A local, domain-general tutoring framework that uses small interactive web assistants as the primary learning surface. It combines immutable curriculum DAGs, conservative learning-evidence projections, adaptive orchestration, private portable workspaces, and a human-owned acceptance workflow.
This repository is a runnable development baseline, not a released unified application. Onboarding/profile/curriculum services and the progress explorer currently need agent or command assistance. Real learner transcript collection, deployment analytics, and public learner-derived issue writes are disabled pending deployment-specific human authority.
Requirements: Node.js 22 or newer. Python 3 is additionally required for the human-acceptance checkpoint helper. The commands below were observed on Windows PowerShell; the automated fixtures are portable, but macOS/Linux have not received human acceptance.
Run all tests (no package install is currently required):
npm testBuild the three synthetic assistant examples and start the loopback-only host:
node examples/interactive-assistants/build.js config examples/interactive-assistants/configs/science-change.json
node examples/interactive-assistants/build.js config examples/interactive-assistants/configs/music-order.json
node examples/interactive-assistants/build.js app examples/interactive-assistants/sources/adult-math-app
node src/interactive-assistant-harness/server.js 41739Open the printed 127.0.0.1 URL with ?card=science_change, ?card=music_order, or ?card=adult_math_recall. These are small science, music, and mathematics fixtures—not a subject-specific architecture or a complete game.
Invoke $run-human-acceptance in Codex for human-only evaluation of UX, aesthetics, comprehension, affordances, accessibility experience, and journey coherence. The facilitator runs technical checks itself, states expectations before each action, accepts feedback at any time, and turns approved actionable feedback into ZzzOps goals before the human-owned release decision.
The visual/non-visual curriculum explorer is runnable with command assistance. A browser-native synthetic tutoring session is also runnable through the tutor skill: the learner opens one local URL, answers the tutor's opening question, and receives validated generated activities inline without returning to the Codex UI. Activities reveal one stage at a time, supporting sources stay under optional More info, and the tutor chooses and explains the next learning step. Session chat/history is memory-only by default, and the harness does not call the OpenAI API directly. Unified non-agent onboarding and real learner/public-feedback scenarios remain unavailable for the reasons recorded in the plan.
- Implementation roadmap: current component/goal traceability and authority gates.
- Learning-design research: evidence basis and context-efficient design handoff.
- Interactive-assistant design and runnable examples.
- Curriculum DAG and progress design and explorer implementation.
- Portable workspace design, repository, and lifecycle operations.
- Safeguarding and privacy and feedback boundary.
Use opaque synthetic users and test_only: true disposable workspaces. Assistants receive a minimized activity slice—not identity, absolute paths, full profiles/history, credentials, or raw chat. Do not aim lifecycle deletion at this repository or any real workspace. External writes and real learner data require explicit reviewed authority that this local baseline does not grant.