메뉴
Review code changes in dotnet/razor for correctness, performance, and consistency with project conventions. Use when reviewing PRs or code changes.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Import Razor formatting log zips that the user has already downloaded from Azure DevOps feedback tickets or GitHub issues into FormattingLogTest, validate whether the captured problem still reproduces, and if needed drive a minimal repro plus fix workflow.
Run the razor-toolset-ci pipeline to validate the current branch against large third-party repositories (MudBlazor, OrchardCore, ASP.NET Core, etc.). Use when asked to run toolset tests, ecosystem tests, or third-party validation.
Analyze CI build and test status from Azure DevOps and Helix for dotnet repository PRs. Use when checking CI status, investigating failures, determining if a PR is ready to merge, or given URLs containing dev.azure.com or helix.dot.net. Also use when asked "why is CI red", "test failures", "retry CI", "rerun tests", "is CI green", "build failed", "checks failing", or "flaky tests".
Analyze VMR codeflow PR status for dotnet repositories. Use when investigating stale codeflow PRs, checking if fixes have flowed through the VMR pipeline, debugging dependency update issues in PRs authored by dotnet-maestro[bot], checking overall flow status for a repo, or diagnosing why backflow PRs are missing or blocked.
Update all Roslyn dependency versions in the razor repository. Use this when asked to update, bump, or change the Roslyn version.
| name | code-review |
| description | Review code changes in dotnet/razor for correctness, performance, and consistency with project conventions. Use when reviewing PRs or code changes. |
Review code changes against conventions and patterns established by dotnet/razor and dotnet/runtime maintainers. The detailed conventions below were originally extracted from 43,000+ maintainer review comments across 6,600+ PRs in dotnet/runtime and represent standards commonly enforced across dotnet repositories.
Reviewer mindset: Be polite but very skeptical. Your job is to help speed the review process for maintainers, which includes not only finding problems the PR author may have missed but also questioning the value of the PR in its entirety. Treat the PR description and linked issues as claims to verify, not facts to accept. Question the stated direction, probe edge cases, and don't hesitate to flag concerns even when unsure.
Use this skill when:
Before analyzing anything, collect as much relevant code context as you can. Critically, do NOT read the PR description, linked issues, or existing review comments yet. You must form your own independent assessment of what the code does, why it might be needed, what problems it has, and whether the approach is sound — before being exposed to the author's framing. Reading the author's narrative first anchors your judgment and makes you less likely to find real problems.
git log --oneline -20 -- <file>). Look for related recent changes, reverts, or prior attempts to fix the same problem. This reveals whether the area is actively churning, whether a similar fix was tried and reverted, or whether the current change conflicts with recent work.Based only on the code context gathered above (without the PR description or issue), answer these questions:
Write down your independent assessment before proceeding. You must produce a holistic assessment (see Holistic PR Assessment) at this stage.
Now read the PR description, labels, linked issues (in full), author information, existing review comments, and any related open issues in the same area. Treat all of this as claims to verify, not facts to accept.
When the environment supports launching sub-agents with different models (e.g., the task tool with a model parameter), run the review in parallel across multiple model families to get diverse perspectives. Different models catch different classes of issues. If the environment does not support this, proceed with a single-model review.
How to execute (when supported):
gpt-5 and gpt-5.1), pick the one with the highest version number.When presenting the final review (whether as a PR comment or as output to the user), use the following structure. This ensures consistency across reviews and makes the output easy to scan.
## 🤖 Copilot Code Review — PR #<number>
### Holistic Assessment
**Motivation**: <1-2 sentences on whether the PR is justified and the problem is real>
**Approach**: <1-2 sentences on whether the fix/change takes the right approach>
**Summary**: <✅ LGTM / ⚠️ Needs Human Review / ⚠️ Needs Changes / ❌ Reject>. <2-3 sentence summary of the overall verdict and key points. If "Needs Human Review," explicitly state which findings you are uncertain about and what a human reviewer should focus on.>
---
### Detailed Findings
#### ✅/⚠️/❌ <Category Name> — <Brief description>
<Explanation with specifics. Reference code, line numbers, interleavings, etc.>
(Repeat for each finding category. Group related findings under a single heading.)
The summary verdict must be consistent with the findings in the body. Follow these rules:
The verdict must reflect your most severe finding. If you have any ⚠️ findings, the verdict cannot be "LGTM." Use "Needs Human Review" or "Needs Changes" instead. Only use "LGTM" when all findings are ✅ or 💡 and you are confident the change is correct and complete.
When uncertain, always escalate to human review. If you are unsure whether a concern is valid, whether the approach is sufficient, or whether you have enough context to judge, the verdict must be "Needs Human Review" — not LGTM. Your job is to surface concerns for human judgment, not to give approval when uncertain. A false LGTM is far worse than an unnecessary escalation.
Separate code correctness from approach completeness. A change can be correct code that is an incomplete approach. If you believe the code is right for what it does but the approach is insufficient (e.g., treats symptoms without investigating root cause, silently masks errors that should be diagnosed, fixes one instance but not others), the verdict must reflect the gap — do not let "the code itself looks fine" collapse into LGTM.
Classify each ⚠️ and ❌ finding as merge-blocking or advisory. Before writing your summary, decide for each finding: "Would I be comfortable if this merged as-is?" If any answer is "no," the verdict must be "Needs Changes." If any answer is "I'm not sure," the verdict must be "Needs Human Review."
Devil's advocate check before finalizing. Re-read all your ⚠️ findings. For each one, ask: does this represent an unresolved concern about the approach, scope, or risk of masking deeper issues? If so, the verdict must reflect that tension. Do not default to optimism because the diff is small or the code is obviously correct at a syntactic level.
Before reviewing individual lines of code, evaluate the PR as a whole. Consider whether the change is justified, whether it takes the right approach, and whether it will be a net positive for the codebase.
Every PR must articulate what problem it solves and why. Don't accept vague or absent motivation. Ask "What's the rationale?" and block progress until the contributor provides a clear answer.
"I am not sure why is this needed. ... It's not immediately obvious whether this happens only for the bridge comparison tests or whether it can happen for real-life scenarios too."
Challenge every addition with "Do we need this?" New code, APIs, abstractions, and flags must justify their existence. If an addition can be avoided without sacrificing correctness or meaningful capability, it should be.
"I don't think we should take this change, at all. A change which makes the VS runner see the same assets as the CLI runner, sure. But random extra hacking on the side, no."
Demand real-world use cases and customer scenarios. Hypothetical benefits are insufficient motivation for expanding API surface area or adding features. Require evidence that real users need this.
"It is not clear to me whether you can hit a real-world scenario on 32-bit platforms where it makes a difference."
Require measurable performance data before accepting optimization PRs. Demand BenchmarkDotNet results or equivalent proof — never accept performance claims at face value.
"Can you please share a benchmark using BenchmarkDotNet against public System.Text.Json APIs that demonstrates the improvement?"
Distinguish real performance wins from micro-benchmark noise. Trivial benchmarks with predictable inputs overstate gains from jump tables, branch elimination, and similar tricks. Require evidence from realistic, varied inputs.
"Try to benchmark it with an input that varies randomly. Jump tables are great for trivial micro-benchmarks, but they are less great for real world code."
Investigate and explain regressions before merging. Even if a PR shows a net improvement, regressions in specific scenarios must be understood and explicitly addressed — not hand-waved.
"Could you please inspect the regressions on why exactly it's an improvement there?"
Check whether the PR solves the right problem at the right layer. Look for whether it addresses root cause or applies a band-aid. Prefer fixing the actual source of an issue over adding workarounds to production code.
"The offset behind
Flags.IndexMaskshould always be correct. Instead of checking that the index is in range in all its usages, we should fix the root cause where the offset wasn't computed/updated correctly."
When a PR takes a fundamentally wrong approach, redirect early. Don't iterate on implementation details of a flawed design. Push back on the overall direction before the contributor invests more time.
"I'm still hesitating whether separating FEATURE_HW_INTRINSICS from SIMD and MASKED_HW_INTRINSICS is the right approach ... An alternative would be to handle them like #113689 and fix the value numbering."
Ask "Why not just X?" — always prefer the simplest solution. When a PR uses a complex approach, challenge it with the simplest alternative that could work. The burden of proof is on the complex solution.
"Wouldn't it be simpler to just do a regular mono stackwalk when we need to record and raise a sample?"
Explicitly weigh whether the change is a net positive. A performance trade-off that shifts costs around is not automatically beneficial. Demand clarity that the change is a win in the typical configuration, not just in a narrow scenario.
"It is a performance trade-off. You will shift the costs around. It is not clear to me whether it would be a win at the end in the typical configuration."
Reject overengineering — complexity is a first-class cost. Unnecessary abstraction, extra indirections, and elaborate solutions for marginal gains are actively rejected.
"This optimization smells funny. It seems overly complicated for little win. Is this path hot? Can we instead store the home directory?"
Every addition creates a maintenance obligation. Long-term maintenance cost outweighs short-term convenience. Code that is hard to maintain, increases surface area, or creates technical debt needs stronger justification.
"The primary goal of this project is to minimize our long-term maintenance costs. Building multiple optimizing code generators would go against that goal."
Require large or mixed PRs to be split into focused changes. Each PR should address one concern. Mixed concerns make review harder and increase regression risk.
"I think I'm going to break this into two pieces, even though that's more work."
Defer tangential improvements to follow-up PRs. Police scope creep by asking contributors to separate concerns. Even good ideas should wait if they're not part of the PR's core purpose.
"Should probably be a separate PR."
Flag breaking changes and require formal process. Any behavioral change that could affect downstream consumers needs documentation, API review, and explicit approval — even when the change improves the codebase internally.
"Introduce the new API in this PR. Remove the old check in another PR, mark it as breaking change and document it (as any other breaking change)."
Assess regression risk proportional to the change's blast radius. High-risk changes to stable code need proportionally higher value and more thorough validation.
"I wanted to backport this change to .NET 10, potentially .NET 9, and wouldn't want to introduce any risky changes."
Ensure new code matches existing patterns and conventions. Deviations from established patterns create confusion and inconsistency. If a rename or restructuring is warranted, do it uniformly in a dedicated PR — not piecemeal.
"This change is inconsistent with the rest of the global pointers. If we want to consider renaming these, I think we should do it in a separate PR and apply it consistently to all global pointers."
Check whether a similar approach has been tried and rejected before. If a prior attempt didn't work, require a clear explanation of what's different this time.
"If it's not worthwhile, especially if it was previously tried and it wasn't obviously beneficial, then we should close the issue."
Use Debug.Assert for internal invariants, not exceptions. For internal-only callers, assert assumptions rather than throwing ArgumentException. Prefer Debug.Assert(value != null) over the null-forgiving operator (!).
"Since there are no public callers, this should be an assert, not an ArgumentException." — bartonjs
Use throw for reachable error paths, UnreachableException for exhaustive switches. When a code path might be hit at runtime, throw an exception rather than asserting. Use throw new UnreachableException() for default cases in exhaustive switches. Use PlatformNotSupportedException (not NotSupportedException) for platform gaps. In native code, use _ASSERTE(!"message").
"We prefer throw rather than asserts so it is more apparent if some scenario makes it here." — VSadov
Include actionable details in exception messages. Use nameof for parameter names. Include the unsupported type or unexpected value. Never throw empty exceptions.
"You should add some message here:
throw new ArgumentException($\"Unknown ArrayFunctionType: {functionType}.\", nameof(method));" — jkoritzinsky
Initialize output parameters in all code paths. When a method has out parameters or pointer outputs (bytesWritten, numLocals), ensure they are initialized to a defined value in all error paths.
"Clear numLocals here (or at start of the method) so that it is initialized in all error cases?" — jkotas
Challenge exception swallowing that masks unexpected errors. When a PR adds try/catch blocks that silently discard exceptions (catch { continue; }, catch { return null; }), question whether the exception represents a truly expected, recoverable condition or an unexpected error signaling a deeper problem (race conditions, memory corruption, build environment issues). Silently catching exceptions that "shouldn't happen" hides root causes and makes debugging harder. The default disposition should be to let unexpected exceptions propagate or fail fast so the real issue gets investigated.
"Why do we want to mask this error? ... Our general strategy in AOT compilers is to fail the compilation when the input is malformed. It is expected that the malformed input can and will cause the compiler to crash or fail." — jkotas
Fix root cause, not symptoms or workarounds. Investigate and fix the root cause rather than adding workarounds or suppressing warnings. Revert broken commits before layering fixes.
"Let's try to investigate the root cause instead of taking this fix as-is since there could be other issues/AVs related to the mangling of the list." — jkotas
Prefer safe code over unsafe micro-optimizations. Do not introduce Unsafe.As, Unsafe.AsRef, or raw pointers without demonstrable performance need. Prefer Span-based APIs. If performance is the issue, prefer fixing the JIT.
"I do not want to introduce unsafe code for things like this. If it's material, the cast should be elided by the JIT." — stephentoub
Delete dead code and unnecessary wrappers. Remove dead code, unnecessary wrappers, obsolete fields, and unused variables when encountered or when the only caller changes.
"Unnecessary wrapper", "Dead code that I happen to notice", "This is the only use of m_canBeRuntimeImpl. It can be deleted." — jkotas
Seal classes when Equals uses exact type matching. If a class implements Equals with GetType() comparison, seal the class to prevent subtle inheritance bugs.
"Is there a reason why this Equals implementation is exact type match only even though ContextHolder isn't sealed? I'd prefer to block this class of failure by sealing the class." — kg
Backport targeted fixes, not refactorings. When backporting to servicing branches, create small targeted fixes. Backporting large refactorings introduces unnecessary risk.
"If we need the fix in .NET 10 for Android, we should do a small targeted fix that just adds the few lines under Android ifdef." — jkotas
Avoid closures and allocations in hot paths. When a lambda captures locals creating a closure, consider using a static delegate with a state parameter (value tuple). Avoid string concatenation; use span-based operations.
"Since this is capturing
dataandcontextin the closure it's allocating a closure for every call. The usual fix is to make the callback take a TState, and just pass through a value-tuple." — bartonjs
Pre-allocate collections when size is known. Pass capacity to Dictionary, HashSet, List constructors when the expected count is available.
"We can pre-allocate the dictionary of the right size." — jkotas
Structs in dictionaries need IEquatable<T> and GetHashCode. Without these, the runtime falls back to boxing allocations for equality comparison.
"If a struct doesn't override equality comparison logic, the runtime will end up using a fallback that boxes the value." — MihaZupan
Place cheap checks before expensive operations. Order conditionals so cheapest/most-common checks come first. Move expensive work after early-exit checks.
"These checks are cached cheap bit tests. We should do them first, and run the more expensive IL header decoder only when the modes do not match." — jkotas
Allocate resources lazily where possible. Allocate expensive resources on first use, not during initialization. Avoid forcing type initialization during startup.
"We try to do things lazily where possible since it is good for performance." — jkotas; "Do not force initialization during runtime startup just to make cDAC work. Startup is our number one perf problem." — jkotas
Extract throw helpers into [DoesNotReturn] methods. Move throwing logic from error paths into separate static local functions or helper methods to allow the JIT to inline the success path.
"Please move the body of this
if (throwOnFailure)block to a separate [DoesNotReturn] throwing static local function." — stephentoub
Avoid O(n²) patterns in collections and hot paths. Watch for linear scans inside loops, repeated RemoveAt in loops. Use RemoveAll, single-pass restructuring, or appropriate data structures.
"Since RemoveParsedValue performs a linear scan, this change makes the setter have a quadratic worst-case complexity." — MihaZupan
Cache repeated accessor calls in locals. Store the result of repeated property/getter calls in a local variable.
"Maybe read it out once into local to reduce number of calls to m_type_data_get_type?" — lateralusX
Consider scalability, not just throughput. Evaluate whether data structures, caches, and locking strategies will hold up at high cardinality or under concurrent load. Watch for unbounded collection growth, lock contention that worsens with core count, and O(1) assumptions that break at scale.
Use well-named constants instead of magic numbers. No raw hex or decimal constants without explanation. Don't duplicate magic constants across files.
"0x7F00 says nothing to the typical reader. Adding a comment that explains it means
(float)Int128.MaxValue... then means a lot." — tannergooding
Use PascalCase for constants; descriptive names for booleans. All constant locals and fields use PascalCase (except interop constants matching external names). Boolean fields should be positive and descriptive (_hasCurrent not valid).
"We use PascalCasing to name all our constant local variables and fields." — bartonjs
Name methods to accurately reflect their behavior. Update names when behavior changes. Get* implies a return value; use Print*/Display* for void. ThrowIf not ThrowExceptionIf.
"This method is not returning anything after this change, so Get... name does not fit." — jkotas
Prefer early return to reduce nesting. Use early returns for short/error cases to avoid unnecessary nesting. Put the error case first, success return last.
"Rather than having the extra layer of indentation for the else block, could we do it as:
if (...) { return ...; }" — stephentoub
Avoid using static and #region in new code. using static is costly when reading code outside IDEs (e.g., GitHub review). #region gets out of date quickly.
"
using staticadds an astronomical cost when an IDE isn't available." — bartonjs
Place local functions at method end, fields first in types. Local functions go at the end of the containing method. Fields are the first members declared in a type.
"The generally agreed upon pattern has been that local functions are placed at the end of the method." — AaronRobinsonMSFT
Narrow warning suppression to smallest scope. Avoid file-wide #pragma suppressions. Disable only around the specific line that triggers the warning.
"Suppressing warnings broadly is generally bad practice." — AaronRobinsonMSFT
Use pattern matching and is/or/and patterns. Prefer is patterns and C# pattern matching over manual type checks and comparisons. Use named parameters for boolean arguments.
"
return !(typeDesc.Category is TypeFlags.Boolean or TypeFlags.Char);" — jkotas
Do not initialize fields to default values (CA1805). The CLR zero-initializes fields. Explicit = false, = 0, = null is redundant.
"CA1805: Do not initialize unnecessarily." — MichalStrehovsky
Sealed classes do not need the full Dispose pattern. A simple Dispose() is sufficient since no derived class can introduce a finalizer.
"Given that the class is marked sealed now, I personally don't think the full dispose pattern is needed." — Youssef1313
Keep PRs focused on their stated scope. No accidental file modifications, no unrelated refactoring, no whitespace noise, no build artifacts. Each PR should serve a single purpose.
"Please be more deliberate about your pull requests... it makes for a very muddy source history." — bartonjs
Do large refactorings and renames in separate PRs. Separate no-diff refactors from functional changes. Mechanical renames should be separate from logic changes.
"I always prefer to do the no-diff refactors first and build the diff changes on top." — AndyAyersMS
Merge to main first, then backport to release branches. Use the /backport command. Backports to servicing are limited to security bugs, regressions, and reliability issues.
"In general, performance related fixes do not meet the bar, unless they are fixing significant regression." — jkotas
Extract duplicated logic into shared helper methods. Fix improvements inside shared helpers so all callers benefit.
"Would it be better to move this to a helper method instead of duplicating it?" — tarekgh
Move shared code to shared files, not duplicated across runtimes. When identical code exists across CoreCLR and NativeAOT, move it to the shared partition (using #if !MONO if needed).
"There is quite a bit of identical code between NativeAOT and CoreCLR. Can we move it into the shared file?" — jkotas
Use existing APIs instead of creating parallel ones. Before introducing new types, enums, or helpers, check if existing ones serve the same purpose. Fix existing utilities rather than introducing duplicates.
"Can you use the existing SignatureAttributes.Instance instead? It means the same thing." — jkotas
Delete dead code and unused declarations aggressively. When removing code, also remove helper methods, enum values, function declarations, and resx strings that are no longer used.
"This function isn't used. Please delete." — davidwrighton
Store error strings in .resx, not inline code. Reference via the SR class. When removing code that uses a resx string, delete the unused string entry.
"We do not store string message in code. Instead, they should be stored in .resx, with optional format argument and referenced with SR." — huoyaoyuan
Don't modify auto-generated files or eng/common manually. Change the generator or source definition instead. Files in eng/common are synced from dotnet/arcade.
"Things in eng/common come from the dotnet/arcade repository. Your fixes here will be undone the next time arcade is synced." — vcsjones
Match existing style in modified files. The existing style in a file takes precedence over general guidelines. Do not change existing code for style alone.
"If a file happens to differ in style from these guidelines, the existing style in that file takes precedence." — huoyaoyuan
Always add regression tests for bug fixes and behavior changes. Prefer adding [InlineData] test cases to existing test files rather than creating new ones. Ensure new test files are included in the csproj.
"The PR needs a regression test added. TypeInfoTests.cs is a good place to add it (add new InlineData)." — jkotas
Test edge cases, error paths, and all affected types. Include empty strings, negative values, boundary conditions, Turkish 'i', surrogate pairs. Test both true and false for boolean options. Choose inputs that can't accidentally pass if output wasn't touched.
"Pick an input that doesn't decode to all 0s so that the test can't pass even if the output wasn't touched at all." — MihaZupan
Test assertions must be specific. Assert exact expected values (exact OperationStatus, exact byte counts), not broad conditions. Ensure tests actually fail when the fix is reverted.
"The current asserts are too broad to be useful." — MihaZupan
Delete flaky and low-value tests rather than patching them. Do not add tests known to be flaky. If a test relies on fragile runtime details and cannot be made reliable, prefer deletion.
"It would be better to delete the test. No point in adding flaky tests." — jkotas
Make test data deterministic and culture-independent. Create CultureInfo with explicit format settings. Use [Theory] with [InlineData] over individual [Fact] methods.
"I would suggest in the test you create a culture like
var culture = new CultureInfo(\"de-DE\"); culture.DateTimeFormat.AbbreviatedMonthGenitiveNames = [...]" — tarekgh
Catch only expected exceptions in fuzz tests. Catching all exceptions masks bugs like undocumented exceptions escaping the API.
"Would it be possible to catch only the exceptions that we expect here?... It has helped me to find that the library was throwing undocumented exceptions." — adamsitnik
Use modern xUnit patterns for xUnit-based tests. In xUnit test projects (for example, most libraries tests), use Assert.* instead of the legacy return 100 == success pattern, use [Fact]/[Theory], prefer ThrowsAnyAsync<OperationCanceledException> for cancellation, and name regression test classes after the issue number (e.g., Runtime_117605). Legacy non-xUnit tests under src/tests may continue to use the existing return 100 convention.
"Can we change the tests here to use Asserts and not the legacy 'return 100 == success' model?" — jkoritzinsky
Comments should explain why, not restate code. Delete comments like // Get the types that just duplicate the code in English. Don't include historical context about why code changed.
"Comments that just duplicate the code in plain English are not very useful. This comment should explain why we are doing this." — jkotas
Delete or update obsolete comments when code changes. Stale comments describing old behavior are worse than no comments.
"The whole comment starting with
Note:can be deleted. It is no longer applicable." — jkotas
Track deferred work with GitHub issues and searchable TODOs. Reference a tracking issue in TODO comments with a consistent prefix (e.g., TODO-Async:). Remove ancient TODOs that will never be addressed.
"Could you please tag all these places that need review with async TODO so that they can be found easily and none of them falls through the cracks?" — jkotas
Don't duplicate comments on interface implementations. Documentation comments belong on the interface definition. Duplicating leads to divergence.
"It is enough to have these comments on the interface. Duplicating them is just going to lead to the comments diverging over time." — jkotas
Add XML doc comments on all new public APIs. These seed the official API documentation on learn.microsoft.com. Properties should start with "Gets the ..." or "Gets or sets the ...". Do not add XML docs to test code.
"Please also include /// comments on all the new APIs (they seed the api docs)." — MihaZupan
Use SHA-specific or commit-based links in documentation. Don't use branch-relative links that break when files move.
"Best to use sha-specific links." — richlander
Use established terminology in user-facing text. Do not expose internal type names, private field names, or codenames like "Roslyn" in public docs or error messages.
"'non-explicit type' is not an established term." — jkotas; "Roslyn is our internal codename. It should not be used in public docs." — jkotas
Retain copyright headers and license information. All C# and C++ source files must include the standard license header, including test files. When porting from other projects, retain original copyright and update THIRD-PARTY-NOTICES.TXT.
"All C# and C++ source files should have license header (incl tests)." — jkotas