[MattyStacks]

Pulls the backlog item "Let the bot process attached or pasted result files" (prompted by #36).

What changes

• Result files through issues. A new Benchmark result (file) form takes one attached or pasted .yaml/.yml/.json result file. Any issue with the benchmark label and one result file (like Add my benchmark run to the benchmarks results #36's Open GPU Benchmarker export) works too.
  • Downloads come only from github.com/user-attachments/files/, with redirects allowed only to *.githubusercontent.com, capped at 1 MB and 20 s, and saved into $OUT, never the checkout.
  • YAML aliases are refused. The file's fields are kept (including detected and diagnostics).
  • submitted_by comes only from the Credit box. testbench_id, source, and synthetic are dropped.
• Submitting is consent. Both forms drop their required CC BY 4.0 and frame-generation checkboxes and show the license as a notice. The PR body says the result was submitted under CC BY 4.0, and the reviewer checklist adds a frame-generation item.
• Fallbacks.
  • An issue with neither the form nor a file gets a reply linking both forms.
  • Attachments that aren't result files (CSVs, screenshots) exit 3. The bot replies and adds needs-maintainer, and the steps a maintainer follows are in the new docs/PROCESSING_ISSUES.md.
• benchmark-processor Copilot agent (.github/agents/benchmark-processor.agent.md). Assign a benchmark issue to Copilot and pick this agent.
  • It writes the result YAML itself from whatever the issue holds: pasted numbers, a plain description, a file, a CSV or Cyberpunk capture (through parse.py / import_cyberpunk.py), or a screenshot.
  • It adds a missing GPU or game with sources, runs validate.py / build.py / test_build.py, and opens a data-only PR listing every assumption. If something can't be worked out, it opens a draft with questions instead.
  • It may only change data/, and it treats the issue and attachments as untrusted.
  • copilot-setup-steps.yml (one read-only job, no secrets) installs its Python dependencies. parse.py gains --date.
• Converting raw captures is still in the backlog, along with the openbenchmarker-repo export change.

Testing

• New check_issue_files() covers:
  • an attached file, a pasted file in the form, and a pasted JSON block;
  • a CSV attachment going to a maintainer;
  • nine kinds of failure;
  • the redirect allowlist.
• It uses the hidden --attachment-dir and never touches the network. check_bot_safety() makes sure the workflow never passes that flag.
• Ran the script live against Add my benchmark run to the benchmarks results #36's real attachment on a temp copy of data/. It wrote data/community/rtx_5080_16gb_fe/result_20261026_openbenchmarker_neon_flythrough.yaml, which passes the checks with two warnings (no proof, no official baseline).
• validate.py, build.py, test_build.py, py_compile, the blog build, and git diff --check all pass.

After merge

• Run "Copilot Setup Steps" once from Actions, then try the agent on a test issue (it only works from main). If downloads are blocked, add objects.githubusercontent.com to Settings → Copilot → Internet access.
• Tag v1.12.0.0.
• Edit or relabel Add my benchmark run to the benchmarks results #36 so the bot reruns.

🤖 Generated with Claude Code

[github-actions]

Cloudflare Pages preview

• This commit (2f51055): https://2a74f413.open-gpu-benchmarks.pages.dev
• Latest on this branch: https://feat-issue-attachments.open-gpu-benchmarks.pages.dev