fix(ci): secrets context not allowed in step if, use outputs

.github/workflows/ai-fix.yml

@@ -68,15 +68,29 @@ jobs:
             echo "::warning::No CPA credentials; CPA fallback channel disabled (gateway only)."
           fi
 
+      - name: Detect gateway source
+        id: gw
+        shell: bash
+        env:
+          GATEWAY_BASE_URL: ${{ secrets.GATEWAY_BASE_URL }}
+        run: |
+          set -euo pipefail
+          # secrets context is not allowed in step `if:`; expose via outputs.
+          if [[ -n "${GATEWAY_BASE_URL}" ]]; then
+            echo "external=true" >> "$GITHUB_OUTPUT"
+          else
+            echo "external=false" >> "$GITHUB_OUTPUT"
+          fi
+
       - name: Set up Node
-        if: secrets.GATEWAY_BASE_URL == ''
+        if: steps.gw.outputs.external != 'true'
         uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4
         with:
           node-version: 24
           cache: npm
 
       - name: Start in-job gateway (free models, no login needed)
-        if: secrets.GATEWAY_BASE_URL == ''
+        if: steps.gw.outputs.external != 'true'
         shell: bash
         run: |
           set -euo pipefail

.github/workflows/ai-review.yml

@@ -102,15 +102,30 @@ jobs:
             echo "head_sha=${head_sha}"
           } >> "$GITHUB_OUTPUT"
 
+      - name: Detect gateway source
+        id: gw
+        shell: bash
+        env:
+          GATEWAY_BASE_URL: ${{ secrets.GATEWAY_BASE_URL }}
+        run: |
+          set -euo pipefail
+          # secrets context is not allowed in step `if:`; expose via outputs.
+          # Always runs (trivial check) so downstream never reads an absent output.
+          if [[ -n "${GATEWAY_BASE_URL}" ]]; then
+            echo "external=true" >> "$GITHUB_OUTPUT"
+          else
+            echo "external=false" >> "$GITHUB_OUTPUT"
+          fi
+
       - name: Set up Node
-        if: steps.gate.outputs.skip != 'true' && secrets.GATEWAY_BASE_URL == ''
+        if: steps.gate.outputs.skip != 'true' && steps.gw.outputs.external != 'true'
         uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4
         with:
           node-version: 24
           cache: npm
 
       - name: Start in-job gateway (free models, no login needed)
-        if: steps.gate.outputs.skip != 'true' && secrets.GATEWAY_BASE_URL == ''
+        if: steps.gate.outputs.skip != 'true' && steps.gw.outputs.external != 'true'
         shell: bash
         run: |
           set -euo pipefail

.github/workflows/ai-scan.yml

@@ -59,15 +59,29 @@ jobs:
             echo "::warning::No CPA credentials; CPA fallback channel disabled (gateway only)."
           fi
 
+      - name: Detect gateway source
+        id: gw
+        shell: bash
+        env:
+          GATEWAY_BASE_URL: ${{ secrets.GATEWAY_BASE_URL }}
+        run: |
+          set -euo pipefail
+          # secrets context is not allowed in step `if:`; expose via outputs.
+          if [[ -n "${GATEWAY_BASE_URL}" ]]; then
+            echo "external=true" >> "$GITHUB_OUTPUT"
+          else
+            echo "external=false" >> "$GITHUB_OUTPUT"
+          fi
+
       - name: Set up Node
-        if: secrets.GATEWAY_BASE_URL == ''
+        if: steps.gw.outputs.external != 'true'
         uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4
         with:
           node-version: 24
           cache: npm
 
       - name: Start in-job gateway (free models, no login needed)
-        if: secrets.GATEWAY_BASE_URL == ''
+        if: steps.gw.outputs.external != 'true'
         shell: bash
         run: |
           set -euo pipefail

github_bot/src/agent_orchestrator.py

@@ -414,11 +414,13 @@ def run_multi_agent_review(self, ctx: ReviewContext) -> str:
             "response_ids": [metas[r].get("response_id", "n/a") for r in pipeline if r in metas],
             "roles": {r: metas.get(r, {}) for r in pipeline},
         }
-        # Only the VERDICT: line counts. Prose mentions of NEEDS_CHANGES must
-        # not flip the outcome; unparseable outputs degrade to COMMENT.
+        # Only the VERDICT: line counts (tolerating markdown bold like
+        # **VERDICT: APPROVE**). Prose mentions of NEEDS_CHANGES must not flip
+        # the outcome; unparseable outputs degrade to COMMENT.
         verdict = "APPROVE"
         for text in results.values():
-            m = re.search(r"(?m)^\s*VERDICT\s*:\s*(APPROVE|NEEDS_CHANGES|COMMENT)", text.upper())
+            m = re.search(r"(?m)^\s*\*{0,3}\s*VERDICT\s*:\s*(APPROVE|NEEDS_CHANGES|COMMENT)\b",
+                          text.upper())
             parsed = m.group(1) if m else "COMMENT"
             if parsed == "NEEDS_CHANGES":
                 verdict = "NEEDS_CHANGES"

tests/test_ai_bot.py

@@ -14,6 +14,7 @@
 
 from agent_orchestrator import (  # noqa: E402
     AgentOrchestrator,
+    ReviewContext,
     is_triage_report_publishable,
     scrub_internal_names,
     validate_triage_report,
@@ -284,8 +285,7 @@ def test_verdict_aggregation_verdict_line_only(self):
         orch = AgentOrchestrator.__new__(AgentOrchestrator)
         orch.config = load_json(BOT_DIR / "config" / "bot_config.json")
         ctx = AgentOrchestrator.__new__(AgentOrchestrator)
-        import agent_orchestrator as ao
-        review_ctx = ao.ReviewContext(title="t", body="b")
+        review_ctx = ReviewContext(title="t", body="b")
         # Prose mention of NEEDS_CHANGES must not flip an APPROVE verdict.
         orch._run_role = lambda role, prompt, **kw: (
             "VERDICT: APPROVE\nAll good, no NEEDS_CHANGES here.", {"response_id": "r1"})
@@ -306,6 +306,21 @@ def mixed(role, prompt, **kw):
         report3 = AgentOrchestrator.run_multi_agent_review(orch, review_ctx)
         self.assertIn("FINAL_VERDICT: NEEDS_CHANGES", report3)
 
+    def test_verdict_aggregation_bold_markdown(self):
+        # Live models wrap the verdict in bold: **VERDICT: APPROVE**.
+        # Bold-italic (***) is tolerated too.
+        orch = AgentOrchestrator.__new__(AgentOrchestrator)
+        orch.config = load_json(BOT_DIR / "config" / "bot_config.json")
+        review_ctx = ReviewContext(title="t", body="b")
+        orch._run_role = lambda role, prompt, **kw: (
+            "**VERDICT: APPROVE**\n\nAll good.", {"response_id": "r1"})
+        report = AgentOrchestrator.run_multi_agent_review(orch, review_ctx)
+        self.assertIn("FINAL_VERDICT: APPROVE", report)
+        orch._run_role = lambda role, prompt, **kw: (
+            "***VERDICT: APPROVE***\n\nAll good.", {"response_id": "r2"})
+        report2 = AgentOrchestrator.run_multi_agent_review(orch, review_ctx)
+        self.assertIn("FINAL_VERDICT: APPROVE", report2)
+
 
 class TestScanAndRunner(unittest.TestCase):
     def test_fingerprint_stable(self):