qa-researchlisted

# QA Research Gathers current, sourced best practices relevant to a branch's changes and converts them into **concrete checks**, not decorative notes. Standalone, and also used by `test-iteration`. ## 1. Identify topics From the branch diff (use `"${CLAUDE_PLUGIN_ROOT}/scripts/branch-diff.sh" <branch> [base]` if available, otherwise `git diff --name-only origin/<base>...HEAD` inline), identify the affected technologies / behaviours worth researching (e.g. "JWT refresh rotation", "idempotent payment callbacks", "React Server Components caching"). **Done when:** you have a short, deduplicated topic list. ## 2. Research — tool-agnostic with fallbacks Pick the first available search tool. The last option works on a bare Claude Code install, so research is never silently skipped unless truly nothing is available. - Exa MCP (`mcp__exa__web_search_exa`, `mcp__exa__web_fetch_exa`) → `context7` (for library/framework docs) → built-in **`WebSearch`** → only if none exist, mark **"research skipped: no search tool"** and continue. **Loading Exa for subagents.** Exa tools may be deferred — a subagent first calls `ToolSearch query:"select:mcp__exa__web_search_exa,mcp__exa__web_fetch_exa"`. If a custom subagent doesn't receive MCP (a known Claude Code bug), run a `general-purpose` agent and name the full tool `mcp__exa__web_search_exa` in its prompt. **Scale effort.** 1-2 topics → 1 agent; 3-5 topics → one agent per topic; don't spawn multiple agents for a single topic. **Per-agent