feat: implement issue #2169 — pr-review reaches approve but cannot post it (approval marker missing), and the same head gets approve/LOW in one run and escalate/HIGH in the next

prompts/cascade-action.md

@@ -48,6 +48,9 @@ ESCALATE_TO_AI=$(jq -r 'if .decision == "escalate" and .risk != "HIGH" then "tru
 # a metadata digest so the demanded metadata fix (which mints no commit) re-arms the
 # re-review; a false value keeps commit-only re-arm. Default false.
 METADATA_ONLY=$(jq -r 'if .decision == "escalate" and (.metadata_only == true) then "true" else "false" end' "$FINAL_RESULT")
+# Marker vocabulary: "approved"/"escalated" — the posting guard and review-cycle
+# consumers expect these exact strings, never the raw verdict word (#2169)
+DECISION_MARKER=$([ "$DECISION" = "approve" ] && echo "approved" || echo "escalated")
 ```
 
 2. Compose the review body. Write it to a temp file to avoid shell quoting issues:
@@ -80,7 +83,7 @@ Replace each PLACEHOLDER with the actual values using sed. Then substitute the r
 HEADING=$([ "$DECISION" = "approve" ] && echo "APPROVED ✓" || echo "NEEDS HUMAN REVIEW")
 sed -i \
   -e "s|PLACEHOLDER_SHA|$PR_HEAD_SHA|g" \
-  -e "s|PLACEHOLDER_DECISION|$DECISION|g" \
+  -e "s|PLACEHOLDER_DECISION|$DECISION_MARKER|g" \
   -e "s|PLACEHOLDER_RISK|$RISK|g" \
   -e "s|PLACEHOLDER_HEADING|$HEADING|g" \
   -e "s|PLACEHOLDER_TIER|$FINAL_TIER|g" \

scripts/post-pr-review.sh

@@ -130,6 +130,31 @@ BODY=$(jq -r '.body // ""' "$VERDICT_JSON")
 # body edit (AC3).
 METADATA_ONLY=$(jq -r '.metadata_only // false' "$VERDICT_JSON")
 
+# Derive the approval marker word from the verdict, not from model-written text
+# (issue #2169). The verdict vocabulary is `approve`, but every marker consumer
+# (the #1754 guard below, review-one-pr.sh, carry-forward, the miss-rate metric)
+# expects the past tense `decision=approved`. cascade-action.md substituted the
+# raw verdict word, so a valid approve verdict failed the guard on formatting
+# alone and nothing was posted. Normalize only the present-tense word inside the
+# head-SHA v1 marker; a missing or contradictory decision still fails closed.
+if [ "$DECISION" = "approve" ]; then
+  # -z treats the body as one record so a marker wrapped across lines still
+  # matches. Detect changes without command substitution stripping newlines: write
+  # to temp files and compare via cmp, so a trailing-newline-only difference
+  # doesn't falsely trigger a substitution when sed made no actual change (#2169).
+  BODY_FILE="$(mktemp)"
+  NORMALIZED_FILE="$(mktemp)"
+  printf '%s' "$BODY" > "$BODY_FILE"
+  sed -E -z \
+    "s/(<!-- pr-review-agent v1 sha=${PR_HEAD_SHA}[[:space:]][^>]*decision=)approve([[:space:]]|-->)/\1approved\2/" \
+    "$BODY_FILE" > "$NORMALIZED_FILE"
+  if ! cmp -s "$BODY_FILE" "$NORMALIZED_FILE"; then
+    echo "  approval marker said decision=approve — normalized to decision=approved from the verdict (#2169)"
+    BODY="$(cat "$NORMALIZED_FILE")"
+  fi
+  rm -f "$BODY_FILE" "$NORMALIZED_FILE"
+fi
+
 # Marker keying the single human-escalation artifact (issue #1754 AC2). An
 # escalation must leave a visible, updatable comment on the PR — not just a
 # label. The comment is upserted (created once, then patched in place) so a

tests/test_post_pr_review_escalation.bats

@@ -353,6 +353,79 @@ write_approve_verdict() {
   [ "$(cat "$POSTED_REVIEWS_FILE")" = "[]" ]
 }
 
+# ── #2169: the marker word is derived from the verdict, not model text ──────
+# prompts/cascade-action.md substituted the verdict's raw `decision` value
+# (`approve`) into the marker, while the #1754 guard requires `decision=approved`.
+# A valid approve verdict then failed on formatting alone and nothing was posted.
+
+@test "#2169: approve verdict whose marker says decision=approve is submitted as decision=approved" {
+  local body="<!-- pr-review-agent v1 sha=${SHA} decision=approve risk=MEDIUM -->
+
+## Automated review — APPROVED ✓"
+  local vf; vf=$(write_approve_verdict "$body")
+  run bash "$POST_SCRIPT" "$PR_URL" "$vf" "false"
+  echo "$output" >&2
+  [ "$status" -eq 0 ]
+  [[ "$output" != *"lacks a complete approval marker"* ]]
+  grep -qF "<!-- pr-review-agent v1 sha=${SHA} decision=approved risk=MEDIUM -->" "$REVIEW_OUT"
+  ! grep -qE 'decision=approve([[:space:]]|-->)' "$REVIEW_OUT"
+}
+
+@test "#2169: decision=approve directly before the marker close is normalized too" {
+  local body="<!-- pr-review-agent v1 sha=${SHA} risk=LOW decision=approve-->
+
+## Automated review — APPROVED ✓"
+  local vf; vf=$(write_approve_verdict "$body")
+  run bash "$POST_SCRIPT" "$PR_URL" "$vf" "false"
+  echo "$output" >&2
+  [ "$status" -eq 0 ]
+  grep -qF "decision=approved-->" "$REVIEW_OUT"
+}
+
+@test "#2169: decision=approve for a STALE sha still fails closed" {
+  local old="aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"
+  local body="<!-- pr-review-agent v1 sha=${old} decision=approve risk=LOW -->
+
+## Automated review — APPROVED ✓"
+  local vf; vf=$(write_approve_verdict "$body")
+  run bash "$POST_SCRIPT" "$PR_URL" "$vf" "false"
+  echo "$output" >&2
+  [ "$status" -eq 1 ]
+  [ ! -s "$REVIEW_OUT" ]
+  [ "$(cat "$POSTED_REVIEWS_FILE")" = "[]" ]
+}
+
+@test "#2169: an approve verdict whose marker contradicts it (decision=escalate) still fails closed" {
+  local body="<!-- pr-review-agent v1 sha=${SHA} decision=escalate risk=HIGH -->
+
+## Automated review — NEEDS HUMAN REVIEW"
+  local vf; vf=$(write_approve_verdict "$body")
+  run bash "$POST_SCRIPT" "$PR_URL" "$vf" "false"
+  echo "$output" >&2
+  [ "$status" -eq 1 ]
+  [[ "$output" == *"lacks a complete approval marker"* ]]
+  [ ! -s "$REVIEW_OUT" ]
+  [ "$(cat "$POSTED_REVIEWS_FILE")" = "[]" ]
+}
+
+@test "#2169: decision=approve outside the v1 marker (prose) is not rewritten and still fails closed" {
+  local body="<!-- pr-review-agent v1 sha=${SHA} risk=LOW -->
+
+The verdict was decision=approve in prose only."
+  local vf; vf=$(write_approve_verdict "$body")
+  run bash "$POST_SCRIPT" "$PR_URL" "$vf" "false"
+  echo "$output" >&2
+  [ "$status" -eq 1 ]
+  [ ! -s "$REVIEW_OUT" ]
+}
+
+@test "#2169: cascade-action prompt maps the decision word before substituting the marker" {
+  local prompt="$REPO_ROOT/prompts/cascade-action.md"
+  grep -qF 'DECISION_MARKER=$([ "$DECISION" = "approve" ] && echo "approved" || echo "escalated")' "$prompt"
+  grep -qF 's|PLACEHOLDER_DECISION|$DECISION_MARKER|g' "$prompt"
+  ! grep -qF 's|PLACEHOLDER_DECISION|$DECISION|g' "$prompt"
+}
+
 @test "AC2: exit 100 (not 101) when no escalation comment can be confirmed" {
   export COMMENT_FAILS=1
   local vf; vf=$(write_escalate_verdict)