Problem
Required Strix review on .github#2106 exact head 611ccd73460ab0188e0085956ade6180bb28a91a reported source findings against scripts/ci/pingora_edge_policy.py and scripts/ci/contextual_orchestrator_review_policy.py as part of that PR review, but neither file is in GitHub's canonical changed-file set for #2106.
This is not a claim that the underlying repository-wide concerns are harmless. It is an evidence-scope defect: a PR security verdict must distinguish an exact base→head delta finding from debt found in unchanged protected-base source.
Executable evidence
Authenticated PR identity:
- PR:
.github#2106
- base: protected
main@64f483db9d052322c65bcdf1675d66138156f306
- head:
611ccd73460ab0188e0085956ade6180bb28a91a
- GitHub changed-file inventory: exactly seven paths —
codeql-scan-dispatch.yml, CHANGELOG, ADR/doctoring/gap docs, and two CodeQL contract tests. Neither reported source file appears.
Blob identity proves the two reported files are unchanged across the exact tuple:
scripts/ci/pingora_edge_policy.py: base blob = head blob = eb0e3a741a93a89c0364af07ff963920d82f9684
scripts/ci/contextual_orchestrator_review_policy.py: base blob = head blob = 241f84bef79259087085099f16b86b2580d0d207
Nevertheless Strix job 103684345530 produced findings against both paths and caused the PR review lane to fail.
The Pingora finding has a real canonical owner: #2149. Current #2149 closes quoted image/absolute command/package-option/line-continuation cases but still lacks ARG/ENV variable-indirection coverage; exact-head review 5190657129 records that remaining owner-local finding. Do not repair it in #2106.
The review-policy endpoint finding also needs provenance-aware classification before being called credential exfiltration. The .github launcher copies chat_base_url / credential metadata from CO DiscoveredModel; at the currently proposed CO pin 012beaac…, discovered rows populate chat_base_url and auth_scheme from the code-owned ProviderModelSource, not directly from remote model-list response fields. The leaf bridge still duplicates routing authority and is scheduled for retirement under #1759/#2042, but the Strix report did not establish an attacker-controlled path for this exact execution.
Required owner repair
- Before publishing a PR source finding, bind repository, PR number, live open state, exact base ref/SHA and exact head SHA.
- Materialize the complete canonical GitHub changed-file inventory for that exact tuple, with rename status and pagination.
- A finding presented as a PR-delta finding must map to an authenticated changed path and, where line-level evidence is claimed, to a changed hunk/side. Base-identical paths must not be attributed to the PR delta.
- If bounded dependency/context closure legitimately exposes a defect in unchanged source, publish it separately as
repository_baseline / context_dependency evidence with explicit provenance. It may remain non-passing if policy requires that, but it must not be described as introduced or changed by the PR.
- Preserve security sensitivity: do not narrow the scanner workspace to changed files only. The model may need unchanged dependencies to judge changed code; the repair is evidence attribution, not context starvation.
- Fail closed if the authenticated changed-file set, base/head tuple, or finding path provenance cannot be established.
- Add deterministic RED fixtures for: changed source finding, unchanged dependency context, completely base-identical source finding, rename, stacked base, and stale-head report.
Acceptance
Replay an unchanged #2106-equivalent seven-path canary. A real defect in one of those seven paths still blocks. The Pingora/review-policy repository-baseline observations may be retained as separate owner debt, but they cannot masquerade as exact PR-delta findings. No synthetic success, source-neutral retrigger, scanner suppression, provider fallback, or branch-protection weakening is acceptance evidence.
Problem
Required Strix review on
.github#2106exact head611ccd73460ab0188e0085956ade6180bb28a91areported source findings againstscripts/ci/pingora_edge_policy.pyandscripts/ci/contextual_orchestrator_review_policy.pyas part of that PR review, but neither file is in GitHub's canonical changed-file set for #2106.This is not a claim that the underlying repository-wide concerns are harmless. It is an evidence-scope defect: a PR security verdict must distinguish an exact base→head delta finding from debt found in unchanged protected-base source.
Executable evidence
Authenticated PR identity:
.github#2106main@64f483db9d052322c65bcdf1675d66138156f306611ccd73460ab0188e0085956ade6180bb28a91acodeql-scan-dispatch.yml, CHANGELOG, ADR/doctoring/gap docs, and two CodeQL contract tests. Neither reported source file appears.Blob identity proves the two reported files are unchanged across the exact tuple:
scripts/ci/pingora_edge_policy.py: base blob = head blob =eb0e3a741a93a89c0364af07ff963920d82f9684scripts/ci/contextual_orchestrator_review_policy.py: base blob = head blob =241f84bef79259087085099f16b86b2580d0d207Nevertheless Strix job
103684345530produced findings against both paths and caused the PR review lane to fail.The Pingora finding has a real canonical owner: #2149. Current #2149 closes quoted image/absolute command/package-option/line-continuation cases but still lacks
ARG/ENVvariable-indirection coverage; exact-head review5190657129records that remaining owner-local finding. Do not repair it in #2106.The review-policy endpoint finding also needs provenance-aware classification before being called credential exfiltration. The
.githublauncher copieschat_base_url/ credential metadata from CODiscoveredModel; at the currently proposed CO pin012beaac…, discovered rows populatechat_base_urlandauth_schemefrom the code-ownedProviderModelSource, not directly from remote model-list response fields. The leaf bridge still duplicates routing authority and is scheduled for retirement under #1759/#2042, but the Strix report did not establish an attacker-controlled path for this exact execution.Required owner repair
repository_baseline/context_dependencyevidence with explicit provenance. It may remain non-passing if policy requires that, but it must not be described as introduced or changed by the PR.Acceptance
Replay an unchanged #2106-equivalent seven-path canary. A real defect in one of those seven paths still blocks. The Pingora/review-policy repository-baseline observations may be retained as separate owner debt, but they cannot masquerade as exact PR-delta findings. No synthetic success, source-neutral retrigger, scanner suppression, provider fallback, or branch-protection weakening is acceptance evidence.