build-perf-diagnostics

# Build Performance Diagnostics Agent diagnoses proven MSBuild slowness from a binlog. ## When to Use Agent uses this skill when: - Baseline data already exists - User needs target, task, analyzer, restore, or evaluation root cause - Wall clock time stays high after first-line fixes ## When Not to Use Agent does not use this skill when: - Agent has not captured a baseline yet - Build system is not MSBuild based - User needs pure incremental-skip analysis or pure graph tuning ## Required Inputs | Input | Required | Notes | |---|---|---| | Binlog | Yes | Existing `.binlog` from the slow build | | Build scope | Yes | Project or solution represented by the binlog | | Baseline context | Yes | Cold, warm, or no-op expectation | | Validation command | No | Targeted build command for fix verification | ## Diagnostic Workflow | Step | Agent action | Test | Pass | |---|---|---|---| | 1. Load evidence | Agent opens the binlog with MCP tools when available, else replay workflow | Agent extracts total duration and dominant summaries | Agent has reliable timing evidence | | 2. Rank costs | Agent lists slowest projects, targets, tasks, and analyzer time | Agent sorts costs by cumulative impact | Agent identifies the top bottleneck category | | 3. Classify root cause | Agent maps evidence to RAR, analyzers, copy I/O, restore, evaluation, or graph shape | Agent cites the exact symptom threshold | Agent produces one primary diagnosis | | 4. Recommend minimal fix | Agent selects the smallest change that addresses the primary diagnosis | Agent ties change to one file or command | Agent avoids unrelated tuning | | 5. Verify | Agent reruns the targeted command with the same scope | Agent compares timing and summary output | Agent confirms measurable improvement or records no gain | ## Bottleneck Thresholds | Category | Signal | Action path | |---|---|---| | RAR | `ResolveAssemblyReference` stays above 5s per project | Reduce reference load, network paths, or transitive closure | | Analyzers | Analyzer cost exceeds 30% of `Csc` time | Condition analyzers for inner loop, trim redundant packages | | Single target domination | One target exceeds 50% of total build time | Optimize or split the dominating step | | Low parallel efficiency | Wall clock nearly matches total project time sum | Use `build-parallelism` guidance | | Copy-heavy build | `Copy` time is materially high | Use hardlinks, skip unchanged files, trim output copy set | | Restore in every build | Restore appears in warm or no-op builds | Separate restore or enable static graph restore | | Evaluation overhead | Build starts slowly before execution | Use `eval-performance` guidance | ## Workflow Commands | Purpose | Command | |---|---| | Replay binlog | `dotnet msbuild build.binlog -noconlog -fl -flp:"v=diag;logfile=full.log;performancesummary"` | | Read summaries | `grep "Target Performance Summary\|Task Performance Summary" -A 50 full.log` | | Check per-project cost | `grep "done building project\|Project Performance Summary" full.log` | | Check analyzers | `grep -i "Total analyzer execution time\|analyzer.*elapsed\|CompilerAnalyzerDriver" full.log` | | Check node use | `grep -i "node.*assigned\|Building with\|scheduler" full.log` | ## Validation | Check | Test | Pass | |---|---|---| | Diagnosis quality | Agent ties one recommendation to one dominant cost | Evidence points to a single primary bottleneck | | Fix scope | Agent changes only the files or commands required | No unrelated build churn appears | | Timing proof | Agent reruns the targeted build | Duration or dominant summary improves measurably | | Escalation | Agent routes leftover issues to a specialist skill | Follow-up matches the remaining bottleneck | ## Common Pitfalls | Pitfall | Agent correction | |---|---| | Agent skips baseline context | Agent starts from measured cold, warm, or no-op data | | Agent reports every slow line item equally | Agent ranks by dominant cost first | | Agent removes analyzers outright | Agent preserves CI enforcement and narrows inner-loop cost | | Agent blames parallelism for single-project slowness | Agent separates intra-project cost from graph-level scheduling |