fix(strix): route direct OpenAI fallback through its API base (replacement)

.github/workflows/opencode-review-dispatch.yml

@@ -4477,6 +4477,10 @@ jobs:
           # so the OpenRouter slots use cheap paid models billed against the
           # org's OpenRouter credits), then the full-size GPT-4.1 long-context
           # endpoint and provider-specific GPT/o3 fallbacks.
+          # The direct-OpenAI slot runs GPT-5.4: the current valid model name
+          # on the direct OpenAI platform (the retired gpt-5.6-luna catalog
+          # entry returned 404s), kept cost-efficient against the org's
+          # OpenAI budget between top-ups.
           OPENCODE_MODEL_CANDIDATES: "${{ needs.validate-pr-metadata.outputs.is_private == 'false' && 'nvidia-nim/nvidia/llama-3.3-nemotron-super-49b-v1.5 nvidia-nim/nvidia/llama-3.1-nemotron-ultra-253b-v1 nvidia-nim/nvidia/nemotron-3-super-120b-a12b nvidia-nim/nvidia/nemotron-3-ultra-550b-a55b nvidia-nim/meta/llama-3.3-70b-instruct nvidia-nim/deepseek-ai/deepseek-v4-pro nvidia-nim/mistralai/codestral-22b-instruct-v0.1 opencode-free/nemotron-3-ultra-free opencode-free/deepseek-v4-flash-free opencode-free/north-mini-code-free opencode-free/laguna-s-2.1-free opencode-free/ling-3.0-flash-free opencode-free/big-pickle opencode-free/mimo-v2.5-free opencode-free/hy3-free opencode-free/minimax-m3-free opencode-free/glm-5-free opencode-free/kimi-k2.5-free opencode-free/qwen3.6-plus-free ' || '' }}opencode/gpt-5.6-terra github-models/deepseek/deepseek-v3-0324 openai/gpt-5.4 openrouter/deepseek/deepseek-v3.2 openrouter/qwen/qwen3-coder github-models/openai/gpt-4.1 github-models/openai/gpt-5 github-models/openai/gpt-5-chat github-models/openai/o3 github-models/deepseek/deepseek-r1-0528 github-models/deepseek/deepseek-r1"
           # One attempt per model, then fall through to the next model. Retrying
           # the SAME model 5x let a rate-limited/hung leader consume the whole

.github/workflows/strix.yml

@@ -693,7 +693,10 @@ jobs:
           echo "LLM_API_BASE_FILE=$llm_api_base_file" >> "$GITHUB_ENV"
 
       - name: Prepare GitHub Models fallback credentials
-        if: steps.gate.outputs.provider_mode == 'openai_direct' || steps.gate.outputs.provider_mode == 'openrouter' || steps.gate.outputs.provider_mode == 'nvidia_nim'
+        # github_models is included because its STRIX_FALLBACK_MODELS chain
+        # ends in openai-direct/gpt-5.4, which needs the direct-OpenAI key and
+        # API base to authenticate and route after the primary is exhausted.
+        if: steps.gate.outputs.provider_mode == 'openai_direct' || steps.gate.outputs.provider_mode == 'openrouter' || steps.gate.outputs.provider_mode == 'nvidia_nim' || steps.gate.outputs.provider_mode == 'github_models'
         env:
           GITHUB_MODELS_FALLBACK_TOKEN: ${{ secrets.STRIX_GITHUB_MODELS_TOKEN || github.token }}
           OPENAI_FALLBACK_KEY: ${{ secrets.STRIX_OPENAI_API_KEY || secrets.OPENAI_API_KEY }}
@@ -725,6 +728,9 @@ jobs:
             openai_fallback_key_file="$RUNNER_TEMP/openai_fallback_key.txt"
             printf '%s' "$openai_trimmed" > "$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"
           fi
 
       - name: Prepare Vertex AI credentials
@@ -857,6 +863,7 @@ jobs:
           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_KEY_FILE: ${{ env.STRIX_OPENAI_FALLBACK_KEY_FILE }}
+          STRIX_OPENAI_FALLBACK_API_BASE_FILE: ${{ env.STRIX_OPENAI_FALLBACK_API_BASE_FILE }}
           STRIX_FAIL_ON_PROVIDER_SIGNAL: "1"
           STRIX_VERTEX_FALLBACK_MODELS: ""
           NPM_CONFIG_IGNORE_SCRIPTS: "true"

CHANGELOG.md

@@ -5,7 +5,17 @@ this file. The format follows Keep a Changelog, and versioned releases follow
 Semantic Versioning where the repository publishes a release.
 
 ## [Unreleased]
-
+- Route Strix cross-provider fallbacks to explicit direct-OpenAI models
+  (`openai-direct/...`) through the OpenAI inference endpoint instead of
+  inheriting the primary provider's `LLM_API_BASE`: the workflow now
+  provisions `STRIX_OPENAI_FALLBACK_API_BASE_FILE`
+  (`https://api.openai.com/v1`), the gate prefers it for explicit OpenAI
+  fallback models and otherwise resolves no override so litellm defaults to
+  OpenAI, and a non-https override fails configuration. This removes the
+  NVIDIA-NIM-edge `404 page not found` that made the contracted final
+  fallback unreachable after NIM exhaustion.
+- Align stale `gpt-5.6-luna` test expectations with the valid `gpt-5.4`
+  contract left behind by the earlier model rename.
 - Honor each trusted base project's exact, integrity-bearing pnpm
   `packageManager` specification in OpenCode coverage images through the pinned
   Node distribution's Corepack runtime, instead of admitting the specification

docs/doctoring/strix-openai-fallback-api-base-routing.md

@@ -0,0 +1,74 @@
+# Strix direct-OpenAI fallback API-base routing: evidence and design record
+
+## Decision
+
+A Strix cross-provider fallback to an explicit direct-OpenAI model
+(`openai-direct/...` or `openai_direct/...`) must route through the OpenAI
+inference endpoint, never through the primary provider's `LLM_API_BASE`. The
+gate now prefers an explicit `STRIX_OPENAI_FALLBACK_API_BASE_FILE` for such
+models and, when the override is absent, resolves no API-base override so
+litellm uses its default `https://api.openai.com/v1` endpoint.
+
+The central workflow writes `https://api.openai.com/v1` into
+`$RUNNER_TEMP/openai_fallback_api_base.txt` and exports
+`STRIX_OPENAI_FALLBACK_API_BASE_FILE` whenever it publishes the OpenAI
+fallback key file, so every provider chain that ends in
+`openai-direct/gpt-5.4` (NVIDIA NIM primary, OpenRouter primary,
+GitHub Models primary) inherits correct routing automatically.
+
+## Failure this fixes
+
+Required-CI evidence (BandScope PR #1021 strix run 32800796577, 2026-08-25)
+showed the NVIDIA NIM primary and first fallback exhausting provider
+availability, then the contracted final fallback `openai-direct/gpt-5.4`
+failing with a plain-text gateway error:
+
+```text
+LLM CONNECTION FAILED
+Could not establish connection to the language model.
+Error: 404 page not found
+```
+
+Root cause: with `provider_mode=nvidia_nim`, the workflow sets
+`LLM_API_BASE_FILE=https://integrate.api.nvidia.com/v1`. The gate reused that
+base for the openai-direct fallback child, so litellm sent OpenAI requests to
+the NVIDIA NIM edge, whose Go gateway answered `404 page not found`. The
+fallback key was already routed correctly (`STRIX_OPENAI_FALLBACK_KEY_FILE`);
+only the base URL leaked from the primary provider. Because no vulnerability
+report artifact was produced, the gate failed closed — correct policy on an
+incomplete scan, but caused by routing rather than by any repository finding.
+
+## Trust boundary
+
+The override is a runner-provisioned regular file under `$RUNNER_TEMP`,
+resolved through the same `resolve_trusted_input_file` boundary as the other
+API-base files: it must be a regular non-symlink file inside the trusted input
+root, must trim to a single `https://` URL, and must not contain whitespace or
+control characters. Absent or empty overrides fall back to litellm's default
+endpoint instead of failing configuration, preserving standalone local gate
+runs where no workflow provisioning exists.
+
+## Verification contract
+
+Regression evidence proves that:
+
+1. with a NVIDIA NIM primary base configured, `openai-direct/gpt-5.4`
+   resolves through the explicit OpenAI fallback base when provided;
+2. without an override, the resolver returns no base so litellm defaults to
+   `https://api.openai.com/v1`;
+3. NVIDIA NIM primary attempts keep resolving through the NIM edge;
+4. `github_models/*` fallbacks keep their dedicated GitHub Models endpoint;
+5. a non-https override fails configuration (exit 2) instead of scanning;
+6. the workflow provisions the override file and passes it into the gate env;
+7. the required-workflow smoke contract pins both sides of the wiring; and
+8. the stale `gpt-5.6-luna` expectations left behind by the model rename are
+   aligned with the valid `gpt-5.4` contract in queue-contract tests.
+
+## Limitations
+
+This change restores reachability of the final fallback; it does not create
+OpenAI quota. If the OpenAI key is absent or exhausted after NIM exhaustion,
+the gate still fails closed as provider-unavailable — by design, because no
+complete authoritative scan exists. Hosted model catalogs may also change
+independently of this repository; model-name updates remain manual contract
+changes reviewed through CI.

scripts/ci/strix_quick_gate.sh

@@ -2426,15 +2426,20 @@ resolved_llm_api_base_for_model() {
 	if is_vertex_model "$model"; then
 		return 0
 	fi
-	if is_explicit_openai_model "$model" && ! is_explicit_openai_model "$PRIMARY_MODEL"; then
-		# A direct-OpenAI fallback must not inherit a foreign primary provider's
-		# endpoint (for example NVIDIA NIM or OpenRouter).
-		return 0
-	fi
-
-	local api_base_file="$LLM_API_BASE_FILE"
+	local api_base_file="${LLM_API_BASE_FILE:-}"
 	local api_base_file_name="LLM_API_BASE_FILE"
-	if is_github_models_model "$model" && [ -n "${STRIX_GITHUB_MODELS_API_BASE_FILE:-}" ]; then
+	if is_explicit_openai_model "$model" && [ -n "${STRIX_OPENAI_FALLBACK_API_BASE_FILE:-}" ]; then
+		# Cross-provider fallback: openai-direct/* candidates must reach the
+		# direct OpenAI API even when the primary provider selected a
+		# different LLM_API_BASE_FILE endpoint (e.g. NVIDIA NIM). Without
+		# this the fallback hits the primary gateway and 404s.
+		api_base_file="$STRIX_OPENAI_FALLBACK_API_BASE_FILE"
+		api_base_file_name="STRIX_OPENAI_FALLBACK_API_BASE_FILE"
+		# The workflow always provisions this file for cross-provider fallbacks.
+		# In standalone runs, an explicitly supplied LLM_API_BASE_FILE remains
+		# a caller-owned custom OpenAI-compatible endpoint rather than being
+		# silently discarded.
+	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.

scripts/ci/test_strix_quick_gate.sh

@@ -374,6 +374,8 @@ assert_strix_workflow_pr_trigger_hardened() {
 	assert_file_contains "$workflow_file" "steps.gate.outputs.provider_mode == 'nvidia_nim' && 'nvidia_nim/nvidia/llama-3.3-nemotron-super-49b-v1.5 openai-direct/gpt-5.4'" "strix workflow gives NVIDIA NIM scans contracted fallbacks"
 	assert_file_not_contains "$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" "STRIX_OPENAI_FALLBACK_API_BASE_FILE" "strix workflow routes direct-OpenAI fallbacks through a trusted API base file"
+	assert_file_contains "$workflow_file" "https://api.openai.com/v1" "strix workflow uses the OpenAI platform endpoint for direct fallbacks"
 	assert_file_contains "$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_pr_review_autofix_nvidia_nim_contract.py

@@ -19,7 +19,7 @@
 DOCTORING_RECORD = Path("docs/doctoring/hourly-nvidia-nim-autofix.md")
 CHANGELOG = Path("CHANGELOG.md")
 REVIEW_DISPATCH_WORKFLOW = Path(".github/workflows/opencode-review-dispatch.yml")
-REVIEW_DISPATCH_BLOB_SHA = "ed3f7b44f9afdd6ab295426e5d0440aeca6bdfb5"
+REVIEW_DISPATCH_BLOB_SHA = "478788d41ed3e0d8389292a059d225336a925830"
 
 
 def _workflow_text(path: Path) -> str: