Agent catalogue

Eight agents. One audit trail. One human-in-the-loop discipline.

Every XueDao Lao Shi agent runs on the ClairVAS S240–S247 line, grounded in your own course materials via S241. Each one has a defined scope, a defined hand-off back to a human (teacher, grader, or learner), and a defined record in the audit log. Mix and match; bring your own LMS.

S240

AI Tutor (Lao Shi)

Socratic · Bloom-tagged · impartiality wall

Student-facing tutor that teaches by asking questions, not by giving away answers. Tags every exchange with the Bloom level being exercised, infers the suspected gap, and refuses to write the student's assignment for them. The pedagogy your teachers want, without the supervision overhead.

  • Returns: tutor_response, bloom_level, suspected_gap, suggested_followup, mastery_delta, refused_to_answer_directly.
  • Impartiality wall — refuses to do the student's homework even if asked directly.
  • Available via the lao_shi_chat workflow on ClairVAS — single entry point for Moodle integrations.
S241

Course Retriever

Course-scoped RAG · powers S242 / S244 / S245 / S246 / S247

Course-scoped RAG adapter over the ClairVAS B004 Knowledge Engine. Indexes your Moodle PDFs, lesson notes, and assessment rubrics; serves every other agent on this page their grounding passages. The agent your teachers don't see but everyone depends on.

  • Sits on top of ClairVAS B004 Knowledge Engine; one Chroma collection per course.
  • Used as a sub-agent by S242 / S244 / S245 / S246 / S247 — your tenant only pays for it once.
  • Course-scoped — passages from course A never bleed into a chat about course B.
S242

Free Chat (grounded)

Course Q&A · cited · Bloom-aware

Looser-format student chat for "what does this term mean," "give me an example," "explain again differently." Grounded in S241 so answers cite a specific page in the actual course material — no general-internet drift.

  • For the "I just want a clean answer right now" student moments.
  • Every answer cites the originating PDF + page so the student can verify.
  • Bloom-aware so the response complexity matches the question complexity.
S243

Course Indexer

Moodle PDF ingest · GCS-backed · idempotent

Watches a Moodle course folder for new PDFs and indexes them into S241's retrieval store. Teachers upload normally; the AI catches up in the background. The same uploaded original is preserved on GCS so the audit log can prove what was actually indexed.

  • POST /api/v1/courses/index — pulls Moodle course folders, chunks PDFs, embeds, persists.
  • Original GCS object preserved so the audit log can prove what was actually indexed.
  • Idempotent — re-running on the same course refreshes only what changed.
S244

Design Guidance

Teacher copilot · curriculum-aware

Helps a teacher plan a lesson, scope an assessment, or sequence a unit. Grounded in the same course passages S241 serves to students, so the design fits the materials students will actually see. Teacher-in-the-loop: drafts, never publishes.

  • Teacher-side complement to S240 — same retrieval store, different reasoning scaffold.
  • Returns draft lesson plans + assessment rubrics + sequencing suggestions.
  • Teacher-in-the-loop: drafts, never publishes; signs off via the LMS's normal flow.
S245

Skills Coach

Next-best-skill · target role / cert · effort estimate

Recommends the single next-best skill for a learner given their recent activity, target role / cert, and the tenant's catalogue. Bloom-tags the target level and estimates effort. Grounded in catalogue passages via S241.

  • Pulls recent activity from the LMS, matches against catalogue + target role, returns one recommendation.
  • Effort estimate in hours; Bloom level the next skill is targeting.
  • Cites the catalogue passages it used so the recommendation isn't a black box.
S246

Assessment Grader

Essay / short-answer · rubric · human-review flag

Grades short-answer and essay assessments against the course's authored materials and an optional rubric. Returns score, band, Bloom level demonstrated, strengths, improvements, and a human-review flag for borderline cases. Human grader retains the final word.

  • Returns: score, band, Bloom level demonstrated, strengths, improvements, rubric breakdown.
  • human_review_required flag for borderline cases — never auto-publishes a grade.
  • Grounded in the course's own materials, not the model's general knowledge of the subject.
S247

Content Generator

Lesson plan · Bloom-spread quiz · source-cited

Drafts a lesson plan plus a Bloom-spread quiz from a topic, target audience, and the course's existing materials. Source-cites each quiz item and flags gaps honestly when the corpus is thin. Instructor-in-the-loop: drafts, never publishes.

  • Inputs: topic, target audience, hours available, optional Bloom-spread target.
  • Outputs: lesson plan, quiz items each cited to a source passage, gap flags when corpus is thin.
  • Instructor-in-the-loop: drafts, never publishes; flagged corpus gaps surface honestly.

Pick the agents that solve your bottleneck.

Most schools start with S240 + S241 + S243 (Tutor + Retriever + Indexer) and add S246 / S247 once their teachers want grading assist or content drafts. The S241 / S242 path covers free-form Q&A from day one.