fix(strix): make Azure and cross-provider fallbacks executable

[seonghobae]

Current exact-head recovery (2026-08-24)

• Head: ab3d764547082e1b55b6257cc1cd9aa5d951fa30
• Protected base: main@613a33e0cb1c6db9790fae99f6253445712ac37a
• Adopted protected-main PR fix(strix): unblock trusted smoke during provider-prefix migration #1286 as the second parent without force-push. Its temporary either-prefix smoke compatibility was resolved in favor of this owner branch's completed openai_direct/* migration and explicit rejection of retired openai-direct/*.
• Recreated GitHub tree ed281e2f4a50bad277524b7282ab490d1661e466 byte-for-byte from the locally verified merge, then advanced the branch by compare-and-set fast-forward.
• Local GREEN: required-workflow smoke PASS, full test_strix_quick_gate.sh PASS, and git diff --check PASS.
• Hosted exact-head GREEN: Strix Changed Path Quality run 32712853098, job 97387791739, checked out exact SHA ab3d764547082e1b55b6257cc1cd9aa5d951fa30, passed 1,417 tests with 1 skip and 16 subtests, and recorded test_strix_quick_gate: PASS.
• Organization Commercial Readiness, Exact Artifact SBOM Attestation, and OpenCode Rust Coverage Toolchain quality workflows also completed successfully on the exact head.
• CodeQL and eight security/SBOM workflows remain pending or queued and are non-passing until terminal.
• Unresolved review threads: 0 at publication.
• Exact-current-head qualifying formal verdict: absent.

Decision: WAIT_FOR_TERMINAL_EXACT_HEAD_CHECKS_AND_INDEPENDENT_FORMAL_REVIEW.

Previous exact-head recovery (historical)

• Head: 7011fee275eaa257ce491efb4812dd3e98ed649e
• Protected base: main@0c6b9a6459c9dbdf5e23fb01df7a32a8a14964b3
• Causal repair chain: the owner tree restores exact provider credential/endpoint isolation, authenticated failure classification, severity-boundary parsing, typed non-passing incomplete scans, and trusted-lock installation. It removes the duplicate direct-OpenAI fallback that normalized to primary gpt-5.6-luna; bounded same-model retries remain, and exhausted direct-primary execution fails closed. The current test makes the final empty-default fallback behavior explicit.
• Local owner-tree GREEN: 1,407 passed, 1 skipped, 16 subtests; full test_strix_quick_gate.sh: PASS.
• Hosted exact-head GREEN: Strix Changed Path Quality run 32693699962, job 97331828910, checked out exact SHA 7011fee, ran 1,407 tests·16 subtests, and recorded test_strix_quick_gate: PASS.
• Unresolved review threads: 0.
• Security/SBOM workflows remain queued and therefore non-passing; exact-current-head formal Reviews API verdict is absent.

Decision: WAIT_FOR_TERMINAL_EXACT_HEAD_CHECKS_AND_INDEPENDENT_FORMAL_REVIEW.

Supersedes #1213 with the protected-main replacement for the unsupported-temperature and cross-provider Strix failure. It also absorbs the two unique quality-path triggers from closed duplicate #1262.

Root causes

• Azure rejected the implicit temperature=0.2 sent for gpt-5.6-sol, while LiteLLM had no fallback model group for that selector.
• The outer fallback passed the workflow-facing openai-direct/ alias through without a valid LiteLLM provider, inherited the NVIDIA endpoint and credential, and failed before authoritative vulnerability evidence existed.
• A privileged same-repository Strix PR could replace the trusted workflow dependency lock with its PR-head lock before provider credentials reached the installed scanner.

Repair

• Recognize only the known full-line unsupported-temperature capability failure and reject split-line, target-prefixed, numeric-continuation, and standalone-brownout imitations.
• Require a non-identifier leading boundary on every severity marker so configuration identifiers such as STRIX_FAIL_ON_MIN_SEVERITY cannot be promoted into vulnerability findings; identifier-only reports remain non-authoritative and fail closed.
• Normalize openai-direct/ to trusted selector openai_direct/, then dispatch through LiteLLM provider form openai/ with the established direct-OpenAI credential and no inherited NVIDIA API base.
• Skip candidates lacking their trusted key, continue bounded fallback, and remain fail closed when exhausted.
• Keep only exact console/report advisories on the bounded allowlist; altered warning text remains blocking.
• Install only from the trusted workflow checkout lock, reject missing/symlinked locks, compare exact on-disk bytes with HEAD:requirements-strix-ci-hashes.txt via git hash-object --no-filters, and require hash-pinned wheels.
• Pin the smallest compatible LiteLLM wheel (1.94.2) because 1.94.1 has no Python 3.13 manylinux wheel.
• Retain the current default-branch smoke marker solely for bootstrap compatibility while executable contracts forbid a PR-head lock source.

Predecessor exact-head verification (non-passing after head movement)

• Head: 0445447310257bbd5eef1975c5dbcdf649f311df
• Protected base: main@93b9cbb145bfe03453c9421dcf7e0668da0f8050
• RED (provider alias): all four active fallback branches used openai-direct/gpt-5.6-luna; the focused contract failed because LiteLLM routing requires canonical openai_direct/*.
• RED (history integrity): required OpenCode run 32683332756, coverage job 97303853223, correctly rejected predecessor 05feeb1… because its post-merge restoration looked like a stale replay after the damaged 844cef4… merge tree.
• Causal history repair: merge anchor f428b3e553b8e52c4b0eacc46bc2941bef0aa231 keeps the complete owner tree and rebinds protected main as the second parent without changing content; the current doc-only descendant does not revert merged base work.
• Focused GREEN: 16 deterministic NVIDIA/provider fallback regressions, required-workflow smoke, shell syntax, and git diff --check pass; every active fallback branch uses openai_direct/gpt-5.6-luna, while only a scoped frozen-smoke compatibility comment retains the historical alias.
• Full Python: 1,407 passed, 1 skipped, 16 subtests.
• Full Strix quick-gate harness: PASS.
• Hosted exact-head run 32684026983, job 97305671238: exact checkout SHA attested, 1,407 tests·16 subtests and full quick-gate PASS.
• All 13 substantive exact-head repository workflows completed successfully; the repair-only workflow was skipped and is not counted as passing evidence.
• Duplicate fix(strix): correct STRIX_FALLBACK_MODELS provider prefix (hyphen -> underscore) #1285 was revalidated through its raced final head and closed after its provider-prefix and credential/endpoint-isolation intent was proven already present under this canonical owner.
• Unresolved review threads: 0.
• Exact-current-head formal Reviews API verdict: absent; predecessor CHANGES_REQUESTED, COMMENTED, and status-only evidence are non-passing.

No secret, raw-provider, approval, protection, model-selection, severity threshold, or incomplete-scan success boundary is weakened.

Decision: WAIT_FOR_INDEPENDENT_EXACT_HEAD_FORMAL_REVIEW.

Summary by CodeRabbit

• 보안
  • 신뢰된 검증 결과와 잠금 파일만 사용해 PR 기반 변조 위험을 줄였습니다.
  • 샌드박스 자격 증명, 실행 파일 설치, 액션 권한 및 커밋 고정 검사를 강화했습니다.
  • 심볼릭 링크, 범위를 벗어난 취약점 결과와 불완전한 증거를 차단합니다.
• 개선
  • Strix 모델 대체 경로와 오류 처리를 개선했습니다.
  • 직접 OpenAI 대체 실행을 지원하고 중복 대체 구성을 방지합니다.
  • 유효한 보고서 또는 취약점 없음 결과가 없으면 실패로 처리합니다.
• 문서
  • 의존성 보안 경계와 모델 대체 동작 안내를 추가했습니다.

[coderabbitai]

Note

Reviews paused

It looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the reviews.auto_review.auto_pause_after_reviewed_commits setting.

Use the following commands to manage reviews:

• @coderabbitai resume to resume automatic reviews.
• @coderabbitai review to trigger a single review.

Use the checkboxes below for quick actions:

• ▶️ Resume reviews
• 🔍 Trigger review

📝 Walkthrough

Walkthrough

Strix 워크플로가 신뢰된 의존성 잠금, 검증된 dispatch 식별자, 격리된 상태 게시 및 Docker 샌드박스를 사용합니다. 게이트가 공급자별 폴백, 모델 정규화, 보고서 검증, advisory 필터링 및 authoritative 실행 evidence를 검사합니다. 관련 smoke 테스트와 회귀 테스트를 확장했습니다.

Changes

Strix 보안 및 폴백 처리

Layer / File(s)	Summary
신뢰된 워크플로 경계와 상태 게시 .github/workflows/strix.yml, scripts/ci/strix_required_workflow_smoke.sh, tests/test_required_workflow_queue_contract.py, tests/test_strix_workflow_dependency_hashes.py, requirements-strix-ci*.txt	검증된 dispatch 메타데이터를 job output으로 전달하고 상태 게시를 별도 job으로 분리합니다. PR head 잠금 설치를 제거하고 신뢰된 잠금의 blob·해시와 바이너리 전용 설치를 검증합니다. 권한, action SHA, 모델 입력 및 Docker 자격 증명 경계를 검사합니다.
공급자 폴백과 모델 라우팅 scripts/ci/strix_model_utils.sh, scripts/ci/strix_quick_gate.sh, scripts/ci/test_strix_quick_gate.sh, tests/test_strix_nvidia_nim_not_found_fallback.py, docs/doctoring/strix-unsupported-sampling-fallback.md, requirements-strix-ci*.txt	openai-direct/ 별칭을 openai_direct/ 형식으로 정규화합니다. 공급자별 키와 API base를 선택합니다. 지원되지 않는 sampling parameter 오류와 외부 fallback 경로를 처리합니다. LiteLLM을 1.94.2로 고정합니다.
실행 증거와 보고서 판정 scripts/ci/strix_quick_gate.sh, scripts/ci/test_strix_quick_gate.sh, tests/test_strix_severity_identifier_boundary.py, docs/doctoring/strix-nvidia-nim-not-found-fallback.md	심볼릭 링크 보고서와 대상 경로 이탈을 거부합니다. Markdown·JSON 보고서와 severity 표식을 검사합니다. 허용된 advisory를 필터링합니다. 구조화된 severity 또는 zero-findings evidence가 없으면 성공 실행을 실패 처리합니다. threshold vulnerability와 인프라 오류는 fail-closed로 처리합니다.
문서와 품질 게이트 docs/doctoring/strix-trusted-dependency-lock.md, docs/doctoring/strix-unsupported-sampling-fallback.md, CHANGELOG.md, .github/workflows/strix-changed-path-quality-ci.yml, tests/test_strix_workflow_dependency_hashes.py	신뢰된 잠금, 샌드박스, 토큰 격리, sampling fallback 및 fail-closed 경계를 문서화합니다. 관련 문서와 모델 유틸리티 변경 시 품질 검사가 실행되도록 경로 계약을 갱신합니다.

Estimated code review effort: 4 (Complex) | ~75 minutes

Merge Risk:🟡 Moderate · up to 026c2

The PR makes Azure and cross-provider fallback routing executable and constrains dependency installation to the trusted lock, reducing fallback failures and untrusted dependency risk. Functional checks pass, but current-head security/SBOM validation is still queued and one documentation reference is invalid, so merge should wait for terminal validation and the small documentation fix.

Sequence Diagram(s)

sequenceDiagram
participant GitHubWorkflow
participant DispatchMetadata
participant StrixQuickGate
participant LiteLLM
participant FallbackProvider
participant StatusPublisher
GitHubWorkflow->>DispatchMetadata: 저장소와 HEAD 검증
DispatchMetadata-->>StrixQuickGate: 검증된 식별자 전달
GitHubWorkflow->>StrixQuickGate: 신뢰된 잠금과 fallback key 전달
StrixQuickGate->>LiteLLM: 정규화된 모델 요청
LiteLLM-->>StrixQuickGate: 실행 결과 또는 fallback 가능 오류 반환
StrixQuickGate->>FallbackProvider: 조건부 fallback 요청
StrixQuickGate-->>StatusPublisher: 검증 결과와 scan enabled 전달
StatusPublisher->>StatusPublisher: 성공 evidence 확인 후 상태 게시

Loading

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	Docstring check was indeterminate for this PR — some files could not be analyzed in time. Not blocking.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	제목은 PR의 주요 목표인 Azure 및 제공자 간 Strix fallback 실행 수정을 정확히 설명합니다.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch codex/strix-provider-failover-main-20260823

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[seonghobae]

Independent current-head review request for exact HEAD b05edc069cba461ad1d641ce8809027ad909c14b.

Gap G-03: required Strix still fail-closes 0-vuln complete scans on MODEL QUALITY WARNING and openai-direct/ LiteLLM LLM Provider NOT provided until #1263 (b05edc069cba461ad1d641ce8809027ad909c14b) is independently approved and merged. Predecessor-head reviews are not merge evidence. Do not self-approve. COPILOT_GITHUB_TOKEN is not used.

[seonghobae]

@opencode-agent review

Independent exact-current-head review requested for 1fd718f3177d3e8ffe908aed38b50dc94e926f8d (base main@885f2cd251999f21cf562cab3e2d9cc3cc3ec737).

Verification on this exact head:

• RED reproduced repository-controlled TARGET OUTPUT: Error: litellm.BadRequestError ... incorrectly authorizing a cross-provider fallback; the anchored provider-line matcher and target-prefix fixture are now GREEN.
• Focused provider/fallback suite: 16 passed.
• Full local suite: 1,395 passed, 1 skipped, 16 subtests.
• Hosted Strix quality run 32648578834/job 97216588527 checked out this exact SHA, passed the same 1,395-test breadth, and completed test_strix_quick_gate: PASS.
• Hosted organization policy run 32648578846/job 97216556061 checked out this exact SHA and passed 34 tests.
• Current review threads are resolved. Pending/queued checks and predecessor reviews are not counted; please issue a formal Reviews API verdict only for this exact head.

[seonghobae]

Exact-head 1fd718f3177d3e8ffe908aed38b50dc94e926f8d remains blocked on two independent human approvals under ruleset 18156473. Bot reviews (Devin/CodeRabbit) are present but do not satisfy the ruleset. Requesting independent review so the required-workflow SHA can pick up the openai-direct rewrite. Do not weaken scanner gates.

[seonghobae]

Strix default-branch dispatch exposed a real privileged-workflow supply-chain defect: the scan replaced the trusted workflow dependency lock with a same-repository PR-head lock before provider credentials later reached the installed scanner. Exact head de3abcd removes that materialization and makes both bounded smoke contracts plus pytest enforce trusted-source-only dependencies. Validation: 1396 passed, 1 skipped, 16 subtests; focused Strix set 40 passed plus 16 subtests; required-workflow smoke and shell syntax passed; git diff --check passed. No gate was weakened: incomplete provider scans and reported vulnerabilities remain non-passing.

[seonghobae]

Strix default-branch dispatch exposed a real privileged-workflow supply-chain defect: the scan replaced the trusted workflow dependency lock with a same-repository PR-head lock before provider credentials later reached the installed scanner. Exact head 975a9e4 removes that operation, keeps installation on the trusted-source lock, and retains only the current default-branch smoke marker needed for bootstrap compatibility. Both old-main and exact-head bounded smoke contracts pass. Validation: 1396 passed, 1 skipped, 16 subtests; focused Strix set 40 passed plus 16 subtests; shell syntax and git diff --check passed. No gate was weakened: incomplete provider scans and reported vulnerabilities remain non-passing.

[seonghobae]

Exact current head 7011fee275eaa257ce491efb4812dd3e98ed649e remains source-clean and all review threads are resolved. I downloaded the failed Strix artifact from run 32693700056:

• primary nvidia_nim/nvidia/nemotron-3-super-120b-a12b returned HTTP 429 rate limit on retries;
• NIM fallback nvidia_nim/nvidia/llama-3.3-nemotron-super-49b-v1.5 produced a zero-finding report but the gate retained the earlier provider-failure signal and correctly failed closed;
• direct fallback openai-direct/gpt-5.6-luna then returned HTTP 404.

The artifact proves provider availability/quota failure, not a source vulnerability. No scan result was promoted to success. A failed-job rerun was requested; its new queue is tracked separately. No merge/readiness claim until a fresh same-head authoritative Strix result and protected approvals exist.

[seonghobae]

Exact-head update: 94f56fd2cc269fffcc55c453e22eca5676f97b1b

Root-cause repair for the observed cross-provider Strix failure:

• NIM rate-limit/provider exhaustion remains fail-closed.
• Direct OpenAI fallback now receives a trusted explicit https://api.openai.com/v1 API-base file, alongside the already-separated OpenAI fallback credential, so it cannot inherit the exhausted NIM/OpenRouter/GitHub Models endpoint.
• Added an exact resolver regression test and updated doctoring/CHANGELOG evidence.

Exact-head local verification on this tree: 17 Strix fallback tests; 98 related Python contract tests; Bash syntax; actionlint; and git diff --check passed. The full shell harness was also attempted and exposes a pre-existing nested PR-head fixture/path-safety mismatch; no test or gate was weakened.

@opencode-agent Please perform a new substantive review against this exact head only, including changed-file walkthrough, fallback credential/API-base boundary, security impact, and test evidence. Do not reuse predecessor-head evidence.

[seonghobae]

Exact-head repair update: d6c34c59f18e959596ddc18f2c169d312e571758 (parent 94f56fd2cc269fffcc55c453e22eca5676f97b1b).

• RED: a complete zero-finding scan whose target narrative mentioned OpenAI on one line and documented Fatal/Denied/Warn/Warning labels on another was falsely classified as provider infrastructure failure.
• Fix: require an LLM-provider marker and the failure word on the same physical non-allowlisted line.
• GREEN: the new negative control and existing Fatal/Warning/Denied positive controls pass the focused shell harness; genuine same-line provider signals remain fail-closed.
• Syntax and whitespace checks pass.

All predecessor reviews/checks are historical. Current hosted runs are queued/in progress and therefore non-passing.

@opencode-agent Please perform a new substantive formal review against exact head d6c34c59f18e959596ddc18f2c169d312e571758 only. Validate the same-line authentication boundary, negative target-output spoof control, genuine provider-failure controls, and complete changed diff; do not reuse predecessor verdicts.

[seonghobae]

Final exact-head follow-up: a73831f60f83a61df95dbcb6999094fab80bb352.

The first full-harness pass correctly exposed a regression in the initial narrowing: an exact HF advisory with a malicious appended Fatal suffix could pass because HF Hub was not in the same-line provider context. That predecessor head is rejected.

The final minimal repair includes the exact HF Hub advisory context in the same-line authentication boundary. Focused GREEN now covers: clean target narrative, genuine litellm Fatal, openai Warning, anthropic Denied, and altered HF-advisory suffix. The clean narrative passes; all four authenticated/altered provider signals remain fail-closed.

All hosted checks and reviews for earlier heads are historical. New exact-head workflows are pending. @opencode-agent review this exact head only, including the altered-advisory negative control.

[seonghobae]

Disposition for current informational threads on exact head a73831f60f83a61df95dbcb6999094fab80bb352:

• PRRT_kwDOS_C14s6bnVv- (console warning context): confirmed intentional. Bare target/repository narrative is no longer provider evidence; the same physical line must carry an authenticated LLM-provider marker (or exact HF Hub advisory context) and the failure word. Specific timeout/rate-limit/token/connection/service/model classifiers and report-artifact detection remain unchanged. Focused TDD covers clean split-line narrative, genuine provider signals, and altered HF-advisory suffix.
• PRRT_kwDOS_C14s6bnVxn (status follow-up token): confirmed intentional credential isolation. The follow-up first uses the exchanged target-app token, then the two explicit central credentials. Its GITHUB_TOKEN is read-only, so it cannot accidentally broaden target writes. If none can publish, no strix commit status is fabricated; target branch protection remains blocked/authoritative. A complete successful scan run remains source evidence, while an absent required status is still non-passing governance evidence. test_strix_scan_cannot_read_target_pr_or_publish_status and the trusted required-workflow smoke pin this boundary.

No source change is justified by either informational observation.

[seonghobae]

Exact-head remediation on 4fb737a: fixed the Strix Changed Path Quality CI trigger so changes to requirements-strix-ci.txt, requirements-strix-ci-hashes.txt, and tests/test_strix_severity_identifier_boundary.py rerun the exact-head quality contract. Added regression assertions for all four contract inputs. Focused verification: 16 passed; Strix shell syntax; interrogate docstring coverage 100 percent; git diff --check passed. Please publish a fresh independent substantive review and required hosted checks for this exact head. Previous-head evidence is stale; no gate was weakened and no bypass was used.

[seonghobae]

Exact-head review disposition for 4fb737a6c44ad3fc73270f55ff8a5abb415f1e3c:

• requirements-strix-ci.txt and requirements-strix-ci-hashes.txt are now explicit changed-path quality trigger inputs.
• tests/test_strix_severity_identifier_boundary.py is now an explicit trigger input.
• tests/test_strix_workflow_dependency_hashes.py pins all three paths so this cannot silently drift.
• Focused current-tree verification: 16 tests and 16 subtests passed; both contract files compile.

This addresses threads PRRT_kwDOS_C14s6bnfnn and PRRT_kwDOS_C14s6bnfpK. Hosted exact-head workflows remain authoritative and non-passing until terminal.

[seonghobae]

Exact-head local verification completed for 4fb737a: full pytest 1408 passed, 1 skipped, 16 subtests; focused trigger contract 16 passed; interrogate docstring coverage 100 percent; shell syntax, compileall, and git diff --check passed. Hosted current-head workflows and independent approval remain pending, so this PR stays on the normal protected path.

[seonghobae]

Hosted exact-head verification is now available for 4fb737a6c44ad3fc73270f55ff8a5abb415f1e3c: Strix Changed Path Quality run 32705249918, job 97364828647, checked out the exact submitted SHA, ran 1,408 tests·16 subtests, and recorded test_strix_quick_gate: PASS. All review threads are resolved.

The remaining queued security/SBOM workflows and absence of an exact-current-head qualifying formal verdict are non-passing; no merge is requested until those become authoritative.

[seonghobae]

Protected-main synchronization evidence for exact head ab3d764547082e1b55b6257cc1cd9aa5d951fa30:

• non-force merge parents: prior owner head 4fb737a6c44ad3fc73270f55ff8a5abb415f1e3c and protected main@613a33e0cb1c6db9790fae99f6253445712ac37a;
• exact tree: ed281e2f4a50bad277524b7282ab490d1661e466, matching the locally verified merge tree byte-for-byte;
• resolved the only semantic conflict by retaining the completed openai_direct/* migration contract: fix(strix): unblock trusted smoke during provider-prefix migration #1286's temporary either-prefix smoke compatibility is no longer needed on this successor, and active openai-direct/* remains rejected;
• scripts/ci/strix_required_workflow_smoke.sh: PASS;
• full scripts/ci/test_strix_quick_gate.sh: PASS;
• git diff --check: PASS;
• unresolved review threads: 0.

New hosted quality, security, SAST, OSV, SBOM, provenance, and CodeQL runs are queued/in progress and remain non-passing until terminal success. No predecessor check or review is transferred to this head, and no merge is requested yet.

[seonghobae]

Current-head review refresh requested.

• Exact head: ab3d764547082e1b55b6257cc1cd9aa5d951fa30
• Base: 613a33e0cb1c6db9790fae99f6253445712ac37a
• Current review threads are resolved; older CHANGES_REQUESTED evidence is not being treated as a current verdict.
• All 11 required Checks are newly queued for this exact head.

Please run a new independent substantive review against this exact head before merge.

[seonghobae]

Current-head review loop

• Pull request: fix(strix): make Azure and cross-provider fallbacks executable #1263
• Exact head checked: ab3d764547082e1b55b6257cc1cd9aa5d951fa30
• Current source inspection confirms the earlier cross-provider fallback finding is resolved: the trusted quick gate validates and reads STRIX_OPENAI_FALLBACK_KEY_FILE, selects the dedicated key for non-direct primaries, and uses the separately provisioned direct OpenAI API-base file. The changed-path trigger also contains the severity-boundary regression test.
• Current hosted feature Checks are successful; the remaining review state is stale/absent current-head approval rather than an unaddressed source defect.

Please publish a fresh substantive review for this exact head and replace stale review state through the normal review workflow. Do not merge until the protected approval gate is current.

[seonghobae]

Exact-head Strix RCA for ab3d764547082e1b55b6257cc1cd9aa5d951fa30:

• The source checks completed the scan with Vulnerabilities 0, but no authoritative report artifact was produced.
• NVIDIA NIM emitted provider infrastructure/failure-signal output; the configured openai-direct/gpt-5.6-luna fallback then failed with 404 page not found.
• This is provider/control-plane evidence, not a confirmed source vulnerability in this PR. The direct-OpenAI API-base repair is being handled by the related exact-head PR fix(strix): route direct-OpenAI fallback models to the OpenAI platform endpoint #1295; predecessor failure evidence is not reused as approval.
• Required checks are otherwise substantially green, but Strix is failing closed, coverage-source-tree and remaining security checks are pending, and current-head independent approval is absent.

Decision: WAIT_AND_REMEDIATE. No force merge and no merge until the provider route is repaired and fresh exact-head Strix evidence plus substantive approval are available.

[opencode-agent]

Pull request overview

OpenCode could not approve from deterministic current-head evidence because GitHub Checks have failed.

Findings

1. HIGH Current-head GitHub Checks - Fix failed required checks before approval

• Problem: Failed same-head checks remain for ab3d764547082e1b55b6257cc1cd9aa5d951fa30.
• Root cause: The model-unavailable evidence fallback is allowed only when peer GitHub Checks are complete and clean.
• Fix: Read and fix the failed check logs below, then rerun the current-head checks.
• Regression test: Keep the model-unavailable fallback gated on an empty failed-check rollup.

Failed checks:

• Strix Security Scan/strix: FAILURE (https://github.com/ContextualWisdomLab/.github/actions/runs/32712851121/job/97387875273)
• Strix Security Scan/strix: failure (https://github.com/ContextualWisdomLab/.github/actions/runs/32712851121/job/97387875273)

Changed-File Evidence Map

flowchart LR
PR["PR changed files"] --> Evidence["OpenCode bounded evidence"]
Evidence --> S1["Workflow (2 files)"]
S1 --> I1["GitHub Actions review job"]
I1 --> R1["Review risk: Workflow (2 files)"]
R1 --> V1["actionlint plus required checks"]
Evidence --> S2["Changed file (3 files)"]
S2 --> I2["repository behavior"]
I2 --> R2["Review risk: Changed file (3 files)"]
R2 --> V2["required checks"]
Evidence --> S3["Docs (3 files)"]
S3 --> I3["operator or user guidance"]
I3 --> R3["Review risk: Docs (3 files)"]
R3 --> V3["docs review"]
Evidence --> S4["CI script (4 files)"]
S4 --> I4["review and security gate shell path"]
I4 --> R4["Review risk: CI script (4 files)"]
R4 --> V4["bash -n plus Strix self-test"]
Evidence --> S5["Test (4 files)"]
S5 --> I5["regression suite"]
I5 --> R5["Review risk: Test (4 files)"]
R5 --> V5["targeted test run"]

Loading

[seonghobae]

Current-head owner-path revalidation for the downstream Inkspan gate failure:

• canonical owner PR: fix(strix): make Azure and cross-provider fallbacks executable #1263ab3d764547082e1b55b6257cc1cd9aa5d951fa30
• independently resolved protected base / trusted Strix source: main@613a33e0cb1c6db9790fae99f6253445712ac37a
• current-head required Strix failure: run 32712851121, job 97387875273
• downstream .github#897@6406b4663b216e4a85fbf96a7cbd33bf9c0a990e: run 32710533347, job 97380758535
• downstream Inkspan stack affected: ContextualWisdomLab/inkspan#299@1efbd632719b0e54c5a33e3a95224a4bd2f869d3

The current required run reconfirms the previously identified self-hosting boundary rather than a new branch defect. pull_request_target checks out the target PR head for scanning, but executes TRUSTED_STRIX_GATE from protected main@613a33e…. In job 97387875273, the primary NVIDIA model is 429-limited; the NVIDIA fallback then reaches Penetration test completed with Vulnerabilities 0 (No exploitable vulnerabilities detected), after which the protected-main gate still emits Strix run emitted provider infrastructure or failure-signal output; failing closed. The final direct-OpenAI fallback then returns 404 page not found. Therefore an unchanged rerun cannot prove this branch's #1291 repair: the required gate source is the predecessor protected-main implementation.

Falsifiable acceptance remains: a completed zero-finding Strix run carrying only the normal MODEL QUALITY WARNING advisory must not be reclassified as provider failure; genuinely authenticated provider failure signals and incomplete scans must remain fail-closed per #891. The branch's exact-head quality run already exercises its changed gate, but that evidence cannot substitute for the required protected-source security check.

Smallest safe owner-path is controlled integration of this canonical gate repair under live governance, followed by a required Strix run whose trusted workflow/gate source SHA actually contains the repair. After that, revalidate .github#897@6406b466… and Inkspan #299 from their own exact heads; do not transfer predecessor or branch-only evidence and do not weaken the security requirement.

[opencode-agent]

Pull request overview

OpenCode could not approve from deterministic current-head evidence because GitHub Checks have failed.

Findings

1. HIGH Current-head GitHub Checks - Fix failed required checks before approval

• Problem: Failed same-head checks remain for ab3d764547082e1b55b6257cc1cd9aa5d951fa30.
• Root cause: The model-unavailable evidence fallback is allowed only when peer GitHub Checks are complete and clean.
• Fix: Read and fix the failed check logs below, then rerun the current-head checks.
• Regression test: Keep the model-unavailable fallback gated on an empty failed-check rollup.

Failed checks:

• Strix Security Scan/strix: FAILURE (https://github.com/ContextualWisdomLab/.github/actions/runs/32712851121/job/97387875273)
• Strix Security Scan/strix: failure (https://github.com/ContextualWisdomLab/.github/actions/runs/32712851121/job/97387875273)

Changed-File Evidence Map

flowchart LR
PR["PR changed files"] --> Evidence["OpenCode bounded evidence"]
Evidence --> S1["Workflow (2 files)"]
S1 --> I1["GitHub Actions review job"]
I1 --> R1["Review risk: Workflow (2 files)"]
R1 --> V1["actionlint plus required checks"]
Evidence --> S2["Changed file (3 files)"]
S2 --> I2["repository behavior"]
I2 --> R2["Review risk: Changed file (3 files)"]
R2 --> V2["required checks"]
Evidence --> S3["Docs (3 files)"]
S3 --> I3["operator or user guidance"]
I3 --> R3["Review risk: Docs (3 files)"]
R3 --> V3["docs review"]
Evidence --> S4["CI script (4 files)"]
S4 --> I4["review and security gate shell path"]
I4 --> R4["Review risk: CI script (4 files)"]
R4 --> V4["bash -n plus Strix self-test"]
Evidence --> S5["Test (4 files)"]
S5 --> I5["regression suite"]
I5 --> R5["Review risk: Test (4 files)"]
R5 --> V5["targeted test run"]

Loading

[seonghobae]

BandScope consumer canary for the current Strix owner lane (read-only handoff; no BandScope leaf workaround):

• Consumer PR: ContextualWisdomLab/bandscope#783
• Exact BandScope head: 1168c8f4257de5de036ea54bf5ee73edb83e775e
• Protected BandScope base: develop@acdbea6344fe1231c39535b575f4de35e4c607c9
• Failed required Strix run/job: 32696787054 / 97340299862
• Trusted central Strix source used by that run: .github@0c6b9a6459c9dbdf5e23fb01df7a32a8a14964b3
• Exact target-head binding was correct.

First causal sequence from the job log: the primary NVIDIA NIM route hit HTTP 429; a later NIM fallback completed with Vulnerabilities 0, but the gate still emitted Strix run emitted provider infrastructure or failure-signal output; failing closed, then the direct-OpenAI fallback failed routing with HTTP 404. No BandScope source/security finding was produced.

This matches this owner lane's authenticated-failure / provider-alias / fallback-isolation scope and #1291's zero-finding false-positive class. There is no correct BandScope-local repair: #783's repository-local CI, release, security-audit, aggregate Security Scan, SBOM, build-baseline, Bandit, Semgrep, secret-scan plus current central coverage/OpenCode/Noema evidence are already successful on the unchanged head; Strix remains the central security-evidence gate.

I have requested one rerun of the failed exact-head Strix job to distinguish a transient primary 429 from the deterministic fallback/control-plane defect. Do not treat that dispatch as success. Consumer acceptance after this PR integrates: rerun Strix against unchanged bandscope#783@1168c8f...; require exact-head terminal-success with authoritative zero-finding/finding evidence, no inherited prior-attempt infrastructure signal after a clean later attempt, and working provider-specific credential/API-base routing. Preserve fail-closed behavior for genuine provider failures and real findings.