The review does not assume docs will be updated later — a user-visible surface change without matching doc updates is a finding in its own right. Full detection rules live in .claude/agents/pr-code-reviewer.md (Step 7). Short version:
Nine checks that static per-file analysis tends to miss — each requires tracing across multiple files or comparing code against docs/descriptions:
-
Branch-wide stale-mechanism sweep (Rule S): When the staged diff replaces mechanism X with mechanism Y (e.g., PowerShell fallback → az rest, api://{appId} → bare appId GUID, per-app admin-consent URLs → az ad sp create), enumerate the terms being replaced and grep -rn the entire branch (not just the staged delta) for each term. For every hit outside the staged delta, classify:
- Stale comment (FIX — code/comment lie about behavior).
- Stale user-facing log/warning/exception text (FIX — operators read these in production).
- Stale doc comment / XML doc (FIX — IDE surfaces these to consumers).
- Intentional historical reference (KEEP — e.g., commit message archaeology, CHANGELOG
[Released] entries).
This is the analog of "rename hygiene" applied to behavior changes. It catches the family of issues where the implementation moves but the surrounding documentation lies. Treat each surviving reference as MEDIUM if user-facing (LogWarning, LogError, exception messages, Warnings collection entries), LOW if internal-only (private comments).
Six of the nine unique findings in PR #432's Copilot review were of this exact shape — all stale "PowerShell fallback" and "api://{appId}" references on lines outside the staged delta.
-
Test-class parallelism safety (Rule T): For every test class in the diff (or anywhere on the branch), check whether the class reads or writes any static property or field whose name matches the pattern *ForTests* or *TestOverride* (these are the conventional escape hatches in this codebase). If yes, the class must carry either:
[Collection("Sequential")] attribute, OR[CollectionDefinition(..., DisableParallelization = true)] for a class-specific collection it owns.
Without one of these, xUnit may run the class in parallel with other classes that also mutate the same static state, producing flaky cross-class races. Severity: MEDIUM. The fix is one line. Two PR #432 findings (BatchPermissionsOrchestratorTests, BatchPermissionsOrchestratorMissingSpTests) were exactly this — both mutated BypassSpProvisioningForTests without [Collection] until Copilot flagged it.
Implementation: grep -rn "BypassConsentChecksForTests\|BypassSpProvisioningForTests\|OpenUrlOverrideForTests\|<OtherKnownForTestsName>" <test_files>. For each hit, read the enclosing test class and verify the attribute is present.
-
Branch-scope completeness checkpoint (Rule U): The single biggest miss category in the PR #432 review. Before declaring the review complete, list every file that appears in the branch diff (git diff $(git merge-base HEAD origin/main)...HEAD --name-only), not just the staged delta. The review must touch each file in that list at least to the extent of running the other rules against it.
In practice this means: if the staged delta is 4 files but the branch diff is 22 files, the review surface is 22, not 4. Prior /review-staged runs do not absolve the current run of covering the rest of the branch — assume nothing has been covered until you've explicitly read it in this run.
-
Hardening-bypass detection (Rule V): Added 2026-05-29 after PR #432's second Copilot review caught a regression Rules N–U missed. When the staged diff hardens an entry-point helper by making a parameter required (so the compiler forces it through one path), the lower-level function it delegates to typically still keeps the parameter optional — and any direct caller of the lower-level function bypasses the hardening. The hardening only protects the path that goes through the helper.
Detection: when the diff includes a signature change of the shape "parameter X was optional with = null default, now required (no default)", identify the lower-level function the helper delegates to. If that function still has X optional, run grep -rn "<LowerLevelFunctionName>" --include="*.cs" src/<ProductionPath>/ (production callers only) and verify each direct caller passes X. Any caller that doesn't is a recurrence of the exact bug the hardening was meant to prevent.
Concrete PR #432 example: commit 7a1e317 made mcpScopesByAudience required on ExecuteBatchPermissionsStepAsync so the AllSubcommand and NonDwBlueprintSetupOrchestrator entry points couldn't forget it. But ConfigureAllPermissionsAsync (the lower-level orchestrator method) still had knownMcpAudienceAppIds optional, and PermissionsSubcommand.ConfigureMcpPermissionsAsync called it directly — bypassing the hardening and re-introducing the AADSTS500011 V2 routing regression. Copilot flagged it as HIGH; the skill missed it.
Severity: HIGH when the bypassed call site is in production code (live regression risk). LOW when it's a test fixture (tests routinely use null defaults). Treat the parameter-becoming-required signature change in the diff as a trigger: the moment you see one, follow the call chain down and grep direct callers of the lower-level function.
Implementation: at the start of the review, emit the list of branch-level files and treat them as the review surface. At the end, verify each file was read or explicitly justified as "no rule applies."
-
CLI option precondition enforcement (Rule W): Added 2026-06-02 after PR #440's Copilot review. When a new or modified Option<...> has a description that constrains its applicability — phrases like (--X only), use with --Y, Only meaningful with --Z, requires an existing ... — the handler must validate that precondition and fail fast (logger.LogError(...) + context.ExitCode = 1 + return) when the option is supplied outside that context. A flag that is accepted and then silently ignored (because the code path it feeds is skipped) is a bug: the user gets exit 0 and assumes it took effect.
Detection: for every option whose description contains a constraint keyword, grep the handler for a guard that errors when the option is present but the precondition is false. Use context.ParseResult.CommandResult.FindResultFor(<option>) != null to detect "supplied" (not the value, which can't tell omitted from empty). If no such guard exists, flag it. Severity: MEDIUM. PR #440 example: --messaging-endpoint was documented (--m365 only) and with --endpoint-only, but both handlers accepted it unconditionally and dropped it when the messaging step / endpoint-only path was skipped.
-
Option<string?> empty-vs-omitted conflation (Rule X): Added 2026-06-02 after PR #440. When a string option is read as GetValueForOption(opt)?.Trim() and then gated only with string.IsNullOrWhiteSpace(...), an explicitly-passed empty value (--opt "" or --opt " ") is indistinguishable from omitted. This is a defect whenever "omitted" triggers a different behavior than "empty should" — e.g. omitted falls back to config / prompts / defers, so --opt "" silently does the fallback instead of erroring.
Detection: for each Option<string?> whose omitted-path does something non-trivial (prompt, default, defer, config fallback), verify the handler distinguishes "specified" via FindResultFor(opt) != null and emits a targeted error + ExitCode = 1 when specified-but-whitespace. This is the CLAUDE.md "Input Validation" rule applied to the explicit-empty case. Severity: MEDIUM. PR #440 example: --messaging-endpoint "" was .Trim()ed to "", read as omitted, and triggered prompting/deferral instead of a clear error.
-
Ignored result-bearing call / missing exit code (Rule Y): Added 2026-06-02 after PR #440. Extends the exit-code-completeness check (CLAUDE.md item 8) to discarded return values. When a command handler awaits an operation that returns a success/failure-bearing result (an enum like *Result, a bool, a tuple with an outcome) and does not capture or branch on it, any failure that result encodes will still exit 0 — breaking scripting/CI.
Detection: in every command handler, find await SomeOp(...) calls whose return type is non-void/non-Task and whose result is not assigned or checked. For each, confirm the failure values map to context.ExitCode = 1. Severity: MEDIUM (HIGH if the command is commonly scripted). PR #440 example: the --endpoint-only path called RegisterEndpointAndSyncAsync(...) (returns EndpointRegistrationResult) and discarded it, so NotConfigured / contract-mismatch / Failed all exited 0.
-
Dead local / unused assignment the compiler won't flag (Rule Z): Added 2026-06-03 after PR #440's second review. A local that is assigned but never read is dead code. The trap is that TreatWarningsAsErrors does not catch most of these: the C# unused-variable warning (CS0219) fires only for constant initializers. When the initializer is a property access, method call, LINQ expression, or new, the compiler stays silent (it can't assume the access is side-effect-free), so the build is green and the dead local survives. This means a green build/test run — the skill's one objective signal — gives no indication, so the sweep must be done by reading.
Detection: for every local declared in the diff (and the surrounding method), confirm it is read at least once after assignment. Pay special attention to locals whose initializer is x.Count / x.ToList() / SomeMethod(...) / new T(...) — these are exactly the ones the compiler won't warn about. A common shape is a "summary" or "counter" local that was meant to be used in an output/log line but the line ended up using the individual counters instead. Severity: LOW (dead code, not a runtime defect). Fix by deleting the local, unless the intent was to surface it (e.g. a total in a summary line) — in which case wire it in. PR #440 example: var totalChecks = requirementChecks.Count; in RequirementsSubcommand was never read (the summary used passedChecks/warningChecks/failedChecks); the build stayed green because .Count is a property access, and the skill walked past it.
-
CHANGELOG entry crispness (Rule AA): CHANGELOG.md [Unreleased] feeds straight into the nuget.org release notes, so each entry must be one crisp sentence about the user-visible change — not an essay. Flag entries that explain internal mechanism, name classes/methods/internals, restate rationale, or run multiple sentences of background. Fix by cutting to what a package consumer needs to know. Severity: LOW.
-
Code comment crispness (Rule BB): Comments added in the diff must be crisp and follow .NET conventions — /// XML doc on public members, short // only on non-obvious logic. Flag multi-line rationale essays, comments that restate what the code already says, and commit-message-style narration in source. A comment states why in one line; it does not retell the change (that belongs in the commit/PR). Severity: LOW.
-
Cancellation reachability — Ctrl+C must reach every blocking await (Rule CC): The CLI uses cooperative cancellation (CancelOnProcessTermination in Program.cs), so Ctrl+C only works on a path that observes the command's token. Existing checks (.None in a handler, Task.Delay without a token, retry loops that don't check IsCancellationRequested) assume a token is already in scope. This rule adds the missing-parameter case: a service/helper method whose body awaits a blocking operation — an HTTP send (PostAsync/SendAsync/GetAsync), a retry wrapper (ExecuteWithRetryAsync), a Task.Delay backoff, or a network poll loop — but exposes no CancellationToken parameter at all, so callers cannot thread one even though a token is available upstream (SetupContext.CancellationToken, context.GetCancellationToken()). Symptom: Ctrl+C appears to hang for the full retry/timeout window. Detection: for each such method added/modified in the diff, confirm its signature takes a CancellationToken AND forwards it into every blocking await (cancellationToken: ct, ..., ct)); if absent, trace the call chain to the command handler and confirm a token exists to thread. When adding the parameter, verify every intermediate helper between the handler and the leaf is updated (partial threading leaves a gap), and append Arg.Any<CancellationToken>() to that method's NSubstitute setups/Received/DidNotReceive matchers — otherwise they silently stop matching once the real token flows. Severity: MEDIUM (HIGH when the operation has a long retry+timeout window, e.g. HTTP with exponential backoff, where the hang is most acute). Real example (PR adding retries to TeamsGraphBackendConfigurator): SetBackendConfigurationAsync / ClearBackendConfigurationAsync wrapped the HTTP send in RetryHelper but took no CancellationToken, so the retry + backoff ran with CancellationToken.None and Ctrl+C did nothing until the attempts were exhausted.
-
Unconditional cancellation rethrow swallows library timeouts (Rule DD): The mirror of Rule P. A catch (OperationCanceledException) / catch (TaskCanceledException) that rethrows unconditionally in code that awaits HttpClient (or any API that surfaces its own timeout as TaskCanceledException) cannot tell a user Ctrl+C from a client/library timeout — both throw TaskCanceledException, but only user cancellation sets ct.IsCancellationRequested. Rethrowing both makes a genuine timeout/failure propagate as a silent cancellation abort instead of surfacing as a failure. Detection: for every catch (OperationCanceledException)/catch (TaskCanceledException) that does throw; in a method awaiting an HTTP call or a retry wrapper, confirm the catch is guarded with when (ct.IsCancellationRequested) (or otherwise distinguishes the user token from a timeout). If it rethrows unconditionally, flag it — a timeout will exit silently. Severity: MEDIUM. Real example (same PR as Rule CC): TeamsGraphBackendConfigurator added catch (OperationCanceledException) { throw; } after wrapping the send in RetryHelper; an HttpClient 2-min timeout (which RetryHelper surfaces after exhausting retries with IsCancellationRequested == false) would have aborted setup silently instead of returning Failed. Fixed by guarding with when (ct.IsCancellationRequested).
-
Claude Code reads .claude/agents/pr-code-reviewer.md for review process guidelines
-
Claude Code reads .github/copilot-instructions.md for coding standards
-
Claude Code gets staged files: git diff --staged --name-only
-
Claude Code gets staged changes: git diff --staged
-
Claude Code gets the full branch diff against main:
git diff $(git merge-base HEAD origin/main)...HEAD --name-only
git diff $(git merge-base HEAD origin/main)...HEAD
This covers ALL commits on the branch since it diverged from origin/main — including prior commits that are no longer in the staged diff. The union of staged files and branch-diff files is the full review surface. This is the same view Copilot and PR reviewers see.
When there are no staged files, skip step 4 and use only the branch diff. When both exist, label findings clearly — [staged] for changes only in the staged diff, [branch] for changes from prior commits, so the developer knows which issues are still ahead of them vs. already committed.
-
Claude Code reads the complete current content of every file that appears in either diff (staged or branch) to enable full-file semantic analysis.
6a. If any file in the combined diff is a test file (path contains .Tests. or filename ends in Tests.cs), Claude Code MUST also invoke the pr-test-analyzer agent as a subagent to review test coverage quality. Pass the combined diff and list of test files as context. The pr-test-analyzer agent focuses on:
- NSubstitute predicate precision (predicates on mutable reference types evaluated at assertion time vs. call time)
- Missing test scenarios for new orchestration code paths (e.g., 409 idempotent path, tri-state verification results)
- Assertions that track implementation rather than requirements
- Test method names that accurately describe the contract being verified
Include the
pr-test-analyzer findings in the review document under a dedicated Test Coverage section. This is critical for catching issues that exist in unchanged sections of modified files, such as:
- Duplicate hardcoded constants or magic values that already exist elsewhere
- Parallel code structures that should be consolidated (e.g., a method building the same spec list as a shared helper)
- Unused or dead code that was already there but not touched by the diff
- Missing calls to shared helpers — where the diff adds a new use but existing code still has the old duplicate pattern
-
Claude Code performs semantic analysis using its own capabilities
-
Claude Code identifies specific issues with line numbers and code references
-
Claude Code runs the full test suite with per-test timing:
cd src && dotnet test tests.proj --configuration Release --logger "console;verbosity=normal" 2>&1
Parse the output for lines matching [X s] or [X,XXX ms] patterns. Extract test class name, method name, and duration. Flag any test method taking > 1 second. Group findings by test class and include the measured times in the review.
-
Claude Code writes markdown file to .codereviews/claude-staged-<timestamp>.md
