Curator Agent Capabilities

Curator Agent: Capabilities

The Curator Agent is ColabEdu’s AI-powered content ingestion pipeline. It transforms curricular documents (PDFs, Canvas/Moodle LMS packages) and open educational resources (CK-12, Khan Academy, Procomún, Europeana) into structured pedagogical specs ready for the platform.

What the Curator Does

Capability	Description
Source Discovery	Identifies canonical URLs of official curricular documents and OER/LMS repository APIs.
Document Download	Downloads PDFs and LMS packages (`.imscc` Canvas, `.mbz` Moodle) to the local specs repository.
AI Analysis & Generation	Extracts pedagogical criteria, generates slides/widgets or quality rubrics using Gemini.
Gem Routing	`curator_router` classifies each input and delegates to the specialized Gem (`lesson`, `resource`, `c2`…).
YAML Scaffolding	Writes complete OAS `InteractiveLesson`, `ResourceLearning`, or `ExerciseSpec` specs.
LMS Package Ingestion	Unzips `.imscc`/`.mbz`, uploads assets to GCS, and generates `InteractiveLesson` with slides.
OER Indexing	Indexes open source resources with a 3-dimension pedagogical `qualityScore`.
Bulk Ingestion	Spec sync via `sync_specs.sh` or `curator batch`.
Seed & Bloom	Derives new exercises from ingested seeds (source docs, rubrics) via configurable Bloom strategies. Seeds may be private; derived specs are public.

Unified Pipeline

Document (PDF / GCS URI) — or LMS Package (.imscc / .mbz) — or OER Resource (JSON/API)
        │
        ▼
POST /api/v1/curator/curate
  { source, sourceType, outputType, courseId?, pathwayId?, sandboxMode,
    ownerEmail?, shared }
        │
        ▼
CuratorWorkflow (LangGraph4j)
  ├─ [fetch_source]    → download + SHA-256 hash
  │         │
  │   sourceType == LMS_PACKAGE?
  │         ├─ YES ─► [unzip_lms]          → LmsPackageProcessor.unzipPackage()
  │         │             └─► [upload_lms_assets] → GCS + assetUrlMap
  │         │                     └─► [gather_context]
  │         └─ NO  ─► [gather_context]
  ├─ [gather_context]   → HierarchyContextResolver → CurricularContext
  ├─ [propose_content]  → Gem selected by curator_router
  │       ├─ ExerciseSpec      ← curator_c2      (exam PDFs)
  │       ├─ InteractiveLesson ← curator_lesson  (LMS packages)
  │       └─ ResourceLearning  ← curator_resource (OER APIs)
  ├─ [wait_for_review]  → headless=true → auto-approve
  ├─ [register_specs]   → IngestionService (visibility: PUBLIC | PRIVATE)

SEED & BLOOM PIPELINE (Content Derivation)
  ──────────────────────────────────────────
  curator curate --source rubric.pdf --visibility private   ← ingest as seed
       → seed ingested with visibility=PRIVATE
  curator bloom derive --seed ib.spa.textual_analysis_rubric \
                       --strategy SOURCE_DERIVATION|PROGRESSION|MISCONCEPTION_REMEDIATION
       → derived specs generated with visibility=PUBLIC + provenance.seedRef

Curricular Context Resolution

The same HierarchyContextResolver is shared with the Exercise Creation Wizard:

Explicit override (--standard / --subject)
    └─ courseId  →  CourseClient (standard, level, subject from the course)
        └─ pathwayId  →  PathwayClient (standard, level from the pathway)
            └─ referenceCode in manifest (field parsing)
                └─ generic prompt (no context)

Curator Gems

The gem.curator_router classifies each input and delegates to the specialized Gem:

Gem	Output	Triggers when
`gem.curator_router`	— (routes)	Always — classifies input and delegates
`gem.curator_c0`	`BlockRubric`	Normative curricular documents
`gem.curator_c1`	`Recipe`	Complete didactic units
`gem.curator_c2`	`ExerciseSpec`	Exam PDFs, FRQ
`gem.curator_lesson`	`InteractiveLesson`	LMS packages (`.imscc` Canvas, `.mbz` Moodle)
`gem.curator_resource`	`ResourceLearning`	OER APIs: CK-12, Khan Academy, Procomún, Europeana

Gem YAML files are stored in ce-specs/catalog/personas/curators/.

Generated Spec Types

The ColabEdu curation pipeline can generate the following OAS-compliant spec types. Each maps to a Gem and a pipeline mode:

Spec Types by Pipeline

Spec Type	OAS Kind	Gem / Mode	Description
`ExerciseSpec`	`BLOCK_CONTENT`	`gem.curator_c2` / Classic	Exercises, FRQ, text analysis with achievement bands
`BlockRubric`	`BLOCK_RUBRIC`	`gem.curator_c0` / Classic	Normative rubrics from curricular documents
`Recipe`	`RECIPE`	`gem.curator_c1` / Classic	Complete didactic units with full pedagogical scaffolding
`PathwayTemplate`	`PATHWAY_TEMPLATE`	`gem.curator_c1` / ACA	Learning pathways for a curriculum standard or theme
`InteractiveLesson`	`INTERACTIVE_LESSON`	`gem.curator_lesson` / LMS	Interactive slides from LMS packages (Canvas, Moodle)
`ResourceLearning`	`RESOURCE_LEARNING`	`gem.curator_resource` / OER	Indexed OER resources with 3-dimensional quality score
`CurriculumRequirement`	spec-as-code YAML	`CurriculumGapAnalyzer`	Requirement definition files driving gap analysis

ExerciseSpec (`C2_EXERCISES`)

Classic Curator spec: exercises, FRQ, text analysis with achievement bands. Stored as BLOCK_CONTENT kind: ExerciseSpec.

BlockRubric (`BLOCK_RUBRIC`, C0)

Generated by gem.curator_c0 from normative curricular documents (official syllabi, standards frameworks). The C0 rubric defines the competency structure that C1 recipes and C2 exercises must map to.

kind: BLOCK_RUBRIC
referenceCode: us.ap.spanish_language.hl.rubric_c0
metadata:
  standard: AP_SPANISH
  country: US
  competencies:
    - id: simulated_conversation
      weight: 0.30
    - id: email_reply
      weight: 0.25

Recipe / PathwayTemplate (`RECIPE`, C1)

Generated by gem.curator_c1 from complete didactic units. Contains full pedagogical scaffolding: objectives, activities, assessment criteria, and content sequencing. Used as a template for generating targeted C2 exercises.

kind: RECIPE
referenceCode: us.ap.spanish_language.hl.recipe_simulated_conversation
metadata:
  theme: simulated_conversation
  specType: RECIPE
  provenance:
    seedRef: us.ap.spanish_language.hl.rubric_c0

InteractiveLesson

Generated from LMS packages (Canvas Commons, Moodle Hub). Contains slides, interactive widgets, and learning objectives mapped to curricular competencies. See InteractiveLesson Reference.

ResourceLearning

Indexes OER resources with a 3-dimension qualityScore: alignment, pedagogicalRichness, accessibility. See ResourceLearning Reference.

CurriculumRequirement (spec-as-code YAML)

Not an OAS spec — a requirement definition file stored in ce-specs/catalog/requirements/. Read by the CurriculumGapAnalyzer to determine what specs are needed per standard + country.

standard: AP_SPANISH
country: US
themes:
  - id: simulated_conversation
    name: Simulated Conversation
    specTypes:
      C2_EXERCISES: 10
      INTERACTIVE_LESSON: 5
  - id: email_reply
    name: Email Reply
    specTypes:
      C2_EXERCISES: 8

These YAML files are the spec-as-code curriculum definition — they encode institutional curriculum requirements so the ACA can measure and close coverage gaps automatically.

Supported Sources

The KnownSourcesRegistry registers 14 active sources:

ID	Type	Standard	Scope
`madrid`, `mec`, `ebau.cat`	EBAU Exams	LOMLOE	🇪🇸
`collegeboard`, `apcentral`	AP Exams	AP	🇺🇸
`web`	Google CSE	—	🌐
`mec-procomun`, `mec-recursos`	OER	LOMLOE	🇪🇸
`khan-academy`, `ck12`	OER	—	🌐
`europeana-edu`	OER	—	🇪🇺
`canvas-commons`	LMS	—	🌐
`moodle-hub`, `moodle-net`	LMS	—	🌐

# List all registered sources
docs sources

# Filter by scope
docs sources --scope es
docs sources --source canvas-commons

Operation Modes

Mode 1 — Headless / Automated (CI/CD)

curator curate \
  --source exam.pdf \
  --courseId 142
# → fetch → context → propose → register (auto-approve)
# job.shared = true → visible to all CONTENT_CREATOR/TEACHER

Mode 2 — Sandbox + Manual Review

curator curate \
  --source canvas_module.imscc \
  --source-type LMS_PACKAGE \
  --output-type INTERACTIVE_LESSON \
  --sandbox \
  --output ./draft_specs/
curator validate --spec ./draft_specs/lesson.yaml --ingest

Mode 3 — LMS/OER Batch

docs discover --source canvas-commons --subject biology --output canvas.yaml
docs download --list canvas.yaml --local-dir ce-specs/sources/lms/canvas/
curator batch --manifest ce-specs/tests/lms_batch_manifest.yaml --parallel 2
curator test  --lms --all --threshold 70

Mode 4 — Shared Pedagogical Review Queue

Allows any pedagogue with CONTENT_CREATOR, TEACHER, or ADMIN role to view and approve jobs launched from the CLI, regardless of who launched them.

# Without --user: job is shared (visible to all pedagogues)
curator curate \
  --source exam.pdf \
  --referenceCode us.ap.spanish_language.hl
# → job.shared = true, job.owner_email = null
# → 🟡 "Shared" badge in the Job Command Center

# With --user: job is assigned to the specified pedagogue (only they see it)
curator curate \
  --source exam.pdf \
  --referenceCode us.ap.spanish_language.hl \
  --user pedagogue@colabedu.org
# → job.shared = false, job.owner_email = "pedagogue@colabedu.org"
# → visible only for that user in their panel

RBAC Visibility Strategy

User Type	Own Jobs	Shared Jobs (CLI)
Standard user	✅ Yes	❌ No
`CONTENT_CREATOR`	✅ Yes	✅ Yes
`TEACHER`	✅ Yes	✅ Yes
`ADMIN`	✅ Yes	✅ Yes

Mode 5 — Autonomous Curator Agent (ACA)

The ACA is a fully autonomous pipeline that orchestrates discover → fetch → curate → ingest for an entire curriculum standard in a single command, with automatic recovery from crashes and optional human-in-the-loop quality gates.

# 1. Create an autonomous plan for AP Spanish (US)
aca plan create \
  --name "AP Spanish US Sprint" \
  --country US --standard AP --level HS --subject spanish \
  --gates after_c0,after_c1
# → returns planId: a3f9b821-...

# 2. Start (fully async — ACA handles everything)
aca plan start a3f9b821

# 3. Monitor
aca plan status a3f9b821

# 4. Approve gate after_c0 (review C0 specs before generating C1/C2)
aca plan approve a3f9b821 --gate after_c0

# 5. Handle items requiring human review
aca plan list a3f9b821 --status NEEDS_REVIEW
aca plan approve a3f9b821 --item <itemId>

Plan lifecycle states:

State	Description
`DRAFT`	Created, awaiting `start`
`RUNNING`	Actively processing items
`WAITING_REVIEW`	Paused at an approval gate
`PAUSED`	Stall detected or manually paused — resume with `aca plan start`
`COMPLETED`	All items processed
`FAILED`	Critical error or cancelled

Quality gating: Items with confidence_score < threshold (default 0.75) are marked NEEDS_REVIEW and routed to the HITL queue instead of being auto-ingested.

Recovery: A scheduled job (checkAndRecoverStalledPlans) detects plans with no last_heartbeat update for > 30 minutes and sets them to PAUSED for safe restart.

Architecture

ce-svc-cli
├── CuratorCommands.java      — curate (--source-type, --output-type, --user)
├── CuratorTestCommands.java  — test (--lms), batch
├── DocDiscoveryCommands.java — discover, download, catalog, search, stats
└── LmsPackageProcessor.java  — IMSCC/MBZ unzip + GCS upload

ce-svc-ai-services
├── CuratorController         — POST /api/v1/curator/curate + /batch
├── CuratorWorkflow           — LangGraph4j (conditional LMS branch)
├── LmsPackageDiscoveryStrategy — Canvas Commons, MoodleNet APIs
├── HierarchyContextResolver  — curricular context shared with Wizard
└── IngestionService          — BLOCK_CONTENT kind: Lesson | Resource | Exercise

ce-svc-model-api (DDL v9)
├── jobs.shared       BOOLEAN DEFAULT TRUE  — shared queue visibility
├── jobs.owner_id     NULLABLE              — allows ownerless jobs
├── jobs.owner_email  VARCHAR(255) NULLABLE — CLI launcher email
├── curator_plans.*   — ACA plan state: status, last_heartbeat, config_json
└── curator_plan_items.* — per-item: source_url, confidence_score, attempts, status

ce-specs/
├── catalog/personas/curators/  — Gems: router, lesson, resource, c2
├── sources/lms/                — .imscc / .mbz packages
├── sources/oer/                — CK-12, Khan JSON resources
└── tests/lms_curator_test_manifest.yaml

Seed & Bloom — Content Derivation Engine

The Seed & Bloom engine generates original pedagogical exercises from ingested source materials without exposing proprietary content. Source docs (books, rubrics, exam PDFs) are ingested as private seeds; the derived exercises are public specs with full provenance traceability.

Architecture

ce-svc-content-engine
├── BloomStrategy (interface)
│   ├── SourceDerivationBloom    — derives exercises from a private/public source seed
│   ├── ProgressionBloom         — generates 4 difficulty tiers from any seed
│   ├── MisconceptionRemediationBloom — remediation exercises for known misconceptions
│   └── CurricularPipelineBloom  — C0 rubric → C2 exercise set (planned)
└── SeedBloomService              — dispatches to the active strategy

ce-svc-ai-services
├── BloomController               — REST: POST /api/v1/bloom/derive
└── EduWorkflowTools              — MCP: bloomContent tool

ce-svc-cli
└── CuratorCommands               — curator bloom derive

Bloom Strategies

Strategy	Trigger	Output
`SOURCE_DERIVATION`	Teacher provides a seed doc (book, rubric)	N original exercises derived from the source
`PROGRESSION`	Any exercise seed	SIMPLIFIED / STANDARD / EXTENDED / CHALLENGE tiers
`MISCONCEPTION_REMEDIATION`	Grader identifies a misconception	Targeted remediation exercise
`CURRICULAR_PIPELINE`	C0 rubric seed	Full C2 exercise set (planned)

Provenance Model

Every derived spec includes a metadata.provenance block:

metadata:
  referenceCode: ib.spa.derived.textual_analysis_ex_001
  visibility: PUBLIC
  provenance:
    seedRef: ib.spa.textual_analysis_rubric     # parent seed
    bloomStrategy: SOURCE_DERIVATION
    bloomedAt: "2025-06-11T15:00:00Z"

Entry Points

Entry Point	How
CLI	`curator bloom derive --seed <ref> --strategy SOURCE_DERIVATION`
REST	`POST /api/v1/bloom/derive`
MCP tool	`bloomContent` (Teacher Companion, Exercise Wizard)
Grading	`RemediationEventOrchestrator` → `MISCONCEPTION_REMEDIATION`

CLI Tutorial — Complete end-to-end walkthrough
Autonomous Curator Agent — ACA plan lifecycle, recovery, and quality gates
Taxonomy Reference — Bundle and standard schemas
InteractiveLesson Reference — Spec generated from LMS packages
ResourceLearning Reference — Spec generated from OER resources
Gem Reference — Curator Gems: router, lesson, resource
Seed & Bloom Engine — Content derivation concepts and YAML provenance reference

Curator Agent Capabilities

Curator Agent: Capabilities

What the Curator Does

Unified Pipeline

Curricular Context Resolution

Curator Gems

Generated Spec Types

Spec Types by Pipeline

ExerciseSpec (C2_EXERCISES)

BlockRubric (BLOCK_RUBRIC, C0)

Recipe / PathwayTemplate (RECIPE, C1)

InteractiveLesson

ResourceLearning

CurriculumRequirement (spec-as-code YAML)

Supported Sources

Operation Modes

Mode 1 — Headless / Automated (CI/CD)

Mode 2 — Sandbox + Manual Review

Mode 3 — LMS/OER Batch

Mode 4 — Shared Pedagogical Review Queue

RBAC Visibility Strategy

Mode 5 — Autonomous Curator Agent (ACA)

Architecture

Seed & Bloom — Content Derivation Engine

Architecture

Bloom Strategies

Provenance Model

Entry Points

ExerciseSpec (`C2_EXERCISES`)

BlockRubric (`BLOCK_RUBRIC`, C0)

Recipe / PathwayTemplate (`RECIPE`, C1)