graph-new

Read the "Building a graph" and "Four ways to build a graph that blows up" sections of the graph-engineering skill first.

Requirement: $ARGUMENTS

Do this:

1. Decompose the requirement into nodes. Hard rules while decomposing:
   • Reproducible work — tests, builds, lints, formatters, migrations — must be a command node. Do not wrap it in an ai_task.
   • Every point that needs a human decision (spec choice, naming policy, deploy approval) is a user_input or a checkpoint.
   • Always put a checkpoint in front of anything irreversible (deploy, migration, force push).
   • A node whose failure should still flow downstream needs an explicit on_failure: "continue". With the default halt, a failing test kills the run before it can reach the node that fixes it.
   • Express a backwards path as a kind: "loop" edge. A cycle made of forward edges is rejected at save.
2. Check yourself before registering:
   • Does the path that decides success or failure have at least one command or condition? Branching on an ai_task's report about itself produces well-organized nonsense.
   • Did you create branches that all merge back into the same node?
   • If there is a loop, does something inside it change state, and can the exit condition ever become true?
3. Register it with graph_create. If validation rejects it, fix it and register again.
4. Summarize the nodes and edges in a table, and point out where a human will be asked to step in.

Do not run it. End with "review the graph, edit it if needed, lock it, then run /graph-engineering:graph-run <name>".