AGENT SKILL

resharper-clt

Use when running JetBrains ReSharper Command Line Tools over a .NET solution — the CleanupCode/format fix pass an agent runs before claiming C# work complete, or the once-per-review-cycle InspectCode gate whose output a code review triages. Do NOT use for non-.NET languages, for judging design or correctness by reading code, as a substitute for the compiler and test gates, or to run InspectCode from inside a per-task completion gate.

License 许可
MIT
最近更新
2026-11-15
Stars
★ 23
Forks
3

01 前言 Front Matter

name
resharper-clt
license
MIT
description
Use when running JetBrains ReSharper Command Line Tools over a .NET solution — the CleanupCode/format fix pass an agent runs before claiming C# work complete, or the once-per-review-cycle InspectCode gate whose output a code review triages. Do NOT use for non-.NET languages, for judging design or correctness by reading code, as a substitute for the compiler and test gates, or to run InspectCode from inside a per-task completion gate.

03 SKILL.md 正文

ReSharper Command Line Tools

InspectCode runs ReSharper's inspection set over a solution and emits a machine-readable report. CleanupCode applies ReSharper's formatting and cleanup profiles to source files. Both are mechanical: the same solution at the same revision produces the same output, which is what makes their result citable as evidence rather than an opinion.

That property is also the boundary. This skill produces and remediates tool output. It does not judge design, correctness, or intent — those belong to the reviewer and its lenses.

Setup

The tools ship as a .NET global tool, pinned in the repository's tool manifest at .config/dotnet-tools.json:

{
  "version": 1,
  "isRoot": true,
  "tools": {
    "jetbrains.resharper.globaltools": {
      "version": "2026.2.1",
      "commands": ["jb"],
      "rollForward": false
    }
  }
}

ALWAYS restore before the first invocation in a session:

dotnet tool restore

If the repository has no manifest entry, say so and stop rather than installing a tool version nobody pinned. A floating analyzer version makes every later report incomparable.

Commands

Resolve the solution path from the repository rather than assuming one, and write reports to the path declared as .agent_scratchpad where the repository declares one. Substitute <solution> and <scratchpad> below.

Purpose Command
Full-solution inspection dotnet jb inspectcode <solution> --output=<scratchpad>/inspectcode.xml --format=Xml
Project-scoped inspection dotnet jb inspectcode <solution> --project="<ProjectName>" --output=<scratchpad>/inspect-<project>.xml --format=Xml
SARIF for CI or automated scanning dotnet jb inspectcode <solution> --output=<scratchpad>/results.sarif --format=Sarif --severity=WARNING
Formatting and cleanup dotnet jb cleanupcode <solution>

Scope the run to the affected projects when the solution is large and you are iterating by hand. The once-per-review-cycle review gate takes the full solution: it is the only inspection pass in the run, so narrowing it drops coverage nothing else replaces.

Two cadences, and which one you are in

This skill has two commands and they run at different times. Getting them the wrong way round is the mistake this section exists to prevent.

Command Who runs it When
Fix pass dotnet format, dotnet format analyzers, cleanupcode the agent that wrote the code every task, before READY_FOR_REVIEW
Inspection inspectcode the review gate runner once per review cycle, over the whole accumulated change

The reason is contention and cost. inspectcode loads the entire solution; run at baseline and again at the end by every worker, it was both the slowest part of a completion gate and the part most likely to collide with a concurrent worker's build output. Run once, on a quiescent tree, it costs a single solution load and reports against a change that is finished. A worker that runs inspectcode is doing the council's job at N times the price.

Running the fix pass before you claim work is complete

Any agent implementing, modifying, or refactoring C# MUST run this before reporting READY_FOR_REVIEW.

Fix mechanically, then stop. A defect a machine can fix is not a defect worth a reviewer's attention. Every mechanical finding that reaches a reviewer costs a review pass to report, a rework cycle to fix, and another pass to confirm — to arrive at the same edit cleanupcode would have made for free. Erasing them here is the whole point.

  1. Baseline. Before the first edit, collect language diagnostics for the files you are permitted to change and keep the output. Without it you cannot distinguish a diagnostic you introduced from one that was already there, and "pre-existing" is a claim that needs proof. Do not take an inspectcode baseline; that comparison now belongs to the review gate.

  2. Fix deterministically, scoped to the files you changed. Three commands, in this order, each applying rather than verifying:

    dotnet format <solution> --include <changed files>
    dotnet format <solution> analyzers --include <changed files>
    dotnet jb cleanupcode <solution> --profile="<cleanup profile>" --include="<changed files>" --no-build
    

    --no-build skips the tool's own restore and build, so the solution must already be restored into its default obj/: run dotnet restore <solution> first when nothing has. A restore redirected with BaseIntermediateOutputPath does not count.

    When the host's C# coding standard gives its own cleanupcode command in its ## Completion section, run that command instead of this line.

    --include is not optional. Unscoped, these rewrite files the task never touched, and the change under review stops being the change you made. The pass is idempotent: running it twice produces no second diff, so a re-run after a rework cycle is safe.

    All three load MSBuild and write into obj/. Where concurrent workers may run them against the same projects, give each an isolated artifacts path — see the host's per-agent completion gate.

  3. Re-collect diagnostics over the files you changed — whole file, not only the changed members. Not workspace-wide: while other workers are in flight, a workspace-wide pass reads their half-finished state and attributes it to you.

  4. Fix every finding in your change set that step 2 could not. What survives the fix pass is the genuinely non-mechanical remainder, and that is the part worth your judgement.

  5. Report the result. The fix pass applied, diagnostics clean on your files, and the isolated artifacts path you used — or an explicit list of what remains with baseline proof.

What the cleanup profile should and should not do

Fix what the repository's own standards and analyzers already require: predefined type keywords, explicit types where the standard forbids var, redundant qualifiers and parentheses, unused usings, fields that can be readonly, formatting. Those are settled decisions, and re-arguing one in review is waste.

Leave everything the repository has not declared. Brace style with no .editorconfig rule behind it, member reordering, and file layout are the tool's opinion rather than the project's standard — and reordering members in particular buries the actual change under churn a reviewer then has to read past. A cleanup profile that imposes undeclared preferences trades review cost for diff noise, which is not the trade this gate exists to make.

A green build is not the same measurement. The compiler and ReSharper's inspection set overlap partially and disagree at the edges by design — a suggestion-severity inspection is invisible to dotnet build and still a real finding. That is why the review gate pins InspectCode at --severity=SUGGESTION rather than trusting the build.

Frequently introduced inspections, and what to do

Inspection Remediation
PossibleMultipleEnumeration Materialize the IEnumerable with .ToList() or .ToArray() before iterating it more than once.
ConditionIsAlwaysTrueOrFalseAccordingToNullableAPIContract Fix the contradictory null check, or correct the nullable annotation on the DTO or API model it disagrees with.
InheritdocInvalidUsage Replace <inheritdoc /> on a type that inherits no documentation with a real <summary>.
UseCollectionExpression Modernize the array or list initializer to a collection expression [...].
CheckNamespace Align the file's namespace to its directory hierarchy.
UnusedParameter.Local, PrivateFieldCanBeConvertedToLocalVariable Remove the dead parameter, or narrow the single-assignment field to a local.
RedundantSuppressNullableWarningExpression Remove the unnecessary !. Suppressing a warning the compiler no longer raises hides the next real one.

NEVER fix an inspection by breaking a public contract. Deleting a public API action method, an interface member, or a DTO serialization property to silence a "never used" inspection trades a warning for a breaking change. Suppress it with a stated reason instead, or leave it and say why.

Running it as a review gate

In a code review this is a gate, not a lens: it is a repeatable command whose output the static-analysis-triage lens attributes to the change. Declare it under review.gates in .kyber-weave/kyber-weave.yml so the review runner executes it rather than the reviewer improvising a command:

review:
  gates:
    - id: dotnet-tool-restore
      run: [dotnet, tool, restore]
      blocking: true
    - id: inspectcode
      run: [dotnet, jb, inspectcode, <solution>, --output=<scratchpad>/inspectcode.xml, --format=Xml, --severity=SUGGESTION, --caches-home=<scratchpad>/gates/inspectcode-cache]
      blocking: true

dotnet-tool-restore provisions the pinned jb command; without it the inspection gate fails on a missing tool rather than on a finding. --severity=SUGGESTION states the tool's current default rather than changing it — pinned, because the suggestion-level inspections are exactly the ones the compiler cannot see, and a default that moves would silently drop them.

run is argv, never a command line — the gate runner refuses a shell.

The gate result records the id, exit code, and duration — not the report. The static-analysis-triage lens reads the report itself, from the --output path the gate declaration names, so that path is part of the contract between them: change it in one place and the lens finds nothing.

InspectCode exits zero whether or not it finds anything. blocking: true here means the analyzer must have run, not that it found nothing. The judgement about what it found is the lens's, and it is made against the diff — a solution-wide report is expected to carry pre-existing findings, and attributing those to the change under review is the specific error that lens exists to avoid.

The reviewer's own obligations around that output:

  • Do not re-derive it. The lens attributes diagnostics to the diff and reports them by rule identifier. Reading the code to form a second opinion about a rule is a different lens's job.
  • Report by rule identifier and location. "Analyzer warning" is not a finding.
  • A clean run is a recorded result, not silence. State the gate ran and returned zero findings.
  • An unavailable tool is a stated skip. If the manifest has no entry or the restore fails, record the gate as skipped with the reason. A review that quietly omits a gate reads as a review that passed one.

Worked example

A change adds a service that filters an IEnumerable<Load> and iterates the result twice. dotnet build is green — nothing here is a compiler error.

At the task. The worker runs the fix pass, scoped to its own files and its own build output:

dotnet format Hauling.sln --include 2-Application/Hauling.Application/Dispatch/LoadPlanner.cs \
  -p:BaseIntermediateOutputPath=.agent_scratchpad/T3/artifacts/obj/
dotnet format Hauling.sln analyzers --include 2-Application/Hauling.Application/Dispatch/LoadPlanner.cs \
  -p:BaseIntermediateOutputPath=.agent_scratchpad/T3/artifacts/obj/
dotnet restore Hauling.sln
dotnet jb cleanupcode Hauling.sln --profile="<cleanup profile>" --include="2-Application/Hauling.Application/Dispatch/LoadPlanner.cs" --no-build

Formatting, redundant qualifiers, and unused usings are gone. The multiple enumeration is not — no formatter can decide that materializing the sequence is the right fix. The worker reports:

DIAGNOSTICS: clean on 2-Application/Hauling.Application/Dispatch/LoadPlanner.cs | fix pass: format, format analyzers, cleanupcode — all applied | artifacts: .agent_scratchpad/T3/artifacts | baseline: .agent_scratchpad/T3/diagnostics-baseline.md | remaining: none

At the end of the run. The inspectcode gate runs once per review cycle over the whole solution:

<Issue TypeId="PossibleMultipleEnumeration"
       File="2-Application/Hauling.Application/Dispatch/LoadPlanner.cs" Line="47"
       Message="Possible multiple enumeration" />

static-analysis-triage checks that line against the accumulated diff, finds it inside T3's change, and reports it by rule id and location. It becomes a review finding routed back to a worker of that type — .ToList() before the first iteration — rather than something every worker paid a solution load to look for. Findings on lines the run did not touch are pre-existing and the lens drops them; that attribution, not the worker's baseline, is what "pre-existing" now means.

Boundaries

  • .NET solutions only. There is no ReSharper CLT for other languages; say so rather than substituting a different tool under this skill's name.
  • cleanupcode rewrites source. It belongs to the completion gate above, run by the agent that wrote the code, before any reviewer sees it — and NEVER as part of a review itself. A reviewer that edits the code it is judging has stopped being a reviewer.
  • inspectcode is the mirror image: it belongs to the review gate, once per review cycle, and NEVER to a per-task completion gate. A worker that runs it loads the whole solution to analyze a tree its peers are still editing, and pays that cost once per task instead of once per review cycle.
  • Formatter output is not a review finding. Whitespace differences belong to cleanupcode, not to the findings list. A reviewer that finds mechanical issues is looking at a change whose completion gate did not run: the correct report is that single fact, not a list of nits the gate would have erased.

04 同仓库更多 Skills