Agent
Deep-reasoning subagent pinned to Opus — architecture design, root-cause analysis for complex bugs, algorithm design, trade-off analysis. Dispatched by the orchestrator for reasoning-heavy phases it should not burn its own context on. 触发词：深度分析、根因分析、架构方案、方案权衡、root cause、architecture design、trade-off analysis、deep…
Agent
Mechanical-execution subagent pinned to Sonnet — boilerplate, well-specified routine changes, running tests/lint/build and fixing trivial failures, formatting, batch replacements, routine test cases. Any implementation work with a clear spec and little judgment required goes here.…
Agent
Independent verification subagent, Sonnet by default (the orchestrator passes model:opus for critical changes) — verifies a change or conclusion with fresh context by running the tests itself, checking each acceptance criterion, and actively constructing counterexamples. Dispatched after an implementer reports "done"…