luminal-challenge

Imported from albertoHdzE/CausalBool (luminal-challenge/AGENTS.md). Install upstream with npx skills add albertoHdzE/CausalBool --skill luminal-challenge. Copyright stays with the author.

Instructions for agents working on the Luminal challenge

Read the canonical direct-index plan and task status before implementing or reviewing this challenge. These documents are self-contained; do not depend on conversation history.

For the runtime optimization assignment, also read OPTIMIZATION_PHASE_PLAN.md, version 1.1. It defines worker scope, executable benchmark requirements, and final lead review. V3 remains the frozen baseline. V4 repair2 received lead acceptance with limitations: see final review.

• Follow the definitions, task ownership, interfaces, budgets, acceptance gates, and implementation/reviewer prompts in the plan.
• The new solution uses direct decimal-anchor/free-mask/sumandos operations. BDDs, classical/serial seeding, and classical allocation are prohibited in its production path. The existing hybrid prototype is historical, not the target.
• Read ../GOVERNANCE/GLOSSARY.md for terminology and the pinned .reference/README.md and .reference/machine.py for machine semantics.
• Free coordinates are per schema. Engine limits govern issues per cycle, not in-flight duration. Scratch is 256 words with inclusive live intervals.
• Preserve reference files, frozen classical code, historical results, and unrelated repository edits. Never claim UNKNOWN is UNSAT.
• Other agents may be working simultaneously. Edit only assigned files and do not revert their changes. Only the lead accepts a completed task.
• Retain real test output, commands, exit codes, and hashes. Use the plan's independent review checklist and rerun matrix; do not trust success banners.
• Prefer codebase-memory MCP graph tools for code discovery when available; use rg for non-code searches or unavailable/insufficient graph results.
• Do not publish or submit externally as part of implementation.

Start implementation with plan section 10. Start independent review with section 11. If no delegation tools exist, follow the same task dependencies sequentially and preserve the review gates.