Pattern miner: Add rejection-learning and cross-repo pattern tips from upstream mining

patterns/archetypes/code-health-auditor.json

@@ -28,7 +28,8 @@
     "Use tracker-id and skip-if-match so a scheduled run doesn't reopen the same finding every day",
     "Set expires on the created issue or pull request so stale findings don't accumulate",
     "Assign findings to copilot so remediation can be picked up automatically",
-    "Persist the history of previously-audited structural dimensions in cache-memory and weight selection toward ones not covered in recent runs, so scheduled runs rotate through file size, duplication, and complexity instead of repeating the same dimension"
+    "Persist the history of previously-audited structural dimensions in cache-memory and weight selection toward ones not covered in recent runs, so scheduled runs rotate through file size, duplication, and complexity instead of repeating the same dimension",
+    "Before creating a new finding, search the most recently closed items from this workflow and read the close reason and any review feedback — treat a rejecting comment or 'not planned' close as a negative signal and avoid re-proposing the same fix"
   ],
   "anti_patterns": []
 }

patterns/archetypes/code-improvement.json

@@ -59,7 +59,8 @@
     "Use explicit DO NOT constraints to keep changes scoped to the requested improvement",
     "Use protected-files: fallback-to-issue for manifests, CI configuration, and agent instructions",
     "Add a pre-activation step that skips scheduled runs once too many open PRs share your title prefix, to avoid spamming maintainers",
-    "When rotating across several improvement categories (e.g. code quality, testing, docs, performance), persist the history of previously-selected categories in cache-memory and weight selection toward ones not covered in recent runs, so scheduled runs don't repeatedly re-select the same category"
+    "When rotating across several improvement categories (e.g. code quality, testing, docs, performance), persist the history of previously-selected categories in cache-memory and weight selection toward ones not covered in recent runs, so scheduled runs don't repeatedly re-select the same category",
+    "Cap the schedule to one open pull request at a time with skip-if-match (max: 1) and expires, and before proposing a new fix search the most recently closed pull requests from this workflow — treat a 'not planned' close or rejecting review comment as a negative signal and avoid re-proposing the same change"
   ],
   "anti_patterns": [
     "ci-coach",

patterns/archetypes/contribution-guidelines-checker.json

@@ -29,7 +29,8 @@
     "Label the pull request ready when it complies and leave one constructive comment when it does not — never both",
     "Use a small model since this runs on every pull request open/synchronize event and cost adds up",
     "For high-volume repositories, pre-filter and fetch PR data in a steps: block, then dispatch each PR to an inline sub-agent for evaluation so the parent workflow only compiles results",
-    "Use DO NOT constraints to avoid merging, approving, or closing contributions — label and comment only"
+    "Use DO NOT constraints to avoid merging, approving, or closing contributions — label and comment only",
+    "To monitor a different repository than the one the workflow runs in, enable the GitHub toolsets needed for the external repo, set target-repo on add-labels and add-comment, and document the required token scopes in the prompt"
   ],
   "anti_patterns": []
 }