feat: add OpenCode harness running Kimi K3 through the Vercel AI Gateway

.env.example

@@ -1,3 +1,10 @@
-# AI SDK Core direct-provider credentials.
+# Claude Code harness and Anthropic ai-sdk experiments.
 ANTHROPIC_API_KEY=
+
+# Codex harness and OpenAI ai-sdk experiments. The default LLM judge is an
+# OpenAI model, so rubric-scored evals need this key regardless of which
+# harness produced the run.
 OPENAI_API_KEY=
+
+# Vercel AI Gateway. The opencode harness routes through this.
+AI_GATEWAY_API_KEY=

.github/workflows/eval-refresh.yml

@@ -259,6 +259,7 @@ jobs:
     env:
       ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
       OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+      AI_GATEWAY_API_KEY: ${{ secrets.AI_GATEWAY_API_KEY }}
     steps:
       - name: Checkout
         uses: actions/checkout@9f698171ed81b15d1823a05fc7211befd50c8ae0 # v6.0.3
@@ -286,6 +287,7 @@ jobs:
           {
             echo "ANTHROPIC_API_KEY=${ANTHROPIC_API_KEY}"
             echo "OPENAI_API_KEY=${OPENAI_API_KEY}"
+            echo "AI_GATEWAY_API_KEY=${AI_GATEWAY_API_KEY}"
           } > .env
 
       - name: Run evals
@@ -300,6 +302,12 @@ jobs:
             --runs "${{ needs.prepare.outputs.runs }}" \
             --timeout-sec "${{ needs.prepare.outputs.timeout_sec }}"
 
+          # A skipped experiment (e.g. missing API key) exits 0 without writing results.
+          if [ ! -f "results/${{ matrix.experiment }}/${{ matrix.eval_id }}.json" ]; then
+            echo "::error::no result written for ${{ matrix.experiment }} x ${{ matrix.eval_id }} (experiment skipped?)"
+            exit 1
+          fi
+
       - name: Upload raw results
         uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
         with: