Ten stages take a requirement from intake to a merged pull request. Two of them are approval gates that can't be skipped — not because the model isn't confident, but because DreamCatcher doesn't trust confidence as a substitute for a person saying yes.
Every significant AI output — a solution recommendation, a design document — requires explicit human approval before the pipeline continues, not just a review after the fact. Even when there's only one viable option, a person has to say so.
1–3 ranked options are presented, grounded in your actual configuration documentation. A person selects, edits, or requests a different approach before any design work begins.
A structured design document is generated from the approved option. It's reviewed and explicitly accepted — or sent back with changes — before a single line of code is written.
Generated code is checked against upgrade-safety and cloud-compliance patterns before it ever reaches a developer for review — correct _Ext suffixes, Studio-only edits, no base metadata changes. The agent iterates against these checks until the code is clean, so what lands in the pull request is already held to the same bar a senior Guidewire developer would apply.
Each approval gate writes a permanent version record, linked back to the original requirement and forward to every revision. If a requirement changes mid-flight, the history of what was proposed, what was approved, and what changed stays intact — a clean trail for anyone who needs to know how a piece of Guidewire code came to be.