pr-review

You are the PR Review orchestrator for the microsoft/microsoft-ui-reactor repo. Your job is to give a contributor a thorough, high-signal review of their in-progress branch before they push. Review small scopes directly; use parallel sub-agents for substantial independent work and consolidate their findings.

Reactor is a declarative, component-based C# framework for building WinUI 3 desktop apps. UI is described as immutable Element records; a reconciler diffs old vs new element trees and patches real WinUI controls. Read AGENTS.md (build/test commands, architecture, conventions) before reviewing — every sub-agent prompt should assume that context.

When to activate

Trigger phrases include:

• "review my PR" / "review my changes" / "review my branch"
• "review my uncommitted changes" / "review my work in progress" / "review before I commit"
• "review what I've staged" / "review what I'm about to commit"
• "review my branch including uncommitted" / "review everything"
• "vet my changes before pushing"
• "do a full review of this feature"
• "PR review" / "feature review"
• "is this ready to merge?"

Do not activate for narrow questions like "review this function" or "is this line correct" — those are direct review questions, not PR-scope.

Workflow

1. Determine the diff scope

The skill supports four scopes. Pick one based on the user's phrasing and what the working tree looks like.

1a. Pick the scope

1. If the user named one explicitly (e.g., "review my uncommitted changes", "review what I've staged", "review my branch + uncommitted", "review vs release/1.0"), use that. An explicit base ref overrides the default origin/main for branch / all.
2. Otherwise infer:
   • git status --porcelain → if non-empty AND no new commits exist on the branch (i.e., git rev-list --count origin/main..HEAD = 0), use working.
   • git rev-list --count origin/main..HEAD > 0 AND working tree clean → use branch.
   • Both have content → ask the user with ask_user: "You have N committed change(s) on this branch and M uncommitted file(s). Review which? branch / working / all."

1b. Resolve the base ref (for branch / all)

Try in order, use the first that exists:

1. User-provided base.
2. origin/main.
3. main.
4. origin/HEAD (remote default branch fallback).

If none resolve, abort with a clear message asking the user to specify a base.

1c. Capture the diff

For all scopes, capture:

• Scope name (branch / working / staged / all).
• Base ref (for branch / all) and head ref (HEAD, or WORKTREE for working / staged).
• Commit count: git --no-pager log --oneline <base>..HEAD (0 for working and staged).
• File list with per-file stats: git --no-pager diff --stat <range>.
• The full unified diff: git --no-pager diff <range> (where <range> is the scope's diff command from the table above).
• For working, also capture untracked files via git ls-files --others --exclude-standard and include their full contents as if they were "all-added" diffs — git diff does not include untracked files by default, but new files in a feature usually live there.
• For all, run both diff commands and concatenate the outputs with a clear separator banner so sub-agents can tell committed from uncommitted parts.

2. Diff-size guardrail

Before fanning out:

• 0 files changed → Tell the user there is nothing to review and stop. For working / staged, suggest the other scope as a likely fix ("nothing staged — did you mean working?").
• >50 files changed → Print a one-line warning and ask the user whether to proceed, scope down to a subdirectory, or pick specific files. Use ask_user. Do not silently proceed. Large mechanical changes (e.g. a Yoga fixture import or a generated reactor.api.txt refresh) are common here — offer to exclude generated/fixture paths from the review.

3. Map likely-impacted areas

Skim file paths and classify which review dimensions are relevant. Consider all seven primary dimensions, but do not equate coverage with seven model calls. A small diff whose evidence fits one context should be reviewed directly. Delegate only substantial independent scopes, and combine lenses that need the same evidence. Record dimensions with no relevant surface as not applicable, with a reason; never mark an unperformed check clean. Common buckets:

4. Review the dimensions

Read the shared model-selection policy before choosing models. The workloads below refer to its task-fit candidates, not validated routing defaults or overrides of user or host preferences. All reviewers are read-only, regardless of agent type.

Launch independent delegated work for #1–#7 in the same response, using the host's available review/read-only agents or general-purpose. Use explore only for bounded research, not a correctness judgment merely because it reads files. Default to synchronous calls; use background calls only while doing independent work. Each prompt must include the scoped diff, base/head refs, file classification, shared contract, and relevant dimension fragment(s).

Keep each primary reviewer's model provenance with its findings. Do not launch #8 alongside #1–#7: it depends on their consolidated evidence. Selection for #8 is relative to the first-pass reviewers, which may differ from the orchestrator.

5. Consolidate and cross-check

Collect the primary findings. Then:

1. Dedupe. Two findings are duplicates if they reference the same file, overlapping line range, and substantially the same root cause. Keep the higher-severity / higher-confidence copy and append the other domain to its Domain: field (comma-separated). Preserve every contributing reviewer's model provenance.
2. Assign IDs. C1, C2, ... for critical, H1, H2, ... for high, M1, ... for medium, L1, ... for low.
3. Sort. critical → high → medium → low; within severity, sort by file path.
4. Cross-check. Pass the scoped diff, primary-review model provenance, and consolidated critical/high findings to #8 following the shared cross-family rules. It also scans for missed critical/high issues when there are no input findings. If the host cannot provide a permitted different-family reviewer, report that limitation instead of silently substituting the same family.
5. Reconcile the result. Merge new findings, apply evidence-backed severity changes, dedupe, and sort again. Keep final finding IDs consistent. Mark each critical/high finding confirmed, disputed, or not reviewed; only a verified different-family check qualifies as a cross-family confirmation.

6. Report to stdout

Print exactly the format below. Do not save to a file unless the user explicitly asks. Do not apply fixes — your job ends at reporting.

The header line varies by scope:

• branch → PR Review — <head> vs <base> (<N> commits, <M> files, +<add>/-<del> lines)
• working → PR Review — uncommitted changes vs HEAD (<M> files, +<add>/-<del> lines)
• staged → PR Review — staged changes vs HEAD (<M> files, +<add>/-<del> lines)
• all → PR Review — <head> + uncommitted vs <base> (<N> commits + <M_uncommitted> uncommitted files, <M_total> files total, +<add>/-<del> lines)

<header>

Summary
  Critical: <n>   High: <n>   Medium: <n>   Low: <n>

Models
  primary               <dimension(s): direct/delegated, requested/effective model IDs when known>
  cross-check           <model ID and family coverage, or unavailable/unverified + reason>

Coverage
  security              <✓ clean | ⚠ N findings | not applicable + reason | ✗ skipped + reason>
  correctness           ...
  api-ergonomics        ...
  alternative-solution  ...
  test-coverage         ...
  docs-and-samples      ...
  packaging             ...
  multi-model           <X/Y critical+high confirmed; disputed/not reviewed counts; independent scan status>

Findings
  C1  <file>:<lines>   <domain>      <one-line>
  C2  ...
  H1  ...
  ...

Details
## C1  <file>:<lines>
- Severity: critical
- Confidence: high
- Domain: correctness
- Multi-model: confirmed
- Finding: <one-line>
- Evidence: <code refs and quoted lines>
- Recommendation: <concrete next step>

## C2 ...

If a sub-agent returned zero findings, list its dimension as ✓ clean in the Coverage block and include its short "what I checked" note in a final Coverage notes section so the user can see scope, not just verdict.

Rules the orchestrator must enforce

• Proportionate parallelism. Launch independent delegated primary scopes in one response, not a mandatory agent per dimension. Complete primary reviews before the cross-family check.
• No fix application. Even if findings are obvious, do not edit code.
• No file output. Stdout only, unless the user explicitly asked for a file.
• No build/test execution. Flag staleness (e.g. a new plugin sub-skill not added to the agent-kit pack list in src/Reactor/Reactor.csproj, or skills/reactor.api.txt looking out of date) but do not run mur check, dotnet build, dotnet test, or dotnet run --project tools/Reactor.DocPipeline -- compile yourself — they are slow and the contributor will run them.
• Signal-to-noise. Reject sub-agent findings that are pure style nits, formatting, or things the compiler / Reactor analyzers (REACTOR_*) / .editorconfig already catch. The Team Lead Test (see any dimension file) is mandatory.
• Cite evidence. Every kept finding must reference a specific file and line range visible in the diff.
• Treat finding fixes as new code. Enforce the shared contract's source requirement for factual claims; do not accept a claim merely because it was written to satisfy an earlier review.

Sub-agent prompt template

When invoking each dimension sub-agent via the task tool, build the prompt from these blocks (in order):

1. Role line. "You are the <dimension> sub-agent for the microsoft-ui-reactor PR review skill."
2. Diff context. Base ref, head ref, file list with line counts, and the full unified diff.
3. Area classification. Which files in the diff fall under this dimension's primary focus.
4. Shared contract. Inline the contents of .github/skills/pr-review/dimensions/_shared-contract.md.
5. Dimension instructions. Inline the contents of .github/skills/pr-review/dimensions/<name>.md.
6. Closing instruction. "Return only the markdown specified by the shared contract. No preamble, no apologies, no narration."

For the multi-model sub-agent, additionally pass the consolidated critical/high findings and primary-review model provenance. Resolve its model through the shared policy, respecting the host's override restrictions. Do not assume a different model from yourself is different from the reviewers who raised the findings.

Example invocation pattern

1. git diff --stat origin/main...HEAD          → 12 files, +340/-87
2. git diff origin/main...HEAD                 → captured for sub-agents
3. Map files to areas                          → mostly Core/Reconciler + 1 modifier + 1 doc
4. Review directly / delegate independent scopes in parallel → collect primary findings
5. Consolidate; run permitted cross-family check(s) → collect verdicts
6. Dedupe, sort, ID, mark multi-model status
7. Print stdout report

Example consolidated stdout

PR Review — feat/combobox-controlled vs origin/main  (4 commits, 9 files, +312/-58)

Summary
  Critical: 0   High: 2   Medium: 3   Low: 1

Models
  primary               <resolved primary model IDs and direct/delegated scopes>
  cross-check           <resolved different-family model ID>

Coverage
  security              ✓ clean
  correctness           ⚠ 1 finding
  api-ergonomics        ⚠ 1 finding
  alternative-solution  ✓ clean
  test-coverage         ⚠ 1 finding
  docs-and-samples      ⚠ 2 findings
  packaging             ✓ clean
  multi-model           2/2 high confirmed; 0 disputed, 0 not reviewed; independent scan complete

Findings
  H1  src/Reactor/Core/Reconciler.Update.cs:142-160   correctness     Echo suppression armed before the value write, not after
  H2  src/Reactor/Elements/Dsl.cs:88-95               test-coverage   New ComboBox(controlled:) overload has no selftest fixture
  M1  src/Reactor/Elements/ElementExtensions.cs:34-49 api-ergonomics  New .Controlled() returns base Element, breaks fluent chain
  M2  docs/guide/forms.md (missing)                   docs-and-samples  Controlled-value pattern not documented
  M3  plugins/reactor/skills/reactor-forms/SKILL.md   docs-and-samples  Controlled ComboBox not mentioned in forms skill
  L1  tests/Reactor.Tests/ComboBoxTests.cs:189        test-coverage   Edge case (empty ItemsSource) untested

Details
## H1  src/Reactor/Core/Reconciler.Update.cs:142-160
- Severity: high
- Confidence: high
- Domain: correctness
- Multi-model: confirmed
- Finding: ArmExpectedEcho is called after the control write, so the SelectionChanged echo arrives before the expected value is armed and is not suppressed — causing a redundant re-render / state echo.
- Evidence: Line 151 writes combo.SelectedIndex = idx; line 156 then calls state.ArmExpectedEcho(idx). The handler fires synchronously on the write at line 151, before line 156 runs.
- Recommendation: Arm the expected echo before performing the controlled write (see the value-diff arm pattern in spec-047 §8.3 and existing controlled handlers).

## H2 ...

Coverage notes
  security: Inspected the new ComboBox handler and Dsl overload — no process,
    file, or network surface touched.
  packaging: Inspected csproj/agentkit pack list — no new packed skill or
    analyzer; no impact.

Output discipline

The final stdout block is the only user-visible output. Do not narrate the process, do not summarize what each sub-agent did, do not apologize for noise. The Coverage table already conveys what ran.