improve: better line anchoring with code content in prompt

Include actual code content with each anchor line so the LLM
can match findings to specific changed lines. Mark added vs
context lines and instruct the model to prefer added lines.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: anakrish

.github/scripts/perspective-review.sh

@@ -33,37 +33,42 @@ gh api "repos/${REPO}/pulls/${PR_NUMBER}" \
   | head -c 60000 > /tmp/pr_diff.txt
 echo "Diff size: $(wc -c < /tmp/pr_diff.txt) bytes"
 
-# Parse diff to extract valid RIGHT-side line numbers per file.
-# These are the only lines the PR Review API will accept for inline comments.
+# Parse diff to extract valid RIGHT-side line anchors with code content.
+# Format: file:line:type:code  (type is "added" for + lines, "context" for unchanged)
 echo "Extracting valid line anchors from diff..."
 awk '
-  /^diff --git/ {
-    # Extract filename from +++ line (next after ---)
-    file = ""
-  }
-  /^\+\+\+ b\// {
-    file = substr($0, 7)  # strip "+++ b/"
-  }
+  /^diff --git/ { file = "" }
+  /^\+\+\+ b\// { file = substr($0, 7) }
   /^@@ / {
-    # Parse new-file line number from @@ -old,len +new,len @@
     match($0, /\+([0-9]+)(,([0-9]+))?/, arr)
     start = arr[1] + 0
-    count = (arr[3] != "") ? arr[3] + 0 : 1
     line = start
   }
   file != "" && !/^diff --git/ && !/^---/ && !/^\+\+\+/ && !/^@@/ {
     if (/^-/) {
-      # Deleted line: not on RIGHT side, skip
+      # Deleted line: skip (not on RIGHT side)
+    } else if (/^\+/) {
+      # Added line
+      code = substr($0, 2)  # strip leading +
+      gsub(/\t/, "    ", code)
+      print file ":" line ":added:" code
+      line++
     } else {
-      # Added (+) or context ( ) line: valid on RIGHT side
-      if (file != "" && line > 0) {
-        print file ":" line
-      }
+      # Context line
+      code = substr($0, 2)  # strip leading space
+      gsub(/\t/, "    ", code)
+      print file ":" line ":context:" code
       line++
     }
   }
-' /tmp/pr_diff.txt > /tmp/valid_anchors.txt
-echo "Valid anchors: $(wc -l < /tmp/valid_anchors.txt)"
+' /tmp/pr_diff.txt > /tmp/valid_anchors_full.txt
+
+# Also create a plain file:line list for validation
+awk -F: '{print $1 ":" $2}' /tmp/valid_anchors_full.txt > /tmp/valid_anchors.txt
+
+ADDED_COUNT=$(grep -c ':added:' /tmp/valid_anchors_full.txt || echo 0)
+CONTEXT_COUNT=$(grep -c ':context:' /tmp/valid_anchors_full.txt || echo 0)
+echo "Valid anchors: ${ADDED_COUNT} added, ${CONTEXT_COUNT} context"
 
 # Step 3: Select perspectives based on changed paths
 PERSPECTIVES="reliability-engineer,test-engineer"
@@ -103,8 +108,27 @@ done
 
 DIFF_CONTENT=$(cat /tmp/pr_diff.txt)
 
-# Build the anchor list for the prompt (file:line pairs the LLM can reference)
-ANCHOR_LIST=$(cat /tmp/valid_anchors.txt)
+# Build a structured anchor table for the prompt.
+# Show added lines prominently, include some context lines for reference.
+ANCHOR_TABLE=$(awk -F: '
+  {
+    file = $1; line = $2; type = $3
+    # Rejoin remaining fields as code (code may contain colons)
+    code = ""
+    for (i = 4; i <= NF; i++) {
+      if (i > 4) code = code ":"
+      code = code $i
+    }
+    if (type == "added") {
+      printf "  + %s:%s  %s\n", file, line, code
+    }
+  }
+' /tmp/valid_anchors_full.txt)
+
+# Also list context lines but more compactly (just file:line ranges)
+CONTEXT_SUMMARY=$(awk -F: '
+  $3 == "context" { print $1 ":" $2 }
+' /tmp/valid_anchors_full.txt | head -50)
 
 # Step 5: Review with each perspective, posting one PR review per perspective
 TOTAL_FINDINGS=0
@@ -145,20 +169,32 @@ ${KNOWLEDGE_CTX}
 PR Diff:
 ${DIFF_CONTENT}
 
-IMPORTANT: You must anchor findings to exact lines from this list of valid diff lines.
-Each entry is file:line. Only use lines from this list:
+=== ANCHORING INSTRUCTIONS ===
+
+Each finding MUST be anchored to a specific line in the diff.
+Below are the ADDED lines (marked with +) that you can reference.
+Pick the most relevant added line for each finding.
+
+ADDED LINES (preferred — use these):
+${ANCHOR_TABLE}
+
+CONTEXT LINES (also valid, but prefer added lines above):
+${CONTEXT_SUMMARY}
+
+For each finding, set "file" and "line" to an EXACT file:line pair from the lists above.
+Do NOT invent line numbers. Do NOT use line numbers that aren't listed.
 
-${ANCHOR_LIST}
+=== OUTPUT FORMAT ===
 
-Respond with a JSON array of findings. Each finding must have:
-- "severity": one of "critical", "important", "suggestion"
-- "title": a single sentence suitable as a heading
-- "file": exact file path from the valid lines list above
-- "line": exact line number from the valid lines list above, or null if no suitable anchor
-- "body": 2-4 sentences explaining the issue in markdown
+Respond with a JSON array. Each finding:
+- "severity": "critical" | "important" | "suggestion"
+- "title": one-sentence heading
+- "file": exact file path from the anchor lists
+- "line": exact line number from the anchor lists
+- "body": 2-4 sentence explanation in markdown
 
-If you find no issues, return an empty array: []
-Return ONLY valid JSON — no markdown fences, no extra text.
+If no issues found, return: []
+Return ONLY valid JSON — no markdown fences, no commentary.
 PROMPT
 
   PROMPT_CONTENT=$(cat /tmp/review_prompt.txt)