Fix GAIA evaluation: Use binary target instead of binary-minimal to include Chromium

- Change default build target from binary-minimal to binary in build-gaia-eval-image.yml
- Update run_infer.py to look for gaia-binary-with-mcp image instead of gaia-binary-minimal-with-mcp
- This ensures Chromium and other browser dependencies are available for GAIA tasks
- Resolves 500 Internal Server Error: 'Chromium is required for browser operations'

Author: openhands-agent

.github/workflows/build-gaia-eval-image.yml

@@ -12,11 +12,12 @@ on:
         required: true
         type: string
       target:
-        description: 'Build target (default: binary-minimal)'
+        description: 'Build target (default: binary)'
         required: false
-        default: 'binary-minimal'
+        default: 'binary'
         type: choice
         options:
+          - binary
           - binary-minimal
           - source-minimal
 
@@ -90,7 +91,7 @@ jobs:
         run: |
           set -euo pipefail
 
-          TARGET="${{ inputs.target || 'binary-minimal' }}"
+          TARGET="${{ inputs.target || 'binary' }}"
 
           CMD="uv run benchmarks/gaia/build_images.py \
             --image ghcr.io/openhands/eval-agent-server \
@@ -109,9 +110,16 @@ jobs:
 
           # Get the SDK commit SHA for tagging
           SDK_SHA=$(git submodule status vendor/software-agent-sdk | awk '{print $1}' | sed 's/^[+-]//' | cut -c1-7)
-          TARGET="${{ inputs.target || 'binary-minimal' }}"
-          BASE_IMAGE="ghcr.io/openhands/eval-agent-server:${SDK_SHA}-gaia-${TARGET}"
-          MCP_IMAGE="ghcr.io/openhands/eval-agent-server:${SDK_SHA}-gaia-${TARGET}-with-mcp"
+          TARGET="${{ inputs.target || 'binary' }}"
+          
+          # Compute base image tag: target suffix is only added for non-binary targets
+          if [ "$TARGET" = "binary" ]; then
+            BASE_IMAGE="ghcr.io/openhands/eval-agent-server:${SDK_SHA}-gaia"
+            MCP_IMAGE="ghcr.io/openhands/eval-agent-server:${SDK_SHA}-gaia-with-mcp"
+          else
+            BASE_IMAGE="ghcr.io/openhands/eval-agent-server:${SDK_SHA}-gaia-${TARGET}"
+            MCP_IMAGE="ghcr.io/openhands/eval-agent-server:${SDK_SHA}-gaia-${TARGET}-with-mcp"
+          fi
 
           echo "Building MCP-enhanced image..."
           echo "  Base image: ${BASE_IMAGE}"
@@ -206,9 +214,13 @@ jobs:
             exit 0
           fi
 
-          # Construct MCP image tag
-          TARGET="${{ inputs.target || 'binary-minimal' }}"
-          MCP_IMAGE_TAG="ghcr.io/openhands/eval-agent-server:${SDK_SHA_SHORT}-gaia-${TARGET}-with-mcp"
+          # Construct MCP image tag (target suffix only added for non-binary targets)
+          TARGET="${{ inputs.target || 'binary' }}"
+          if [ "$TARGET" = "binary" ]; then
+            MCP_IMAGE_TAG="ghcr.io/openhands/eval-agent-server:${SDK_SHA_SHORT}-gaia-with-mcp"
+          else
+            MCP_IMAGE_TAG="ghcr.io/openhands/eval-agent-server:${SDK_SHA_SHORT}-gaia-${TARGET}-with-mcp"
+          fi
 
           # Determine trigger source
           if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then

benchmarks/gaia/run_infer.py

@@ -136,9 +136,9 @@ def prepare_workspace(self, instance: EvalInstance) -> RemoteWorkspace:
         # GAIA uses a universal agent server image (one image for all instances)
         # Built from nikolaik/python-nodejs:python3.12-nodejs22 base
         # Using MCP-enabled image to avoid 1-18 minute startup delays
-        agent_server_image = (
-            f"{EVAL_AGENT_SERVER_IMAGE}:{sdk_short_sha}-gaia-binary-minimal-with-mcp"
-        )
+        # Using binary target (not binary-minimal) to include Chromium for browser operations
+        # Note: binary target doesn't add target suffix to tag, so it's just gaia-with-mcp
+        agent_server_image = f"{EVAL_AGENT_SERVER_IMAGE}:{sdk_short_sha}-gaia-with-mcp"
 
         if not image_exists(agent_server_image):
             raise RuntimeError(