sweep

# Codebase Sweep Deploy 4 parallel review agents to scan the entire SwarmLLM codebase for issues. Each agent focuses on a different category. All findings are collected, deduplicated, and presented as a prioritized action list. ## Pre-Sweep: Load Prior Findings Before launching agents, check if `.claude/sweep-log.jsonl` exists. If it does: 1. Read its contents — each line is a JSON object: `{"file":"...","line":N,"kind":"...","summary":"...","status":"fixed|wontfix|deferred","date":"YYYY-MM-DD"}` 2. Extract all entries where `status` is `"fixed"` or `"wontfix"` — these are KNOWN issues 3. Pass the known-issues list to EACH agent with explicit instructions: "Do NOT re-report any of these known issues. Focus on finding NEW issues not in this list." If the file doesn't exist, proceed normally (first sweep). ## File Rotation Strategy To avoid always scanning files in the same order (which causes convergence): 1. Get the list of all `.rs` files: `find src/ -name '*.rs' | sort` 2. Get the list of all `.js` files: `find frontend/js/ -name '*.js' | sort` 3. Pick a rotation offset based on the current sweep count (line count of sweep-log.jsonl ÷ 10, modulo file count) 4. Tell Agent 1+3 to start from offset N in the Rust file list, wrapping around 5. Tell Agent 2+4 to start from offset N in the JS/frontend file list, wrapping around This ensures each sweep round examines files in a different order. ## Agents to Deploy (in parallel, with isolation: "worktree") IMPORTANT: Launc...