ContextQMD
Back to Skiasharp

Anti-Patterns (Full Reference)

.agents/skills/issue-repro/references/anti-patterns.md

3.119.44.7 KB
Original Source

Anti-Patterns (Full Reference)

The critical rules are in SKILL.md. This file has the complete list with examples.

Step Result Decision Table

You RunWhat HappensStep ResultWhy
dotnet buildBuild succeeds, exits 0successCommand succeeded
dotnet buildBuild fails with CS0117failureCommand failed
dotnet build (reporter said it fails)Build fails with CS0117failureCommand still failed — matching the report doesn't change the technical outcome
Render imageExits 0 but pixels wrongwrong-outputProcess succeeded, output incorrect

Full Anti-Pattern List

  1. Inline binary content. Use artifacts array with URLs, not inline data.
  2. Unlimited output. Truncate: 2KB success, 4KB failure, 50 lines stack trace.
  3. Source code investigation. Stop at "did it reproduce." Fixes are issue-fix's job.
  4. Prompting the user. Auto-proceed with logged warning. Skill must be non-interactive.
  5. Absolute paths in output. Redact /Users/{name}/ → $HOME/.
  6. Giving up too early. Try multiple approaches, versions, platforms before not-reproduced.
  7. Building from source first. Start with NuGet packages. Source is Phase 3C only.
  8. Editorial judgment in conclusion. reproduced = reported behavior occurred, even if by-design.
  9. Mismarking step results. result = technical outcome, not expectation match.
  10. Pre-emptive version assumptions. Inspect the nupkg before claiming incompatibility.
  11. Abandoning on environment issues. Missing workloads, sudo prompts are fixable — not blockers.
  12. Skipping Docker for Linux bugs. Try Docker before concluding needs-platform.
  13. Assuming TFM incompatibility. .NET is forward-compatible. net8.0 works on net10.0.
  14. Stopping at build success for WASM. Serve the app, check browser console with Playwright.
  15. Retrying without changing variables. Change platform, version, or approach — not retry count.
  16. Setup failures ≠ "not reproduced." Docker timeout, missing Playwright = blocker, not conclusion.
  17. Testing WASM for every bug. Only when issue signals suggest browser/web.
  18. Silently skipping cross-platform. Record why in notes, set scope to "unknown".
  19. Reusing build artifacts across versions. Fresh project dirs or rm -rf bin/ obj/ between versions.

Output Limits

See SKILL.md for the authoritative output limits table.

FieldMax Size
reproductionSteps[].output (success)2KB
reproductionSteps[].output (failure)4KB
errorMessages.stackTrace5KB / 50 lines
File contentInline for small source files; omit binaries

Redaction: /Users/{name}/ → $HOME/, tokens → [REDACTED]

Additional Anti-Patterns

  1. Never use sudo. If a command requires sudo, find an alternative. sudo prompts for a password interactively, which blocks the session indefinitely. Use user-local installs, Docker, or different approaches.

  2. No repo markdown artifacts. NEVER create markdown summary files (e.g., REPRO_SUMMARY.md, COMPLETION_REPORT.md) in the repository working tree. All working files belong in the session workspace (~/.copilot/session-state/). Clean up any accidentally created files before persisting.

  3. Fabricating output or evidence. NEVER use echo/print statements to simulate command output, claim "reproduced" from static code analysis without executing the code, or report environment details (OS version, SDK version) without actually running the diagnostic commands. If you cannot run it, report needs-platform.

  4. Modifying product source during repro. Reproduction ONLY creates new test projects in /tmp/skiasharp/repro/{timestamp}/. NEVER edit binding/, externals/, samples/, source/, tests/, utils/, or any other product source — that is the issue-fix skill's job. Revert immediately if done accidentally.

  5. Skipping validation. NEVER skip the validation script (validate-repro.ps1 / validate-repro.py). NEVER assume the JSON is valid without running it. NEVER persist to the data cache without seeing ✅ from the validator. Mentally reviewing the JSON is not a substitute for the script.

  6. Using reproduced/not-reproduced for non-bug issues. For enhancements, feature requests, and documentation issues, use confirmed/not-confirmed instead. reproduced means "reported misbehavior was observed" — it doesn't make sense for "the feature is missing."

  7. Using confirmed/not-confirmed for bugs. For bugs, always use reproduced/not-reproduced. confirmed means "reporter's non-bug claim was verified" — it doesn't apply to behavioral bugs.