Imported from JimmyCII/nova-teachingassistant (AGENTS.md). Install upstream with npx skills add JimmyCII/nova-teachingassistant. Copyright stays with the author.

AGENTS.md — Working Agreement for AI Agents in This Repo

This is the canonical instruction file for any AI coding agent (Claude Code, Codex, etc.) working in TeacherMind. CLAUDE.md and CODEX_BOOTSTRAP.md point here.

What this project is

TeacherMind — an AI concierge agent (Gemini + Google ADK, Cloud Run) that reduces a 6th-grade Arizona math teacher's admin burden. Canonical spec: CAPSTONE_SPEC.md. Scope/non-goals/acceptance: docs/requirements.md. The agent persona is Nova (built in teacher Karrie's spirit — not a clone). Kaggle Concierge-Agents capstone, due 2026-07-06.

Follow the playbook

Operate per C:\Users\jimco\Dev\Project-Playbook\PROJECT-PLAYBOOK.md. Key rules:

• Plan before code. Lock the data model / core logic before any UI. No app code until requirements + data model are settled.
• Branch discipline. git init + initial commit exists. Do all work on a feature branch per increment (feat/v1-increment-N). Never build or commit on main without explicit consent. Keep main releasable.
• Checkpoint loop: branch → build → verify (tests + type-check + lint) → fix inline → commit (feat(scope): / fix(scope): / docs(scope):) → merge after acceptance → one-line report.
• Dependencies: permissive licenses only (MIT/BSD/Apache-2.0); none new without a yes; record in DEPENDENCIES.md.
• Destructive ops: preview first; never rm -rf, git push -f, or git reset --hard without approval.

⚠ PII / FERPA — non-negotiable

This project derives from a real teacher's private files. Never commit student PII. The Karrie research under docs/karrie_profile/, docs/karrie_profile.md, docs/karrie_exploration_prompt.md, and any *_raw.csv are gitignored / local-only because they reference real student-name filenames and IEP/SPED/SEI paths. Do not un-ignore or commit them until fully PII-scrubbed (student names → [STUDENT], IEP/SPED/SEI references removed). When in doubt, treat any individual student data as untouchable: no names, IDs, grades, attendance, IEP/medical.

Multi-agent / model routing (cost tiers)

Stay model-agnostic; route by capability tier (Project-Playbook §4): Low (Haiku-class) for volume/scaffolding, Medium (Sonnet-class) for normal feature work, High (Opus-class) only for architecture/security/high-impact + final pre-commit review. Default to the lowest sufficient tier; record decisions in agents/logs/routing-audit-log.md.

Where things live

• agents/prompts/ — multi-agent role prompts · agents/logs/ — routing + decision logs
• prompts/ — session/build prompts (start with prompts/session-1-kickoff.md)
• docs/ — requirements + (local-only) persona research · agent/ — ADK agent code (no src/) · tests/ — tests