ci: read integration test default models from checked-out branch

.github/run-eval/resolve_model_config.py

@@ -43,6 +43,17 @@ def _sigterm_handler(signum: int, _frame: object) -> None:
 # Keep this list in sync with SDK LLM config parameters that are SDK-internal.
 SDK_ONLY_PARAMS = {"disable_vision"}
 
+# Default model IDs for scheduled/label-triggered integration test runs.
+# This list is the single source of truth; the integration-runner workflow
+# reads it from the checked-out branch via `resolve_model_config.py`, so
+# changes on release branches take effect even though `pull_request_target`
+# evaluates workflow-level env vars from the default branch.
+DEFAULT_INTEGRATION_MODEL_IDS = [
+    "claude-sonnet-4-6",
+    "deepseek-v4-flash",
+    "kimi-k2.6",
+    "gemini-3.1-pro",
+]
 
 # Model configurations dictionary
 MODELS = {

.github/workflows/integration-runner.yml

@@ -49,8 +49,6 @@ on:
 
 env:
     N_PROCESSES: 4 # Global configuration for number of parallel processes for evaluation
-    # Default models for scheduled/label-triggered runs (subset of models from resolve_model_config.py)
-    DEFAULT_MODEL_IDS: claude-sonnet-4-6,deepseek-v4-flash,kimi-k2.6,gemini-3.1-pro
 
 jobs:
     setup-matrix:
@@ -75,27 +73,26 @@ jobs:
               id: resolve-models
               env:
                   MODEL_IDS_INPUT: ${{ github.event.inputs.model_ids || '' }}
-                  DEFAULT_MODEL_IDS: ${{ env.DEFAULT_MODEL_IDS }}
               run: |
-                  # Use input model_ids if provided, otherwise use defaults
-                  if [ -z "$MODEL_IDS_INPUT" ]; then
-                    MODEL_IDS="$DEFAULT_MODEL_IDS"
-                    echo "No model_ids specified, using defaults: $MODEL_IDS"
-                  else
-                    MODEL_IDS="$MODEL_IDS_INPUT"
-                    echo "Using specified model_ids: $MODEL_IDS"
-                  fi
-
-                  # Resolve model configs using resolve_model_config.py
-                  # Transform output to matrix format for integration tests
-                  MATRIX=$(python3 << EOF
+                  # Resolve model configs using resolve_model_config.py from the
+                  # checked-out branch.  DEFAULT_INTEGRATION_MODEL_IDS lives in
+                  # that file so release branches can update the default set
+                  # without waiting for main to be updated (pull_request_target
+                  # evaluates workflow-level env from the default branch).
+                  MATRIX=$(python3 << 'EOF'
                   import json
+                  import os
                   import sys
                   sys.path.insert(0, '.github/run-eval')
-                  from resolve_model_config import MODELS
+                  from resolve_model_config import DEFAULT_INTEGRATION_MODEL_IDS, MODELS
 
-                  model_ids = "$MODEL_IDS".split(",")
-                  model_ids = [m.strip() for m in model_ids if m.strip()]
+                  model_ids_input = os.environ.get("MODEL_IDS_INPUT", "").strip()
+                  if model_ids_input:
+                      model_ids = [m.strip() for m in model_ids_input.split(",") if m.strip()]
+                      print(f"Using specified model_ids: {','.join(model_ids)}", file=sys.stderr)
+                  else:
+                      model_ids = list(DEFAULT_INTEGRATION_MODEL_IDS)
+                      print(f"No model_ids specified, using defaults: {','.join(model_ids)}", file=sys.stderr)
 
                   matrix = []
                   for model_id in model_ids:
@@ -104,7 +101,6 @@ jobs:
                           print(f"Error: Model ID '{model_id}' not found. Available: {available}", file=sys.stderr)
                           sys.exit(1)
                       model = MODELS[model_id]
-                      # Create run-suffix from model id (replace special chars with underscore)
                       run_suffix = model_id.replace("-", "_").replace(".", "_") + "_run"
                       matrix.append({
                           "id": model_id,

tests/cross/test_resolve_model_config.py

@@ -14,6 +14,7 @@
 run_eval_path = Path(__file__).parent.parent.parent / ".github" / "run-eval"
 sys.path.append(str(run_eval_path))
 from resolve_model_config import (  # noqa: E402  # type: ignore[import-not-found]
+    DEFAULT_INTEGRATION_MODEL_IDS,
     MODELS,
     check_model,
     find_models_by_id,
@@ -659,3 +660,12 @@ def test_deepseek_v4_flash_config():
     assert model["id"] == "deepseek-v4-flash"
     assert model["display_name"] == "DeepSeek V4 Flash"
     assert model["llm_config"]["model"] == "litellm_proxy/deepseek/deepseek-v4-flash"
+
+
+def test_default_integration_model_ids_all_exist():
+    """Every ID in DEFAULT_INTEGRATION_MODEL_IDS must exist in MODELS."""
+    for model_id in DEFAULT_INTEGRATION_MODEL_IDS:
+        assert model_id in MODELS, (
+            f"DEFAULT_INTEGRATION_MODEL_IDS contains '{model_id}' "
+            f"which is not in MODELS"
+        )