Codebase Sweep

Four parallel sonnet feature-dev:code-reviewer agents, one category each. They only READ, so no worktree isolation (every subagent already starts with a fresh context; a worktree only isolates files).

Prior findings — by grep, never pasted

.claude/sweep-log.jsonl holds every finding ever made (~470 KB, 1.5k lines). Never paste it into a prompt — that is ~110K tokens per agent. Tell each agent: before reporting a finding, search .claude/sweep-log.jsonl for the file or symbol with the Grep tool; drop it if a fixed or wontfix entry covers it. (The reviewer agents have Grep, Glob and Read, but no Bash.)

Rotation

YOU compute each agent's file list and paste it into its prompt — the agents cannot run find. Offset = (line count of the sweep log ÷ 10) mod file count; agents 1 and 3 take find src/ -name '*.rs' | sort from that offset, agents 2 and 4 find frontend/js/ -name '*.js' | sort, wrapping around.

Docs drift — run it yourself first (seconds, no agent)

python3 examples/docs_drift.py --memory checks mechanically what agents read past: every backticked name a doc cites still exists, a module::path::item is defined under that path, each DIAGNOSTICS DIAG row's level and fields match its tracing call, every file.md § "Heading" resolves. It is a report — read each line (proposals and other projects' names are legitimately absent), fix the real ones, log the rest wontfix. Agent 4 then spends its budget on prose and numbers, not names.

Agents (launch all four in one message, model: sonnet, effort: high)

1. Dead code + stale references — pub items with no external caller, unreachable arms, comments naming removed code, #[allow(dead_code)] hiding a real warning.
2. Duplication — near-identical blocks (>5 lines), one transformation done in several places, duplicate fetches or DOM patterns in the frontend.
3. Consistency — an error type chosen at a call site, unbounded collections, unvalidated API input, magic numbers, unstructured tracing calls.
4. Frontend + i18n + docs — English bypassing I18n.t(), dead CSS/JS, doc comments or docs/ text that no longer match the code.

Each finding: file, line, what is wrong, confidence ≥ 80%. Not test-only code, not an item listed in docs/ARCHITECTURE.md § "Deferred Items", and "0 new issues" is a valid result — never manufacture one.

After the agents return

1. Deduplicate, then verify every finding yourself before acting — see "Verify before deleting" below.
2. Fix what is obvious and low-risk immediately, committing as you go. For the rest, research (diagnosis rule 0) and decide — you manage this project; raise with the user only what needs their hands.
3. Append one line per addressed finding to .claude/sweep-log.jsonl: {"file":"…","line":N,"kind":"…","summary":"…","status":"fixed|wontfix|deferred","date":"YYYY-MM-DD"}

Verify before deleting

Sweep agents report dead code / orphaned keys with confidence ≥80%, but their grep may miss call sites in adjacent directories (R120 caught Agent 4 missing 6 enc.* callers in init.js, core/utils.js, chat.js). Before deleting anything an agent flagged:

grep -rn "thing_name" frontend/js/ frontend/index.html frontend/css/   # full frontend
grep -rn "thing_name\b" src/ tests/ crates/                            # word-boundary; catches re-exports

Cheap to verify; expensive to mis-restore. If callers exist, log as wontfix in .claude/sweep-log.jsonl to prevent re-report.