fix(strix): route the openai-direct terminal fallback to OpenAI credentials

.github/workflows/strix.yml

@@ -708,6 +708,29 @@ jobs:
             echo "STRIX_OPENAI_FALLBACK_KEY_FILE=$openai_fallback_key_file" >> "$GITHUB_ENV"
           fi
 
+      - name: Prepare direct-OpenAI fallback credentials
+        # The nvidia_nim, github_models, and openrouter fallback matrices all
+        # end with openai-direct/gpt-5.6-luna. Without OpenAI credentials those
+        # candidates previously ran against the primary provider endpoint with
+        # the wrong key and failed closed on every provider outage.
+        if: steps.gate.outputs.provider_mode == 'nvidia_nim' || steps.gate.outputs.provider_mode == 'openrouter' || steps.gate.outputs.provider_mode == 'github_models'
+        env:
+          OPENAI_FALLBACK_KEY: ${{ secrets.STRIX_OPENAI_API_KEY || secrets.OPENAI_API_KEY }}
+        run: |
+          umask 077
+          sanitized="$(printf '%s' "$OPENAI_FALLBACK_KEY" | tr -d '\r\n')"
+          trimmed="$(printf '%s' "$sanitized" | sed 's/^[[:space:]]*//;s/[[:space:]]*$//')"
+          if [ -z "$trimmed" ]; then
+            echo '::notice::No direct-OpenAI key available; Strix scans run without the openai-direct terminal fallback.'
+            exit 0
+          fi
+          openai_fallback_key_file="$RUNNER_TEMP/openai_fallback_key.txt"
+          printf '%s' "$trimmed" > "$openai_fallback_key_file"
+          echo "STRIX_OPENAI_FALLBACK_KEY_FILE=$openai_fallback_key_file" >> "$GITHUB_ENV"
+          openai_fallback_api_base_file="$RUNNER_TEMP/openai_fallback_api_base.txt"
+          printf '%s' 'https://api.openai.com/v1' > "$openai_fallback_api_base_file"
+          echo "STRIX_OPENAI_FALLBACK_API_BASE_FILE=$openai_fallback_api_base_file" >> "$GITHUB_ENV"
+
       - name: Prepare Vertex AI credentials
         if: steps.gate.outputs.provider_mode == 'vertex_ai'
         env:
@@ -837,6 +860,7 @@ jobs:
           STRIX_FALLBACK_MODELS: ${{ steps.gate.outputs.provider_mode == 'github_models' && 'openai-direct/gpt-5.6-luna' || steps.gate.outputs.provider_mode == 'openai_direct' && 'openai-direct/gpt-5.6-luna' || steps.gate.outputs.provider_mode == 'openrouter' && 'openai-direct/gpt-5.6-luna' || steps.gate.outputs.provider_mode == 'nvidia_nim' && 'nvidia_nim/nvidia/llama-3.3-nemotron-super-49b-v1.5 openai-direct/gpt-5.6-luna' || '' }}
           STRIX_GITHUB_MODELS_API_BASE_FILE: ${{ env.STRIX_GITHUB_MODELS_API_BASE_FILE }}
           STRIX_GITHUB_MODELS_KEY_FILE: ${{ env.STRIX_GITHUB_MODELS_KEY_FILE }}
+          STRIX_OPENAI_FALLBACK_API_BASE_FILE: ${{ env.STRIX_OPENAI_FALLBACK_API_BASE_FILE }}
           STRIX_OPENAI_FALLBACK_KEY_FILE: ${{ env.STRIX_OPENAI_FALLBACK_KEY_FILE }}
           STRIX_FAIL_ON_PROVIDER_SIGNAL: "1"
           STRIX_VERTEX_FALLBACK_MODELS: ""

CHANGELOG.md

@@ -60,6 +60,17 @@ Semantic Versioning where the repository publishes a release.
 
 ### Fixed
 
+- Recognize the hyphenated `openai-direct/` fallback alias (pinned verbatim by
+  protected main's own trusted `strix_required_workflow_smoke.sh`, so the
+  `STRIX_FALLBACK_MODELS` string itself cannot change) in
+  `child_model_for_api_base`, alongside the existing underscored
+  `openai_direct/` form. Previously the hyphenated alias passed through
+  unrecognized and unrewritten, so a NIM-exhaustion fallback to
+  `openai-direct/gpt-5.6-luna` reached LiteLLM as a literal, unrecognized
+  provider string (`litellm.BadRequestError: LLM Provider NOT provided`)
+  instead of the intended `openai/gpt-5.6-luna`, observed after NVIDIA NIM
+  rate-limited both the primary and first fallback model in three
+  consecutive Strix runs.
 - Publish only the sanitized cumulative Strix report tree, avoiding a later
   copy of relative scanner output that could reintroduce known internal warning
   text into uploaded security evidence.

scripts/ci/strix_quick_gate.sh

@@ -380,12 +380,12 @@ if [ -n "$STRIX_GITHUB_MODELS_KEY_FILE" ]; then
 	fi
 fi
 
-# Optional cross-provider fallback credentials for direct-OpenAI fallback
-# models (openai-direct/... or openai_direct/...). When the primary model runs
-# against NVIDIA NIM, OpenRouter, or GitHub Models, its LLM_API_KEY cannot
-# authenticate a direct-OpenAI fallback; this file carries the OpenAI key.
-# Optional: without it, explicit direct-OpenAI models keep using LLM_API_KEY,
-# which is correct whenever the primary already runs against direct OpenAI.
+# Optional cross-provider fallback credentials for openai-direct/ (and the
+# underscore alias openai_direct/) fallback models. When the primary provider
+# is NVIDIA NIM, GitHub Models, or OpenRouter, those fallbacks must
+# authenticate against the direct OpenAI API with an OpenAI key instead of
+# reusing LLM_API_KEY_FILE / LLM_API_BASE_FILE. Both files are optional;
+# without them openai-direct fallbacks keep requiring LLM_API_BASE_FILE.
 STRIX_OPENAI_FALLBACK_KEY_FILE="${STRIX_OPENAI_FALLBACK_KEY_FILE:-}"
 if [ -n "$STRIX_OPENAI_FALLBACK_KEY_FILE" ] && { [ ! -f "$STRIX_OPENAI_FALLBACK_KEY_FILE" ] || [ -L "$STRIX_OPENAI_FALLBACK_KEY_FILE" ]; }; then
 	echo "ERROR: STRIX_OPENAI_FALLBACK_KEY_FILE must reference a regular file containing the API key." >&2
@@ -403,16 +403,14 @@ if [ -n "$STRIX_OPENAI_FALLBACK_KEY_FILE" ]; then
 	fi
 fi
 
-is_explicit_openai_model() {
-	case "$1" in
-	openai_direct/* | openai-direct/*)
-		return 0
-		;;
-	*)
-		return 1
-		;;
-	esac
-}
+STRIX_OPENAI_FALLBACK_API_BASE_FILE="${STRIX_OPENAI_FALLBACK_API_BASE_FILE:-}"
+if [ -n "$STRIX_OPENAI_FALLBACK_API_BASE_FILE" ] && { [ ! -f "$STRIX_OPENAI_FALLBACK_API_BASE_FILE" ] || [ -L "$STRIX_OPENAI_FALLBACK_API_BASE_FILE" ]; }; then
+	echo "ERROR: STRIX_OPENAI_FALLBACK_API_BASE_FILE must reference a regular file containing the API base URL." >&2
+	exit 2
+fi
+if [ -n "$STRIX_OPENAI_FALLBACK_API_BASE_FILE" ] && ! STRIX_OPENAI_FALLBACK_API_BASE_FILE="$(resolve_trusted_input_file "STRIX_OPENAI_FALLBACK_API_BASE_FILE" "$STRIX_OPENAI_FALLBACK_API_BASE_FILE")"; then
+	exit 2
+fi
 
 require_non_negative_integer() {
 	local value="$1"
@@ -789,6 +787,20 @@ is_github_models_model() {
 	esac
 }
 
+## True when the model routes to the direct OpenAI API through one of the two
+## accepted spellings. The workflow's fallback matrices use the hyphen alias
+## while some self-tests and callers pass the underscore form.
+is_openai_direct_model() {
+	case "$1" in
+	openai-direct/* | openai_direct/*)
+		return 0
+		;;
+	*)
+		return 1
+		;;
+	esac
+}
+
 is_github_models_api_compatible_model() {
 	case "$1" in
 	openai/openai/* | github_models/* | \
@@ -2417,7 +2429,13 @@ resolved_llm_api_base_for_model() {
 
 	local api_base_file="$LLM_API_BASE_FILE"
 	local api_base_file_name="LLM_API_BASE_FILE"
-	if is_github_models_model "$model" && [ -n "${STRIX_GITHUB_MODELS_API_BASE_FILE:-}" ]; then
+	if is_openai_direct_model "$model" && [ -n "${STRIX_OPENAI_FALLBACK_API_BASE_FILE:-}" ]; then
+		# Cross-provider fallback: openai-direct/* (and openai_direct/*)
+		# candidates must reach the direct OpenAI API even when the primary
+		# provider selected a different LLM_API_BASE_FILE endpoint.
+		api_base_file="$STRIX_OPENAI_FALLBACK_API_BASE_FILE"
+		api_base_file_name="STRIX_OPENAI_FALLBACK_API_BASE_FILE"
+	elif is_github_models_model "$model" && [ -n "${STRIX_GITHUB_MODELS_API_BASE_FILE:-}" ]; then
 		# Cross-provider fallback: when the active primary provider uses a
 		# different API base (for example OpenRouter), github_models/* fallback
 		# attempts must still route through the GitHub Models inference endpoint.
@@ -2482,8 +2500,8 @@ child_model_for_api_base() {
 	fi
 
 	case "$model" in
-	openai_direct/*)
-		printf 'openai/%s\n' "${model#openai_direct/}"
+	openai_direct/* | openai-direct/*)
+		printf 'openai/%s\n' "${model#*/}"
 		return 0
 		;;
 	# The workflow contract spells the direct-OpenAI fallback with a hyphen
@@ -2540,7 +2558,12 @@ run_strix_once() {
 	local child_llm_api_key=""
 	if ! is_vertex_model "$(normalize_model "$model")"; then
 		child_llm_api_key="$LLM_API_KEY"
-		if is_github_models_model "$(normalize_model "$model")" && [ -n "$STRIX_GITHUB_MODELS_KEY" ]; then
+		if is_openai_direct_model "$model" && [ -n "$STRIX_OPENAI_FALLBACK_KEY" ]; then
+			# Cross-provider fallback: openai-direct/* (and openai_direct/*)
+			# models authenticate with the direct-OpenAI key, not the primary
+			# provider key in LLM_API_KEY_FILE.
+			child_llm_api_key="$STRIX_OPENAI_FALLBACK_KEY"
+		elif is_github_models_model "$(normalize_model "$model")" && [ -n "$STRIX_GITHUB_MODELS_KEY" ]; then
 			# Cross-provider fallback: github_models/* models authenticate
 			# with the GitHub Models token, not the direct-OpenAI key.
 			child_llm_api_key="$STRIX_GITHUB_MODELS_KEY"

scripts/ci/test_strix_quick_gate.sh

@@ -365,6 +365,7 @@ assert_strix_workflow_pr_trigger_hardened() {
 	assert_file_contains "$workflow_file" "steps.gate.outputs.provider_mode == 'nvidia_nim' && 'nvidia_nim/nvidia/llama-3.3-nemotron-super-49b-v1.5 openai-direct/gpt-5.6-luna'" "strix workflow gives NVIDIA NIM scans contracted fallbacks"
 	assert_file_not_contains "$workflow_file" "STRIX_FALLBACK_MODELS: \${{ steps.gate.outputs.provider_mode == 'github_models' && 'github_models/openai/o3" "strix workflow fallback list must not depend on GitHub Models, which is in platform-wide retirement"
 	assert_file_contains "$workflow_file" "Prepare GitHub Models fallback credentials" "strix workflow provisions GitHub Models fallback credentials for direct-OpenAI scans"
+	assert_file_contains "$workflow_file" "if: steps.gate.outputs.provider_mode == 'nvidia_nim' || steps.gate.outputs.provider_mode == 'openrouter' || steps.gate.outputs.provider_mode == 'github_models'" "strix workflow provisions direct-OpenAI fallback credentials for GitHub Models scans"
 	assert_file_contains "$GATE_SCRIPT" "STRIX_GITHUB_MODELS_KEY_FILE" "strix gate reads the optional GitHub Models fallback key file"
 	assert_file_contains "$GATE_SCRIPT" "STRIX_GITHUB_MODELS_API_BASE_FILE" "strix gate routes github_models fallback models through the GitHub Models endpoint"
 	assert_file_not_contains "$workflow_file" 'github_models/deepseek/deepseek-r1-0528 | github_models/deepseek/deepseek-v3-0324)' "strix workflow keeps DeepSeek GitHub Models restricted to fallback-only routing"

tests/test_strix_nvidia_nim_not_found_fallback.py

@@ -73,6 +73,34 @@ def _classifies_as_nvidia_not_found(log_text: str) -> bool:
     return completed.returncode == 0
 
 
+def _child_model_for_api_base(model: str, llm_api_base_value: str) -> str:
+    """Execute the production model-alias normalizer against one input pair."""
+
+    gate_source = STRIX_GATE.read_text(encoding="utf-8")
+    function_source = "\n".join(
+        _function_block(gate_source, name)
+        for name in (
+            "is_github_models_api_base",
+            "is_github_models_model",
+            "child_model_for_api_base",
+        )
+    )
+    script = "\n".join(
+        (
+            "set -euo pipefail",
+            function_source,
+            'child_model_for_api_base "$1" "$2"',
+        )
+    )
+    completed = subprocess.run(
+        ["bash", "-c", script, "strix-normalizer", model, llm_api_base_value],
+        check=True,
+        capture_output=True,
+        text=True,
+    )
+    return completed.stdout.strip()
+
+
 def _workflow_signal_pattern(workflow: str, variable_name: str) -> str:
     """Extract one single-quoted POSIX ERE assigned in the Strix workflow."""
 
@@ -205,6 +233,12 @@ def test_workflow_uses_available_free_first_nvidia_plan(self) -> None:
             f"'{FREE_NVIDIA_FALLBACK} openai-direct/gpt-5.6-luna'",
             workflow,
         )
+        self.assertIn(
+            "steps.gate.outputs.provider_mode == 'nvidia_nim' || "
+            "steps.gate.outputs.provider_mode == 'openrouter' || "
+            "steps.gate.outputs.provider_mode == 'github_models'",
+            workflow,
+        )
 
         default_gate = workflow.split("- name: Gate Strix secrets", maxsplit=1)[1]
         default_gate = default_gate.split(
@@ -213,6 +247,29 @@ def test_workflow_uses_available_free_first_nvidia_plan(self) -> None:
         )[0]
         self.assertNotIn(RETIRED_PRIMARY_MODEL, default_gate)
 
+    def test_gate_normalizes_hyphenated_openai_direct_fallback_alias(self) -> None:
+        """Route the NIM-exhaustion fallback alias to a real LiteLLM provider.
+
+        `STRIX_FALLBACK_MODELS`' NVIDIA NIM entry ends in the hyphenated
+        `openai-direct/gpt-5.6-luna` alias (the workflow's user-facing input
+        spelling, also pinned verbatim by protected main's own trusted
+        `strix_required_workflow_smoke.sh`, so this exact string cannot
+        change). The gate must still resolve it to LiteLLM's `openai/`
+        provider -- the same target the underscored `openai_direct/` alias
+        already reaches -- or NVIDIA NIM rate-limiting the primary and first
+        fallback model leaves the run one hop from
+        `litellm.BadRequestError: LLM Provider NOT provided`.
+        """
+
+        self.assertEqual(
+            _child_model_for_api_base("openai-direct/gpt-5.6-luna", ""),
+            "openai/gpt-5.6-luna",
+        )
+        self.assertEqual(
+            _child_model_for_api_base("openai_direct/gpt-5.6-luna", ""),
+            "openai/gpt-5.6-luna",
+        )
+
     def test_outer_workflow_requires_litellm_context_for_nvidia_404(self) -> None:
         """Reject provider-like target text in the outer neutralization gate."""