From 5ecbebb0d4bee769826302c24ac00a00c799521b Mon Sep 17 00:00:00 2001
From: seonghobae <seonghobae@users.noreply.github.com>
Date: Mon, 24 Aug 2026 18:56:58 +0900
Subject: [PATCH 01/11] fix(strix): route openai-direct fallback to the OpenAI
 API base

0c6b9a64 mapped the openai-direct prefix and routed the OpenAI key,
but the child scan still read LLM_API_BASE_FILE -- the primary
provider's endpoint. Observed on LineageWeave#570 (run 32701426812):
after both NVIDIA models 429'd, the openai-direct/gpt-5.6-luna
fallback reached integrate.api.nvidia.com with an unknown model path
and died in 4s with '404 page not found'.

When the candidate is explicit-openai and
STRIX_OPENAI_FALLBACK_API_BASE_FILE is configured, select that file as
the api-base source, exactly mirroring the existing GitHub Models
cross-provider routing. Key routing was already correct.
---
 scripts/ci/strix_quick_gate.sh | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/scripts/ci/strix_quick_gate.sh b/scripts/ci/strix_quick_gate.sh
index 36ec3e5f8..fa676a0d5 100755
--- a/scripts/ci/strix_quick_gate.sh
+++ b/scripts/ci/strix_quick_gate.sh
@@ -2417,7 +2417,14 @@ 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 "${STRIX_GITHUB_MODELS_API_BASE_FILE:-}" ]; then
+	if is_explicit_openai_model "$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"
+	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.

From 84eb37943d2b8ad57fd405c858721f71e3144ca6 Mon Sep 17 00:00:00 2001
From: Seongho Bae <me@seonghobae.me>
Date: Mon, 24 Aug 2026 19:15:47 +0900
Subject: [PATCH 02/11] fix(strix): wire direct OpenAI fallback endpoint

---
 .github/workflows/strix.yml         | 4 ++++
 scripts/ci/test_strix_quick_gate.sh | 2 ++
 2 files changed, 6 insertions(+)

diff --git a/.github/workflows/strix.yml b/.github/workflows/strix.yml
index c8c054e42..8eb5be301 100644
--- a/.github/workflows/strix.yml
+++ b/.github/workflows/strix.yml
@@ -706,6 +706,9 @@ 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
@@ -838,6 +841,7 @@ 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: ${{ env.STRIX_OPENAI_FALLBACK_API_BASE_FILE }}
           STRIX_FAIL_ON_PROVIDER_SIGNAL: "1"
           STRIX_VERTEX_FALLBACK_MODELS: ""
           NPM_CONFIG_IGNORE_SCRIPTS: "true"
diff --git a/scripts/ci/test_strix_quick_gate.sh b/scripts/ci/test_strix_quick_gate.sh
index 945eb3fb3..cc601e5dc 100755
--- a/scripts/ci/test_strix_quick_gate.sh
+++ b/scripts/ci/test_strix_quick_gate.sh
@@ -365,6 +365,8 @@ assert_strix_workflow_pr_trigger_hardened() {
 	assert_file_contains "$workflow_file" "steps.gate.outputs.provider_mode == 'nvidia_nim' && 'nvidia_nim/nvidia/llama-3.3-nemotron-super-49b-v1.5 openai-direct/gpt-5.6-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 "$workflow_file" "STRIX_OPENAI_FALLBACK_API_BASE_FILE" "strix workflow routes direct-OpenAI fallbacks through a trusted API base file"
+	assert_file_contains "$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"

From 689212bebd5e84d67f263ea04648df516b238507 Mon Sep 17 00:00:00 2001
From: Seongho Bae <me@seonghobae.me>
Date: Tue, 25 Aug 2026 15:20:26 +0900
Subject: [PATCH 03/11] test(strix): pin direct-OpenAI fallback base routing
 and default-endpoint contract

Adds a regression contract for the openai-direct fallback API-base fix:
override file routes to https://api.openai.com/v1, absence resolves no
override (litellm default endpoint) so explicit OpenAI models can never
inherit a foreign primary gateway, NVIDIA/GitHub Models primaries keep
their bases, and non-https overrides fail configuration. Also records the
doctoring evidence and CHANGELOG entry for the routing fix.
---
 CHANGELOG.md                                  |  34 +++
 .../strix-openai-fallback-api-base-routing.md |  74 ++++++
 scripts/ci/strix_quick_gate.sh                |   7 +
 tests/test_strix_openai_fallback_api_base.py  | 211 ++++++++++++++++++
 4 files changed, 326 insertions(+)
 create mode 100644 docs/doctoring/strix-openai-fallback-api-base-routing.md
 create mode 100644 tests/test_strix_openai_fallback_api_base.py

diff --git a/CHANGELOG.md b/CHANGELOG.md
index 1630c32d4..00c525c80 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -5,6 +5,40 @@ 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
+  during materialization and then rejecting every version except pnpm 11.5.3;
+  route generic coverage and docstring package scripts through the same
+  Corepack boundary instead of invoking a removed bare `pnpm` binary.
+- Keep `--trust-lockfile` only for pnpm 11.3 and newer
+  (`trustLockfile` landed in pnpm 11.3). pnpm 9, 10, and 11.0–11.2 reject
+  that flag and previously failed LineageWeave JavaScript coverage before
+  tests could run. Jest test scripts still receive `--coverage` because Jest
+  documents a native coverage flag.
+- Run declared JavaScript test scripts without synthesizing `--coverage` when
+  the package does not declare a compatible coverage command, but keep the
+  coverage result failed until the repository adds a lock-pinned provider and
+  owned coverage command. A generic `c8`, `nyc`, or Istanbul dependency no
+  longer makes an unrelated test runner receive an unsupported flag.
+- Fix OpenCode coverage evidence for exact-base, organization-owned Python VCS
+  dependencies without weakening registry hashes or the networkless PR sandbox,
+  reject namespace, ambiguous, linked, native-extension, and installed-metadata
+  layouts, and make exact roots readable by the unprivileged coverage user.
+
+
 
 - Honor each trusted base project's exact, integrity-bearing pnpm
   `packageManager` specification in OpenCode coverage images through the pinned
diff --git a/docs/doctoring/strix-openai-fallback-api-base-routing.md b/docs/doctoring/strix-openai-fallback-api-base-routing.md
new file mode 100644
index 000000000..4f7112303
--- /dev/null
+++ b/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
+`$RUNNER_TEMP/openai_fallback_api_base.txt` and exports
+`STRIX_OPENAI_FALLBACK_API_BASE_FILE` whenever it publishes the OpenAI
+fallback key file, so every provider chain that ends in
+`openai-direct/gpt-5.4` (NVIDIA NIM primary, OpenRouter primary,
+GitHub Models primary) inherits correct routing automatically.
+
+## Failure this fixes
+
+Required-CI evidence (BandScope PR #1021 strix run 32800796577, 2026-08-25)
+showed the NVIDIA NIM primary and first fallback exhausting provider
+availability, then the contracted final fallback `openai-direct/gpt-5.4`
+failing with a plain-text gateway error:
+
+```
+LLM CONNECTION FAILED
+Could not establish connection to the language model.
+Error: 404 page not found
+```
+
+Root cause: with `provider_mode=nvidia_nim`, the workflow sets
+`LLM_API_BASE_FILE=https://integrate.api.nvidia.com/v1`. The gate reused that
+base for the openai-direct fallback child, so litellm sent OpenAI requests to
+the NVIDIA NIM edge, whose Go gateway answered `404 page not found`. The
+fallback key was already routed correctly (`STRIX_OPENAI_FALLBACK_KEY_FILE`);
+only the base URL leaked from the primary provider. Because no vulnerability
+report artifact was produced, the gate failed closed — correct policy on an
+incomplete scan, but caused by routing rather than by any repository finding.
+
+## Trust boundary
+
+The override is a runner-provisioned regular file under `$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.
diff --git a/scripts/ci/strix_quick_gate.sh b/scripts/ci/strix_quick_gate.sh
index fa676a0d5..a7493192e 100755
--- a/scripts/ci/strix_quick_gate.sh
+++ b/scripts/ci/strix_quick_gate.sh
@@ -2415,6 +2415,13 @@ resolved_llm_api_base_for_model() {
 		return 0
 	fi
 
+	if is_explicit_openai_model "$model" && [ -z "${STRIX_OPENAI_FALLBACK_API_BASE_FILE:-}" ]; then
+		# Without workflow provisioning, an explicit direct-OpenAI model must
+		# still not inherit a foreign primary base: resolve no override so
+		# litellm defaults to https://api.openai.com/v1.
+		return 0
+	fi
+
 	local api_base_file="$LLM_API_BASE_FILE"
 	local api_base_file_name="LLM_API_BASE_FILE"
 	if is_explicit_openai_model "$model" && [ -n "${STRIX_OPENAI_FALLBACK_API_BASE_FILE:-}" ]; then
diff --git a/tests/test_strix_openai_fallback_api_base.py b/tests/test_strix_openai_fallback_api_base.py
new file mode 100644
index 000000000..f2c4237a7
--- /dev/null
+++ b/tests/test_strix_openai_fallback_api_base.py
@@ -0,0 +1,211 @@
+"""Regression contract for direct-OpenAI fallback API-base routing.
+
+When the Strix primary provider is NVIDIA NIM (or OpenRouter / GitHub Models),
+the workflow's ``LLM_API_BASE_FILE`` points at that provider's endpoint. A
+cross-provider fallback to ``openai-direct/gpt-5.4`` must never inherit that
+base: routing an OpenAI model through the NVIDIA NIM edge yields a plain-text
+gateway 404 ("404 page not found") instead of OpenAI responses, so the final
+contracted fallback could never complete a scan.
+
+The gate must therefore prefer an explicit
+``STRIX_OPENAI_FALLBACK_API_BASE_FILE`` for explicit direct-OpenAI models, and
+fall back to litellm's default OpenAI endpoint when the override is absent.
+"""
+
+from __future__ import annotations
+
+import re
+import subprocess
+import tempfile
+import unittest
+from pathlib import Path
+
+
+REPOSITORY_ROOT = Path(__file__).resolve().parents[1]
+STRIX_GATE = REPOSITORY_ROOT / "scripts" / "ci" / "strix_quick_gate.sh"
+STRIX_MODEL_UTILS = REPOSITORY_ROOT / "scripts" / "ci" / "strix_model_utils.sh"
+STRIX_WORKFLOW = REPOSITORY_ROOT / ".github" / "workflows" / "strix.yml"
+OPENAI_FALLBACK_BASE = "https://api.openai.com/v1"
+
+
+def _function_block(source: str, function_name: str) -> str:
+    """Return one top-level Bash function, including its closing brace."""
+
+    match = re.search(
+        rf"(?ms)^{re.escape(function_name)}\(\) \{{\n.*?^\}}\n",
+        source,
+    )
+    if match is None:
+        raise AssertionError(f"missing Bash function: {function_name}")
+    return match.group(0)
+
+
+def _resolver_helpers() -> list[str]:
+    """Collect every helper the production API-base resolver depends on."""
+
+    gate_source = STRIX_GATE.read_text(encoding="utf-8")
+    utils_source = STRIX_MODEL_UTILS.read_text(encoding="utf-8")
+    helper_names = (
+        ("is_vertex_model", gate_source),
+        ("is_github_models_model", gate_source),
+        ("is_github_models_api_base", gate_source),
+        ("is_github_models_api_compatible_model", gate_source),
+        ("is_explicit_openai_model", gate_source),
+        ("resolve_trusted_input_file", gate_source),
+        ("trim_whitespace", utils_source),
+    )
+    helpers: list[str] = []
+    for name, source in helper_names:
+        try:
+            helpers.append(_function_block(source, name))
+        except AssertionError as exc:  # pragma: no cover - shape drift guard
+            raise AssertionError(f"resolver helper missing: {name}") from exc
+    return helpers
+
+
+def _resolve_api_base(env: dict[str, str], model: str) -> tuple[int, str]:
+    """Execute the production API-base resolver for one model and env."""
+
+    gate_source = STRIX_GATE.read_text(encoding="utf-8")
+    resolver_source = _function_block(gate_source, "resolved_llm_api_base_for_model")
+    helper_sources = _resolver_helpers()
+    with tempfile.TemporaryDirectory(prefix="strix-openai-fallback-base-") as temp_dir:
+        base_path = Path(temp_dir) / "primary_base.txt"
+        if "LLM_API_BASE_FILE" in env:
+            base_path.write_text(env["LLM_API_BASE_FILE"], encoding="utf-8")
+            env = {**env, "LLM_API_BASE_FILE": str(base_path)}
+        fallback_path = Path(temp_dir) / "openai_fallback_api_base.txt"
+        if "STRIX_OPENAI_FALLBACK_API_BASE_FILE" in env:
+            fallback_path.write_text(
+                env["STRIX_OPENAI_FALLBACK_API_BASE_FILE"],
+                encoding="utf-8",
+            )
+            env = {
+                **env,
+                "STRIX_OPENAI_FALLBACK_API_BASE_FILE": str(fallback_path),
+            }
+        github_models_path = Path(temp_dir) / "github_models_api_base.txt"
+        if "STRIX_GITHUB_MODELS_API_BASE_FILE" in env:
+            github_models_path.write_text(
+                env["STRIX_GITHUB_MODELS_API_BASE_FILE"],
+                encoding="utf-8",
+            )
+            env = {
+                **env,
+                "STRIX_GITHUB_MODELS_API_BASE_FILE": str(github_models_path),
+            }
+        script_lines = [
+            "set -euo pipefail",
+            f'STRIX_INPUT_FILE_ROOT="{temp_dir}"',
+            *helper_sources,
+            resolver_source,
+        ]
+        command_env = {
+            key: value
+            for key, value in env.items()
+            if key in {
+                "LLM_API_BASE_FILE",
+                "STRIX_GITHUB_MODELS_API_BASE_FILE",
+                "STRIX_OPENAI_FALLBACK_API_BASE_FILE",
+            }
+        }
+        completed = subprocess.run(
+            [
+                "bash",
+                "-c",
+                "\n".join([*script_lines, 'resolved_llm_api_base_for_model "$1"']),
+                "strix-resolver",
+                model,
+            ],
+            check=False,
+            capture_output=True,
+            text=True,
+            env={
+                "PATH": "/usr/bin:/bin:/usr/local/bin",
+                "HOME": temp_dir,
+                **command_env,
+            },
+        )
+    return completed.returncode, completed.stdout.strip()
+
+
+class ExplicitOpenAIFallbackRouting(unittest.TestCase):
+    """Direct-OpenAI fallbacks must not inherit the primary provider base."""
+
+    def test_nvidia_primary_with_override_routes_to_openai(self) -> None:
+        """openai-direct/gpt-5.4 uses the explicit OpenAI API base file."""
+
+        rc, api_base = _resolve_api_base(
+            {
+                "LLM_API_BASE_FILE": "https://integrate.api.nvidia.com/v1",
+                "STRIX_OPENAI_FALLBACK_API_BASE_FILE": OPENAI_FALLBACK_BASE,
+            },
+            "openai-direct/gpt-5.4",
+        )
+        self.assertEqual(rc, 0)
+        self.assertEqual(api_base, OPENAI_FALLBACK_BASE)
+
+    def test_nvidia_primary_without_override_uses_default_openai(self) -> None:
+        """Without an override the resolver returns no base (litellm default)."""
+
+        rc, api_base = _resolve_api_base(
+            {"LLM_API_BASE_FILE": "https://integrate.api.nvidia.com/v1"},
+            "openai-direct/gpt-5.4",
+        )
+        self.assertEqual(rc, 0)
+        self.assertEqual(api_base, "")
+
+    def test_primary_provider_models_keep_their_base(self) -> None:
+        """NVIDIA NIM primary attempts still resolve through the NIM edge."""
+
+        rc, api_base = _resolve_api_base(
+            {"LLM_API_BASE_FILE": "https://integrate.api.nvidia.com/v1"},
+            "nvidia_nim/nvidia/nemotron-3-super-120b-a12b",
+        )
+        self.assertEqual(rc, 0)
+        self.assertEqual(api_base, "https://integrate.api.nvidia.com/v1")
+
+    def test_github_models_fallback_keeps_github_models_base(self) -> None:
+        """github_models/* fallbacks keep their dedicated inference endpoint."""
+
+        rc, api_base = _resolve_api_base(
+            {
+                "LLM_API_BASE_FILE": "https://integrate.api.nvidia.com/v1",
+                "STRIX_GITHUB_MODELS_API_BASE_FILE": "https://models.github.ai/inference",
+            },
+            "github_models/openai/gpt-5.4",
+        )
+        self.assertEqual(rc, 0)
+        self.assertEqual(api_base, "https://models.github.ai/inference")
+
+    def test_invalid_https_override_is_configuration_failure(self) -> None:
+        """A non-https override fails configuration instead of scanning."""
+
+        rc, _ = _resolve_api_base(
+            {
+                "LLM_API_BASE_FILE": "https://integrate.api.nvidia.com/v1",
+                "STRIX_OPENAI_FALLBACK_API_BASE_FILE": "http://api.example.com/v1",
+            },
+            "openai-direct/gpt-5.4",
+        )
+        self.assertEqual(rc, 2)
+
+
+class WorkflowProvisionsFallbackBase(unittest.TestCase):
+    """The workflow must publish and pass the explicit OpenAI fallback base."""
+
+    def test_workflow_writes_openai_fallback_api_base_file(self) -> None:
+        workflow = STRIX_WORKFLOW.read_text(encoding="utf-8")
+        self.assertIn("STRIX_OPENAI_FALLBACK_API_BASE_FILE=", workflow)
+        self.assertIn(OPENAI_FALLBACK_BASE, workflow)
+
+    def test_workflow_passes_override_into_gate_environment(self) -> None:
+        workflow = STRIX_WORKFLOW.read_text(encoding="utf-8")
+        self.assertIn(
+            "STRIX_OPENAI_FALLBACK_API_BASE_FILE: ${{ env.STRIX_OPENAI_FALLBACK_API_BASE_FILE }}",
+            workflow,
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()

From 7c596f38eb2630328f0cee7fd050a5219408820e Mon Sep 17 00:00:00 2001
From: Seongho Bae <me@seonghobae.me>
Date: Tue, 25 Aug 2026 15:49:06 +0900
Subject: [PATCH 04/11] test(strix): align remaining gpt-5.6-luna expectations
 with the gpt-5.4 contract

The model rename in a724582 missed two test expectations that still
asserted the nonexistent gpt-5.6-luna default and fallback names; align
them with the shipped gpt-5.4 contract so required CI passes.
---
 tests/test_required_workflow_queue_contract.py    | 2 +-
 tests/test_strix_nvidia_nim_not_found_fallback.py | 4 ++--
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/tests/test_required_workflow_queue_contract.py b/tests/test_required_workflow_queue_contract.py
index e58f5e6c0..1d79f1daa 100644
--- a/tests/test_required_workflow_queue_contract.py
+++ b/tests/test_required_workflow_queue_contract.py
@@ -505,7 +505,7 @@ def test_nvidia_nim_defaults_preserve_existing_fallbacks_without_secret(
     assert strix.returncode == 0, strix.stderr
     assert {
         "provider_mode=openai_direct",
-        "strix_model=gpt-5.6-luna",
+        "strix_model=gpt-5.4",
     } <= set(strix_output.read_text().splitlines())
     assert (
         "STRIX_MODEL: ${{ steps.gate.outputs.strix_model }}"
diff --git a/tests/test_strix_nvidia_nim_not_found_fallback.py b/tests/test_strix_nvidia_nim_not_found_fallback.py
index 990269725..17f0e9a30 100644
--- a/tests/test_strix_nvidia_nim_not_found_fallback.py
+++ b/tests/test_strix_nvidia_nim_not_found_fallback.py
@@ -192,7 +192,7 @@ def test_workflow_uses_available_free_first_nvidia_plan(self) -> None:
         workflow = STRIX_WORKFLOW.read_text(encoding="utf-8")
         default_expression = (
             "steps.target_visibility.outputs.is_private == 'false' && "
-            f"'{DEFAULT_NVIDIA_MODEL}' || 'gpt-5.6-luna'"
+            f"'{DEFAULT_NVIDIA_MODEL}' || 'gpt-5.4'"
         )
         self.assertIn(default_expression, workflow)
         self.assertIn(
@@ -202,7 +202,7 @@ def test_workflow_uses_available_free_first_nvidia_plan(self) -> None:
         )
         self.assertIn(
             "steps.gate.outputs.provider_mode == 'nvidia_nim' && "
-            f"'{FREE_NVIDIA_FALLBACK} openai-direct/gpt-5.6-luna'",
+            f"'{FREE_NVIDIA_FALLBACK} openai-direct/gpt-5.4'",
             workflow,
         )
 

From 2d982374e961ec76c93f07e233655935effbb00d Mon Sep 17 00:00:00 2001
From: Seongho Bae <me@seonghobae.me>
Date: Tue, 25 Aug 2026 16:37:38 +0900
Subject: [PATCH 05/11] fix(opencode-review): complete gpt-5.6-luna to gpt-5.4
 rename across dispatch chain

The a724582 rename updated strix.yml and the smoke contract but left the
opencode-review-dispatch workflow, its agent-contract tests, the
failed-check fallback-findings marker, and the pinned review-dispatch
blob SHAs asserting the retired model name, so required CI failed with
three missing-string markers. Align every remaining reference with the
shipped gpt-5.4 direct-OpenAI contract and refresh the paired blob SHAs.
---
 .github/workflows/opencode-review-dispatch.yml            | 6 +++---
 .../ci/emit_opencode_failed_check_fallback_findings.sh    | 2 +-
 tests/test_opencode_agent_contract.py                     | 8 ++++----
 tests/test_pr_review_autofix_nvidia_nim_contract.py       | 2 +-
 4 files changed, 9 insertions(+), 9 deletions(-)

diff --git a/.github/workflows/opencode-review-dispatch.yml b/.github/workflows/opencode-review-dispatch.yml
index dd65d90e1..33a4fd08e 100644
--- a/.github/workflows/opencode-review-dispatch.yml
+++ b/.github/workflows/opencode-review-dispatch.yml
@@ -4058,8 +4058,8 @@ jobs:
                   "apiKey": "{env:OPENAI_API_KEY}"
                 },
                 "models": {
-                  "gpt-5.6-luna": {
-                    "name": "OpenAI GPT-5.6 Luna (direct)",
+                  "gpt-5.4": {
+                    "name": "OpenAI GPT-5.4 (direct)",
                     "tool_call": true,
                     "reasoning": true,
                     "options": {
@@ -4481,7 +4481,7 @@ jobs:
           # cost-efficient tier, cheaper than the legacy gpt-5 it replaced
           # ($1/$6 vs $1.25/$10 per 1M tokens) so the org OpenAI budget
           # stretches further between top-ups.
-          OPENCODE_MODEL_CANDIDATES: "${{ 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-5.6-luna 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"
+          OPENCODE_MODEL_CANDIDATES: "${{ needs.validate-pr-metadata.outputs.is_private == 'false' && 'nvidia-nim/nvidia/llama-3.3-nemotron-super-49b-v1.5 nvidia-nim/nvidia/llama-3.1-nemotron-ultra-253b-v1 nvidia-nim/nvidia/nemotron-3-super-120b-a12b nvidia-nim/nvidia/nemotron-3-ultra-550b-a55b nvidia-nim/meta/llama-3.3-70b-instruct nvidia-nim/deepseek-ai/deepseek-v4-pro nvidia-nim/mistralai/codestral-22b-instruct-v0.1 opencode-free/nemotron-3-ultra-free opencode-free/deepseek-v4-flash-free opencode-free/north-mini-code-free opencode-free/laguna-s-2.1-free opencode-free/ling-3.0-flash-free opencode-free/big-pickle opencode-free/mimo-v2.5-free opencode-free/hy3-free opencode-free/minimax-m3-free opencode-free/glm-5-free opencode-free/kimi-k2.5-free opencode-free/qwen3.6-plus-free ' || '' }}opencode/gpt-5.6-terra github-models/deepseek/deepseek-v3-0324 openai/gpt-5.4 openrouter/deepseek/deepseek-v3.2 openrouter/qwen/qwen3-coder github-models/openai/gpt-4.1 github-models/openai/gpt-5 github-models/openai/gpt-5-chat github-models/openai/o3 github-models/deepseek/deepseek-r1-0528 github-models/deepseek/deepseek-r1"
           # One attempt per model, then fall through to the next model. Retrying
           # the SAME model 5x let a rate-limited/hung leader consume the whole
           # step, so the pool never reached a healthy fallback model.
diff --git a/scripts/ci/emit_opencode_failed_check_fallback_findings.sh b/scripts/ci/emit_opencode_failed_check_fallback_findings.sh
index ccd35a273..b8a755e6e 100755
--- a/scripts/ci/emit_opencode_failed_check_fallback_findings.sh
+++ b/scripts/ci/emit_opencode_failed_check_fallback_findings.sh
@@ -956,7 +956,7 @@ extract_strix_failed_check_block "$EVIDENCE_FILE" "$strix_evidence_file"
 
 emit_known_missing_string_finding \
 	"$EVIDENCE_FILE" \
-	"steps.target_visibility.outputs.is_private == 'false' && 'nvidia_nim/nvidia/nemotron-3-ultra-550b-a55b' || 'gpt-5.6-luna'" \
+	"steps.target_visibility.outputs.is_private == 'false' && 'nvidia_nim/nvidia/nemotron-3-ultra-550b-a55b' || 'gpt-5.4'" \
 	"Strix public scans must default to NVIDIA NIM while private scans retain the contracted provider" \
 	".github/workflows/strix.yml" \
 	"scripts/ci/test_strix_quick_gate.sh"
diff --git a/tests/test_opencode_agent_contract.py b/tests/test_opencode_agent_contract.py
index 8ee6e86fc..2360fdb26 100644
--- a/tests/test_opencode_agent_contract.py
+++ b/tests/test_opencode_agent_contract.py
@@ -186,7 +186,7 @@ def test_opencode_model_pool_sets_high_effort_for_capable_candidates():
         ["opencode-free", "qwen3.6-plus-free"],
         ["opencode", "gpt-5.6-terra"],
         ["github-models", "deepseek/deepseek-v3-0324"],
-        ["openai", "gpt-5.6-luna"],
+        ["openai", "gpt-5.4"],
         ["openrouter", "deepseek/deepseek-v3.2"],
         ["openrouter", "qwen/qwen3-coder"],
         ["github-models", "openai/gpt-4.1"],
@@ -197,7 +197,7 @@ def test_opencode_model_pool_sets_high_effort_for_capable_candidates():
         ["github-models", "deepseek/deepseek-r1"],
     ]
     assert zen_models == ["gpt-5.6-terra"]
-    assert direct_openai_models == ["gpt-5.6-luna"]
+    assert direct_openai_models == ["gpt-5.4"]
     assert openrouter_models == [
         "deepseek/deepseek-v3.2",
         "qwen/qwen3-coder",
@@ -1740,7 +1740,7 @@ def test_workflow_provisions_sandbox_tool_and_reviewer_agent():
     assert (
         "opencode/gpt-5.6-terra "
         "github-models/deepseek/deepseek-v3-0324 "
-        "openai/gpt-5.6-luna "
+        "openai/gpt-5.4 "
         "openrouter/deepseek/deepseek-v3.2 "
         "openrouter/qwen/qwen3-coder "
         "github-models/openai/gpt-4.1 "
@@ -1887,7 +1887,7 @@ def test_workflow_provisions_sandbox_tool_and_reviewer_agent():
     )
     assert (
         "github-models/deepseek/deepseek-v3-0324 "
-        "openai/gpt-5.6-luna "
+        "openai/gpt-5.4 "
         "openrouter/deepseek/deepseek-v3.2 "
         "openrouter/qwen/qwen3-coder "
         "github-models/openai/gpt-4.1 "
diff --git a/tests/test_pr_review_autofix_nvidia_nim_contract.py b/tests/test_pr_review_autofix_nvidia_nim_contract.py
index 799b9e9fb..3b6c3a054 100644
--- a/tests/test_pr_review_autofix_nvidia_nim_contract.py
+++ b/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 = "dd65d90e10e5040562b501ade1a40f89572f0984"
+REVIEW_DISPATCH_BLOB_SHA = "33a4fd08e93e653e715b380ad9f9523e11d231e6"
 
 
 def _workflow_text(path: Path) -> str:

From 9288031cb4b9f5959b356c4f7407212595a271ef Mon Sep 17 00:00:00 2001
From: Seongho Bae <me@seonghobae.me>
Date: Tue, 25 Aug 2026 18:50:02 +0900
Subject: [PATCH 06/11] fix(strix): provision OpenAI fallback credentials for
 github_models primary; dedupe CHANGELOG

- CodeRabbit major: the github_models primary's STRIX_FALLBACK_MODELS chain
  ends in openai-direct/gpt-5.4, but 'Prepare GitHub Models fallback
  credentials' excluded the github_models mode, so that fallback could not
  authenticate or route after primary exhaustion. Include github_models in
  the provisioning step.
- Devin bug: remove the four accidentally duplicated [Unreleased] bullets
  from CHANGELOG.md, keeping a single copy before '### Added'.
- Add markdownlint MD040 language to the doctoring evidence fence.
---
 .github/workflows/strix.yml                   |  5 ++-
 CHANGELOG.md                                  | 44 -------------------
 .../strix-openai-fallback-api-base-routing.md |  2 +-
 3 files changed, 5 insertions(+), 46 deletions(-)

diff --git a/.github/workflows/strix.yml b/.github/workflows/strix.yml
index 6ba4fb4a8..bb656dc56 100644
--- a/.github/workflows/strix.yml
+++ b/.github/workflows/strix.yml
@@ -674,7 +674,10 @@ 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 }}
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 00c525c80..523766d5e 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -17,50 +17,6 @@ Semantic Versioning where the repository publishes a release.
 - 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
-  during materialization and then rejecting every version except pnpm 11.5.3;
-  route generic coverage and docstring package scripts through the same
-  Corepack boundary instead of invoking a removed bare `pnpm` binary.
-- Keep `--trust-lockfile` only for pnpm 11.3 and newer
-  (`trustLockfile` landed in pnpm 11.3). pnpm 9, 10, and 11.0–11.2 reject
-  that flag and previously failed LineageWeave JavaScript coverage before
-  tests could run. Jest test scripts still receive `--coverage` because Jest
-  documents a native coverage flag.
-- Run declared JavaScript test scripts without synthesizing `--coverage` when
-  the package does not declare a compatible coverage command, but keep the
-  coverage result failed until the repository adds a lock-pinned provider and
-  owned coverage command. A generic `c8`, `nyc`, or Istanbul dependency no
-  longer makes an unrelated test runner receive an unsupported flag.
-- Fix OpenCode coverage evidence for exact-base, organization-owned Python VCS
-  dependencies without weakening registry hashes or the networkless PR sandbox,
-  reject namespace, ambiguous, linked, native-extension, and installed-metadata
-  layouts, and make exact roots readable by the unprivileged coverage user.
-
-
-
-- 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
-  during materialization and then rejecting every version except pnpm 11.5.3;
-  route generic coverage and docstring package scripts through the same
-  Corepack boundary instead of invoking a removed bare `pnpm` binary.
-- Keep `--trust-lockfile` only for pnpm 11.3 and newer
-  (`trustLockfile` landed in pnpm 11.3). pnpm 9, 10, and 11.0–11.2 reject
-  that flag and previously failed LineageWeave JavaScript coverage before
-  tests could run. Jest test scripts still receive `--coverage` because Jest
-  documents a native coverage flag.
-- Run declared JavaScript test scripts without synthesizing `--coverage` when
-  the package does not declare a compatible coverage command, but keep the
-  coverage result failed until the repository adds a lock-pinned provider and
-  owned coverage command. A generic `c8`, `nyc`, or Istanbul dependency no
-  longer makes an unrelated test runner receive an unsupported flag.
-- Fix OpenCode coverage evidence for exact-base, organization-owned Python VCS
-  dependencies without weakening registry hashes or the networkless PR sandbox,
-  reject namespace, ambiguous, linked, native-extension, and installed-metadata
-  layouts, and make exact roots readable by the unprivileged coverage user.
-
 ### Added
 
 - Refresh the live product and technical gap baseline against the current
diff --git a/docs/doctoring/strix-openai-fallback-api-base-routing.md b/docs/doctoring/strix-openai-fallback-api-base-routing.md
index 4f7112303..c7a2a1890 100644
--- a/docs/doctoring/strix-openai-fallback-api-base-routing.md
+++ b/docs/doctoring/strix-openai-fallback-api-base-routing.md
@@ -23,7 +23,7 @@ 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

From 91c79dc882992deae9d2f0118f001c0f77c0f1af Mon Sep 17 00:00:00 2001
From: Seongho Bae <me@seonghobae.me>
Date: Tue, 25 Aug 2026 18:56:53 +0900
Subject: [PATCH 07/11] fix(opencode-review): restore deduplicated changelog
 entries and rename Luna comment

- Devin bug: the CHANGELOG dedup accidentally removed the four legitimate
  [Unreleased] bullets entirely; restore them once alongside the two Strix
  entries so pending history is not dropped.
- Devin bug: the OPENCODE_MODEL_CANDIDATES rationale comment still named
  the retired 'GPT-5.6 Luna' slot with stale pricing; rewrite it for the
  valid gpt-5.4 direct-OpenAI contract.
---
 .../workflows/opencode-review-dispatch.yml    | 10 +++++-----
 CHANGELOG.md                                  | 20 +++++++++++++++++++
 2 files changed, 25 insertions(+), 5 deletions(-)

diff --git a/.github/workflows/opencode-review-dispatch.yml b/.github/workflows/opencode-review-dispatch.yml
index 33a4fd08e..478788d41 100644
--- a/.github/workflows/opencode-review-dispatch.yml
+++ b/.github/workflows/opencode-review-dispatch.yml
@@ -4471,16 +4471,16 @@ jobs:
           # or used for product/model improvement, so private repositories
           # include neither NIM nor anonymous free candidates and start at the
           # existing keyed fallback list: OpenCode Zen GPT-5.6 Terra, DeepSeek
-          # V3, the direct GPT-5.6 Luna slot, and pinned PAID
+          # V3, the direct GPT-5.4 slot, and pinned PAID
           # OpenRouter coder models (free-tier candidates hit the shared
           # free-models-per-day cap and hung for the full candidate timeout,
           # 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-5.6 Luna: the newest family's
-          # cost-efficient tier, cheaper than the legacy gpt-5 it replaced
-          # ($1/$6 vs $1.25/$10 per 1M tokens) so the org OpenAI budget
-          # stretches further between top-ups.
+          # The direct-OpenAI slot runs GPT-5.4: the current valid model name
+          # on the direct OpenAI platform (the retired gpt-5.6-luna catalog
+          # entry returned 404s), kept cost-efficient against the org's
+          # OpenAI budget between top-ups.
           OPENCODE_MODEL_CANDIDATES: "${{ 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
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 523766d5e..056b3cff1 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -16,6 +16,26 @@ Semantic Versioning where the repository publishes a release.
   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
+  during materialization and then rejecting every version except pnpm 11.5.3;
+  route generic coverage and docstring package scripts through the same
+  Corepack boundary instead of invoking a removed bare `pnpm` binary.
+- Keep `--trust-lockfile` only for pnpm 11.3 and newer
+  (`trustLockfile` landed in pnpm 11.3). pnpm 9, 10, and 11.0–11.2 reject
+  that flag and previously failed LineageWeave JavaScript coverage before
+  tests could run. Jest test scripts still receive `--coverage` because Jest
+  documents a native coverage flag.
+- Run declared JavaScript test scripts without synthesizing `--coverage` when
+  the package does not declare a compatible coverage command, but keep the
+  coverage result failed until the repository adds a lock-pinned provider and
+  owned coverage command. A generic `c8`, `nyc`, or Istanbul dependency no
+  longer makes an unrelated test runner receive an unsupported flag.
+- Fix OpenCode coverage evidence for exact-base, organization-owned Python VCS
+  dependencies without weakening registry hashes or the networkless PR sandbox,
+  reject namespace, ambiguous, linked, native-extension, and installed-metadata
+  layouts, and make exact roots readable by the unprivileged coverage user.
 
 ### Added
 

From c5f1c74cda6cfa315f1c69e4680df72c091ce328 Mon Sep 17 00:00:00 2001
From: Seongho Bae <me@seonghobae.me>
Date: Tue, 25 Aug 2026 19:05:17 +0900
Subject: [PATCH 08/11] test: repin review workflow contract to exact blob

---
 tests/test_pr_review_autofix_nvidia_nim_contract.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tests/test_pr_review_autofix_nvidia_nim_contract.py b/tests/test_pr_review_autofix_nvidia_nim_contract.py
index 3b6c3a054..3a8ddcae1 100644
--- a/tests/test_pr_review_autofix_nvidia_nim_contract.py
+++ b/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 = "33a4fd08e93e653e715b380ad9f9523e11d231e6"
+REVIEW_DISPATCH_BLOB_SHA = "478788d41ed3e0d8389292a059d225336a925830"
 
 
 def _workflow_text(path: Path) -> str:

From 3c7f430f59df0eec2b6d70c8272c60b906d071dd Mon Sep 17 00:00:00 2001
From: Seongho Bae <me@seonghobae.me>
Date: Tue, 25 Aug 2026 21:01:24 +0900
Subject: [PATCH 09/11] fix(strix): preserve standalone custom OpenAI bases

---
 scripts/ci/strix_quick_gate.sh               | 13 +++++--------
 tests/test_strix_openai_fallback_api_base.py | 14 +++++++++++---
 2 files changed, 16 insertions(+), 11 deletions(-)

diff --git a/scripts/ci/strix_quick_gate.sh b/scripts/ci/strix_quick_gate.sh
index dd27b63d2..b2ae0fd22 100755
--- a/scripts/ci/strix_quick_gate.sh
+++ b/scripts/ci/strix_quick_gate.sh
@@ -2426,14 +2426,7 @@ resolved_llm_api_base_for_model() {
 	if is_vertex_model "$model"; then
 		return 0
 	fi
-	if is_explicit_openai_model "$model" && [ -z "${STRIX_OPENAI_FALLBACK_API_BASE_FILE:-}" ]; then
-		# Without workflow provisioning, an explicit direct-OpenAI model must
-		# still not inherit a foreign primary base: resolve no override so
-		# litellm defaults to https://api.openai.com/v1.
-		return 0
-	fi
-
-	local api_base_file="$LLM_API_BASE_FILE"
+	local api_base_file="${LLM_API_BASE_FILE:-}"
 	local api_base_file_name="LLM_API_BASE_FILE"
 	if is_explicit_openai_model "$model" && [ -n "${STRIX_OPENAI_FALLBACK_API_BASE_FILE:-}" ]; then
 		# Cross-provider fallback: openai-direct/* candidates must reach the
@@ -2442,6 +2435,10 @@ resolved_llm_api_base_for_model() {
 		# this the fallback hits the primary gateway and 404s.
 		api_base_file="$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
diff --git a/tests/test_strix_openai_fallback_api_base.py b/tests/test_strix_openai_fallback_api_base.py
index f2c4237a7..f472ddeac 100644
--- a/tests/test_strix_openai_fallback_api_base.py
+++ b/tests/test_strix_openai_fallback_api_base.py
@@ -9,7 +9,8 @@
 
 The gate must therefore prefer an explicit
 ``STRIX_OPENAI_FALLBACK_API_BASE_FILE`` for explicit direct-OpenAI models, and
-fall back to litellm's default OpenAI endpoint when the override is absent.
+    fall back to a caller-supplied ``LLM_API_BASE_FILE`` for standalone custom
+    endpoints, or to litellm's default OpenAI endpoint when no base is supplied.
 """
 
 from __future__ import annotations
@@ -145,14 +146,21 @@ def test_nvidia_primary_with_override_routes_to_openai(self) -> None:
         self.assertEqual(rc, 0)
         self.assertEqual(api_base, OPENAI_FALLBACK_BASE)
 
-    def test_nvidia_primary_without_override_uses_default_openai(self) -> None:
-        """Without an override the resolver returns no base (litellm default)."""
+    def test_standalone_custom_base_is_honored_without_override(self) -> None:
+        """A standalone caller's explicit custom endpoint remains effective."""
 
         rc, api_base = _resolve_api_base(
             {"LLM_API_BASE_FILE": "https://integrate.api.nvidia.com/v1"},
             "openai-direct/gpt-5.4",
         )
         self.assertEqual(rc, 0)
+        self.assertEqual(api_base, "https://integrate.api.nvidia.com/v1")
+
+    def test_direct_openai_without_any_base_uses_default_openai(self) -> None:
+        """With no base file, LiteLLM still selects the native OpenAI endpoint."""
+
+        rc, api_base = _resolve_api_base({}, "openai-direct/gpt-5.4")
+        self.assertEqual(rc, 0)
         self.assertEqual(api_base, "")
 
     def test_primary_provider_models_keep_their_base(self) -> None:

From 0547c6b3815865fb1c5d018266c6f033b56fde0e Mon Sep 17 00:00:00 2001
From: Seongho Bae <me@seonghobae.me>
Date: Tue, 25 Aug 2026 21:15:01 +0900
Subject: [PATCH 10/11] fix(strix): reject foreign GitHub Models fallback bases

---
 CHANGELOG.md                                  | 15 +++++----
 .../strix-openai-fallback-api-base-routing.md | 33 +++++++++++--------
 scripts/ci/strix_quick_gate.sh                |  9 +++++
 tests/test_strix_openai_fallback_api_base.py  | 10 ++++++
 4 files changed, 47 insertions(+), 20 deletions(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index c5f5fd856..c6899d38b 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -7,13 +7,14 @@ 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.
+  inheriting a provider-specific primary base: the workflow now provisions
+  `STRIX_OPENAI_FALLBACK_API_BASE_FILE` (`https://api.openai.com/v1`), while
+  standalone caller-supplied `LLM_API_BASE_FILE` values remain honored for
+  OpenAI-compatible endpoints. A GitHub Models base is never inherited, and
+  LiteLLM uses native OpenAI defaults only when no base is supplied. 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
diff --git a/docs/doctoring/strix-openai-fallback-api-base-routing.md b/docs/doctoring/strix-openai-fallback-api-base-routing.md
index c7a2a1890..a51734954 100644
--- a/docs/doctoring/strix-openai-fallback-api-base-routing.md
+++ b/docs/doctoring/strix-openai-fallback-api-base-routing.md
@@ -6,8 +6,11 @@ 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.
+models. In standalone runs, a caller-supplied `LLM_API_BASE_FILE` remains in
+force for an OpenAI-compatible endpoint; a GitHub Models primary endpoint is
+not inherited. Litellm uses its default `https://api.openai.com/v1` endpoint
+only when no base file is supplied (or when that provider-specific base is
+rejected).
 
 The central workflow writes `https://api.openai.com/v1` into
 `$RUNNER_TEMP/openai_fallback_api_base.txt` and exports
@@ -44,9 +47,11 @@ The override is a runner-provisioned regular file under `$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.
+control characters. Absent or empty overrides preserve a caller-supplied
+`LLM_API_BASE_FILE` for standalone local gate runs; when both files are absent,
+litellm selects its default endpoint. A GitHub Models base is explicitly
+rejected for a direct OpenAI model so a missing OpenAI key remains a
+provider-unavailable outcome instead of a configuration error.
 
 ## Verification contract
 
@@ -54,14 +59,16 @@ 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
+2. without either base file, the resolver returns no base so litellm defaults
+   to `https://api.openai.com/v1`;
+3. a standalone caller-supplied custom `LLM_API_BASE_FILE` remains effective;
+4. a GitHub Models primary base is not inherited by a direct-OpenAI fallback;
+5. NVIDIA NIM primary attempts keep resolving through the NIM edge;
+6. `github_models/*` fallbacks keep their dedicated GitHub Models endpoint;
+7. a non-https override fails configuration (exit 2) instead of scanning;
+8. the workflow provisions the override file and passes it into the gate env;
+9. the required-workflow smoke contract pins both sides of the wiring; and
+10. 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
diff --git a/scripts/ci/strix_quick_gate.sh b/scripts/ci/strix_quick_gate.sh
index b2ae0fd22..c5426a6ca 100755
--- a/scripts/ci/strix_quick_gate.sh
+++ b/scripts/ci/strix_quick_gate.sh
@@ -2475,6 +2475,15 @@ resolved_llm_api_base_for_model() {
 		echo "ERROR: LLM_API_BASE must be an https URL when configured." >&2
 		return 2
 	fi
+	# Never let a provider-specific GitHub Models base leak into an explicit
+	# direct-OpenAI fallback when no separate OpenAI override was provisioned.
+	# Other caller-supplied OpenAI-compatible endpoints remain valid standalone
+	# configuration and are intentionally preserved.
+	if is_explicit_openai_model "$model" \
+		&& [ -z "${STRIX_OPENAI_FALLBACK_API_BASE_FILE:-}" ] \
+		&& is_github_models_api_base "$llm_api_base_value"; then
+		return 0
+	fi
 	if is_github_models_api_base "$llm_api_base_value" && ! is_github_models_api_compatible_model "$model"; then
 		echo "ERROR: LLM_API_BASE may route through GitHub Models only when STRIX_LLM uses a GitHub Models-compatible model." >&2
 		return 2
diff --git a/tests/test_strix_openai_fallback_api_base.py b/tests/test_strix_openai_fallback_api_base.py
index f472ddeac..7cdf50f28 100644
--- a/tests/test_strix_openai_fallback_api_base.py
+++ b/tests/test_strix_openai_fallback_api_base.py
@@ -163,6 +163,16 @@ def test_direct_openai_without_any_base_uses_default_openai(self) -> None:
         self.assertEqual(rc, 0)
         self.assertEqual(api_base, "")
 
+    def test_github_models_base_is_not_inherited_without_override(self) -> None:
+        """A cross-provider fallback never inherits GitHub Models routing."""
+
+        rc, api_base = _resolve_api_base(
+            {"LLM_API_BASE_FILE": "https://models.github.ai/inference"},
+            "openai-direct/gpt-5.4",
+        )
+        self.assertEqual(rc, 0)
+        self.assertEqual(api_base, "")
+
     def test_primary_provider_models_keep_their_base(self) -> None:
         """NVIDIA NIM primary attempts still resolve through the NIM edge."""
 

From 4fc261ff03f28b4ab8b44ef6d9f14cdb138eb3ef Mon Sep 17 00:00:00 2001
From: Seongho Bae <me@seonghobae.me>
Date: Tue, 25 Aug 2026 21:28:51 +0900
Subject: [PATCH 11/11] fix(strix): reject known foreign fallback bases

---
 CHANGELOG.md                                   |  6 +++---
 .../strix-openai-fallback-api-base-routing.md  | 18 ++++++++++--------
 scripts/ci/strix_quick_gate.sh                 | 15 +++++++++++++--
 tests/test_strix_openai_fallback_api_base.py   | 13 ++++++++++++-
 4 files changed, 38 insertions(+), 14 deletions(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index c6899d38b..cef0acda6 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -10,9 +10,9 @@ Semantic Versioning where the repository publishes a release.
   inheriting a provider-specific primary base: the workflow now provisions
   `STRIX_OPENAI_FALLBACK_API_BASE_FILE` (`https://api.openai.com/v1`), while
   standalone caller-supplied `LLM_API_BASE_FILE` values remain honored for
-  OpenAI-compatible endpoints. A GitHub Models base is never inherited, and
-  LiteLLM uses native OpenAI defaults only when no base is supplied. A
-  non-https override fails configuration. This removes the NVIDIA-NIM-edge
+  OpenAI-compatible endpoints. Known GitHub Models, NVIDIA NIM, and OpenRouter
+  bases are never inherited, and LiteLLM uses native OpenAI defaults only when
+  no base is supplied. 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`
diff --git a/docs/doctoring/strix-openai-fallback-api-base-routing.md b/docs/doctoring/strix-openai-fallback-api-base-routing.md
index a51734954..d38d222d4 100644
--- a/docs/doctoring/strix-openai-fallback-api-base-routing.md
+++ b/docs/doctoring/strix-openai-fallback-api-base-routing.md
@@ -7,10 +7,10 @@ A Strix cross-provider fallback to an explicit direct-OpenAI model
 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. In standalone runs, a caller-supplied `LLM_API_BASE_FILE` remains in
-force for an OpenAI-compatible endpoint; a GitHub Models primary endpoint is
-not inherited. Litellm uses its default `https://api.openai.com/v1` endpoint
-only when no base file is supplied (or when that provider-specific base is
-rejected).
+force for an OpenAI-compatible endpoint; known GitHub Models, NVIDIA NIM, and
+OpenRouter primary endpoints are not inherited. Litellm uses its default
+`https://api.openai.com/v1` endpoint only when no base file is supplied (or
+when that provider-specific base is rejected).
 
 The central workflow writes `https://api.openai.com/v1` into
 `$RUNNER_TEMP/openai_fallback_api_base.txt` and exports
@@ -49,9 +49,10 @@ 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 preserve a caller-supplied
 `LLM_API_BASE_FILE` for standalone local gate runs; when both files are absent,
-litellm selects its default endpoint. A GitHub Models base is explicitly
-rejected for a direct OpenAI model so a missing OpenAI key remains a
-provider-unavailable outcome instead of a configuration error.
+litellm selects its default endpoint. Known GitHub Models, NVIDIA NIM, and
+OpenRouter bases are explicitly rejected for a direct OpenAI model so a
+missing OpenAI key remains a provider-unavailable outcome instead of a
+configuration error.
 
 ## Verification contract
 
@@ -62,7 +63,8 @@ Regression evidence proves that:
 2. without either base file, the resolver returns no base so litellm defaults
    to `https://api.openai.com/v1`;
 3. a standalone caller-supplied custom `LLM_API_BASE_FILE` remains effective;
-4. a GitHub Models primary base is not inherited by a direct-OpenAI fallback;
+4. known GitHub Models, NVIDIA NIM, and OpenRouter primary bases are not
+   inherited by a direct-OpenAI fallback;
 5. NVIDIA NIM primary attempts keep resolving through the NIM edge;
 6. `github_models/*` fallbacks keep their dedicated GitHub Models endpoint;
 7. a non-https override fails configuration (exit 2) instead of scanning;
diff --git a/scripts/ci/strix_quick_gate.sh b/scripts/ci/strix_quick_gate.sh
index c5426a6ca..c9aa41545 100755
--- a/scripts/ci/strix_quick_gate.sh
+++ b/scripts/ci/strix_quick_gate.sh
@@ -827,6 +827,17 @@ is_github_models_api_base() {
 	esac
 }
 
+is_known_foreign_provider_api_base() {
+	case "$1" in
+	https://models.github.ai/* | https://integrate.api.nvidia.com/* | https://openrouter.ai/*)
+		return 0
+		;;
+	*)
+		return 1
+		;;
+	esac
+}
+
 PRIMARY_MODEL="$(normalize_model "$STRIX_LLM")"
 if [ "$PRIMARY_MODEL" != "$STRIX_LLM" ]; then
 	echo "Normalized STRIX_LLM to provider-qualified model '$PRIMARY_MODEL'."
@@ -2475,13 +2486,13 @@ resolved_llm_api_base_for_model() {
 		echo "ERROR: LLM_API_BASE must be an https URL when configured." >&2
 		return 2
 	fi
-	# Never let a provider-specific GitHub Models base leak into an explicit
+	# Never let a known provider-specific base leak into an explicit
 	# direct-OpenAI fallback when no separate OpenAI override was provisioned.
 	# Other caller-supplied OpenAI-compatible endpoints remain valid standalone
 	# configuration and are intentionally preserved.
 	if is_explicit_openai_model "$model" \
 		&& [ -z "${STRIX_OPENAI_FALLBACK_API_BASE_FILE:-}" ] \
-		&& is_github_models_api_base "$llm_api_base_value"; then
+		&& is_known_foreign_provider_api_base "$llm_api_base_value"; then
 		return 0
 	fi
 	if is_github_models_api_base "$llm_api_base_value" && ! is_github_models_api_compatible_model "$model"; then
diff --git a/tests/test_strix_openai_fallback_api_base.py b/tests/test_strix_openai_fallback_api_base.py
index 7cdf50f28..f078dd2a0 100644
--- a/tests/test_strix_openai_fallback_api_base.py
+++ b/tests/test_strix_openai_fallback_api_base.py
@@ -50,6 +50,7 @@ def _resolver_helpers() -> list[str]:
         ("is_vertex_model", gate_source),
         ("is_github_models_model", gate_source),
         ("is_github_models_api_base", gate_source),
+        ("is_known_foreign_provider_api_base", gate_source),
         ("is_github_models_api_compatible_model", gate_source),
         ("is_explicit_openai_model", gate_source),
         ("resolve_trusted_input_file", gate_source),
@@ -149,12 +150,22 @@ def test_nvidia_primary_with_override_routes_to_openai(self) -> None:
     def test_standalone_custom_base_is_honored_without_override(self) -> None:
         """A standalone caller's explicit custom endpoint remains effective."""
 
+        rc, api_base = _resolve_api_base(
+            {"LLM_API_BASE_FILE": "https://api.example.com/v1"},
+            "openai-direct/gpt-5.4",
+        )
+        self.assertEqual(rc, 0)
+        self.assertEqual(api_base, "https://api.example.com/v1")
+
+    def test_nvidia_base_is_not_inherited_without_override(self) -> None:
+        """A cross-provider fallback never inherits the NVIDIA NIM base."""
+
         rc, api_base = _resolve_api_base(
             {"LLM_API_BASE_FILE": "https://integrate.api.nvidia.com/v1"},
             "openai-direct/gpt-5.4",
         )
         self.assertEqual(rc, 0)
-        self.assertEqual(api_base, "https://integrate.api.nvidia.com/v1")
+        self.assertEqual(api_base, "")
 
     def test_direct_openai_without_any_base_uses_default_openai(self) -> None:
         """With no base file, LiteLLM still selects the native OpenAI endpoint."""