fix(strix): fall back when NVIDIA NIM retires a hosted model

.github/workflows/strix.yml

@@ -443,7 +443,7 @@ jobs:
       - name: Gate Strix secrets
         id: gate
         env:
-          STRIX_MODEL: ${{ github.event.client_payload.strix_llm || (steps.target_visibility.outputs.is_private == 'false' && 'nvidia_nim/nvidia/nemotron-3-ultra-550b-a55b' || 'gpt-5.6-luna') }}
+          STRIX_MODEL: ${{ github.event.client_payload.strix_llm || (steps.target_visibility.outputs.is_private == 'false' && 'nvidia_nim/nvidia/nemotron-3-super-120b-a12b' || 'gpt-5.6-luna') }}
           STRIX_MODEL_REQUESTED: ${{ github.event.client_payload.strix_llm || '' }}
           STRIX_OPENAI_API_KEY: ${{ secrets.STRIX_OPENAI_API_KEY || secrets.OPENAI_API_KEY }}
           STRIX_OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }}
@@ -453,7 +453,7 @@ jobs:
           TARGET_REPOSITORY_PRIVATE: ${{ steps.target_visibility.outputs.is_private }}
         run: |
           strix_model="$(printf '%s' "$STRIX_MODEL" | sed 's/^[[:space:]]*//;s/[[:space:]]*$//')"
-          if [ -z "$STRIX_MODEL_REQUESTED" ] && [ "$strix_model" = "nvidia_nim/nvidia/nemotron-3-ultra-550b-a55b" ] && [ -z "${STRIX_NVIDIA_NIM_API_KEY:-}" ]; then
+          if [ -z "$STRIX_MODEL_REQUESTED" ] && [ "$strix_model" = "nvidia_nim/nvidia/nemotron-3-super-120b-a12b" ] && [ -z "${STRIX_NVIDIA_NIM_API_KEY:-}" ]; then
             strix_model="gpt-5.6-luna"
           fi
           echo "strix_model=$strix_model" >> "$GITHUB_OUTPUT"
@@ -497,7 +497,7 @@ jobs:
                 exit 1
               fi
               ;;
-            nvidia_nim/nvidia/nemotron-3-ultra-550b-a55b)
+            nvidia_nim/nvidia/nemotron-3-super-120b-a12b)
               if [ "$TARGET_REPOSITORY_PRIVATE" != "false" ]; then
                 echo '::error::NVIDIA NIM hosted trial scans are limited to public repositories.'
                 exit 1
@@ -763,7 +763,7 @@ jobs:
             openrouter/free | openrouter/openrouter/free)
               printf '%s' 'openrouter/free' > "$strix_llm_file"
               ;;
-            nvidia_nim/nvidia/nemotron-3-ultra-550b-a55b)
+            nvidia_nim/nvidia/nemotron-3-super-120b-a12b)
               printf '%s' "$strix_model" > "$strix_llm_file"
               ;;
             vertex_ai/gemini-3.1-pro-preview-customtools | vertex_ai/gemini-2.5-flash)
@@ -807,7 +807,7 @@ jobs:
           STRIX_LLM_MAX_RETRIES: 1
           STRIX_TRANSIENT_RETRY_PER_MODEL: 2
           STRIX_TRANSIENT_RETRY_BACKOFF_SECONDS: 60
-          STRIX_FALLBACK_MODELS: ${{ steps.gate.outputs.provider_mode == 'github_models' && 'github_models/openai/o3 github_models/openai/gpt-5-chat' || steps.gate.outputs.provider_mode == 'openai_direct' && 'github_models/openai/o3 github_models/openai/gpt-5-chat' || steps.gate.outputs.provider_mode == 'openrouter' && 'github_models/openai/o3 github_models/openai/gpt-5-chat' || steps.gate.outputs.provider_mode == 'nvidia_nim' && 'github_models/openai/o3 github_models/openai/gpt-5-chat' || '' }}
+          STRIX_FALLBACK_MODELS: ${{ steps.gate.outputs.provider_mode == 'github_models' && 'github_models/openai/o3 github_models/openai/gpt-5-chat' || steps.gate.outputs.provider_mode == 'openai_direct' && 'github_models/openai/o3 github_models/openai/gpt-5-chat' || steps.gate.outputs.provider_mode == 'openrouter' && 'github_models/openai/o3 github_models/openai/gpt-5-chat' || steps.gate.outputs.provider_mode == 'nvidia_nim' && 'nvidia_nim/nvidia/llama-3.3-nemotron-super-49b-v1.5 github_models/openai/o3 github_models/openai/gpt-5-chat' || '' }}
           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_FAIL_ON_PROVIDER_SIGNAL: "1"
@@ -861,7 +861,7 @@ jobs:
           fi
 
           # Recognized signals that the LLM backend was unavailable / starved.
-          backend_unavailable_signal='RateLimitError|Too many requests\. For more on scraping GitHub|exceeded your current quota|insufficient_quota|billing details|"status"[[:space:]]*:[[:space:]]*"RESOURCE_EXHAUSTED"|tokens_limit_reached|Request body too large|Max size:[[:space:]]*[0-9]+[[:space:]]+tokens|Error code:[[:space:]]*413|LLM CONNECTION FAILED|Could not establish connection to the language model|LLM warm-up failed|Configured model and fallback models were unavailable|Configured Vertex model and fallback models were unavailable|emitted provider infrastructure or failure-signal output|before provider infrastructure failure'
+          backend_unavailable_signal='RateLimitError|Too many requests\. For more on scraping GitHub|exceeded your current quota|insufficient_quota|billing details|"status"[[:space:]]*:[[:space:]]*"RESOURCE_EXHAUSTED"|tokens_limit_reached|Request body too large|Max size:[[:space:]]*[0-9]+[[:space:]]+tokens|Error code:[[:space:]]*413|LLM CONNECTION FAILED|Could not establish connection to the language model|LLM warm-up failed|Configured model and fallback models were unavailable|Configured Vertex model and fallback models were unavailable|emitted provider infrastructure or failure-signal output|before provider infrastructure failure|litellm(\.exceptions)?\.NotFoundError[^[:cntrl:]]*Nvidia_nimException[^[:cntrl:]]*Error code:[[:space:]]*404'
           # Any evidence that a vulnerability was actually reported. Its presence
           # forces a hard failure so real findings are NEVER downgraded. Keep the
           # severity branch anchored away from identifiers so environment lines

CHANGELOG.md

@@ -0,0 +1,11 @@
+# Changelog
+
+All notable changes to the organization automation repository are documented in
+this file. The format follows Keep a Changelog, and versioned releases follow
+Semantic Versioning where the repository publishes a release.
+
+## [Unreleased]
+
+### Fixed
+
+- Made Strix treat only a single LiteLLM provider-error line containing NVIDIA NIM context and model-catalog 404 evidence as cross-model fallback evidence, rejecting cross-line signal assembly and provider-like target source literals; moved the public default to Nemotron 3 Super 120B and added a second NVIDIA hosted candidate before GitHub Models without neutralizing reported vulnerabilities.

docs/doctoring/strix-nvidia-nim-not-found-fallback.md

@@ -0,0 +1,76 @@
+# Strix NVIDIA NIM model-catalog fallback: evidence and design record
+
+## Decision
+
+Strix treats an authenticated NVIDIA NIM model-catalog `404 Not Found` as
+provider availability evidence, not as a target-application vulnerability. The
+gate does not retry the same unavailable model. It proceeds to a distinct
+reviewed NVIDIA hosted model and only then to the existing GitHub Models
+candidates.
+
+Public-repository scans now default to
+`nvidia/nemotron-3-super-120b-a12b`. The first fallback is
+`nvidia/llama-3.3-nemotron-super-49b-v1.5`. Private repositories retain the
+contracted provider because NVIDIA hosted trial inputs are restricted to public
+repositories by the central workflow.
+
+## Trust boundary
+
+The NVIDIA classifier accepts only a single bounded log line that contains all
+three signals: a LiteLLM `NotFoundError`, NVIDIA NIM provider context, and
+model-catalog not-found evidence. It does not assemble provider and `404`
+signals from different lines. A bare application `404`, route miss, database
+lookup miss, provider-like source literal, or other target-controlled output is
+not enough to enter model fallback.
+
+This same-line rule matters because scanner stdout can include text derived from
+the repository under review. Requiring the trusted LiteLLM exception marker and
+all provider-availability evidence on one line prevents repository content from
+combining with an unrelated application `404` to spoof infrastructure fallback.
+Provider-side failure also remains a fail-closed incomplete scan until a distinct
+fallback produces complete evidence.
+
+The outer workflow may classify exhausted provider infrastructure as neutral only
+when the run log contains no vulnerability signal. Any reported severity or
+non-zero vulnerability count remains blocking. Scanner reports and attempt logs
+remain available as artifacts.
+
+## Verification contract
+
+Regression evidence proves that:
+
+1. the exact LiteLLM `Nvidia_nimException` 404 observed in required CI is
+   recognized;
+2. an ordinary application 404 is not recognized;
+3. provider context and 404 evidence on different lines are not recognized;
+4. a provider-like source literal on one line without LiteLLM `NotFoundError`
+   context is not recognized;
+5. model-catalog 404s enter cross-model fallback but never same-model retry;
+6. the primary and first fallback are current NVIDIA hosted models;
+7. GitHub Models remain later cross-provider fallbacks;
+8. vulnerability signals prevent neutral infrastructure classification; and
+9. the required-workflow smoke contract pins these properties.
+
+## Limitations
+
+Hosted model catalogs may change independently of this repository. A model-card
+page or supported self-hosted NIM container does not guarantee indefinite hosted
+trial availability. The ordered model plan must therefore be reviewed against
+current NVIDIA documentation whenever a provider returns a catalog 404. This
+change does not treat arbitrary provider errors as success and does not weaken
+Strix severity, changed-file attribution, or independent approval requirements.
+
+## References
+
+Fielding, R., Nottingham, M., & Reschke, J. (2022). *HTTP semantics* (RFC
+9110). Internet Engineering Task Force. https://doi.org/10.17487/RFC9110
+
+NVIDIA Corporation. (2025). *Llama-3.3-Nemotron-Super-49B-v1.5* [Model card].
+NVIDIA NIM. https://build.nvidia.com/nvidia/llama-3_3-nemotron-super-49b-v1_5/modelcard
+
+NVIDIA Corporation. (2026a). *NVIDIA-Nemotron-3-Super-120B-A12B* [Model
+card]. NVIDIA NIM.
+https://build.nvidia.com/nvidia/nemotron-3-super-120b-a12b/modelcard
+
+NVIDIA Corporation. (2026b). *Configuration reference*. NVIDIA AI-Q Blueprint.
+https://docs.nvidia.com/aiq-blueprint/2.2.0-rc1/customization/configuration-reference.html

scripts/ci/strix_quick_gate.sh

@@ -2641,6 +2641,20 @@ is_llm_service_unavailable_error() {
 	return 1
 }
 
+is_nvidia_nim_not_found_error() {
+	# Classify only one bounded LiteLLM provider-error line that also
+	# carries NVIDIA NIM context and model-catalog not-found evidence.
+	# Cross-line signal assembly and provider-like target source text
+	# remain non-retryable so application output cannot spoof fallback.
+	if grep -Ei 'litellm(\.exceptions)?\.NotFoundError' "$STRIX_LOG" |
+		grep -Ei '(Nvidia_nimException|nvidia[_ -]?nim|integrate\.api\.nvidia\.com)' |
+		grep -Eiq '(Error code:[[:space:]]*404|(^|[^0-9])404([^0-9]|$)|model[^[:alnum:]]+not found)'; then
+		return 0
+	fi
+
+	return 1
+}
+
 ## Determines whether the last strix failure is a transient error eligible
 ## for same-model retry (up to STRIX_TRANSIENT_RETRY_PER_MODEL times).
 ## Four error families qualify:
@@ -2909,7 +2923,7 @@ is_midstream_fallback_error() {
 # (httpx, httpcore, requests). Used for generic transport failures where
 # library names alone are insufficient to prove the timeout/connection error
 # originated from an LLM provider rather than the target application.
-LLM_PROVIDER_ONLY_REGEX='(litellm|openai|anthropic|VertexAI|Vertex_ai|vertex\.ai|google\.cloud|GitHub Models|models\.github\.ai|github_models)'
+LLM_PROVIDER_ONLY_REGEX='(litellm|openai|anthropic|VertexAI|Vertex_ai|vertex\.ai|google\.cloud|Nvidia_nimException|nvidia_nim|integrate\.api\.nvidia\.com|GitHub Models|models\.github\.ai|github_models)'
 
 is_llm_token_limit_error() {
 	if grep -Eiq '(tokens_limit_reached|Request body too large|Max size:[[:space:]]*[0-9]+[[:space:]]+tokens|Error code:[[:space:]]*413|(^|[^0-9])413([^0-9]|$))' "$STRIX_LOG" &&
@@ -2953,6 +2967,10 @@ has_detected_infrastructure_error() {
 		return 0
 	fi
 
+	if is_nvidia_nim_not_found_error; then
+		return 0
+	fi
+
 	# Generic strix non-zero exit with known transport/connection errors
 	# that don't fall into the specific categories above.
 	# Use LLM_PROVIDER_ONLY_REGEX (not PROVIDER_CONTEXT_REGEX) to avoid
@@ -3799,6 +3817,10 @@ is_model_retryable_error() {
 		return 0
 	fi
 
+	if is_nvidia_nim_not_found_error; then
+		return 0
+	fi
+
 	if is_github_models_api_compatible_model "$model" && is_github_models_unavailable_model_error; then
 		return 0
 	fi

scripts/ci/strix_required_workflow_smoke.sh

@@ -155,6 +155,11 @@ assert_file_contains "$gate_script" "TARGET_PATH_IS_INTERNAL_PR_SCOPE" "Strix ga
 assert_file_contains "$gate_script" "NPM_CONFIG_IGNORE_SCRIPTS" "Strix gate disables npm lifecycle scripts"
 assert_file_contains "$full_gate_test" "assert_strix_workflow_pr_trigger_hardened" "Full Strix harness remains available outside the required path"
 
+assert_file_contains "$workflow_file" "nvidia_nim/nvidia/nemotron-3-super-120b-a12b" "Strix defaults public scans to the current hosted NVIDIA NIM model"
+assert_file_contains "$workflow_file" "nvidia_nim/nvidia/llama-3.3-nemotron-super-49b-v1.5 github_models/openai/o3 github_models/openai/gpt-5-chat" "Strix tries another NVIDIA hosted model before GitHub Models"
+assert_file_contains "$workflow_file" "Nvidia_nimException" "Strix workflow recognizes provider-scoped NVIDIA NIM failures"
+assert_file_contains "$gate_script" "is_nvidia_nim_not_found_error" "Strix gate classifies NVIDIA NIM model-catalog 404s"
+
 if [ "$failures" -ne 0 ]; then
 	echo "Strix required workflow smoke test failed with $failures failure(s)." >&2
 	exit 1

scripts/ci/test_strix_quick_gate.sh

@@ -289,8 +289,8 @@ assert_strix_workflow_pr_trigger_hardened() {
 	assert_file_not_contains "$workflow_file" "STRIX_TOTAL_TIMEOUT_SECONDS:" "strix workflow must not expose total timeout env names in GitHub logs"
 	assert_file_not_contains "$workflow_file" "STRIX_PR_SCOPE_MAX_FILES_PER_BATCH" "strix workflow must not split Strix PR evidence into separate scanner runs"
 	assert_file_not_contains "$workflow_file" "secrets.STRIX_LLM == 'vertex_ai/gemini-3.1-pro-preview-customtools' && 'vertex_ai/gemini-2.5-flash'" "strix workflow must not quarantine the approved Vertex preview model after organization secret visibility is fixed"
-	assert_file_contains "$workflow_file" "steps.target_visibility.outputs.is_private == 'false' && 'nvidia_nim/nvidia/nemotron-3-ultra-550b-a55b' || 'gpt-5.6-luna'" "strix workflow defaults public scans to NVIDIA NIM and keeps private scans on the contracted provider"
-	assert_file_contains "$workflow_file" 'if [ -z "$STRIX_MODEL_REQUESTED" ] && [ "$strix_model" = "nvidia_nim/nvidia/nemotron-3-ultra-550b-a55b" ] && [ -z "${STRIX_NVIDIA_NIM_API_KEY:-}" ]' "strix workflow falls back to the contracted provider when the NVIDIA secret is absent"
+	assert_file_contains "$workflow_file" "steps.target_visibility.outputs.is_private == 'false' && 'nvidia_nim/nvidia/nemotron-3-super-120b-a12b' || 'gpt-5.6-luna'" "strix workflow defaults public scans to NVIDIA NIM and keeps private scans on the contracted provider"
+	assert_file_contains "$workflow_file" 'if [ -z "$STRIX_MODEL_REQUESTED" ] && [ "$strix_model" = "nvidia_nim/nvidia/nemotron-3-super-120b-a12b" ] && [ -z "${STRIX_NVIDIA_NIM_API_KEY:-}" ]' "strix workflow falls back to the contracted provider when the NVIDIA secret is absent"
 	assert_file_contains "$workflow_file" 'STRIX_MODEL: ${{ steps.gate.outputs.strix_model }}' "strix workflow propagates the gate-selected fallback model to the scanner"
 	assert_file_not_contains "$workflow_file" "secrets.STRIX_LLM ||" "strix workflow must not let the legacy STRIX_LLM secret override PR defaults"
 	assert_file_contains "$workflow_file" "STRIX_LLM must select NVIDIA NIM Nemotron, GitHub Models openai/gpt-5 or newer, direct OpenAI GPT-5.4 or newer, OpenRouter openrouter/free, or an approved organization Vertex AI model" "strix workflow rejects unsupported model inputs"
@@ -348,7 +348,7 @@ assert_strix_workflow_pr_trigger_hardened() {
 	assert_file_not_contains "$workflow_file" '${{ secrets.STRIX_OPENAI_API_KEY || github.token }}' "strix workflow must not use fallback-secret syntax for LLM API keys"
 	assert_file_contains "$workflow_file" "github_models/openai/o3 github_models/openai/gpt-5-chat" "strix workflow keeps GitHub Models fallback on tool-capable OpenAI models without GPT-4.1 downgrade"
 	assert_file_contains "$workflow_file" "steps.gate.outputs.provider_mode == 'openai_direct' && 'github_models/openai/o3 github_models/openai/gpt-5-chat'" "strix workflow gives direct-OpenAI scans GitHub Models fallbacks so provider quota outages degrade instead of skipping"
-	assert_file_contains "$workflow_file" "steps.gate.outputs.provider_mode == 'nvidia_nim' && 'github_models/openai/o3 github_models/openai/gpt-5-chat'" "strix workflow gives NVIDIA NIM scans contracted fallbacks"
+	assert_file_contains "$workflow_file" "steps.gate.outputs.provider_mode == 'nvidia_nim' && 'nvidia_nim/nvidia/llama-3.3-nemotron-super-49b-v1.5 github_models/openai/o3 github_models/openai/gpt-5-chat'" "strix workflow gives NVIDIA NIM scans contracted fallbacks"
 	assert_file_contains "$workflow_file" "Prepare GitHub Models fallback credentials" "strix workflow provisions GitHub Models fallback credentials for direct-OpenAI 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"