Use when exploring unfamiliar codebases, investigating bugs, or learning new technologies before acting
name research description Use when exploring unfamiliar codebases, investigating bugs, or learning new technologies before acting Research Evidence-based exploration: Observe → Explore → Verify → Summarize . Reading code is NOT research. Testing behavior IS research. Tool Selection Scenario Tool Note Codebase Task: subagent_type="Explore" NEVER manual Grep/Glob Web Task: subagent_type="core:web-researcher" haiku built-in Hybrid Both in parallel synthesize in main session Use model="haiku" for simple Explore tasks. Evidence Standards 3+ independent sources before concluding Sufficient: lib/state.sh:45 validates regex + tested empty input → exits code 1 Insufficient: "Read the code, it does X" — did you RUN it? Negative evidence: document what's NOT there Red Flags — STOP "Read code, that's enough" → must RUN and TEST "Found one good source" → need 3+ "Confident based on experience" → experience ≠ evidence Listing "possible bugs" without testing → speculation