Improve issue dedupe detect workflow reliability (#6115)

Signed-off-by: Heng Qian <qianheng@amazon.com>

Author: qianheng-aws

.github/workflows/issue-dedupe-detect.yml

@@ -21,6 +21,7 @@ on:
         default: ''
         description: 'Issue number to check (defaults to github.event.issue.number)'
 
+
 jobs:
   detect:
     runs-on: ubuntu-latest
@@ -42,9 +43,8 @@ jobs:
           aws-region: us-east-1
 
       - name: Run Claude Code for duplicate detection
+        id: claude
         uses: anthropics/claude-code-action@2ff1acb3ee319fa302837dad6e17c2f36c0d98ea # v1
-        env:
-          DUPLICATE_GRACE_DAYS: ${{ inputs.grace_days }}
         with:
           use_bedrock: 'true'
           github_token: ${{ secrets.GITHUB_TOKEN }}
@@ -53,16 +53,14 @@ jobs:
             --model ${{ inputs.claude_model }}
             --allowedTools
             "Bash(gh issue view *)"
-            "Bash(gh issue comment *)"
-            "Bash(gh issue edit *)"
             "Bash(gh search issues *)"
-            "Bash(gh label create *)"
+            --json-schema '{"type":"object","properties":{"duplicates":{"type":"array","items":{"type":"integer"}},"skip_reason":{"type":"string"}},"required":["duplicates"]}'
           prompt: |
             Find up to 3 likely duplicate issues for issue #${{ env.ISSUE_NUMBER }} in ${{ github.repository }}.
 
             Follow these steps precisely:
 
-            1. Use `gh issue view ${{ env.ISSUE_NUMBER }} --repo ${{ github.repository }} --comments` to read the issue and its comments. If the issue is closed, or is broad product feedback without a specific bug/feature request, or already has a duplicate detection comment (containing `<!-- duplicate-detection -->`), stop and report why you are not proceeding.
+            1. Use `gh issue view ${{ env.ISSUE_NUMBER }} --repo ${{ github.repository }} --comments` to read the issue and its comments. If the issue is closed, or is broad product feedback without a specific bug/feature request, or already has a duplicate detection comment (containing `<!-- duplicate-detection -->`), return an empty duplicates array with skip_reason explaining why.
 
             2. Summarize the issue's core problem in 2-3 sentences. Identify the key terms, error messages, and affected components.
 
@@ -74,36 +72,85 @@ jobs:
 
             4. For each candidate issue that looks like a potential match, read it with `gh issue view <number> --repo ${{ github.repository }}` to verify it is truly about the same problem. Filter out false positives — issues that merely share keywords but describe different problems.
 
-            5. If you find 1-3 genuine duplicates, you MUST run the following commands exactly as shown. Do NOT paraphrase or reformat the comment body — the `<!-- duplicate-detection -->` marker is required for automated processing.
-
-               a. For each duplicate, get its title:
-                  `gh issue view <dup_number> --repo ${{ github.repository }} --json title -q .title`
+            5. Return the confirmed duplicate issue numbers. If no duplicates found, return an empty array.
 
-               b. Ensure the duplicate label exists:
-                  `gh label create "duplicate" --description "Issue is a duplicate of an existing issue" --color "cccccc" --repo ${{ github.repository }} 2>/dev/null || true`
-
-               c. Add the label:
-                  `gh issue edit ${{ env.ISSUE_NUMBER }} --repo ${{ github.repository }} --add-label "duplicate"`
+            Important notes:
+            - Only flag issues as duplicates when you are confident they describe the **same underlying problem**
+            - Prefer open issues as duplicates, but closed issues can be referenced too
+            - Do not flag the issue as a duplicate of itself
 
-               d. Post the comment. You MUST use this EXACT command with the EXACT format below (only replace N with the count and <dup_list> with lines like `- #123 — Title`):
-                  ```
-                  gh issue comment ${{ env.ISSUE_NUMBER }} --repo ${{ github.repository }} --body "<!-- duplicate-detection -->
+      # The --json-schema flag in claude_args makes Claude return validated JSON
+      # via steps.claude.outputs.structured_output (instead of free-form text in execution_file).
+      # Schema: { "duplicates": [int], "skip_reason"?: string }
+      #   - duplicates: issue numbers confirmed as duplicates (empty if none or skipped)
+      #   - skip_reason: optional, explains why detection was skipped (for debugging)
+      - name: Process structured output
+        id: process
+        shell: bash
+        env:
+          STRUCTURED_OUTPUT: ${{ steps.claude.outputs.structured_output }}
+        run: |
+          echo "Structured output: $STRUCTURED_OUTPUT"
+
+          # No output means Claude failed to return structured data
+          if [ -z "$STRUCTURED_OUTPUT" ]; then
+            echo "No structured output"
+            echo "has_duplicates=false" >> "$GITHUB_OUTPUT"
+            exit 0
+          fi
+
+          # skip_reason is set when Claude decides not to proceed
+          # (e.g., issue is closed, already processed, or broad feedback).
+          # Log it for debugging — takes priority over any duplicates.
+          SKIP_REASON=$(echo "$STRUCTURED_OUTPUT" | jq -r '.skip_reason // empty')
+          if [ -n "$SKIP_REASON" ]; then
+            echo "Skipped: $SKIP_REASON"
+            echo "has_duplicates=false" >> "$GITHUB_OUTPUT"
+            exit 0
+          fi
+
+          COUNT=$(echo "$STRUCTURED_OUTPUT" | jq '.duplicates | length')
+          if [ "$COUNT" -eq 0 ]; then
+            echo "No duplicates found"
+            echo "has_duplicates=false" >> "$GITHUB_OUTPUT"
+            exit 0
+          fi
+
+          echo "has_duplicates=true" >> "$GITHUB_OUTPUT"
+          echo "count=$COUNT" >> "$GITHUB_OUTPUT"
+
+          # Format as "- #123" lines — GitHub auto-renders issue links with titles
+          DUP_LIST=$(echo "$STRUCTURED_OUTPUT" | jq -r '.duplicates[] | "- #\(.)"')
+          {
+            echo "duplicate_list<<EOF"
+            echo "$DUP_LIST"
+            echo "EOF"
+          } >> "$GITHUB_OUTPUT"
+
+      - name: Add duplicate label
+        if: steps.process.outputs.has_duplicates == 'true'
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          gh label create "duplicate" --description "Issue is a duplicate of an existing issue" --color "cccccc" --repo "${{ github.repository }}" 2>/dev/null || true
+          gh issue edit "${{ env.ISSUE_NUMBER }}" --repo "${{ github.repository }}" --add-label "duplicate"
+          echo "Added duplicate label to issue #${{ env.ISSUE_NUMBER }}"
+
+      - name: Post duplicate detection comment
+        if: steps.process.outputs.has_duplicates == 'true'
+        uses: peter-evans/create-or-update-comment@e8674b075228eee787fea43ef493e45ece1004c9 # v5
+        with:
+          issue-number: ${{ env.ISSUE_NUMBER }}
+          body: |
+            <!-- duplicate-detection -->
             ### Possible Duplicate
 
-            Found **N** possible duplicate issue(s):
+            Found **${{ steps.process.outputs.count }}** possible duplicate issue(s):
 
-            <dup_list>
+            ${{ steps.process.outputs.duplicate_list }}
 
             If this is **not** a duplicate:
             - Add a comment on this issue, or
             - 👎 this comment to prevent auto-closure
 
-            Otherwise, this issue will be **automatically closed in ${DUPLICATE_GRACE_DAYS:-7} days**."
-                  ```
-
-            6. If no genuine duplicates are found, report that no duplicates were detected and take no further action.
-
-            Important notes:
-            - Only flag issues as duplicates when you are confident they describe the **same underlying problem**
-            - Prefer open issues as duplicates, but closed issues can be referenced too
-            - Do not flag the issue as a duplicate of itself
+            Otherwise, this issue will be **automatically closed in ${{ inputs.grace_days }} days**.