cyberneticlibrary

Hunt bugs with adversarial LLM team

nereids-bug-huntworkflowsetup L33
ornlneutronimaging/NEREIDS
What it does

Multi-agent bug hunt: Claude + Codex

Best for

Broad defect sweep of the NEREIDS workspace ahead of a release / paper. RUN FROM THE MAIN REPO ROOT (not a .claude/worktrees/ session) so SAMMY primary sources resolve and subagents are not sandboxed away from them. Complements /review-pipeline (which is per-PR-diff + fix + merge gates); this workflow is whole-crate DETECTION only — it finds and locates defects, cross-verifies them across two LLM families, and reports. It does NOT fix anything.

Inputs
  • · orchestration prompts/script
  • · JSON schema definitions
Outputs
  • · structured report (markdown/JSON)
  • · phased execution plan
Requires
  • · Multi-agent Workflow orchestration
  • · Git version control
Preconditions
  • · Valid git repository or worktree
  • · Phase definitions loaded