fix(ci): harden triage review gates

yiliang114 · Jun 8, 2026 · Jun 8, 2026 · Jun 8, 2026 · Jun 8, 2026 · Jun 8, 2026
commit 0610d61a6c1a4b0acebb5c9212b0b8a37f47c1bc
diff --git a/.github/workflows/qwen-pr-triage.yml b/.github/workflows/qwen-pr-triage.yml
@@ -708,6 +708,18 @@ jobs:
           fi
 
           RESULT="$QWEN_REVIEW_OUTPUT_FILE"
+          if [ -f "$RESULT" ]; then
+            RESULT_BODY="$(jq -r '.body // ""' "$RESULT" 2>/dev/null || echo "")"
+            if [[ "$RESULT_BODY" == *"<!-- qwen-triage:review-unavailable -->"* ]]; then
+              ALT_RESULT="$(find "$GITHUB_WORKSPACE/.qwen/tmp" -maxdepth 1 -name 'qwen-review-*-review.json' 2>/dev/null | head -1 || true)"
+              if [ -n "${ALT_RESULT:-}" ] && [ -f "$ALT_RESULT" ]; then
+                ALT_BODY="$(jq -r '.body // ""' "$ALT_RESULT" 2>/dev/null || echo "")"
+                if [[ "$ALT_BODY" != *"<!-- qwen-triage:review-unavailable -->"* ]]; then
+                  RESULT="$ALT_RESULT"
+                fi
+              fi
+            fi
+          fi
           if [ ! -f "$RESULT" ]; then
             RESULT="$(find "$GITHUB_WORKSPACE/.qwen/tmp" -maxdepth 1 -name 'qwen-review-*-review.json' 2>/dev/null | head -1 || true)"
           fi
@@ -864,7 +876,18 @@ jobs:
           PR_NUMBER: '${{ needs.resolve.outputs.pr_number }}'
         run: |
           set -euo pipefail
-          pr_data="$(gh pr view "$PR_NUMBER" --repo "$GITHUB_REPOSITORY" --json state,isDraft --jq '[.state, .isDraft] | @tsv')"
+          pr_data=""
+          for attempt in 1 2 3; do
+            if pr_data="$(gh pr view "$PR_NUMBER" --repo "$GITHUB_REPOSITORY" --json state,isDraft --jq '[.state, .isDraft] | @tsv')"; then
+              break
+            fi
+            if [ "$attempt" -eq 3 ]; then
+              echo "::error::Failed to read PR state before tmux testing after 3 attempts"
+              exit 1
+            fi
+            echo "::warning::gh pr view failed before tmux testing (attempt ${attempt}/3), retrying..."
+            sleep $((attempt * 15))
+          done
           IFS=$'\t' read -r state is_draft <<< "$pr_data"
           if [ "$state" != "OPEN" ] || [ "$is_draft" = "true" ]; then
             echo "::notice::Skipping tmux testing: PR #${PR_NUMBER} state=${state} draft=${is_draft}."
@@ -967,8 +990,9 @@ jobs:
     #   - review fail        → /review posts CHANGES_REQUESTED, stop
     #   - review pass        → approval-decision posts the final verdict here
     # So this job only runs when product-decision passed (or was explicitly
-    # skipped via a review_only trigger) AND review completed with verdict
-    # pass — never alongside an upstream stage's request-changes.
+    # skipped via a review_only trigger), review completed with verdict pass,
+    # and tmux-testing produced a real verdict — never alongside an upstream
+    # stage's request-changes or a skipped tmux run.
     if: >-
       always() &&
       needs.resolve.outputs.should_run == 'true' &&
@@ -978,7 +1002,10 @@ jobs:
        needs.resolve.outputs.trigger_type == 'review_only') &&
       needs.review.result == 'success' &&
       needs.review.outputs.verdict == 'pass' &&
-      (needs.tmux-testing.result == 'success' || needs.tmux-testing.result == 'skipped')
+      needs.tmux-testing.result == 'success' &&
+      (needs.tmux-testing.outputs.verdict == 'pass' ||
+       needs.tmux-testing.outputs.verdict == 'fail' ||
+       needs.tmux-testing.outputs.verdict == 'timeout')
     timeout-minutes: 10
     concurrency:
       group: 'qwen-pr-triage-pipeline-${{ needs.resolve.outputs.pr_number }}'

diff --git a/.qwen/skills/approval-decision/SKILL.md b/.qwen/skills/approval-decision/SKILL.md
@@ -25,10 +25,11 @@ Final gate in the PR triage pipeline. Read all prior stage results, reflect hone
   `${TRIAGE_RESULTS_DIR:-/tmp/triage-results}`; the CI workflow publishes them
   with the write-capable token.
 - This skill ONLY runs when product-decision passed (or was explicitly
-  skipped by a review-only trigger) AND code review completed without
-  requesting changes. Upstream failures post their own request-changes and
-  stop the pipeline — so never re-litigate an upstream rejection here; this
-  stage weighs review nits + tmux results and issues the final verdict
+  skipped by a review-only trigger), code review completed without requesting
+  changes, and tmux-testing produced a real verdict (`pass`, `fail`, or
+  `timeout`). Upstream failures and tmux skips stop the pipeline — so never
+  re-litigate an upstream rejection here; this stage weighs review nits + tmux
+  results and issues the final verdict.
 
 ## Procedure
 

diff --git a/.qwen/skills/triage/references/pr-workflow.md b/.qwen/skills/triage/references/pr-workflow.md
@@ -73,7 +73,9 @@ When running locally via `/triage <N>`, follow this dependency chain:
     ├─ requested changes → STOP (the review's CHANGES_REQUESTED is the verdict)
     │
     ▼ passed
-tmux-real-user-testing (skip for fork PRs)
+tmux-real-user-testing (internal PRs only)
+    │
+    ├─ skipped → STOP (fork, closed, draft, or no runnable environment)
     │
     ▼
 /approval-decision (reads all prior comments, decides final verdict)
@@ -85,7 +87,8 @@ stage owns the formal review verdict on every path:
 - **product-decision fails** → it posts request-changes; nothing else runs
 - **review requests changes** → that review IS the verdict; tmux and approval skipped
 - **tmux fails** → approval still runs and weighs the failure (review passed, so there is no competing verdict)
-- **approval-decision runs only when product and review both passed** — it issues the single final approve/request-changes
+- **tmux is skipped** → stop; do not approve without real-scenario evidence
+- **approval-decision runs only when product, review, and tmux gates completed** — it issues the single final approve/request-changes
 
 Each stage posts its own comment with a unique marker. Re-runs update in place.
 

@@ -620,17 +620,30 @@ fi
 If there are **no confirmed findings**, submit a single-line review. Use `event=APPROVE` by default; if the presubmit JSON has `downgradeApprove=true`, use `event=COMMENT` and prepend the downgrade reasons to the body. **In emit-only mode**, build the same `{commit_id, event, body}` object and write it to `QWEN_REVIEW_OUTPUT_FILE` instead of calling `gh api`:
 
 ```bash
-# downgradeApprove=false (non-self PR, green CI):
-gh api repos/{owner}/{repo}/pulls/{pr_number}/reviews \
-  -f commit_id="{commit_sha}" \
-  -f event="APPROVE" \
-  -f body="No issues found. LGTM! ✅ _— YOUR_MODEL_ID via Qwen Code /review_"
-
-# downgradeApprove=true (self-PR, CI failing, or CI still running):
-gh api repos/{owner}/{repo}/pulls/{pr_number}/reviews \
-  -f commit_id="{commit_sha}" \
-  -f event="COMMENT" \
-  -f body="No review findings. Downgraded from Approve to Comment: <downgradeReasons joined with '; '>. _— YOUR_MODEL_ID via Qwen Code /review_"
+EVENT="APPROVE"
+BODY="No issues found. LGTM! ✅ _— YOUR_MODEL_ID via Qwen Code /review_"
+
+if [ "<downgradeApprove>" = "true" ]; then
+  EVENT="COMMENT"
+  BODY="No review findings. Downgraded from Approve to Comment: <downgradeReasons joined with '; '>. _— YOUR_MODEL_ID via Qwen Code /review_"
+fi
+
+if [ -n "${QWEN_REVIEW_EMIT_ONLY:-}" ]; then
+  OUTPUT_FILE="${QWEN_REVIEW_OUTPUT_FILE:?set in emit-only mode}"
+  TMP_OUTPUT="${OUTPUT_FILE}.tmp"
+  jq -n \
+    --arg commit_id "{commit_sha}" \
+    --arg event "$EVENT" \
+    --arg body "$BODY" \
+    '{commit_id:$commit_id, event:$event, body:$body}' > "$TMP_OUTPUT"
+  mv "$TMP_OUTPUT" "$OUTPUT_FILE"
+  echo "Emit-only: wrote no-findings review JSON to $OUTPUT_FILE"
+else
+  gh api repos/{owner}/{repo}/pulls/{pr_number}/reviews \
+    -f commit_id="{commit_sha}" \
+    -f event="$EVENT" \
+    -f body="$BODY"
+fi
 ```
 
 Clean up the JSON file in Step 11.