Flake reproducer has been dead since #395 — and its failure looked like a repro by rbuergi · Pull Request #730 · Systemorph/MeshWeaver

rbuergi · 2026-07-30T22:28:08Z

Every dispatch of flake-repro.yml has died before running the test.

It packs src/MeshWeaver.BusinessRules{,.Generator} "exactly as dotnet-test.yml does". #395 removed that step from dotnet-test.yml — the scope generator ships in-process, and legacy #r "nuget:MeshWeaver.BusinessRules.Generator" directives are filtered out of the resolve set at compile — and the projects left src/ with it. This copy kept packing them:

MSBUILD : error MSB1009: Project file does not exist.
Switch: src/MeshWeaver.BusinessRules/MeshWeaver.BusinessRules.csproj

Why it survived this long

The pack step runs before the loop, so:

success  7  Build
failure  8  Pack mesh-local          ← dies here
skipped  9  Loop the suspect test until it fails

The loop shows as skipped while the job reports failure — the exact shape of a successful reproduction, whose entire purpose is to fail. I dispatched it against the TrackActivity flake, saw failure, and was one step from reporting a repro of a test that never executed. What caught it was an artifact list that came back empty.

A reproducer whose did-not-run is indistinguishable from its did-reproduce is worse than no reproducer: it answers the question you asked it, with the wrong answer.

The fix

Remove the dead pack step.
Make that outcome impossible to mistake: the loop step now runs with always() and fails loudly — REPRO DID NOT RUN … the job's failure is the harness, not the flake — when its project is absent.

Deliberately not changed

The log-level guidance. This file already warns that Debug in the CompileWatcher hot path masks the race — repros #1/#2 caught it at the committed level, #3/#4 with Debug missed it across 120+ iterations. Dispatch at the default first; treat Trace as a last resort. Worth flagging that my own dispatch used Trace, which on that evidence may itself have suppressed the race it was hunting.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WJiX3ou4QJnUppszJytrNj

…ke a repro Every dispatch of flake-repro.yml has died before running the test. It packed src/MeshWeaver.BusinessRules{,.Generator} "exactly as dotnet-test.yml does"; #395 removed that step from dotnet-test.yml (the scope generator ships in-process and legacy `#r "nuget:MeshWeaver.BusinessRules.Generator"` directives are filtered out of the resolve set at compile) and the projects left src/ with it. This copy kept packing them: MSBUILD : error MSB1009: Project file does not exist. Switch: src/MeshWeaver.BusinessRules/MeshWeaver.BusinessRules.csproj WHY IT SURVIVED. The pack step runs BEFORE the loop, so the loop showed as `skipped` while the job reported `failure` — the exact shape of a successful reproduction, whose whole purpose is to fail. I dispatched it against the TrackActivity flake, saw `failure`, and was one step away from reporting a repro of a test that never executed. An artifact list that came back empty is what made me look. So this removes the dead step AND makes that outcome impossible to mistake: the loop step now runs with `always()` and fails loudly — "REPRO DID NOT RUN … the job's failure is the harness, not the flake" — when its project is absent. A reproducer whose did-not-run is indistinguishable from its did-reproduce is worse than no reproducer, because it answers the question you asked it with the wrong answer. Not changed, deliberately: the log-level guidance. This file already warns that Debug in the CompileWatcher hot path MASKS the race (repros #1/#2 caught it at the committed level; #3/#4 with Debug missed it across 120+ iterations). Dispatch with the default level first and treat Trace as a last resort — my own dispatch used Trace, which on that evidence may have suppressed the very race it was hunting. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WJiX3ou4QJnUppszJytrNj

Copilot

Pull request overview

Fixes the manual flake-repro.yml workflow so it no longer fails before executing the intended test loop (after #395 removed the BusinessRules pack step), and makes “repro did not run” harder to misinterpret as “flake reproduced”.

Changes:

Removed the stale “pack mesh-local #r packages” step that referenced deleted projects and caused the workflow to fail before the loop.
Ensured the loop step runs with if: always() and added a guard to fail loudly when the target project path is missing.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

      - name: Build
        run: dotnet build --no-restore -c Release -p:CIRun=true -warnaserror


          set -uo pipefail
          iters="$REPRO_ITERS"
          proj="$REPRO_PROJECT"
          filt="$REPRO_FILTER"
+          # A repro run that never executed the test is the one outcome that must NOT be silent:
+          # its exit code is identical to a real reproduction.
+          if [ ! -d "$proj" ]; then
+            echo "::error::REPRO DID NOT RUN — project '$proj' not found. Nothing was reproduced; the job's failure is the harness, not the flake."
+            exit 1
+          fi


github-actions · 2026-07-30T22:44:36Z

Test Results (shard 3)

12 files ±0 12 suites ±0 4m 5s ⏱️ +9s
998 tests ±0 994 ✅ ±0 4 💤 ±0 0 ❌ ±0
1 063 runs ±0 1 059 ✅ ±0 4 💤 ±0 0 ❌ ±0