AI Agent & Automation Workflows

Use when a settled answer needs to become real action and the user wants the steps put into motion through the tools they already have ("make it happen", "set this up", "do it", "get this going", "book it"). Turns the locked answer into the smallest set of concrete steps and runs them through the host's connected tools, asking before anything that writes or sends. Use act after converge or recommend has produced an answer. It acts, so it always shows the plan and gets a go-ahead first.

Use when the user wants to push all in-flight work forward toward review in one pass — closing gaps across Epics, Bugs, and orphan Tasks. NOT for single-ticket work (use engineering) or read-only triage (use my-triage). Trigger: '收斂', 'converge', '推進', '全部推到 review', '把我的單收一收', 'epic 進度', '離 merge 還多遠', '補全'.

Bridge agent that extracts and structures design context from Figma via MCP Server for downstream implementation agents. Use when Figma-to-code bridging, Code Connect management, or design system rule extraction is needed.

Use when a decision was settled earlier and something has changed, and the user wants to know if it still holds rather than redo it from scratch ("does this still work now that X happened", "did the plan survive", "re-check our call"). Re-converges with memory and reports which parts drifted. Use hold to re-check an existing answer against new facts; use converge for a fresh decision, and recommend when you should make the call yourself.

Test data and fixture generation agent. Use when factory pattern design, boundary value data generation, synthetic data generation, or seed data management is needed.

Use when the user asks you to decide for them, or signals they're unsure and want your call ("what should I do", "you pick", "I don't know, recommend something"). Take agency: run the whole loop yourself (explore your own best lines, converge, optimize) and return a clear pick backed by a receipt. The right move when the user is uncertain and wants a decision rather than a list of options. Unlike explore (which only lays out options) and converge (which resolves options the user already supplied), recommend generates its own options and commits to one.