ci: prevent regenerate-models from opening duplicate empty PRs

.github/workflows/manual_regenerate_models.yaml

@@ -63,6 +63,45 @@ jobs:
         with:
           token: ${{ secrets.APIFY_SERVICE_ACCOUNT_GITHUB_TOKEN }}
 
+      # Record master's SHA up front: a later step checks out the auto-update branch, so the change check below
+      # can't compare against HEAD. Resolve master from the remote rather than the checked-out ref: the run may be
+      # dispatched from a non-master ref, but that check and the PR both target master, so master is the baseline.
+      - name: Record master ref
+        id: base
+        run: |
+          git fetch --depth=1 origin master
+          echo "sha=$(git rev-parse FETCH_HEAD)" >> "$GITHUB_OUTPUT"
+
+      # Does the auto-update branch already exist on the remote? If so, a previous dispatch opened the PR
+      # and we append to it. If not, we start it from master (below) so signed-commit creates it there.
+      - name: Determine auto-update branch state
+        id: branch
+        run: |
+          if git ls-remote --exit-code --heads origin "$BRANCH" >/dev/null 2>&1; then
+            echo "exists=true" >> "$GITHUB_OUTPUT"
+            echo "create=false" >> "$GITHUB_OUTPUT"
+          else
+            echo "exists=false" >> "$GITHUB_OUTPUT"
+            echo "create=true" >> "$GITHUB_OUTPUT"
+          fi
+
+      # Check out the existing branch before regenerating so the new models land as a NEW commit on top,
+      # mirroring the commit just pushed to the docs PR. We must switch now, while the tree is clean:
+      # signed-commit's own checkout is a plain `git checkout` that would refuse to overwrite the
+      # regenerated files (which differ between master and the branch) once they're in the working tree.
+      - name: Check out existing auto-update branch
+        if: steps.branch.outputs.exists == 'true'
+        run: |
+          git fetch --depth=1 origin "$BRANCH"
+          git checkout -B "$BRANCH" FETCH_HEAD
+
+      # The branch doesn't exist yet, so regenerate from the recorded master SHA rather than the dispatched
+      # ref (a manual run may be dispatched from a non-master ref). Regenerating on master keeps the codegen
+      # tooling current, and signed-commit then creates the branch (the PR head) on top of master, the PR base.
+      - name: Start the auto-update branch from master
+        if: steps.branch.outputs.exists == 'false'
+        run: git checkout "${{ steps.base.outputs.sha }}"
+
       # Download the pre-built OpenAPI spec artifact from the apify-docs workflow run.
       # Skipped for manual runs — datamodel-codegen will fetch from the published spec URL instead.
       - name: Download OpenAPI spec artifact
@@ -93,46 +132,44 @@ jobs:
             uv run poe generate-models
           fi
 
-      # Proceed only when regeneration actually changes the models relative to the current master.
-      # The job runs on a fresh master checkout, so anything already merged into master (e.g. a
-      # manually merged client PR carrying the same spec change, or a docs PR that merged master in
-      # and re-emits an already-applied change) produces no diff here and we skip — instead of
-      # opening a PR that just duplicates what master already has.
+      # Proceed only when the regenerated models differ from master (compared against the recorded SHA,
+      # since the tree may now sit on the branch), which skips empty-PR runs: a spec change that doesn't
+      # affect the client models, or one already merged into master. Also avoids creating an empty branch.
       - name: Check for model changes
         id: changes
         run: |
-          if git diff --quiet -- src/apify_client/_models.py src/apify_client/_typeddicts.py src/apify_client/_literals.py; then
+          if git diff --quiet "${{ steps.base.outputs.sha }}" -- src/apify_client/_models.py src/apify_client/_typeddicts.py src/apify_client/_literals.py; then
             echo "No model changes relative to master — nothing to regenerate."
             echo "has_changes=false" >> "$GITHUB_OUTPUT"
           else
             echo "has_changes=true" >> "$GITHUB_OUTPUT"
           fi
 
-      # Point the auto-update branch at the current master so the signed-commit step below records
-      # the regenerated models as a single commit on top of it. Resetting to master (instead of
-      # building on a possibly stale existing branch) is what keeps the committed diff relative to
-      # the current master: an already-merged change can never reappear. Any previous content on the
-      # branch is intentionally replaced, so the PR always reflects "current master + freshly
-      # regenerated models" — and it still updates whenever the source docs PR gets new commits.
-      - name: Point branch at current master
-        if: steps.changes.outputs.has_changes == 'true'
-        run: |
-          git checkout -B "$BRANCH"
-          git push --force origin "HEAD:refs/heads/$BRANCH"
-
-      - name: Commit model changes
+      # Append the regenerated models to the auto-update branch as a single signed ("Verified") commit
+      # via apify/actions/signed-commit (GitHub's createCommitOnBranch GraphQL mutation). It's added on
+      # top of the branch tip, never resetting to master: a fresh docs PR creates the branch
+      # (create-branch) and its first commit, and each later docs-PR commit triggers a dispatch that
+      # appends another, so the client PR mirrors the docs PR and stays open.
+      #
+      # Appending (rather than force-pushing the branch to master, as before) is what keeps the PR open:
+      # a "branch == master" tip makes the PR head equal its base, which GitHub auto-closes, and each
+      # dispatch then opens a duplicate PR. signed-commit also sets committed=false when the staged
+      # files match the tip, so a repeated dispatch regenerating identical models adds no commit.
+      - name: Commit regenerated models
         id: commit
         if: steps.changes.outputs.has_changes == 'true'
         uses: apify/actions/signed-commit@v1.2.0
         with:
           message: ${{ env.TITLE }}
-          add: 'src/apify_client/_models.py src/apify_client/_typeddicts.py src/apify_client/_literals.py'
+          add: "src/apify_client/_models.py src/apify_client/_typeddicts.py src/apify_client/_literals.py"
           github-token: ${{ secrets.APIFY_SERVICE_ACCOUNT_GITHUB_TOKEN }}
           branch: ${{ env.BRANCH }}
-          create-branch: 'false'
+          create-branch: "${{ steps.branch.outputs.create }}"
 
+      # Ensure exactly one PR exists for this branch. It's no longer auto-closed, so an existing PR is
+      # reused. Only the first run (or one whose PR step a concurrent dispatch cancelled) creates it.
       - name: Create or update PR
-        if: steps.commit.outputs.committed == 'true'
+        if: steps.changes.outputs.has_changes == 'true'
         id: pr
         env:
           GH_TOKEN: ${{ secrets.APIFY_SERVICE_ACCOUNT_GITHUB_TOKEN }}
@@ -164,9 +201,11 @@ jobs:
             echo "created=true" >> "$GITHUB_OUTPUT"
           fi
 
-      # Post a cross-repo comment on the original docs PR so reviewers know about the corresponding client-python PR.
+      # Post a cross-repo comment on the docs PR pointing reviewers to the companion client-python PR.
+      # Only when something happened: the PR was just created, or a commit was appended. A repeated
+      # dispatch that changes nothing (committed=false) posts no comment, avoiding noise on the docs PR.
       - name: Comment on apify-docs PR
-        if: steps.commit.outputs.committed == 'true' && inputs.docs_pr_number
+        if: inputs.docs_pr_number && (steps.pr.outputs.created == 'true' || steps.commit.outputs.committed == 'true')
         env:
           GH_TOKEN: ${{ secrets.APIFY_SERVICE_ACCOUNT_GITHUB_TOKEN }}
           PR_CREATED: ${{ steps.pr.outputs.created }}