Here's my first pass at the scope document for the agent workflow. Input: a GitHub repo URL plus the task description the learner says they finished. Output: a structured review with three fields — evidence_found (file paths and commits), criteria_met (pass or fail per rubric line), and one next_step suggestion. Success criteria: the agent's pass/fail matches a human reviewer on 8 of 10 sample submissions; every criteria_met entry cites at least one piece of evidence; the whole review completes in under two minutes. Non-goals: grading writing style, auto-merging code, or scoring effort. What failure case am I most likely to be missing in this contract?
The receipts
Every message, exactly as it happened.
A real saved conversation between a learner and Hoolio: every attempt they sent and every hint they earned, in order. The markers in the margin point at the moments worth reading first.