Code Review Skill

Standalone skill that evaluates PR code changes for correctness, safety, performance, and consistency with .NET MAUI conventions. Can be invoked directly by users or by other agents/skills.

Trigger phrases: "review code for PR #XXXXX", "code review PR #XXXXX", "review this PR's code", "analyze code changes in PR", "check PR code quality"

Do NOT use for: "what does PR #XXXXX do?", "summarize PR", "describe the changes", or any informational query — just answer those directly without invoking this skill.

How this differs from other skills:

• pr-review — End-to-end PR workflow (4 phases: pre-flight, gate, try-fix, report). Use when you want the full pipeline including test verification and fix attempts.
• pr-finalize — Verifies PR title/description match implementation + light code review. Use before merging.
• code-review (this skill) — Deep code-only review with MAUI domain rules. Use when you want a thorough code analysis without running tests or modifying the PR.

Core Principles

1. Independence-first — Form your assessment from the code BEFORE reading the PR description. This prevents anchoring on the author's framing
2. Full-context — Read entire source files, not just diffs. Check callers, consumers, and git history
3. Empirical grounding — Reference specific code, line numbers, and call sites. No vague concerns
4. Severity calibration — Distinguish errors from warnings from suggestions. Not everything is critical
5. Failure-mode probing — Challenge your own conclusions with real failure scenarios, not softballs
6. Propagation-aware guards — For an early return, idempotency flag, or latch above downstream side effects, trace every set/clear path and a repeat call after recipients or state change. If that trace exposes concrete misbehavior, do not dismiss it as rare or rationalize it into LGTM: use NEEDS_DISCUSSION while the failure remains unresolved, or NEEDS_CHANGES when the exact state transition verifies a ❌ Error.

Inputs

Exactly one review source is required.

Outputs

Review Workflow

Step 1: Gather Code Context (No PR Narrative)

Do NOT read the PR description or issue yet.

For a materialized review_input, read its candidate diff first, then every supporting source file in full. Trace callers, consumers, and producers available in the snapshot. Do not fetch PR narrative, external pages, or repository history that the fixture does not provide. Then continue at Step 1.5.

For a live pr_number:

If a retrieval command fails, the PR is still available. A failing or unauthenticated command is a fact about that one tool, not about the review. Retry through another read-only route — gh api repos/dotnet/maui/pulls/<PR_NUMBER>/files, or the local checkout with git diff — and report an inability to review only after those also fail. Never ask the caller to paste the diff.

1. Get the diff:

   gh pr diff <PR_NUMBER> --repo dotnet/maui

2. Read full source files for every changed file (not just diff hunks):

   gh pr diff <PR_NUMBER> --repo dotnet/maui --name-only
   # Then read each file in full

3. Check callers and consumers of changed methods/properties:

   • Use LSP findReferences and incomingCalls for modified symbols
   • Understand how the changed code is used

4. Review git history of changed files:

   git log --oneline -10 -- <changed-file>

Step 1.5: Trace External Output Contracts (Always Active)

When changed code classifies external tool output with a regex or string literal:

1. Locate and read the producer, even when it is outside the diff.
2. State the exact condition under which the producer emits each matched token. Confirming that the text exists is not enough: compare the producer's emission condition with the consumer's semantic assumption.
3. Construct an ordinary negative case that must not trip the classifier, then trace it through every downstream guard, cap, veto, or early return.
4. If the ordinary case reaches the restrictive path, report a correctness finding and do not return LGTM unless the over-restriction is explicitly intended and documented. Fail-closed direction does not make the behavior correct.

Before the verdict, include an External Output Contract table with these columns:

A row that only confirms matching text, without comparing the two conditions, is incomplete analysis.

These are the direct-execution form of the always-active Logic/Correctness and Regression Prevention CHECKs in .github/agents/maui-expert-reviewer.md. If the expert agent is unavailable in the current environment, apply these probes yourself rather than skipping them.

Step 2: Delegate to Expert Reviewer

Delegate to the maui-expert-reviewer agent (.github/agents/maui-expert-reviewer.md) which runs per-dimension sub-agent evaluation. The agent's sole output is inline-findings.json — file:line comments in GitHub Review API format.

After the agent finishes:

• If COMMENTS_VIA_FILE=true (CI): Done. The pipeline calls post-inline-review.ps1 to post findings using GH_COMMENT_TOKEN.
• If COMMENTS_VIA_FILE is unset (local): Post inline findings directly:

  COMMIT_SHA=$(gh pr view $PR_NUMBER --repo dotnet/maui --json headRefOid --jq .headRefOid)
  gh api repos/dotnet/maui/pulls/$PR_NUMBER/reviews \
    --method POST \
    --input <(jq -n \
      --arg sha "$COMMIT_SHA" \
      --arg body "Expert review — see inline comments." \
      --argjson comments "$(cat CustomAgentLogsTmp/PRState/$PR_NUMBER/PRAgent/inline-findings.json)" \
      '{commit_id: $sha, body: $body, event: "COMMENT", comments: [$comments[] | {path, line, body, side: "RIGHT"}]}')

Step 3: Form Independent Assessment

Based ONLY on the code (no PR description), answer:

1. What does this change do? Describe the behavioral change in your own words
2. Why might it be needed? Infer motivation from the code
3. Is the approach sound? Would a simpler alternative work?
4. What problems do you see? Run through the agent's dimension CHECKs for matched dimensions

Step 4: Read PR Narrative and Reconcile

Now read the PR description, linked issue, and comments. Treat these as claims to verify, not facts.

1. Where your assessment disagrees with the author's claims, investigate further
2. If the PR claims a bug fix, verify the root cause analysis matches the code
3. Check existing review comments to avoid duplicating feedback

🚨 Prior Review Reconciliation

Check for prior reviews on the same PR — from the Copilot PR reviewer bot, other agents, or human reviewers. You MUST query all THREE surfaces — top-level review bodies, inline review comments, AND PR issue comments. Different reviewers post findings to different surfaces: review-API bots (MauiBot, Copilot bot, this skill's adversarial reviewer) post stubs like "Expert Review — 3 findings, see inline comments" at the top level with the actual ❌/⚠️/💡 markers in inline comments; AI Summary bots and prior-round wall-of-text summaries post to the issue-comments surface (which the review API does NOT return). Querying any subset silently misses findings.

# Surface 1: top-level review bodies (review-API stubs, verdicts, human reviewer prose):
gh pr view <PR_NUMBER> --repo dotnet/maui --json reviews --jq '.reviews[] | select((.body // "") != "") | "Reviewer: \(.author.login) | State: \(.state)\n\(.body)\n---"'

# Surface 2: inline review comments (where MauiBot/Copilot/this skill post ❌/⚠️/💡 findings):
gh api repos/dotnet/maui/pulls/<PR_NUMBER>/comments --paginate \
  --jq '.[] | "\(.user.login) @ \(.path):\(.line // .original_line // 0)\n\(.body)\n---"'

# Surface 3: PR issue comments (where AI Summary bots and prior round wall-of-text summaries post):
gh api repos/dotnet/maui/issues/<PR_NUMBER>/comments --paginate \
  --jq '.[] | "\(.user.login) @ \(.created_at)\n\(.body)\n---"'

Scan all three outputs for ❌ markers, [major]/[moderate] tags, or equivalent severity language from other reviewer formats. Do NOT slice/truncate the body fields — long bot reviews routinely exceed 10K chars and have severity markers in the tail (empirically observed on this skill's own PRs: MauiBot reviews of 26K+ chars with ❌ markers past char 10000); truncating silently drops them and causes false LGTM.

If prior reviews flagged ❌ Error-level issues:

• Verify whether each ❌ Error finding was addressed in subsequent commits
• If unresolved → verdict must be NEEDS_CHANGES
• If status cannot be determined → default to unresolved (caution over optimism)
• NEVER silently drop or contradict a prior ❌ Error finding — confirm it no longer applies to current code before dismissing

Step 5: Check CI Status

Before delivering a verdict, collect the required-check status for the PR. Don't infer CI state from absence of evidence and don't rely on prior commits' status.

gh pr checks <PR_NUMBER> --repo dotnet/maui --required

Exit-code semantics (read this before classifying): gh pr checks --required exit codes are NOT a reliable signal on their own — gh overloads them. Always inspect stdout/stderr.

• Exit 0 is NOT a "clean pass" signal — checks marked skipping (e.g., maui-pr skipping) also exit 0. Read the stdout rows for actual state.
• Exit 1 is overloaded with three cases that look similar but require different responses:
  • (a) Failing required check — stdout lists one or more fail rows.
  • (b) Zero required checks for the branch — stdout is empty and stderr contains the substring checks reported (specifically either no checks reported on the '<branch>' branch when the PR has zero checks of any kind, or no required checks reported on the '<branch>' branch when checks exist but none are required). Both shapes mean the PR has no required gates, NOT a tool failure. Route to the Skipped, pending, or empty result bullet below.
  • (c) gh itself errored — stdout has no check rows and stderr contains GraphQL:, Could not resolve, HTTP 4xx/5xx, or error:. Route to the tool-unavailable fallback at the bottom of this section.
• Exit 8 means required checks are pending and gh is reporting normally.
• Other non-zero exits (e.g., auth failure: gh auth login, network failure, command not found) DO indicate tool unavailability and should trigger the fallback at the bottom of this section.

Classify based on the stdout row content (pass/fail/skipping/pending) and the stderr message, not the exit code alone. If stdout has no check rows and stderr contains a GraphQL: / Could not resolve / error: message, treat as tool-unavailable (fallback). If stdout has no check rows and stderr contains checks reported (either spelling — see (b) above), treat as empty result (not a tool failure).

• PR-caused failing check (compile/build errors, test failures in modified code) → flag as ❌ Error and NEEDS_CHANGES. Surface this in the CI Status / Verdict sections; do NOT also generate per-line inline comments duplicating compiler output (the inline-comment rule in Review Output Format still applies).
• Pre-existing infra flake or known issue (cross-reference with azdo-build-investigator skill if uncertain) → note in summary but still cap confidence per the table in Step 6
• Ambiguous → invoke the azdo-build-investigator skill to determine root cause before finalizing
• PR description acknowledges the failure → note that the author has documented the dependency; the failure still caps confidence
• Skipped, pending, or empty result (required check listed as skipping/pending, or gh exits 1 with stderr containing checks reported — see Exit-1 case (b) — and no stdout rows) → treat CI coverage as undetermined. Do not interpret an empty/skipped result as a passing build. Cap confidence at low and do NOT post LGTM — use NEEDS_DISCUSSION (per Rule #6, which prohibits LGTM on pending/undetermined CI as strictly as on red CI).

Never claim "clean build" or LGTM without running this step. Apply the tool-unavailable fallback when gh cannot determine CI state — either because gh itself is missing/unauthenticated (command not found, gh: To get started with GitHub CLI, please run: gh auth login), or because the command returned a tool/API error instead of check rows (stderr contains GraphQL: / Could not resolve / error: / HTTP 4xx/5xx and stdout has no pass/fail/skipping/pending rows). In any of those cases, record the gap explicitly and cap verdict confidence at low.

Step 6: Blast Radius, Failure-Mode Probing, and Verdict

Blast Radius Assessment

Required when PR modifies: handlers, platform extensions, toolbar/navigation code, page registration, static state, PropertyChanged subscriptions, or startup paths.

Also required — both the assessment below and Failure-Mode Probing — for behavioral changes to these frequently-regressed component families: CollectionView, CarouselView, Image/Graphics, Theme/Style, Gesture/Tap, Button/Entry, Toolbar, and Shell/TabBar. This list is complete and sufficient on its own. The Frequently Regressed Components table in .github/agents/maui-expert-reviewer.md (under the Regression Prevention dimension) mirrors it and adds per-family risk areas; read it for that extra detail when it is present. For these families the usual miss is an untested adjacent scenario: a spacing fix that also runs on scroll-position restoration, a CurrentItem or loop-mode change that also affects ScrollTo, or a touch-handling fix that also affects tap/swipe/gesture.

The Step 2 expert reviewer reports findings only, with no per-dimension activation record, so its output cannot distinguish "Regression Prevention ran and found nothing" from "it never ran" — and a finding from some other dimension is not evidence it ran either. Never claim to have confirmed that a dimension fired. For every family that triggers this section, run the Failure-Mode Probing questions below yourself regardless of what the expert reported.

A prose-only change — documentation or comments — need not carry the family escalation above, provided the edited text is genuinely inert. It is not inert if it alters a public API doc, an analyzer or compiler directive (<auto-generated/>, #pragma warning, suppression attributes), an agent-instruction file this repo executes, or a comment stating a precondition other code relies on without re-verifying ("caller must dispose", "always called on the UI thread", "assumes sorted input"). Non-inert prose still gets a full review, but the Blast Radius table below asks runtime questions — startup ordering, static state, PlatformView nullity — that a text edit cannot answer. Probe the contract the text actually encodes instead: for a documented precondition, whether the code relying on it still holds; for a directive, which warnings or generated-code handling it now suppresses; for an agent-instruction file, whether the new wording fires on invocations it was not meant to reach, contradicts an instruction elsewhere in the same file, or states a condition the agent cannot evaluate from what it already has.

Failure-Mode Probing

Do NOT ask easy rhetorical questions. Probe genuinely challenging failure modes:

• What happens if this code runs on items/pages that DON'T use the new feature?
• What happens during handler disconnect/reconnect (navigation, Shell tab switch)?
• What happens with null Parent, Handler, BindingContext, or PlatformView?
• Can multiple subscriptions accumulate across handler lifecycle (missing unsubscribe)?
• Does static state survive page disposal and get stale?
• When a change adds an early-return guard, idempotency flag, or one-way latch above propagation or other side effects, distinguish the local work the guard suppresses from every downstream effect it also bypasses. List every path that sets the latch and every path that clears it, then trace a subsequent call while the latch is set. Check whether the input, recipient, or downstream state can change while the latch remains set: completing local work once does not prove every recipient has observed the current state. Don't accept "all the scenarios are handled" without tracing that state transition.

Confidence Calibration

Then cap by evidence. The cap and the action required are separate columns — a cap alone is not a verdict, and the action does not change the cap:

Confidence is confidence in the safety recommendation, not confidence that an individual finding exists. Apply the most restrictive applicable cap to the required **Confidence:** field. A reviewer can be certain that a failure mechanism exists while remaining low-confidence that the change is safe to merge.

Do not rationalize away a failure mode you surfaced. If Failure-Mode Probing produces a concrete scenario where the change misbehaves and you cannot disprove it by tracing exact state transitions, you may not downgrade it to 💡 Info or post LGTM. An un-disproven failure mode is an unresolved risk: it caps the required **Confidence:** field at low and the verdict at NEEDS_DISCUSSION. Escalate to NEEDS_CHANGES only when exact state transitions verify a concrete ❌ Error finding; mere plausibility does not establish a defect. High confidence requires the absence of un-disproven failure modes — not a narrative explaining why the one you found is probably fine.

Deliver Verdict

• LGTM — Code is correct, safe, and consistent with MAUI patterns. Ready for human approval.
• NEEDS_CHANGES — Concrete issues found that should be addressed before merge.
• NEEDS_DISCUSSION — Complex tradeoffs or architectural questions that need human judgment.

Review Output Format

Constraints (from Android team's approach):

• Only comment on added/modified lines — don't flag pre-existing code
• One issue per comment. If the same issue appears many times, flag once with a note listing all affected files
• Don't pile on. 3 important comments > 15 nitpicks
• Don't duplicate CI output as inline comments. Skip line-by-line compiler errors and linter findings in inline file:line comments — CI already surfaces those. CI-detected failures must still drive the verdict and appear in the CI Status / Verdict sections per Step 5; this rule only governs the inline-comment surface.
• Avoid false positives. Verify the concern actually applies given full context. If unsure, phrase as a question.

## Code Review — PR #XXXXX

### Independent Assessment
**What this changes:** [Your understanding from code alone]
**Inferred motivation:** [Why this change seems needed]

### Reconciliation with PR Narrative
**Author claims:** [Summary of PR description]
**Agreement/disagreement:** [Where your assessment matches or differs]

### Prior Review Reconciliation
| Prior ❌ Error Finding | Source | Status | Evidence |
|------------------------|--------|--------|----------|
| [finding] | [reviewer] | ✅ Fixed / ❌ Unresolved / 🔄 Obsolete | [evidence] |
*(If no prior reviews with ❌ Error findings, state "No prior ❌ Error findings found.")*

### Blast Radius Assessment
*(Required for infrastructure/handler/platform changes, or for a frequently-regressed component family or non-inert prose per Step 6; omit for simple fixes)*
- Runs for all instances: [yes/no — explanation]
- Startup impact: [yes/no]
- Static/shared state: [yes/no]

### CI Status
*(Required — record what `gh pr checks --required` returned per Step 5)*
- Required-check result: [pass / fail / pending / skipping / no required checks]
- Classification: [PR-caused failure ❌ / pre-existing flake / undetermined / PR-acknowledged]
- Action taken: [none / invoked `azdo-build-investigator` / capped confidence]

### Findings

#### ❌ Error — [Brief description]
[Explanation with specific file:line references]

#### ⚠️ Warning — [Brief description]
[Explanation with specific file:line references]

#### 💡 Suggestion — [Brief description]
[Explanation]

### Failure-Mode Probing
- [Probe]: [Answer — what actually happens in this scenario]
- [Probe]: [Answer]

### Verdict: LGTM / NEEDS_CHANGES / NEEDS_DISCUSSION
**Confidence:** high / medium / low *(justified against calibration table)*
**Summary:** [2-3 sentences explaining the verdict]

Verdict Consistency Rules

1. The verdict must match your most severe finding. If you have any ❌ Error findings, the verdict must be NEEDS_CHANGES. If only ⚠️ Warnings, use judgment but explain.
2. Failure-mode probing before finalizing. Re-read all findings. For each warning, ask: "Would I be comfortable if this merged as-is?"
3. Never approve what you can't verify. If the fix touches platform code you can't fully reason about, say so explicitly and use NEEDS_DISCUSSION.
4. LGTM means no ❌ Errors. You can LGTM with 💡 Suggestions. You can LGTM with ⚠️ Warnings only if you've explained why they're acceptable.
5. Prior ❌ Error findings override. If any prior review flagged an ❌ Error-level issue (using this skill's severity taxonomy) that remains unresolved in the current code, verdict must be NEEDS_CHANGES regardless of your own assessment. Confirm the finding still applies to the current diff before applying the override.
6. Never LGTM if CI is red, pending, or undetermined. If required CI checks are failing, invoke azdo-build-investigator to determine whether failures are PR-caused. Do not post LGTM until CI passes or failures are confirmed PR-unrelated. If required checks are pending, skipping, or absent, use NEEDS_DISCUSSION — code review alone does not warrant LGTM when CI hasn't run. Even when failures are confirmed PR-unrelated, the Step 6 confidence cap still applies (max low).
7. 🚨 NEVER use --approve or --request-changes on GitHub. Only post comments. Approval is a human decision.
8. Findings handling depends on environment. In CI (COMMENTS_VIA_FILE=true), the code-review agent does NOT have the GitHub comment token; the pipeline posts on its behalf. The maui-expert-reviewer sub-agent invoked in Step 2 is the sole producer of CustomAgentLogsTmp/PRState/{PR}/PRAgent/inline-findings.json (structured file:line JSON in GitHub Review API shape), which Review-PR.ps1 posts via post-inline-review.ps1; the code-review skill's wall-of-text summary is posted separately by post-ai-summary-comment.ps1. Do NOT have the wall-of-text-producing code-review agent emit, overwrite, or merge into inline-findings.json itself — overwriting that file with prose findings will corrupt its JSON schema and break post-inline-review.ps1. (When the orchestrator pipeline says "write inline findings to inline-findings.json", it means: ensure the Step 2 expert reviewer ran and produced that file — not that the wall-of-text agent should author the JSON directly.) In local invocation (no COMMENTS_VIA_FILE), the agent may post directly using its own gh credentials per the Step 2 gh api ... reviews --method POST command. In either mode, Rule #7 still applies: never --approve or --request-changes.

Posting the Review

In CI mode (COMMENTS_VIA_FILE=true) the agent writes findings to disk and posting is done separately by Review-PR.ps1. In local invocation (no COMMENTS_VIA_FILE) the agent may post directly per Rule #8 / Step 2.

Inline review comments (preferred — findings at exact file:line):

# Preview first:
pwsh .github/scripts/post-inline-review.ps1 -PRNumber <PR_NUMBER> -DryRun

# Post when ready:
pwsh .github/scripts/post-inline-review.ps1 -PRNumber <PR_NUMBER>

Wall-of-text summary (phase content assembled into a PR review body):

# Called by Review-PR.ps1 automatically:
pwsh .github/scripts/post-ai-summary-comment.ps1

In CI (eng/pipelines/ci-copilot.yml), Review-PR.ps1 calls both post-inline-review.ps1 (for inline findings) and post-ai-summary-comment.ps1 (for the wall-of-text from {phase}/content.md files), using GH_COMMENT_TOKEN. The trusted posting script may submit APPROVE or REQUEST_CHANGES from the final recommendation; the agent itself must not run review commands directly.

Completion Criteria

• Full source files read (not just diffs)
• Independent assessment formed before reading PR narrative
• Prior reviews checked and ❌ Error findings reconciled (Step 4)
• MAUI-specific checklist walked through for each applicable section
• CI status collected via gh pr checks --required and classified (Step 5)
• Blast radius assessed for infrastructure/handler/platform changes (Step 6)
• Failure-mode probing completed with real scenarios, not softballs (Step 6)
• Findings categorized by severity (❌ / ⚠️ / 💡)
• Confidence calibrated against blast radius and evidence tables (Step 6)
• Verdict is consistent with findings AND prior review reconciliation
• Output follows the format above