1
       ;;
     *)
-      echo "unexpected response $code from the github api for gitea#$ref"
-      status=1
+      echo "gitea#$ref: unexpected response $code from the github api, cannot tell whether it is merged"
+      unknown=1
       ;;
   esac
 done
 
-if [ "$status" -ne 0 ]; then
+if [ "$unknown" -eq 1 ] && [ -z "${GITHUB_COM_TOKEN:-}" ]; then
   echo
-  echo "The documentation of an unreleased feature can only be merged once the"
-  echo "feature itself is in gitea main. Keep the \"$LABEL\" label on this pull"
-  echo "request; the scheduled job removes it and reruns this check as soon as"
-  echo "every referenced pull request is merged."
+  echo "The github api allows 60 unauthenticated requests per hour and address,"
+  echo "which the runners share. Set the GITHUB_COM_TOKEN secret to a token with"
+  echo "no scopes to raise that limit."
+fi
+
+if [ "$unmerged" -eq 1 ] || { [ "$unknown" -eq 1 ] && [ "$require_verified" = '1' ]; }; then
+  echo
+  if [ "$labelled" = 'true' ]; then
+    echo "This pull request is labelled \"$LABEL\", so it stays blocked until every"
+    echo "dependency above is merged into gitea main. The scheduled job removes the"
+    echo "label and reruns this check as soon as that is the case."
+  else
+    echo "The documentation of an unreleased feature can only be merged once the"
+    echo "feature itself is in gitea main. Add the \"$LABEL\" label so the pull"
+    echo "request is picked up by the job that watches for it."
+  fi
   exit 1
 fi
 
+if [ "$unknown" -eq 1 ]; then
+  echo
+  echo "Could not verify every dependency, letting the check pass: the pull request"
+  echo "does not carry the \"$LABEL\" label."
+  exit 0
+fi
+
 if [ "$labelled" = 'true' ]; then
   echo
-  echo "every referenced pull request is merged, the \"$LABEL\" label can be removed"
+  echo "every declared dependency is merged, the \"$LABEL\" label can be removed"
 fi

scripts/unlabel-upstream-deps.sh

@@ -54,7 +54,9 @@ for index in $indexes; do
   # the check script exits 0 when every reference is merged; it needs the label
   # to be absent from its input to stay quiet about it, so drop it here
   jq 'del(.labels)' "$work/pull.json" > "$work/pull-unlabelled.json"
-  if ! output="$("$here/check-upstream-deps.sh" "$work/pull-unlabelled.json" 2>&1)"; then
+  # REQUIRE_VERIFIED keeps a dependency whose state could not be determined from
+  # counting as merged, which would drop the label too early
+  if ! output="$(REQUIRE_VERIFIED=1 "$here/check-upstream-deps.sh" "$work/pull-unlabelled.json" 2>&1)"; then
     echo "#$index still waits for upstream:"
     printf '%s\n' "$output" | sed 's/^/    /'
     continue