diff --git a/.github/workflows/pr-iphone-deploy.yml b/.github/workflows/pr-iphone-deploy.yml
index ce991bc..839a916 100644
--- a/.github/workflows/pr-iphone-deploy.yml
+++ b/.github/workflows/pr-iphone-deploy.yml
@@ -215,6 +215,15 @@ jobs:
     timeout-minutes: 120
 
     steps:
+      - name: Checkout
+        # Shallow clone — wait-approval only needs scripts/pr_lgtm_scanner.py,
+        # nothing else from the repo. Same raw-git pattern as build-deploy's
+        # Checkout step (avoids actions/checkout setup).
+        run: |
+          git clone --depth 1 -b ${{ github.head_ref }} https://x-access-token:${PR_TOKEN}@gitea.1000co.fr/${{ github.repository }}.git .
+        env:
+          PR_TOKEN: ${{ secrets.PR_API_TOKEN }}
+
       - name: Poll for LGTM
         env:
           GT_TOKEN: ${{ secrets.PR_API_TOKEN }}
@@ -233,8 +242,8 @@ jobs:
           # edits — e.g. a reviewer changing "KO" → "LGTM". Negligible cost for
           # PRs with <50 comments.
           #
-          # Parsing is done with python3 (preinstalled on the macOS runner and
-          # already used by scripts/ci-status.py) rather than grep over raw JSON:
+          # The LGTM/KO decision is computed by scripts/pr_lgtm_scanner.py
+          # (a standalone Python file) rather than grep over raw JSON:
           #   - the bot's "Ready to test" comment body literally contains the
           #     words LGTM/KO as instructions to the reviewer, so any naive
           #     body grep would self-match and auto-merge ~30s after deploy.
@@ -253,33 +262,13 @@ jobs:
             COMMENTS=$(curl -s -H "Authorization: token ${GT_TOKEN}" \
               "${API}/issues/${PR}/comments?limit=50&page=1")
 
-            DECISION=$(python3 -c '
-              import json, re, sys
-              MARKER = "<!-- tabatago:ready-to-test -->"
-              try:
-                  comments = json.loads(sys.argv[1] or "[]")
-              except Exception:
-                  comments = []
-              lgtm = re.compile(r"\bLGTM\b", re.IGNORECASE)
-              ko    = re.compile(r"\bKO\b",   re.IGNORECASE)
-              hit_lgtm = hit_ko = False
-              for c in comments:
-                  body = c.get("body") or ""
-                  if MARKER in body:
-                      # Skip bot comments — their body contains LGTM/KO as reviewer
-                      # instructions and would otherwise self-trigger a merge.
-                      continue
-                  if ko.search(body):
-                      hit_ko = True
-                  if lgtm.search(body):
-                      hit_lgtm = True
-              if hit_ko:
-                  print("KO")
-              elif hit_lgtm:
-                  print("LGTM")
-              else:
-                  print("PENDING")
-              ' "$COMMENTS")
+            # The LGTM/KO scanner lives in scripts/pr_lgtm_scanner.py rather
+            # than inline here. Embedding multi-line Python in a YAML `run: |`
+            # block scalar is fragile — YAML dedents block-scalar content to
+            # the first line, which produced both YAML parse failures and
+            # Python IndentationErrors in earlier iterations of this workflow.
+            # A standalone script sidesteps all of that and is testable locally.
+            DECISION=$(python3 scripts/pr_lgtm_scanner.py "$COMMENTS")
 
             case "$DECISION" in
               LGTM)