Skip to content

[review-pull-request] Calibrate review-candidate-verifier external/unestablished handling and exclude unestablished DROPs from inline publication #70547

Open

Agent suggestions

Public preview

Description

@PureWeen

Is there an existing issue for this?

  • I have searched the existing issues

Is your feature request related to a problem? Please describe the problem.

Part of EPIC #70231.

Following #70250 (d12d02b143, which added .github/agents/review-candidate-verifier.agent.md) and #70319 (35fbed48c8, which refined its rules), review-candidate-verifier and SKILL.md have exhibited two coupled calibration issues across recent hosted /review runs:

1. Valid findings downgraded to [UNRESOLVED] when in-repo files or standard platform/compiler semantics establish the link

.github/agents/review-candidate-verifier.agent.md:12-20, 24-36 restricts the verifier to the assigned diff, assigned frozen files, and assigned contract artifacts (allowing only an assigned stable BCL contract to establish a link), and :94-98 routes any link depending on an external library, runtime, or transport behavior outside the frozen bundle to DOWNGRADE (downgradeKind: "external", confidence: unresolved). In practice, valid findings have been published as [UNRESOLVED] even when the relevant evidence is in source/<sha> or governed by standard W3C DOM / C# ILLink rules:

  • Preserve FocusOnNavigate focus when interactive component replaces prerendered DOM #70308 (run 43078080293, r4541841629 on src/Components/Web.JS/src/Rendering/FocusOnNavigate.ts:113):
    • Valid bug fixed by @dariatiurina in d2783aa062: restoreNavigationFocus used postorder depthFirstNodeTreeTraversal while onAfterRenderCompleted in the same file used document.querySelector(selector)?.focus(). Published as [UNRESOLVED] because the verifier treated W3C DOM querySelector preorder semantics and selector contract parity within FocusOnNavigate.ts as missing from the assigned bundle.
  • Bound Blazor client validation caching #70453 (run 42926638607, r4532101438 on src/Components/Endpoints/src/Forms/ClientValidationCache.cs:59):
    • Valid bug confirmed by @halter73 in r4532399338 and 5851887134 ("This looks like a good catch. The NativeAOT publish log reports IL2080 here..."): GetValidatableFieldMetadata called unsuppressed fieldKey.ContainerType.GetProperty(...) outside BuildFieldMetadata's [UnconditionalSuppressMessage("Trimming", "IL2070")] scope (ClientValidationCache.cs:215) while fieldKey.ContainerType (Type) lacked [DynamicallyAccessedMembers]. Published as [UNRESOLVED] because the verifier stated the bundle had "no target-version linker contract, linker source, or immutable publish artifact proving that this assembly/method is analyzed and that this call emits IL2070", even though BuildFieldMetadata (and two helpers at :161, :181) in the same file already carry [UnconditionalSuppressMessage("Trimming", "IL2070")] (establishing that the code participates in trimmer/NativeAOT analysis) and C#/ILLink [DynamicallyAccessedMembers] / [UnconditionalSuppressMessage] lexical scoping is a standard compiler/linker rule.
  • Fix Blazor Server navigation during circuit startup #70499 (run 43247376089, r4552534794 on src/Components/Web.JS/src/Platform/Circuits/CircuitManager.ts:544, open PR):
    • NavigationEnhancement.ts:100-108 handles same-page #fragment clicks via history.pushState without calling notifyEnhancedNavigationListeners, leaving lastLocationChange unset when CircuitManager.ts:544 replays initial URI changes on Server renderer attach. Published as [UNRESOLVED] citing browser history.pushState not firing navigation events and SignalR ISingleClientProxy transport, even though history.pushState is standard HTML/DOM behavior and SignalR (src/SignalR/**, src/Components/Server/**) is in source/<sha>.

2. Over-broad external/unestablished classification of in-repo SignalR behavior (#69767) and dropBasis: "unestablished" inline publication (SKILL.md:301-305)

On #69767 (run 43134455604, r4545171667 on src/Components/Server/src/Circuits/RemoteNavigationManager.cs:146 at frozen head 8403c802c0), /review published [P2][UNRESOLVED] Confirm the retained-proxy behavior that produces this timeout, asking to verify from frozen source that target-version SignalR drops messages on ISingleClientProxy while disconnected rather than buffering them — even though src/SignalR/** and src/Components/Server/** are in source/<sha> and the PR includes unit test NavigateTo_WhenInteropCannotComplete_DoesNotRaiseUnhandledException (src/Components/Server/test/Circuits/RemoteNavigationManagerTest.cs:19) and E2E test NavigationIssuedWhileDisconnected_TimesOutWithoutTerminatingReconnectedCircuit (src/Components/test/E2ETest/ServerExecutionTests/NavigationReconnectionTest.cs:35).

Furthermore, .github/skills/review-pull-request/SKILL.md:301-305 currently maps both DOWNGRADE (downgradeKind: "external") and DROP (dropBasis: "unestablished") into NEW_FINDINGS with confidence: unresolved ([UNRESOLVED]). Consequently, even when the verifier rejects an unverified hypothesis with decision: "DROP" (dropBasis: "unestablished"), SKILL.md:301-305 still publishes it as an inline PR comment.

Describe the solution you'd like

  1. Require in-repo inspection before claiming external (.github/skills/review-pull-request/SKILL.md:267-292, .github/agents/review-candidate-verifier.agent.md:12-20, 24-36, 94-98):
    • Require the coordinator's verifier package and the verifier itself to inspect in-repo callers, sibling assemblies (src/SignalR/**, src/Components/**, template callers), existing in-file/in-assembly trimmer/AOT annotations ([UnconditionalSuppressMessage("Trimming" | "Trimmer", ...)], [DynamicallyAccessedMembers], [RequiresUnreferencedCode], [RequiresDynamicCode]), and project/directory MSBuild properties (.csproj, Directory.Build.props, eng/TrimmableProjects.props) inside source/<sha> before classifying a first-party dependency or trimmer/AOT contract as external.
  2. Allow documented, version-invariant platform/language/compiler semantics in links.trigger / links.effect (review-candidate-verifier.agent.md:31-36, 94-98):
    • Extend the existing assigned-BCL allowance (lines 31-34) so standard, version-invariant runtime/platform/compiler contracts — specifically BCL APIs, W3C DOM/HTML specifications (e.g., document.querySelector preorder tree-order matching, history.pushState not firing popstate/navigation events), and C#/ILLink [DynamicallyAccessedMembers] / [UnconditionalSuppressMessage] lexical scoping rules when trimmer/AOT analysis is evidenced by in-file/in-assembly trimmer attributes or project properties — may establish links.trigger and links.effect without requiring an external spec file in the bundle (KEEP -> confidence: verified).
    • Reserve DOWNGRADE (downgradeKind: "external") for genuine external library/runtime implementation details not present in source/<sha> (such as MessagePack internal bounds checks on Reduce allocations when parsing SignalR MessagePack messages #70143, org.msgpack.core.MessageUnpacker state after exceptions on Validate SignalR Java MessagePack binary lengths before allocation #70260, dotnet/runtime getAssemblyExports internals on [blazor] Fix Web Worker initialization on CoreCLR #70504, or MSAL error string formatting on Update azure/msal-browser #66055).
  3. Stop publishing DROP (dropBasis: "unestablished") as inline PR review comments (.github/skills/review-pull-request/SKILL.md:301-305):
  4. Synchronize embedded workflow verifier instructions:
    • Mirror all .github/agents/review-candidate-verifier.agent.md updates in .github/workflows/pull-request-review.md:995-1111 and recompile .github/workflows/pull-request-review.lock.yml (enforced by .github/workflows/tests/pull-request-review-adaptive-validation.test.mjs:49-58).

Additional context

Activity

  1. github-actions commented on Nov 12, 2026

    @github-actions
    Contributor

    Triage Summary

    Area: area-infrastructure (review-pull-request skill / review-candidate-verifier agent tooling)
    Type: Task (bounded calibration of agent behavior)

    Potential Duplicates

    • None found

    Notes

    • Missing: the issue body contains only -; details of the observed verifier behavior, an example PR, and the expected handling are needed to act on it.

    Generated by Issue Triage Agent for dotnet/aspnetcore for #70547 · copilot · auto · 13.4 AIC · ⌖ 1.02 AIC · ⊞ 7.5K · ◷

  2. PureWeen commented on Nov 13, 2026

    @PureWeen
    MemberAuthor

    Additional live occurrence on 2026-11-13: PR #70562 (run 43375184261)

    Run 43375184261 on #70562 (Add the Screen Orientation API to Microsoft.AspNetCore.Components.Platform, frozen head 897ea9d3889cd8c606f41c84f0bd89e75b8c0c1c) published two verifier-dropped (decision: "DROP", dropBasis: "unestablished") candidates as inline [UNRESOLVED] PR review comments:

    1. r4560194274 (src/Components/Platform/src/ScreenOrientation.cs:111) — **[UNRESOLVED] A late successful orientation acquisition is never disposed** (P2):

      • Verifier output: decision: "DROP", dropBasis: "unestablished", noting: "the assigned base-tip patterns deliberately skip pending acquisitions rather than requiring late-result cleanup."
      • Why the claim is also contradicted by repo rules: .github/skills/port-components-platform-api/SKILL.md:113-119 explicitly mandates this exact behavior: "Release the reference only when the acquisition has already completed successfully. A reference that arrives after disposal is never released. This is accepted. It happens only when the scope ends during an acquisition, and it leaves at most one reference for each family. A circuit that waits a minute to close costs more."
    2. r4560194281 (src/Components/Platform/testassets/PlatformTestApp.E2E.Tests/Tests/ScreenTests.cs:54) — **[UNRESOLVED] Orientation coverage is outside the prescribed Components Selenium surface** (P3):

      • Verifier output: decision: "DROP", dropBasis: "unestablished", noting missing red/green evidence for a Selenium test.
      • Why the claim is also contradicted by repo rules: .github/skills/port-components-platform-api/SKILL.md:155-158 explicitly mandates that Microsoft.AspNetCore.Components.Platform E2E coverage live in PlatformTestApp and PlatformTestApp.E2E.Tests (the package's Playwright E2E suite), whereas src/Components/AGENTS.md / docs/BlazorComponentsGuidance.md states the general src/Components/test/E2ETest Selenium rule without mentioning the src/Components/Platform/testassets/PlatformTestApp.E2E.Tests exception.
  3. PureWeen commented on Nov 13, 2026

    @PureWeen
    MemberAuthor

    Additional live occurrence on 2026-11-13: PR #69850 (run 43394407217) — valid in-repo bug unnecessarily downgraded to [UNRESOLVED] (downgradeKind: "external")

    Run 43394407217 on #69850 ([Blazor] Preserve modal dialog state during enhanced SSR DOM synchronization, frozen head 52f54afd090f74264ed0c02e4d158e36983b203b) found a genuine P2 regression at r4561309455 (src/Components/Web.JS/src/Rendering/DomMerging/DomSync.ts:287):

    • Finding: **[UNRESOLVED] Open data-permanent dialogs are no longer preserved across cross-page enhanced navigation** — the new !options.preserveModalState && a instanceof HTMLDialogElement && a.matches(':modal') check in domNodeComparer returns UpdateCost.Infinite before the data-permanent match check (DomSync.ts:325-332), routing matching <dialog data-permanent> elements through treatAsDeletion (nodeToDelete.close()) and replacement instead of treatAsMatch (isDataPermanentElement retention at DomSync.ts:199-206).
    • Verifier output: review-candidate-verifier confirmed all six repository proof links (changedLine, trigger, oldVsNew, effect, contract, intent), yet still returned decision: "DOWNGRADE", downgradeKind: "external" with:

      "missingEvidence": "A target-version-applicable browser/DOM contract or assigned immutable validation artifact proving that showModal() makes the dialog match :modal and that the cited close/removal operations have the claimed observable modal and identity effects."

    • Why this illustrates over-downgrading on standard platform contracts: HTMLDialogElement.showModal() causing <dialog> to match :modal is standard HTML specification behavior and is the exact premise of the PR's own added condition (a instanceof HTMLDialogElement && a.matches(':modal')). Requiring an external browser specification file in the frozen Git bundle to prove standard DOM semantics forces valid browser/JS findings into [UNRESOLVED].

    (Also note on #70562: author @jsakamotoIGJP replied to both [UNRESOLVED] comments r4560194274 and r4560194281 confirming both were intentional per .github/skills/port-components-platform-api/SKILL.md and resolving both threads; see also #70565.)

  4. PureWeen commented on Nov 13, 2026

    @PureWeen
    MemberAuthor

    Updates on 2026-11-13: #70499 (r4552534794) fixed by author + additional downgradeKind: "external" occurrence on #70343 (r4562634162)

    1. Author fix on #70499 (r4552534794):
      • The [P2][UNRESOLVED] finding from run 43247376089 (CircuitManager.ts:544, downgraded to external only because browser history.pushState and SignalR transport were outside the frozen bundle) was confirmed and fixed by @karmegams02 in commit 5633207bdd (r4562990680), which records recordLocationChange(absoluteInternalHref, true) in NavigationEnhancement.ts's hash-only branch and adds E2E coverage.
    2. Additional downgradeKind: "external" occurrence on #70343 (run 43414973483, r4562634162):
      • On frozen head e86b5e3f868e574aa8287057750fa6f6d452fbc4, BlazorComponentsGuidance identified that changing InputSelect.cs:56 to CurrentValueAsString ?? string.Empty leaves <option selected disabled>Choose...</option> (without value="") selected during static SSR (StaticHtmlRenderer.HtmlWriting.cs:113-156,307-339), whereas interactive activation reapplies select.value = "" (BrowserRenderer.ts:303, DomSpecialPropertyUtil.ts:84,135-149) and clears the selection (selectedIndex 0 → -1, already confirmed interactively in r4560077095).
      • static-ssr-verifier returned decision: "DOWNGRADE", downgradeKind: "external" solely because standard HTML <select> parsing and HTMLSelectElement.value = "" DOM semantics are not in the repository source bundle.
  5. PureWeen commented on Nov 17, 2026

    @PureWeen
    MemberAuthor

    Post-#70319 corpus update (2026-11-17)

    Even with #70319 merged on main (.github/agents/review-candidate-verifier.agent.md:32-36), subsequent hosted /review runs continue to show both directions of this calibration gap:

    A. Deterministic in-repo and standard-contract findings still downgraded to [UNRESOLVED] (11 of 14 recent valid findings)

    B. Unverified external BCL/SDK assumptions still published inline as [UNRESOLVED] instead of dropped from NEW_FINDINGS

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions