[morning4coffe-dev]

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!

Description of Change

Adds a maintainer-triggered GitHub Agentic Workflow (gh-aw) for /review performance. The command runs selected managed benchmarks against pinned merge-base/head revisions, asks GPT to interpret the evidence and coverage gaps, and publishes one independently rendered and validated performance report.

It follows the /review tests command entrypoint and #38807's separation between untrusted execution and credentialed reporting. The PR is now scoped to this hosted workflow and its required managed benchmarking/reporting dependencies.

Hosted-only scope

Reduced from 63 changed files to 24, all under .github/. There are no remaining changes to framework source, device tests/runners, eng/, or docs/.

Retained: the workflow source/compiled lock, trusted orchestration, managed selector/ABBA runner/comparator, trusted managed benchmark overlays, coverage catalogs, GPT interpretation skill, deterministic decision/renderer/validator, focused managed/report checks, and command documentation/routing.

Removed: native-device infrastructure and drivers, the local check-pr-performance skill, native request/evidence ingestion, the history writer, duplicate benchmark sources under src/, and their dedicated tests/documentation. This PR no longer depends on #38274. The platform catalog only describes coverage gaps; it does not execute native scenarios or promise an available device driver.

Merged main without rewriting history. Preserved its removal of legacy rerun automation and moved the performance-command exclusion into the new shared recovery helper.

New GitHub Agentic Workflow

• Source: .github/workflows/copilot-review-performance.md.
• Compiled workflow: .github/workflows/copilot-review-performance.lock.yml, generated with gh-aw v0.86.2.
• Trusted orchestrator: .github/scripts/Review-Performance.ps1 (Gather, Measure, Prepare, Render).
• Interpretation skill/model: perf-analysis, using gpt-5.6-sol.

A current write/maintain/admin collaborator requests a review with a newly created comment on an open PR:

/review performance

Only this exact subcommand activates the performance workflow. Other subcommands, unauthorized comments, issues, closed PRs, and changes without product files do not start measurements. The full /review trigger and missed-command recovery explicitly exclude /review performance.

1. Trusted intake: executes from the default branch, pins PR merge-base/head and harness SHAs, and selects coverage without executing PR code.
2. Isolated measurements: a disposable GitHub-hosted Linux job runs managed BenchmarkDotNet benchmarks in ABBA order. Base/head use separate unprivileged users and an explicit child-environment allowlist. No Copilot PAT or posting token reaches benchmark builds/runs.
3. Independent evidence and GPT interpretation: a fresh job imports bounded structured evidence and computes the decision baseline. GPT reads that bundle and the pinned diff and supplies narrative only, with no authority to execute more measurements or select another publication target.
4. Validated gh-aw publication: a separate safe-output job downloads independent evidence, recomputes the decision, renders and validates the report, and rechecks caller permission plus live PR base/head identities before publishing. Older performance reports are hidden. Automation failures produce a run link, not a fabricated performance verdict.

Production uses the existing copilot-pat-pool environment for GPT credentials and the repository's GitHub Actions token for scoped publication. There is no new Azure pipeline, service connection, device queue, or scheduled recovery.

Manual dispatch and retries

After deployment to the default branch:

# Measure, interpret, render, and validate without posting.
gh workflow run copilot-review-performance.lock.yml \
  --repo dotnet/maui --ref main \
  -f pr_number=32382 -f suppress_output=true

# Run the same workflow and publish the validated report.
gh workflow run copilot-review-performance.lock.yml \
  --repo dotnet/maui --ref main \
  -f pr_number=32382 -f suppress_output=false

pr_number must identify an open PR. Dry-run mode suppresses publication, not measurements, GPT interpretation, or report validation. Context, structured evidence, reports, and separate SDK-install/build/benchmark diagnostics are retained for seven days. Diagnostic logs are not admitted as measurement evidence.

Retry with a fresh dispatch or a new command, not Re-run failed jobs. Repeated attempts can leave same-named gh-aw agent artifacts in one run, allowing a downstream job to select an earlier attempt's empty output. The decision resolver recommends a managed retry only for missing/incomplete execution or benchmark data; repeating completed unchanged filters cannot fill excluded benchmark-input or native coverage gaps.

Evidence policy

Evidence	Treatment
Managed allocations	Repeated non-overlapping ranges; default materiality thresholds of 8 bytes and 0.5%
Shared-host timing	Advisory; requires non-overlapping ranges and at least 15% median movement
Sampled benchmark families	Supplemental evidence, not proof the changed path executed
Native paths	Explicit unmeasured coverage gaps; no native execution or local evidence admission
Missing/incompatible evidence	Incomplete or inconclusive, never a whole-PR pass

The deterministic scripts own verdicts, coverage, and permitted actions. GPT cannot erase a measured regression or clear missing/sampled evidence. Recommendations distinguish measurements, static support, and hypotheses. There is no numeric worth-it score, automatic merge decision, code/label mutation, issue closure, or automatic follow-up.

Successful final fork verification

The fresh final fork canary completed successfully against #32382, using the final production implementation from 867ffde2fe.

• Intake, activation, measurements, evidence preparation, GPT interpretation, detection, safe outputs, and conclusion all succeeded.
• Both Core builds succeeded under LinuxUsers isolation with credentialsSanitized: true.
• Seven benchmarks completed in each of the four ABBA runs, with matching benchmark sets and complete measurement data.
• Changed binding/XAML workloads remained excluded. Coverage was explicitly incomplete, with the deterministic next action needs_human_discussion; no native measurements or whole-PR clearance were claimed.
• The downloaded report-validation artifact contains valid: true and errors: [].

Canary commit e2550e87ba differs from production only in the two workflow files: it uses the existing personal Copilot secret, restricts dispatch to the dedicated fork branch, targets the selected upstream PR, and forces staged output. No new upstream report was posted, and no upstream environment protection was changed. Production PAT-pool configuration and normal publication behavior remain intact on the PR branch.

The staged run retains a non-fatal gh-aw conclusion annotation about the absent safe-outputs-items artifact. The actual safe-output rendering/validation job and overall run succeeded; this is not a claim that live upstream publication was exercised by the dry run.

Additional validation and CI caveats

All five retained managed/decision/report suites passed again on the final revision. The 69 existing command-routing/recovery checks passed. PowerShell/JSON syntax, scoped formatting, whitespace checks, strict gh-aw compilation, and native actionlint checks passed. Existing shared-workflow actionlint diagnostics were verified against main and left unchanged.

Normal PR checks are not all green: the Pester check has 2,201 passes and two stale assertions expecting gpt-5.6-sol while Review-Tests.ps1 uses gpt-6-astra. Both test files and their subject are byte-identical to the merged main. Skill Validation fails before inspecting PR content because the unchanged checkout action refuses fork checkout in pull_request_target. These unrelated shared checks were not weakened or modified. maui-pr is neutral on the final GitHub-only revision.

Earlier real hosted runs exposed and verified fixes for assigning to PowerShell's read-only $HOME and using the Windows-only SDK installer on Linux. Those fixes and retained diagnostic artifacts remain in the reduced implementation.

Deployment remains pending: the PR is still a draft and unmerged. The workflow, lock, skill, and trusted scripts must reach dotnet/maui's default branch before the upstream slash command operates. Fork validation does not imply production activation or replace normal CI and human review. Operating instructions are in .github/docs/pr-review-workflow.md.

Issues Fixed

No product issue is closed. This PR adds an explicitly authorized, hosted managed-performance review workflow.

[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 -- 38324

Or

• Run remotely in PowerShell:

iex "& { $(irm https://raw.githubusercontent.com/dotnet/maui/main/eng/scripts/get-maui-pr.ps1) } 38324"

[azure-pipelines]

Azure Pipelines:
Successfully started running 1 pipeline(s).
There may be pipelines that require an authorized user to comment /azp run to run.

[github-actions]

Skill Validation Results

@morning4coffe-dev — new skill validation results are available based on this last commit: 867ffde.
To request a fresh validation after new comments or commits, comment /evaluate-skills.

❌ Skill Validation Results — 867ffde · Add /review performance GitHub Agentic Workflow for PR benchmarks · 2026-09-30T19:22:52Z

❌ Static Checks Failed

Skills: 0 | Eval specs linted:

⏭️ LLM Evaluation: Skipped

💡 LLM evaluation was not run for this external PR.
A repository contributor can post /evaluate-skills on this PR to trigger full evaluation.

🔍 Full results and investigation steps

[dotnet-policy-service]

Hey there @@morning4coffe-dev! Thank you so much for your PR! Someone from the team will get assigned to your PR shortly and we'll get it reviewed.

[dotnet-policy-service]

Hey there @morning4coffe-dev! Thank you so much for your PR! Someone from the team will get assigned to your PR shortly and we'll get it reviewed.