Report only; drop issue creation

Print the report to the job log, the run summary, and an artifact instead of
filing a tracking issue. Removes the label probing, the jq output plumbing, and
the issues: write permission -- the job now needs only contents: read.

This also sidesteps the _claude-code.yml author-association gate, which would
have rejected an issue opened by github-actions[bot] anyway. Wiring up
on-demand root-cause analysis can come later on top of a detector that works.

Author: atalman

.github/workflows/vllm-torch-nightly-triage.yml

@@ -6,10 +6,12 @@ name: vLLM torch-nightly triage
 # "Full CI run - nightly" from the same HEAD in the same cron slot, so the pair is a
 # controlled A/B. This job reports jobs that fail in the former and pass in the latter.
 #
-# ClickHouse stores job metadata only -- log contents are not ingested -- so this
-# identifies *which* jobs regressed, not *why*. The issue body ends with @claude so a
-# maintainer gets on-demand root-cause analysis (which reads the linked Buildkite logs
-# via .claude/skills/vllm-pytorch-ci-triage) without this job needing a Buildkite token.
+# Reports only: the result goes to the job log, the run summary, and an artifact.
+# Nothing is filed anywhere.
+#
+# ClickHouse stores job metadata only (the tables carry log_url pointers, not log
+# bodies), so this answers *which* jobs regressed, not *why*. Root-causing means
+# reading the linked Buildkite logs, e.g. with .claude/skills/vllm-pytorch-ci-triage.
 
 on:
   # TEMPORARY (DO NOT MERGE): run on PRs touching this workflow so it can be
@@ -36,7 +38,6 @@ jobs:
     runs-on: ubuntu-24.04
     permissions:
       contents: read
-      issues: write
     steps:
       - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
 
@@ -59,59 +60,25 @@ jobs:
             --lookback-days "${{ inputs.lookback_days || '14' }}" \
             --output "${RUNNER_TEMP}/report.md" \
             --json-output "${RUNNER_TEMP}/report.json"
-          # No pair found => no report written => nothing to post.
-          if [[ -s "${RUNNER_TEMP}/report.json" ]]; then
+          # No torch-nightly build with a same-commit baseline => no report written.
+          if [[ -s "${RUNNER_TEMP}/report.md" ]]; then
             echo "has_report=true" >> "${GITHUB_OUTPUT}"
-            jq -r '"build=\(.torch_nightly_build)"' "${RUNNER_TEMP}/report.json" >> "${GITHUB_OUTPUT}"
-            jq -r '"regressed=\(.regressed | length)"' "${RUNNER_TEMP}/report.json" >> "${GITHUB_OUTPUT}"
           else
             echo "has_report=false" >> "${GITHUB_OUTPUT}"
           fi
 
-      - name: Summarise
+      - name: Print report
         if: steps.report.outputs.has_report == 'true'
-        run: cat "${RUNNER_TEMP}/report.md" >> "${GITHUB_STEP_SUMMARY}"
-
-      # Never file issues from a PR run -- a test of this workflow must not create
-      # noise. The report is still visible in the job summary above.
-      - name: Open triage issue
-        if: |
-          github.event_name != 'pull_request' &&
-          steps.report.outputs.has_report == 'true' &&
-          steps.report.outputs.regressed != '0'
-        env:
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          BUILD: ${{ steps.report.outputs.build }}
-          COUNT: ${{ steps.report.outputs.regressed }}
         run: |
           set -euo pipefail
-          TITLE="[vllm] torch-nightly regressions in Buildkite #${BUILD} (${COUNT} jobs)"
-          {
-            cat "${RUNNER_TEMP}/report.md"
-            echo
-            echo "---"
-            echo
-            echo "@claude please root-cause these regressions using the"
-            echo "\`vllm-pytorch-ci-triage\` skill. Fetch the Buildkite logs for one"
-            echo "representative job per cluster above, group the failures by root cause,"
-            echo "and draft (do not file) one upstream issue per distinct cause."
-          } > "${RUNNER_TEMP}/issue.md"
-          # Labels are applied only if they already exist -- gh fails the whole
-          # command on an unknown label, and the issue matters more than the label.
-          LABEL_ARGS=()
-          for L in "vllm" "ci: torch-nightly"; do
-            if gh label list --repo "${GITHUB_REPOSITORY}" --search "${L}" \
-                 --json name --jq '.[].name' | grep -qxF "${L}"; then
-              LABEL_ARGS+=(--label "${L}")
-            else
-              echo "note: label '${L}' does not exist, skipping"
-            fi
-          done
-          gh issue create \
-            --repo "${GITHUB_REPOSITORY}" \
-            --title "${TITLE}" \
-            --body-file "${RUNNER_TEMP}/issue.md" \
-            "${LABEL_ARGS[@]}"
+          cat "${RUNNER_TEMP}/report.md"
+          cat "${RUNNER_TEMP}/report.md" >> "${GITHUB_STEP_SUMMARY}"
+
+      - name: No comparable build
+        if: steps.report.outputs.has_report != 'true'
+        run: |
+          echo "No torch-nightly build with a same-commit baseline in the lookback window." \
+            | tee -a "${GITHUB_STEP_SUMMARY}"
 
       - name: Upload report
         if: steps.report.outputs.has_report == 'true'