Eliminate Redundant Restores

Eliminate Redundant Restores

A no-op restore still evaluates every project and checks restore state, so each extra restore in a multi-step build costs seconds to tens of seconds in a large repo. The cure is easy to get wrong: --no-restore on a step that no earlier restore actually covered turns a slow build into a failed one. This skill finds which invocations restore, proves which are already covered, and changes only those.

When to Use

• A pipeline or script runs several dotnet/msbuild steps and the log shows restore output in more than one of them

• A restore step exists, yet later steps still print Determining projects to restore...

• A repository Directory.Build.rsp or the active MSBuild installation's MSBuild.rsp contains -restore

• Someone asks where --no-restore or -restore:false can safely be added

• NETSDK1004 or NETSDK1047 appeared after --no-restore was added

When Not to Use

• The first restore is slow (feed latency, missing package cache): that is caching, not redundancy

• There is only one restore across all invocations: nothing to deduplicate

• Restore errors about versions, lock files or auditing: fix the restore itself

• Unranked "why is my build slow": start with build-perf-baseline, then build-perf-diagnostics

Inputs

Discover repository inputs in the working directory; do not ask the user for paths.

• CI definitions (*.yml, .github/workflows, azure-pipelines*.yml, Jenkinsfile) and build scripts (build.cmd, build.sh, build.ps1, eng/)

• Directory.Build.rsp, Directory.Build.props/.targets, NuGet.config

• Auto-loaded MSBuild.rsp beside the active MSBuild executable, or response files passed explicitly with @file. A repository-local MSBuild.rsp is not auto-loaded

• Restore evidence if present: step logs or a binlog

Workflow

Step 1: Confirm restore is repeated and worth removing

• Log evidence: Determining projects to restore... or All projects are up-to-date for restore. once per invocation.

• Binlog evidence: restore evaluations carry the global properties MSBuildRestoreSessionId and ExcludeRestorePackageImports=true; the Target Performance Summary lists Restore with N calls.

• Quantify each no-op restore. If there is one restore, or the repeated ones total about a second, report "nothing worth removing" and stop without editing.

Step 2: Inventory every invocation in execution order

Starting with .NET SDK 9.0.1xx, --no-restore forwards -restore:false to override rsp flags; older SDKs can still inherit -restore despite skipping the CLI's implicit restore. Whenever an rsp enables restore, use the version-independent --no-restore -restore:false with dotnet build, test, publish or pack, even with --no-build. dotnet msbuild and msbuild.exe take -restore:false, not --no-restore.

dotnet run does not expose the MSBuild -restore:false override. With an rsp on older SDKs, first build using dotnet build --no-restore -restore:false, then run using dotnet run --no-build with the same project, configuration, framework and runtime inputs (or execute the built DLL directly). Passing -restore:false to dotnet run is not a restore-suppression strategy.

Step 3: Decide per invocation

Keep the first restore. Skip a later one only when all four checks pass:

Step 4: Apply the smallest change

• One explicit restore carrying every restore input, then --no-restore, -restore:false or --no-build on later steps.

• Several dotnet test steps over one solution: build once, then dotnet test --no-build on each. Add --no-restore -restore:false if an rsp enables restore, including on .NET 8 and earlier. --no-restore alone still rebuilds.

• Keep self-contained publish settings. A publish --no-build still needs a prior build with matching configuration, framework, RID and self-contained output; restore compatibility does not prove that build outputs exist.

• Scope the change to CI and scripts. Leave -restore in Directory.Build.rsp, because developers rely on it; pass -restore:false per invocation instead.

• Never remove the first restore. If it is a --locked-mode gate, keep it there.

• Do not bundle static graph restore or package caching into this change.

# One restore carries every input later steps use
dotnet restore MyApp.sln -r linux-x64
# Explicit overrides also work when an rsp enables restore on older SDKs
dotnet build MyApp.sln -c Release --no-restore -restore:false
dotnet test MyApp.sln -c Release --no-build --no-restore -restore:false
dotnet publish src/Api -c Release -r linux-x64 --no-restore -restore:false
# msbuild invocations need the explicit switch if an rsp file adds -restore
dotnet msbuild Packaging.proj -t:Pack -restore:false

Step 5: Validate and report truthfully

1. Run the changed pipeline or script end to end in a clean temporary copy (no obj/, no project.assets.json). A project that was never restored only fails there. YAML syntax or flag-presence checks do not validate restore coverage or test execution. If you cannot run it, say "not validated" and list the steps and projects to confirm.

2. Success evidence: every later step passed; Restore calls (binlog) or Determining projects to restore lines fell from N to the expected count; for a time claim, before/after medians of at least three runs.

3. Report a restore failure (NETSDK1004, NETSDK1047) separately from a build failure, and never claim success after either.

Output

Explain which steps restored, which earlier restore covers them, what changed, and what remains unvalidated. State whether redundant restores were removed or no change was needed, keeping the required first restore. Use prose or a findings table as appropriate; no fixed table headings or verdict syntax are required.

Common Pitfalls