What problem does it solve?
Launching a Claude Managed Agent spans multiple sessions and phases (interview, staging, grading, scheduling), and it is easy to lose track of where you are or what shape the work should take. This Skill keeps a single session goal on disk, deterministically routes it to the right phase sub-skill, and compiles the goal plus phase into a concrete execution plan.
Core Features & Use Cases
- Deterministic goal routing: goal_router.py scores the goal against keyword signals and returns exit codes (0 route / 3 ask / 4 refuse) so the orchestrator never guesses a lane.
- Persistent goal state: goal_state.py owns ./my-agent/goal.json with init, set, status, and advance commands across the five phases from interview to wrap-up.
- Loop compilation: loop_compiler.py emits a plan.v1 execution shape — single-pass workflow, bounded grade-iterate loop (max_iterations clamped 1-20), or recurring cron deployment loop with optional nested self-grading outcomes.
- Use Case: A user says "grade my support-triage agent until the rubric passes, then run it every morning." The router selects the grade-iterate lane, the compiler emits a bounded loop capped at 5 iterations, and the phase later advances to run-without-you with a 9am cron schedule.
Quick Start
Ask the agent to set up a managed agent goal by saying something like "init a goal to triage my inbox every morning and route me to the right phase."
