Blueprint

Step 3 of the ai-env task workflow. After you've oriented around the problem, this runs one brainstorming round, reads the template from your repo, and writes a structured plan file to ai-workspace/plans with all the required fields (branch, target, threat model, concrete steps). It auto-detects which branch you forked from using git reflog and commits the plan before handing off to review. The threat model table is smart: advisory for internal stuff, adversarial for anything touching auth or input validation. One thing I appreciate is the hard guardrail against skipping the commit step, since losing brainstorming work to a session crash is annoying. Does one job, doesn't try to review or implement.