ci-copilot: support net11-targeting PRs (auto-detect base branch + #35677 on net11.0)

[kubaflo]

Note

Are you waiting for the changes in this PR to be merged?
It would be very helpful if you could test the resulting artifacts from this PR and let us know in a comment if this change resolves your issue. Thank you!

Summary

Enables the Copilot PR-review pipeline (maui-copilot) to review net11.0-targeting PRs, not just main/net10 ones — without merging anything to main. This is the net11 sibling of #35677.

Two pieces, both on this branch (based on origin/net11.0):

1. Ports Enhance MauiBot AI summary review output + support net11-targeting PRs #35677 (enhanced MauiBot AI review) onto net11.0 — the 12-file content of Enhance MauiBot AI summary review output + support net11-targeting PRs #35677 applied cleanly on top of net11.0 (zero conflicts), so the net11 line carries the enhanced AI Summary review output, token-usage aggregation, and deep-UI-test review-id patching.

2. ci-copilot.yml now auto-detects the PR's base branch. Previously the pipeline installed workloads and squash-merged the PR using whatever branch it was launched from (self). A net11-targeting PR reviewed from a main-based pipeline branch (or vice-versa) therefore got the wrong .NET SDK band and hit cross-branch squash-merge conflicts.

   A new "Resolve PR base branch" step (in both the CopilotReview and DeepUITests stages):

   • reads the PR's baseRefName via gh,
   • allowlist-validates it (^(main|net[0-9]+\.0)$, fail-closed),
   • and switches the worktree (detached) to origin/<baseRef> before workload install and the squash-merge.

   build.ps1 --target=dotnet then provisions the correct band (net11 11.0.100 / main 10.0.100), and Review-PR.ps1 -Phase Setup squash-merges onto the PR's real base — so one pipeline branch handles both main- and net11-targeting PRs.

   Trusted scripts are captured from the pipeline ref before the base checkout, so later tasks keep running the reviewed pipeline-branch scripts (security rule 3); Task 1 Setup now sources Review-PR.ps1 from $TRUSTED like the later tasks.

Intentionally NOT changed

• review-trigger.yml — unchanged (so this is testable without a main merge; point the pipeline at this branch via /review --branch feature/net-11-ai-review or workflow_dispatch).
• Review-PR.ps1 — unchanged (its CI path already "stay on current branch and squash-merge"; that branch is now the PR's base).

Validation

• Template-aware YAML parse of ci-copilot.yml.
• bash -n on both new "Resolve PR base branch" steps.
• System.Management.Automation.Language.Parser on all 8 changed PowerShell files.
• Invoke-Pester on Aggregate-CopilotTokenUsage.Tests.ps1, Post-AISummaryComment.Tests.ps1, Review-PR.Tests.ps1 → 40/40 passed.
• Allowlist dry-run (accepts main/net11.0/net10.0; rejects injection, refs/heads/..., net11.1, empty).
• Security checklist greps from ci-copilot-pipeline-security.instructions.md — no new violations.

Relationship to #35677

#35677 targets main. This PR targets net11.0 and applies the same #35677 content plus the base-branch auto-detection. The reviewer infrastructure (#35677) is carried via the trusted-script copy; the reviewed tree is always the PR's real base (main or net11.0) + the PR — so the separation between reviewer and reviewed code is preserved.

[github-actions]

🚀 Dogfood this PR with:

⚠️ WARNING: Do not do this without first carefully reviewing the code of this PR to satisfy yourself it is safe.

curl -fsSL https://raw.githubusercontent.com/dotnet/maui/main/eng/scripts/get-maui-pr.sh | bash -s -- 35994

Or

• Run remotely in PowerShell:

iex "& { $(irm https://raw.githubusercontent.com/dotnet/maui/main/eng/scripts/get-maui-pr.ps1) } 35994"

[kubaflo]

Synthesized Review — PR #35994

Title: ci-copilot: support net11-targeting PRs (auto-detect base branch + #35677 on net11.0)
Author: kubaflo (Jakub Florkowski) · Base: net11.0 · Head: feature/net-11-ai-review
HEAD SHA: 9ae99c1e3fa875feae81dc2813f6acc9e2b169c3
Labels: area-infrastructure · Changed files: 12 (+1800 / −161)

Verdict: NEEDS_DISCUSSION (confidence: medium)

Verdicts disagreed across the four model reviews, so per the conservative-resolution rule
(NEEDS_CHANGES > NEEDS_DISCUSSION > LGTM) the single NEEDS_DISCUSSION governs. The
discussion point is not a code defect — it is solely about CI status at review time (see
"CI status" below). Both inline findings are non-blocking suggestions.

Per-model verdicts

Model	Verdict	Confidence	Findings
gemini-3.1-pro-preview	LGTM	high	0
gpt-5.5	NEEDS_DISCUSSION	medium	0 (CI-not-green concern)
claude-opus-4.6	LGTM	high	1 (Review-PR.ps1:748)
claude-opus-4.8	LGTM	high	1 (ci-copilot.yml:949)

What the PR does

Infrastructure-only change to the Copilot PR-review pipeline so it can review
net11.0-targeting PRs (not just main/net10). Two pieces:

1. Ports #35677 (enhanced MauiBot AI Summary review, token-usage aggregation, deep-UI-test
   review-id patching) onto net11.0.
2. ci-copilot.yml auto-detects the PR base branch — a new "Resolve PR base branch" step
   (in both CopilotReview and DeepUITests stages) reads baseRefName via gh,
   allowlist-validates it (^(main|net[0-9]+\.0)$, fail-closed), and switches the detached
   worktree to origin/<baseRef> before workload install and the squash-merge. Trusted
   scripts are captured from the pipeline ref before the base checkout (security rule 3).

Reviewers concurred the security model is sound: trusted-script snapshot ordering, anchored
allowlist regex, AzDO log-command-injection defense (ConvertTo-AzdoSafeConsole), and
token-usage paths scoped to exclude PR-controllable directories.

Synthesized findings (2 — both non-blocking suggestions)

Both were validated against HEAD code (git show 9ae99c1e…:<path>); both anchor on added (+)
lines that are inside the PR diff, so they are valid for inline posting. They are distinct
file/concern pairs — no clustering required.

1. .github/scripts/Review-PR.ps1:748 — token telemetry (from claude-opus-4.6)

The $outputFields filter (HEAD lines 745–749) keeps any path containing output/completion
(minus cache/total). Since Select-CanonicalTokenFields keeps all shallowest matches,
a future Copilot payload exposing both outputTokens and reasoningOutputTokens at the same
depth would sum both and double-count reasoning tokens. The inputFields filter already
excludes cache variants; adding -and $_.Path -notmatch '(?i)reasoning' restores symmetry.
Forward-looking only — current payloads are unaffected.

2. eng/pipelines/ci-copilot.yml:949 — CI security control duplication (from claude-opus-4.8)

The PR-base allowlist ^(main|net[0-9]+\.0)$ + PRNumber re-guard is duplicated verbatim
across the two "Resolve PR base branch" steps (CopilotReview line 161, DeepUITests line 949
— both confirmed present at HEAD). This allowlist is the control deciding which branch the
worktree switches to before workload install + squash-merge, so silent divergence is the risk.
Extracting a shared steps template (the CopilotReview copy additionally captures $TRUSTED)
prevents drift.

Dropped / invalid findings

None. Every finding raised by any model was verified to reference code present at HEAD. gemini
and gpt-5.5 raised no inline findings.

CI status (basis for the NEEDS_DISCUSSION)

Check	Result
Post results comment (GitHub Actions)	fail (5s)
maui-pr · Run Helix Unit Tests Windows Helix Unit Tests (Release)	fail (57m)
maui-pr · Windows Helix Unit Tests (Debug)	pass
maui-pr · Build macOS/Windows (Debug & Release), Pack macOS	pass
maui-pr (overall), Build Windows (Release), Pack Windows	pending
Copilot review agent / detection / gate / static validation	pass

gpt-5.5 declined LGTM until those failures are confirmed unrelated. opus-4.6 and opus-4.8
assessed them as unrelated: no product or test code changed; only the Release Helix leg
failed while Debug passed; maui-pr is a different pipeline from the edited ci-copilot.yml.

To resolve to LGTM: confirm/re-run the Windows Helix (Release) leg and the "Post results
comment" job clean, and let the pending maui-pr legs finish.

[kubaflo]

🤖 AI-Generated Review (multi-model: GPT-5.5, Claude Opus 4.8, Claude Opus 4.6, Gemini 3.1 Pro)

Non-blocking suggestion (token telemetry): This $outputFields filter keeps any field whose path contains output or completion (excluding cache/cached and total). Because Select-CanonicalTokenFields retains all shallowest matches at the same depth, a future Copilot CLI usage payload that exposes both outputTokens and reasoningOutputTokens side-by-side would sum both — double-counting reasoning tokens inside outputTokens. The inputFields filter already guards its cache variants; consider excluding reasoning here for symmetry:

$outputFields = @($tokenFields | Where-Object {
    $_.Path -match '(?i)(output|completion)' -and
    $_.Path -notmatch '(?i)(cache|cached)' -and
    $_.Path -notmatch '(?i)total' -and
    $_.Path -notmatch '(?i)reasoning'
})

Forward-looking only — current payloads are unaffected. (Flagged by claude-opus-4.6.)

[kubaflo]

🤖 AI-Generated Review (multi-model: GPT-5.5, Claude Opus 4.8, Claude Opus 4.6, Gemini 3.1 Pro)

Non-blocking suggestion (CI security control): The PR-base allowlist ^(main|net[0-9]+\.0)$ plus the PRNumber re-guard is duplicated verbatim between the two "Resolve PR base branch" steps — the CopilotReview stage (line 161) and this DeepUITests copy (line 949). This allowlist is the control that decides which branch the worktree is switched to before workload install and the squash-merge, so silent divergence is the real risk: fixing or tightening one copy could leave the other permissive. Consider extracting a shared steps template (the only real difference is that the CopilotReview copy also captures $TRUSTED) so the validation cannot drift. (Flagged by claude-opus-4.8.)

[kubaflo]

Consolidated into #35677. The net11 enablement (runtime PR-base auto-detection for workloads + merge base, and branch-aware deep-UI/device-test TFM) was cherry-picked onto `feature/enhanced-reviewer` (#35677), which now handles both `main`/net10 and `net11.0` PRs from a single pipeline branch — so this separate net11 branch is no longer needed.

Closing in favor of #35677.