graph-engineering-workflowlisted

# Graph Engineering Workflow ## Purpose Use this skill to turn a coding request into a task graph made of bounded agent loops, independent workers, fresh verifiers, evidence anchors, and explicit merge and repair rules. It is a graph-orchestration discipline, not a fixed implementation checklist and not a reason to spawn agents for every task. The graph should make independent work wider, not make sequential work more complicated. A worker is allowed to contain its own loop: do the work, check it against a real anchor, repair it, and report an inspectable artifact. ## Task Routing Use when the user asks to: - implement a feature across independent files or components; - fix or migrate many independent units; - audit a repository, route set, module set, or data set; - perform parallel research with evidence checking; - run security, privacy, test, input, dependency, or code-quality audits; - coordinate coding agents from Hermes, Codex, Claude Code, or another host. Do not force this graph for a small, clear, single-file task or for work whose steps have real sequential dependencies. Use one agent loop when there is no meaningful fan-out. ## Operating Workflow 1. **Architect before fan-out.** Identify the units of work, their inputs and outputs, and the exact dependencies between them. 2. **Delete fake edges.** An edge exists only when downstream work must read upstream output. The phrase “then” is not evidence of a dependency. 3. **Fan out only independent work.** Par