CF Worker: LeadIntakeAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Normalizes inbound school leads from operator imports, CSV uploads, or web forms by cleaning names, emails, phones, board types, and website domains. Computes a completeness score and dedupe key for each lead, segments schools into premium/standard/emerging/unknown tiers, and surfaces operator-reviewable enrichment research suggestions where data is missing.

Tools: read_crm_leads  ·  read_campaign_status  ·  get_school_status  ·  read_registration_exceptions  ·  render_table

Outcome: CrmAgentDecision with normalized leads array, importSummary (totalLeads, averageCompletenessScore, warningCount, suggestedNextStep), webEnrichmentBoundary statement, confidenceScore, and full CrmDecisionLogShape. humanReviewRequired is true when warningCount > 0.

Guardrails: Hard-blocked list (all CRM agents): post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. No arbitrary outbound web fetch. mutatesBusinessTruth: false; externalCommunicationSent: false.

CF Worker: ReplyClassificationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Classifies inbound school coordinator replies (email, WhatsApp, phone notes, web forms) into one of 7 CRM intents: interested, not_interested, need_callback, need_pricing, need_brochure, registration_sent, or unclear. Scores signal strength via keyword matching, assigns SLA priority, and recommends the next operator action.

Tools: read_crm_leads  ·  read_campaign_status  ·  get_school_status  ·  read_registration_exceptions  ·  render_card

Outcome: CrmAgentDecision with channel, intent (7 ReplyIntent values), displayIntent, matchedSignals, nextAction, slaPriority, suggestedResponseTone, confidenceScore (0.52-0.94), and CrmDecisionLogShape.

Guardrails: Advisory-only; classification requires operator confirmation before updating CRM intent fields. mutatesBusinessTruth: false; externalCommunicationSent: false.

CF Worker: SchoolIntelligenceAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Builds a comprehensive intelligence profile for a specific school computing reliability score (0-100), conversion likelihood (0-100), payment behavior label, participation pattern label, and CRM risk flags from historical participation, payment, completion, certificate delivery, and reply intent data. humanReviewRequired is always forced true.

Tools: read_crm_leads  ·  read_campaign_status  ·  get_school_status  ·  read_registration_exceptions  ·  render_card

Outcome: CrmAgentDecision with reliabilityScore (0-100), conversionLikelihood (0-100), paymentBehavior label, participationPattern label, riskFlags array, rationale array (two sentences), confidenceScore, humanReviewRequired: true.

Guardrails: humanReviewRequired forced true unconditionally. schoolId redacted in all decision log outputs. mutatesBusinessTruth: false; externalCommunicationSent: false.

CF Worker: OpportunityScoringAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Scores a school's conversion opportunity on a 0-100 scale and assigns a pipeline stage (hot/warm/nurture/cold) based on reply intent, reliability score, previous registrations, campaign touches, days since last touch, coordinator engagement level, and exam window urgency.

Tools: read_crm_leads  ·  read_campaign_status  ·  get_school_status  ·  read_registration_exceptions  ·  render_card

Outcome: CrmAgentDecision with conversionScore (0-100), stage (hot/warm/nurture/cold), reasons array, recommendedAction, freshness label, confidenceScore, humanReviewRequired: true.

Guardrails: humanReviewRequired forced true — stage must be operator-reviewed before queue re-prioritization. mutatesBusinessTruth: false; externalCommunicationSent: false.

CF Worker: FollowUpAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Drafts a follow-up plan for a school lead based on reply intent, no-response count, last touch timestamp, preferred channel, and school segment. Calculates cadence gap (1-7 days), selects template key, and flags whether plan should pause for operator review. Never sends any message.

Tools: read_crm_leads  ·  read_campaign_status  ·  get_school_status  ·  read_registration_exceptions  ·  draft_follow_up  ·  render_card

Outcome: CrmAgentDecision with follow-up plan (recommendedAction, cadence, nextTouchAt, templateKey, pauseForReview, channel). draft_follow_up carries requiresHumanApproval: true. humanReviewRequired: true always.

Guardrails: draft_follow_up requires human approval before any message is sent. not_interested schools automatically suppressed. After 4+ no-responses, plan is paused and routed to operator review queue. mutatesBusinessTruth: false; externalCommunicationSent: false.

CF Worker: CampaignAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Designs outreach campaign sequences, channel mixes, target segment logic, and risk control plans for awareness, registration, payment, or reengagement goals across premium, standard, emerging, or mixed school segments. Incorporates exam window urgency, open/reply rate signals, and PII minimization rules.

Tools: read_crm_leads  ·  read_campaign_status  ·  get_school_status  ·  read_registration_exceptions  ·  draft_follow_up  ·  render_table

Outcome: CrmAgentDecision with targetSegmentLogic, channelMix array, 3-step sequence (each requiresApprovalBeforeSend: true), riskControls array, humanReviewRequired: true.

Guardrails: Every sequence step requiresApprovalBeforeSend: true — no message dispatched. Exclusion list (withdrawn_consent, do_not_contact, active_payment_dispute) must be applied. Student/parent PII explicitly excluded from campaign prompts. mutatesBusinessTruth: false.

CF Worker: MeetingSchedulerAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Suggests up to 3 ranked callback time slots by matching availability signals (day-of-week, start/end time, timezone, confidence) against caller preference (morning/afternoon/any) and blackout dates. Never creates calendar events, sends invites, or books meetings.

Tools: read_crm_leads  ·  read_campaign_status  ·  get_school_status  ·  read_registration_exceptions  ·  draft_callback_schedule  ·  render_card

Outcome: CrmAgentDecision with suggestedSlots array (up to 3, each with date/dayOfWeek/startTime/endTime/timezone/confidence/reason), fallbackAction, schedulingBoundary confirming no calendar event is created. humanReviewRequired: true.

Guardrails: No calendar event ever created. draft_callback_schedule requiresHumanApproval: true. humanReviewRequired forced true. schoolId redacted. mutatesBusinessTruth: false; externalCommunicationSent: false.

CF Worker: SchoolMappingAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Fuzzy-matches an imported school name variant to a list of canonical school records (up to 500) using Jaccard token similarity and Dice coefficient bigram matching, boosted by city and state agreement. Returns ranked list of up to 5 candidate matches. Never links or merges records directly.

Tools: read_crm_leads  ·  read_campaign_status  ·  get_school_status  ·  read_registration_exceptions  ·  read_school_match_candidates  ·  render_table

Outcome: CrmAgentDecision with candidateName (normalized), bestMatch object or null, topMatches (up to 5), needsHumanReview: true always, mappingBoundary statement confirming no records are linked.

Guardrails: mappingBoundary embedded in every output: agent recommends only, never links records. needsHumanReview hardcoded true. Deterministic dedupe policy must confirm no conflict before any canonical link. mutatesBusinessTruth: false; arbitraryWebFetchBlocked: true.

CF Worker: CampaignPlanningAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Designs multi-step outreach campaign sequences for school registration, awareness, payment, and re-engagement goals. Used by the operator portal (phase2 CRM) to produce structured campaign plans with channel mix, template sequencing, risk controls, and segment-targeting logic before any message is scheduled.

Tools: read_crm_leads  ·  read_campaign_status  ·  get_school_status  ·  read_registration_exceptions  ·  draft_follow_up  ·  render_table

Outcome: Structured campaign sequence object with target-segment logic, channel mix, step-by-step template keys, risk controls, and recommendation string — all pending operator approval before any message is dispatched.

Guardrails: All campaign drafts require operator approval before send. Student and parent PII suppressed from all campaign prompts. External messages never sent by agent. Advisory output only.

CF Worker: LeadImportValidationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Validates and quality-checks bulk lead imports before they enter the CRM pipeline, identifying missing fields, formatting errors, duplicate keys, and low-confidence records. Returns a validation report with per-row warnings and a recommended next step.

Tools: read_crm_leads  ·  get_school_status  ·  read_registration_exceptions  ·  render_table

Outcome: Validation report containing normalized lead rows, per-row completeness scores and warning codes, duplicate-risk flags, and recommended next step (operator_quality_review or dedupe_and_segment).

Guardrails: Advisory output only — CRM records are not written by this agent. Duplicate resolution requires confirmed human action. No arbitrary outbound web fetch performed.

CF Worker: ReplyFollowupAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Combines reply classification intent with follow-up plan generation for inbound school replies, producing a tailored next-touch recommendation that respects channel preference, cadence rules, and no-response chain limits. Used immediately after a reply is classified.

Tools: read_crm_leads  ·  read_campaign_status  ·  get_school_status  ·  draft_follow_up  ·  render_card

Outcome: Follow-up plan containing recommended action, cadence label, next-touch ISO timestamp, template key, channel, and pauseForReview flag — all pending operator confirmation before any outbound message is drafted or sent.

Guardrails: No external message is sent — draft only. not_interested schools are automatically suppressed from follow-up. Requires human approval before any outbound communication.

CF Worker: RegistrationExceptionAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Triages open registration exceptions by severity, groups them by entity and exception type, and recommends resolution priorities and operator actions. Surfaces critical and high-severity issues that block downstream invoice and exam workflows.

Tools: read_registration_exceptions  ·  read_crm_leads  ·  get_school_status  ·  render_table  ·  render_card

Outcome: Triage report listing open exceptions grouped by severity with recommended resolution actions, blocking-status flags, and priority ordering for operator review.

Guardrails: Exception resolution performed through deterministic platform workflows, not by this agent. Advisory output only — exception status is not updated by this agent.

CF Worker: InvoiceBatchReadinessAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Assesses whether a school or cohort of schools is ready for invoice batch generation by checking registration completeness, exception clearance, and payment setup. Operates at the phase2/phase3 boundary to prevent premature invoice creation.

Tools: read_crm_leads  ·  read_registration_exceptions  ·  get_school_status  ·  render_table

Outcome: Per-school readiness report with invoice-batch-ready boolean, list of blockers, and recommended operator actions — all advisory, with no invoice records created by the agent.

Guardrails: Invoice creation is performed by deterministic finance services, not by this agent. Advisory output only — no financial records are written.

CF Worker: SchoolOnboardingAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Guides operators through the school onboarding checklist for new schools joining the Finverse Academy Olympiad platform, verifying profile data, coordinator contact, board details, and consent records are complete before school activation.

Tools: read_crm_leads  ·  get_school_status  ·  read_registration_exceptions  ·  render_card  ·  render_table

Outcome: Onboarding readiness report with per-field completion status, a list of missing or unverified data points, and recommended next step — advisory, with school activation performed through the deterministic school-status workflow.

Guardrails: School status activation is performed by deterministic services, not by this agent. Consent records must be verified by a human before proceeding.

CF Worker: RegistrationConsentAuditAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Audits the consent and data-collection basis for school registration batches, verifying that coordinator consent has been captured, that student-data collection is within declared scope, and that no consent-gated fields are populated without a valid basis. Supports DPDP Act obligations.

Tools: read_crm_leads  ·  get_school_status  ·  read_registration_exceptions  ·  render_table

Outcome: Consent-audit report listing per-school consent status, non-compliant or missing consent records, and recommended operator actions — advisory only, with no consent records written by this agent.

Guardrails: Consent records are written only through service-role RPCs, not by this agent. Advisory output only — findings must be actioned by a human operator.

CF Worker: RegistrationDataMinimizationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Reviews registration intake batches and CRM records to identify fields that exceed the minimum data necessary for Olympiad operations, flagging excessive collection of student or school data in support of DPDP Act data-minimization obligations. All findings advisory.

Tools: read_crm_leads  ·  get_school_status  ·  read_registration_exceptions  ·  render_table

Outcome: Data-minimization report identifying excess fields per school or batch, with recommended operator actions — advisory only, with no data deleted or redacted by this agent.

Guardrails: Data deletion or redaction must go through approved platform workflows, not this agent. Advisory output only — human confirmation required before any data changes.

CF Worker: OutreachQualityAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Reviews outbound campaign messages and follow-up drafts for quality, tone, compliance with consent and opt-out obligations, and alignment with Finverse Academy brand guidelines before messages are approved for dispatch.

Tools: read_campaign_status  ·  read_crm_leads  ·  render_card  ·  render_table

Outcome: Quality-review report with dimension scores (tone, clarity, CTA, consent-respect, opt-out presence, PII exposure), flagged issues, improvement suggestions, and overall pass/flag decision — all advisory, with message dispatch gated on human approval.

Guardrails: No message is sent by this agent — dispatch requires human approval. Student and parent PII suppressed from all quality-review prompts. Advisory output only.

CF Worker: SchoolProfileEnrichmentAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Enriches existing school CRM profiles with additional metadata derived from CRM signals, campaign responses, and supplied enrichment hints, producing operator-reviewable research plans and field-update suggestions. No enrichment is written to the database without operator confirmation.

Tools: read_crm_leads  ·  get_school_status  ·  read_campaign_status  ·  render_card

Outcome: Enrichment suggestion set with per-field recommended values, confidence scores, operator-research query strings, and requiredReview: true flags — no CRM profile is updated until a human operator confirms each suggestion.

Guardrails: No arbitrary outbound web fetch — enrichment suggestions are research plans for human action. All profile updates require operator confirmation through deterministic school-update workflows. Advisory output only.

CF Worker: RegistrationParsingAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Parses uploaded CSV and XLSX registration intake files and maps raw column headers to canonical schema fields (studentName, email, phone, grade, schoolName, rollNumber, subject, dob, gender, subjectCount) using dice-coefficient fuzzy matching. Scores each header mapping by confidence, identifies unmappable or low-confidence columns, and surfaces missing critical fields.

Tools: get_registration_errors  ·  render_table

Outcome: AgentDecision (decisionType: column_mapping_suggestion) containing per-header mappings array (rawHeader, canonicalField, confidence, alternativeSuggestions), unmappedHeaders list, missingCanonicalFields list, and a mappingSummary (totalHeaders, mappedCount, unmappedCount, averageConfidence, maxRows). Sensitive text is redacted in summary.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Agent maps column headers only; no row data is mutated. All output is advisory, logged, and schema-validated.

CF Worker: DeduplicationAssistantAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Scores a candidate registration row against up to 500 existing rows to detect probable duplicate student entries using normalized name similarity (dice-coefficient), exact email/phone matching, school name similarity, and grade confirmation. Returns confidence-ranked match candidates and a recommendation of accept, manual_review, or merge. Never merges or deletes canonical records automatically.

Tools: get_registration_errors  ·  render_card

Outcome: AgentDecision (decisionType: duplicate_recommendation) with recommendation (accept/manual_review/merge), topMatches array (up to 5 entries with index/studentName/schoolName/confidence/matchReasons), hasDuplicate boolean, forceHumanReview flag, confidenceScore, and dedupeBoundary statement. Student names redacted in summary.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Deduplication never merges or deletes canonical registration records. forceHumanReview is set to true for any non-accept recommendation. NEVER bypass RLS.

CF Worker: RegistrationValidationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Validates up to 200 parsed registration rows per invocation against a fixed rule set covering required-field completeness (studentName, grade, schoolName), numeric grade range (1-12), email format, phone format, and school name minimum length. Classifies each violation as blocker or warning and aggregates counts.

Tools: get_registration_errors  ·  render_table

Outcome: AgentDecision (decisionType: exception_priority) with rowResults array (rowNumber, isValid, blockerCount, warningCount, errors per violation), summary object (totalRows, validRows, invalidRows, totalBlockers, totalWarnings, schoolTier), validationBoundary statement, and recommendation.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Agent classifies errors only; no rows are deleted or corrected automatically. NEVER bypass RLS.

CF Worker: FeeCalculationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Calculates advisory invoice line items for up to 200 students per invocation by applying base fee per subject, school-tier discounts (emerging: 10%, premium: 5%, standard: 0%), special-category discounts (scholarship: 15%, staff_ward: 10%, sibling: 5%), a configurable operator-level discount, and an 18% tax rate by default. All outputs are recommendations only.

Tools: get_invoice_status  ·  render_table

Outcome: AgentDecision (decisionType: payment_match_recommendation) with lineItems array (redacted studentName, subjectCount, schoolTier, specialCategory, grossFee, discountPercent, discountAmount, netFee, taxPercent, taxAmount, totalFee) and summary object (totalRows, grandTotal, totalNetFee, totalTaxAmount, totalDiscountAmount). Total discount clamped to 50% max. feeBoundary notice included.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Fee calculation outputs are recommendations only; invoice records are created by approved platform workflows after operator confirmation. NEVER bypass RLS.

CF Worker: ExceptionResolutionAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Triages individual registration row exceptions across 5 categories (missing_required_field, invalid_format, duplicate_detected, fee_mismatch, school_not_found) and produces a structured plain-English resolution guide with step-by-step coordinator instructions and a corrective action code. Flags every exception for mandatory human review.

Tools: get_registration_errors  ·  render_card

Outcome: AgentDecision (decisionType: exception_priority) with exceptionType, rowNumber, fieldName, plainEnglishExplanation, coordinatorInstructions (4 steps), correctiveAction code, optionally a redacted suggestedValue, forceHumanReview always true, and resolutionBoundary statement. No records are corrected automatically.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. forceHumanReview is always set to true for every exception. Agent provides resolution guidance only; no records are corrected automatically. suggestedValue is always redacted. NEVER bypass RLS.

CF Worker: SchoolOperationsAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Provides a conversational interface and school-level registration health dashboard for school coordinators, allowing them to complete all registration, invoicing, exam-material, and results tasks through conversation. Computes readiness status (invoice_ready, needs_validation, or blocked) from validated-student ratio and pending exception count. Exposes a broad toolset covering registration summaries, invoice management, payment initiation, answer sheet processing, results, certificates, and team management.

Tools: get_registration_summary  ·  get_invoice_status  ·  get_exam_materials_status  ·  create_invite_link  ·  send_invite_emails  ·  get_results_summary  ·  trigger_certificate_generation  ·  get_invoice_download_url  ·  initiate_payment  ·  process_excel_registration  ·  process_payment_receipt  ·  process_answer_sheets  ·  get_question_paper_download  ·  get_answer_sheet_download  ·  create_invigilator  ·  get_certificate_download  ·  get_student_ranking  ·  get_team_members  ·  get_exam_schedule  ·  request_reevaluation

Outcome: AgentDecision (decisionType: school_health_score) with redacted schoolId, readiness status (invoice_ready/needs_validation/blocked), completionPercent, pendingExceptionCount, totalRegisteredStudents, validatedStudents, invoiceReadyBatches, nextSteps array, forceHumanReview flag, operationsBoundary statement.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Agent provides advisory summary only; no school state is mutated. All data scoped to the requesting school only — cross-school data access not permitted. Approval-required tools (initiate_payment, create_invigilator, trigger_certificate_generation, request_reevaluation, process_answer_sheets) must be described to coordinator before execution. NEVER bypass RLS.

CF Worker: InvoiceAssistantAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Reviews invoice line items for amount anomalies, over-billing risks, and subject-count mismatches. Computes a variance percentage between actual and expected invoice amounts and flags invoices where the variance exceeds 10%. Used by finance operators and ops admins before any invoice is approved.

Tools: get_invoice_status  ·  render_card

Outcome: AgentDecision with invoice summary, variancePercent, hasAnomaly flag, confidence score (0.72 for anomaly, 0.88 for clean), humanReviewRequired=true, and approvalBoundary statement confirming the agent cannot approve invoices.

Guardrails: Hard-blocked: post_ledger_entry, approve_refunds, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory-only: agent cannot modify invoice amounts, post ledger entries, or trigger approval workflows directly. forceHumanReview is always true. Approval roles required: ops_admin, finance_user.

CF Worker: PaymentReconciliationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Performs fuzzy matching of incoming payments against open invoice candidates using amount, UTR reference, and payer-name signals. Returns ranked match candidates with per-signal scores and a composite confidence score. Used by finance operators before any ledger posting is confirmed.

Tools: get_invoice_status  ·  render_table

Outcome: AgentDecision with top-5 ranked invoice match candidates, per-signal scores (amount 60%, UTR 30%, payer 10%), bestMatchConfidence, redacted PII flags (utrProvided, payerNameProvided), and reconciliationBoundary stating no ledger entry is performed.

Guardrails: Hard-blocked: post_ledger_entry, approve_refunds, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory-only: agent cannot modify invoice amounts, post ledger entries, or trigger approval workflows directly. forceHumanReview is always true. Approval roles required: ops_admin, finance_user.

CF Worker: FinanceCollectionsAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Prioritises overdue school balances by computing a risk score from days-overdue, invoice amount, and contact-attempt history. Assigns escalation tiers (standard/elevated/critical) and recommends collection action templates for operator review. Drafts collection plans without sending any external messages.

Tools: get_invoice_status  ·  render_card

Outcome: AgentDecision with riskScore (0-100), escalationTier (standard/elevated/critical), templateKey for recommended communication template, schoolId (redacted in logs), daysOverdue, contactAttempts, and collectionsBoundary stating that sending requires human approval.

Guardrails: Hard-blocked: post_ledger_entry, approve_refunds, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory-only: agent cannot modify invoice amounts, post ledger entries, or trigger approval workflows directly. forceHumanReview is always true. Approval roles required: ops_admin, finance_user.

CF Worker: RefundReviewAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Evaluates refund eligibility by checking whether the request is within the 180-day window and does not exceed the original payment amount. Produces an eligibility determination with flags for partial-refund pro-rata checks and policy exceptions.

Tools: get_invoice_status  ·  render_card

Outcome: AgentDecision with eligible boolean, eligibilityScore (0.4 or 0.8), isPartialRefund flag, policy violation flags array, refundRequestId, amounts, daysSincePayment, and approvalBoundary confirming the agent cannot approve or process refunds.

Guardrails: Hard-blocked: post_ledger_entry, approve_refunds, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory-only: agent cannot modify invoice amounts, post ledger entries, or trigger approval workflows directly. forceHumanReview is always true. Approval roles required: ops_admin, finance_user.

CF Worker: AccountingExplanationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Translates individual ledger entries into plain-English explanations for finance coordinators and auditors. Maps account codes to asset/liability/revenue/expense categories and generates a human-readable description of the entry's purpose. Never modifies ledger records.

Tools: get_invoice_status  ·  render_card

Outcome: AgentDecision with entryId, entryType (debit/credit), amount, accountCode, accountCategory, entryDate, relatedInvoiceId, plainEnglishExplanation string, confidenceScore (0.86), and ledgerBoundary confirming no posting or modification is performed.

Guardrails: Hard-blocked: post_ledger_entry, approve_refunds, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory-only: agent cannot modify invoice amounts, post ledger entries, or trigger approval workflows directly. forceHumanReview is always true. Approval roles required: ops_admin, finance_user.

CF Worker: CaReviewAssistant extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Summarises ledger anomaly flags, GST mismatches, and suspicious rounding patterns for a given review period to assist chartered accountants before period close. Computes anomaly rates, identifies top anomaly types by frequency, and raises priority flags for CA sign-off.

Tools: render_table  ·  render_card

Outcome: AgentDecision with reviewPeriod, totalEntries, anomalyFlagCount, anomalyRate, gstMismatchCount, roundNumberCount, priorityFlags list, top-5 anomaly type frequency table, confidence score (0.55-0.85), and reviewBoundary stating professional sign-off is required before corrective action.

Guardrails: Hard-blocked: post_ledger_entry, approve_refunds, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory-only: agent cannot modify invoice amounts, post ledger entries, or trigger approval workflows directly. forceHumanReview is always true. Approval roles required: ops_admin, finance_user.

CF Worker: FinancialReviewAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Produces high-level financial health summaries for CEO-level review dashboards by comparing invoice and payment aggregates for a period, computing reconciliation gaps, and detecting duplicate-amount risks. Flags significant gaps (>5% of invoice total) and duplicate payment patterns.

Tools: get_invoice_status  ·  render_table

Outcome: AgentDecision with reconciliationGap, gapPercent, hasSignificantGap, duplicateAmountCount, duplicateRiskScore, anomalyFlags list, confidence score (0.67-0.87), and reviewBoundary confirming no ledger corrections or payment postings are performed.

Guardrails: Hard-blocked: post_ledger_entry, approve_refunds, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory-only: agent cannot modify invoice amounts, post ledger entries, or trigger approval workflows directly. forceHumanReview is always true. Approval roles required: ops_admin, finance_user.

CF Worker: ReminderAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Drafts tiered payment reminder messages for school coordinators with overdue invoices, selecting the appropriate communication template (gentle nudge through final escalation warning) based on days overdue and prior reminder history. All drafts require operator approval before outbound send.

Tools: get_invoice_status  ·  render_card

Outcome: AgentDecision with daysOverdue, amount, schoolTier, previousReminderCount, tier (tier1-tier5), templateKey, needsHumanEscalation, isFinalEscalation, confidence score (0.72 or 0.86), and reminderBoundary confirming human approval is required before any communication is sent.

Guardrails: Hard-blocked: post_ledger_entry, approve_refunds, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory-only: agent cannot modify invoice amounts, post ledger entries, or trigger approval workflows directly. forceHumanReview is always true. Approval roles required: ops_admin, finance_user.

CF Worker: PaymentReconciliationAssistantAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Assists finance operators in matching incoming bank payments to outstanding invoices using UTR reference, amount, payer name, and school signals. Runs fuzzy-match scoring and returns ranked invoice candidates with confidence scores. Used during daily reconciliation workflows.

Tools: get_invoice_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Ranked list of invoice match candidates with confidence scores, match explanations, and a human-review flag. No ledger entry or payment posting is performed by the agent.

Guardrails: Hard-blocked: post_ledger_entry, approve_refunds, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory-only: agent cannot modify invoice amounts, post ledger entries, or trigger approval workflows directly. forceHumanReview is always true. Approval roles required: ops_admin, finance_user.

CF Worker: FinanceInvoiceReviewAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Reviews invoice batches generated for school registration cycles, comparing billed amounts against expected amounts derived from student count, subject count, and fee schedule. Flags line-item anomalies, over-billing risks, and subject count mismatches for finance_user review before invoice batches are released to schools.

Tools: get_invoice_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Structured invoice review report with per-invoice variance, anomaly flags, confidence scores (minimum 0.85), and a mandatory human-review flag. Agent cannot approve, modify, or release invoices.

Guardrails: Hard-blocked: post_ledger_entry, approve_refunds, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory-only: agent cannot modify invoice amounts, post ledger entries, or trigger approval workflows directly. forceHumanReview is always true. Approval roles required: ops_admin, finance_user.

CF Worker: FinancePaymentIntakeAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Validates and triages incoming payment records at the point of intake — bank statement CSV imports, manual payment proof uploads, and payment gateway webhooks. Checks for required fields (UTR, amount, date, school reference), identifies structurally malformed or incomplete records, and queues them for operator review.

Tools: get_invoice_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Validated payment intake report with field-level errors for invalid records, a school-match confidence score for valid records, and routing recommendations. Agent does not post or mutate payment records.

Guardrails: Hard-blocked: post_ledger_entry, approve_refunds, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory-only: agent cannot modify invoice amounts, post ledger entries, or trigger approval workflows directly. forceHumanReview is always true. Approval roles required: ops_admin, finance_user.

CF Worker: FinanceDuplicatePaymentAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Detects potential duplicate payment submissions by comparing incoming payment records against existing payments using UTR, amount, school, and date proximity signals. Surfaces high-confidence duplicate candidates with evidence for finance operator review to prevent double-counting in reconciliation.

Tools: get_invoice_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Duplicate detection report listing candidate duplicate payments with similarity scores and evidence. Agent does not suppress or delete any payment record.

Guardrails: Hard-blocked: post_ledger_entry, approve_refunds, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory-only: agent cannot modify invoice amounts, post ledger entries, or trigger approval workflows directly. forceHumanReview is always true. Approval roles required: ops_admin, finance_user.

CF Worker: FinanceRefundTriageAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Triages incoming refund requests by evaluating eligibility against policy rules: 180-day payment window, requested amount not exceeding original paid amount, and valid documented reason. Classifies requests as eligible, partial-eligible, or ineligible and routes them to the correct approval workflow.

Tools: get_invoice_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Refund eligibility report with eligibility score, policy flags, and routing recommendation. Agent cannot approve, process, or reject refunds.

Guardrails: Hard-blocked: post_ledger_entry, approve_refunds, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory-only: agent cannot modify invoice amounts, post ledger entries, or trigger approval workflows directly. forceHumanReview is always true. Approval roles required: ops_admin, finance_user.

CF Worker: FinanceCreditNoteAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Reviews and recommends credit note issuance for over-billed invoices, cancelled registrations, or fee adjustment scenarios. Validates that the credit note amount does not exceed the original invoice, checks for GST reversal requirements, and prepares a structured recommendation for finance_user approval.

Tools: get_invoice_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Credit note recommendation with eligibility assessment, GST reversal flag, and evidence refs. Agent does not generate or post credit note documents. GST reversal flag must be reviewed by CA before period close.

Guardrails: Hard-blocked: post_ledger_entry, approve_refunds, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory-only: agent cannot modify invoice amounts, post ledger entries, or trigger approval workflows directly. forceHumanReview is always true. Approval roles required: ops_admin, finance_user.

CF Worker: FinanceRevenueForecastAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Produces short-term revenue forecast signals by analysing registration pipeline depth, historical payment conversion rates, overdue aging buckets, and current-cycle invoiced totals. Surfaces projected collection shortfalls and upside scenarios for CEO-level review dashboards. Outputs are advisory estimates and do not update any accounting record.

Tools: get_invoice_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Revenue forecast report with scenario projections (base/bull/bear), collection gap estimates, risk flags, and confidence score. Agent cannot post revenue or modify any financial record. DCF or valuation output requires CEO confirmation before external sharing.

Guardrails: Hard-blocked: post_ledger_entry, approve_refunds, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory-only: agent cannot modify invoice amounts, post ledger entries, or trigger approval workflows directly. forceHumanReview is always true. Approval roles required: ops_admin, finance_user.

CF Worker: FinanceDunningAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Orchestrates multi-stage dunning sequences for overdue school accounts by computing escalation tier based on days overdue, contact attempt history, and school tier. Drafts graduated reminder copy (gentle to final escalation) and recommends when to route cases to legal escalation. All draft messages require operator approval.

Tools: get_invoice_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Dunning recommendation with escalation tier, draft template key, risk score (0-100), and human-review flag. Agent does not send messages, modify invoice status, or create legal filings. School ID redacted in audit log output.

Guardrails: Hard-blocked: post_ledger_entry, approve_refunds, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory-only: agent cannot modify invoice amounts, post ledger entries, or trigger approval workflows directly. forceHumanReview is always true. Approval roles required: ops_admin, finance_user.

CF Worker: FinanceBankReferenceAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Resolves and validates bank reference numbers (UTR, NEFT/RTGS/IMPS transaction IDs) against the payment intake records to confirm uniqueness and format compliance. Surfaces references that appear malformed, duplicated, or mismatched against the declared payment mode.

Tools: get_invoice_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Bank reference validation report with per-record format compliance status, duplicate flags, and exception routing recommendations. Agent does not modify payment records. UTR and bank reference values are redacted in decision audit logs.

Guardrails: Hard-blocked: post_ledger_entry, approve_refunds, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory-only: agent cannot modify invoice amounts, post ledger entries, or trigger approval workflows directly. forceHumanReview is always true. Approval roles required: ops_admin, finance_user.

CF Worker: FinanceSettlementAuditAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Performs a post-settlement audit of payment batches by comparing settled amounts against reconciled invoice totals, flagging short settlements, over-settlements, and unexplained residual amounts. Packages audit evidence including payment records, invoice matches, and audit trail entries for CA review.

Tools: get_invoice_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Settlement audit report with variance analysis, per-school settlement status, and a packaged audit evidence bundle. Agent does not post corrective entries. CA portal access is strictly read-only.

Guardrails: Hard-blocked: post_ledger_entry, approve_refunds, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory-only: agent cannot modify invoice amounts, post ledger entries, or trigger approval workflows directly. forceHumanReview is always true. Approval roles required: ops_admin, finance_user.

CF Worker: FinanceTaxInvoiceAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Reviews tax invoices for GST compliance — validates GSTIN presence, applicable rate (18% inclusive), CGST/SGST split accuracy, and HSN/SAC code correctness before invoices are dispatched to schools. Surfaces mismatches between computed GST and declared tax amounts for finance_user correction.

Tools: get_invoice_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: GST compliance review report with computed vs declared tax breakdown, GSTIN validation status, and correction recommendations. Agent does not modify or dispatch tax invoices. GST computation follows acc_gst_config rules; agent never hardcodes tax rates.

Guardrails: Hard-blocked: post_ledger_entry, approve_refunds, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory-only: agent cannot modify invoice amounts, post ledger entries, or trigger approval workflows directly. forceHumanReview is always true. Approval roles required: ops_admin, finance_user.

CF Worker: FinanceVendorPaymentAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Reviews vendor payment requests for TDS applicability and deduction thresholds under sections 194C, 194J, and 194T before payment runs are initiated. Computes cumulative YTD vendor payments to detect threshold breaches, flags TDS shortfalls, and recommends TDS amounts for deduction.

Tools: get_invoice_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: TDS advisory report with threshold breach status, recommended deduction amount, section applicability, and human-review flag. Agent cannot initiate vendor payments or post TDS entries. TDS thresholds are always read from acc_settings; agent never hardcodes threshold amounts.

Guardrails: Hard-blocked: post_ledger_entry, approve_refunds, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory-only: agent cannot modify invoice amounts, post ledger entries, or trigger approval workflows directly. forceHumanReview is always true. Approval roles required: ops_admin, finance_user.

CF Worker: FinanceRbiComplianceAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Monitors payment transactions and collection activities for RBI regulatory compliance — digital payment mode compliance, KYC linkage checks, and large transaction reporting thresholds (e.g., cash transaction reporting above ₹10 lakh). Surfaces compliance gaps for ops_admin review.

Tools: get_invoice_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: RBI compliance assessment report with per-transaction exception flags, severity ratings, and escalation recommendations. Agent cannot file regulatory reports or modify payment records. PII fields (KYC identifiers) subject to DPDP data minimisation.

Guardrails: Hard-blocked: post_ledger_entry, approve_refunds, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory-only: agent cannot modify invoice amounts, post ledger entries, or trigger approval workflows directly. forceHumanReview is always true. Approval roles required: ops_admin, finance_user.

CF Worker: FinanceLedgerExplainabilityAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Explains ledger entries, journal postings, and reconciliation decisions in plain language for finance coordinators, school operators, and auditors. Maps account codes to categories, narrates debit/credit movements, and surfaces related invoice and payment records as cross-references.

Tools: get_invoice_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Plain-English ledger explanation with account category, entry narrative, related record cross-references, and confidence score. Agent cannot post, modify, or approve ledger entries. COA code classification is advisory; authoritative codes are determined by acc_classification_rules rules engine only.

Guardrails: Hard-blocked: post_ledger_entry, approve_refunds, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory-only: agent cannot modify invoice amounts, post ledger entries, or trigger approval workflows directly. forceHumanReview is always true. Approval roles required: ops_admin, finance_user.

CF Worker: ProductionPlanningAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Estimates OMR sheet counts, print schedules, and vendor capacity requirements from registration headcounts and center allocations. Produces a regional batch plan with recommended print volumes and sequencing notes. Used by ops_admin and exam_ops roles before committing any print orders.

Tools: get_production_batch_status  ·  get_vendor_delay_report  ·  get_shipment_tracker  ·  get_omr_intake_status  ·  get_attendance_mismatches  ·  get_evaluation_progress  ·  freeze_registrations  ·  approve_emergency_reprint  ·  assign_vendor  ·  confirm_shipment_delivery  ·  read_exam_status  ·  render_table

Outcome: AgentDecision (decisionType: orchestration_intent) with totalSheets, bufferQuantity, per-region batch recommendations, totalRecommendedBatches, sequencingNote, planBoundary caveat, confidenceScore (0.77-0.82), and human-review flag. freeze_registrations (critical) and approve_emergency_reprint (high) require Approval gate before execution.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory; vendor and center identifiers are redacted in decision logs via redactSensitiveText. No mutation tools registered unless noted. Plan/manifest/recommendation outputs require ops_admin approval before action.

CF Worker: PrintBatchGenerationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Generates print batch manifests partitioned by exam center (school), grade, and subject with per-school OMR sheet counts and barcode range assignments. Takes batchId, list of schools with student counts and subjects, OMR sheets per student, and spare-sheet percentage to produce packaging instructions for each school.

Tools: read_exam_status  ·  render_table

Outcome: AgentDecision (decisionType: orchestration_intent) with per-school manifest list (schoolId, studentCount, subjects, spareSheets, totalSheetsForSchool, packagingInstruction), batch summary (schoolCount, totalSheets), manifestBoundary caveat, confidenceScore (0.85). Vendor print commitment requires separate ops_admin approval.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory; vendor and center identifiers are redacted in decision logs via redactSensitiveText. No mutation tools registered unless noted. Plan/manifest/recommendation outputs require ops_admin approval before action.

CF Worker: OmrAllocationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Assigns OMR serial number ranges to exam centers with overlap detection and allocation conflict flagging. Accepts a batch of school allocations with student counts, a starting serial number, and any existing allocation ranges to check against, then proposes non-overlapping serial ranges.

Tools: read_exam_status  ·  render_table

Outcome: AgentDecision (decisionType: anomaly_detection) with proposedAllocations list, duplicateDetected boolean, conflicts array, totalSerialsProposed, allocationBoundary caveat. confidenceScore 0.55 if duplicate detected, 0.88 otherwise. Writing allocations to the database requires an approved deterministic workflow.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory; vendor and center identifiers are redacted in decision logs via redactSensitiveText. No mutation tools registered unless noted. Plan/manifest/recommendation outputs require ops_admin approval before action.

CF Worker: VendorCoordinationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Reviews vendor SLA compliance against contracted timelines and flags at-risk print or packaging jobs for escalation. Accepts vendor ID, task type, days to deadline, order quantity, and current status to generate a draft communication and urgency classification. Exam_ops operators must review and approve the draft before any message is sent.

Tools: draft_email  ·  render_card

Outcome: AgentDecision (decisionType: communication_draft) with vendorId, taskType, urgencyScore (critical/high/normal), slaRisk (high/medium), draftCommunication (subject, body, requiresOperatorApprovalBeforeSend flag), commercialTermsBoundary caveat. The operator must approve before any communication is dispatched.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory; vendor and center identifiers are redacted in decision logs via redactSensitiveText. No mutation tools registered unless noted. Plan/manifest/recommendation outputs require ops_admin approval before action.

CF Worker: VendorIntelligenceAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Scores vendor reliability from SLA adherence, dispute history, and on-time delivery records to surface procurement risk. Computes a delay risk score and reprint risk score from historical on-time rate, defect rate, order size, days to deadline, and previous delays, then classifies the vendor as low/medium/high risk.

Tools: read_exam_status  ·  render_card

Outcome: AgentDecision (decisionType: vendor_risk_score) with vendorId, delayRiskScore (0-100), reprintRiskScore (0-100), prediction (low/medium/high_risk), factor breakdown, predictionBoundary caveat. Vendor decisions require ops_admin approval.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory; vendor and center identifiers are redacted in decision logs via redactSensitiveText. No mutation tools registered unless noted. Plan/manifest/recommendation outputs require ops_admin approval before action.

CF Worker: LogisticsTrackingAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Monitors dispatch and transit status across exam centers, flags delivery exceptions, and recommends follow-up actions. Computes an impact score and escalation level (critical/high/medium) from student count and days to exam. autoEscalate=true if status is delayed or exception.

Tools: read_exam_status  ·  render_card  ·  escalate_conversation

Outcome: AgentDecision (decisionType: exception_priority) with shipmentId, schoolId, currentStatus, impactScore, escalationLevel, autoEscalate flag, trackingBoundary caveat. escalate_conversation tool requires operator confirmation before action. Actual rerouting or vendor reorder requires an approved ops workflow.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory; vendor and center identifiers are redacted in decision logs via redactSensitiveText. No mutation tools registered unless noted. Plan/manifest/recommendation outputs require ops_admin approval before action.

CF Worker: SchoolConfirmationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Tracks school acknowledgment of received exam material and flags missing confirmations before the exam window. Classifies delivery status as confirmed, partial, or missing and enumerates exception types. Used by exam_ops to trigger re-dispatch or exception workflows when discrepancies are detected.

Tools: read_exam_status  ·  render_card  ·  escalate_conversation

Outcome: AgentDecision with deliveryStatus (confirmed/partial/missing), packagesExpected/Received, missingPackages, omrSheetsExpected/Received, omrShortfall, damagedCount, hasException, exceptionTypes list, confirmationBoundary caveat. Actual logistics record updates require an approved update workflow.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory; vendor and center identifiers are redacted in decision logs via redactSensitiveText. No mutation tools registered unless noted. Plan/manifest/recommendation outputs require ops_admin approval before action.

CF Worker: PackagingManifestAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Validates packaging manifests for per-center completeness and flags missing or surplus item counts. Produces a complete per-school package manifest with contents, extras, and labeling instructions for exam_ops and warehouse operators before initiating packing at the fulfillment center.

Tools: read_exam_status  ·  render_table

Outcome: AgentDecision with per-school package list (schoolId, studentCount, subjectNames, spareSheets, contents with omrSheets per subject, extras, totalItemCount, packagingLabel), batch summary, manifestBoundary caveat, confidenceScore (0.86). Actual warehouse packing must be initiated through an approved ops workflow.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory; vendor and center identifiers are redacted in decision logs via redactSensitiveText. No mutation tools registered unless noted. Plan/manifest/recommendation outputs require ops_admin approval before action.

CF Worker: ProductionFreezeReadinessAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Assesses whether the current registration and production state is safe to freeze for print-run finalization. Checks that registration batches are reconciled, payment statuses confirmed, and no open exceptions would invalidate a freeze. Does not execute the freeze itself — that requires a separate critical-risk approval gate.

Tools: read_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Freeze-readiness advisory report with a pass/fail verdict, list of blocking items, and recommended next action for the ops_admin approval workflow.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory; vendor and center identifiers are redacted in decision logs via redactSensitiveText. No mutation tools registered unless noted. Plan/manifest/recommendation outputs require ops_admin approval before action.

CF Worker: PrintOrderPlanningAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Translates confirmed registration headcounts and center allocation data into print order quantities per grade, subject, and region. Used by the Operator portal in phase3 to generate the input manifest that vendors use to begin printing, ensuring buffer quantities and wastage margins are correctly factored.

Tools: read_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Structured print order manifest with per-center, per-subject quantities including buffer margins — advisory payload ready for ops_admin approval. Buffer and wastage percentages are read from configuration; never hardcoded.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory; vendor and center identifiers are redacted in decision logs via redactSensitiveText. No mutation tools registered unless noted. Plan/manifest/recommendation outputs require ops_admin approval before action.

CF Worker: ShipmentTrackingAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Monitors the real-time dispatch and transit status of OMR packets across all exam centers, surfacing delayed or missing shipments requiring follow-up. Provides a consolidated view of which packets have been dispatched, are in transit, have been received, or are exceptions.

Tools: read_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Shipment status dashboard payload grouped by status and region, with overdue packets flagged and escalation recommendations attached. Delivery confirmation uses the deterministic confirm_shipment_delivery tool.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory; vendor and center identifiers are redacted in decision logs via redactSensitiveText. No mutation tools registered unless noted. Plan/manifest/recommendation outputs require ops_admin approval before action.

CF Worker: VendorSlaRiskAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Evaluates in-flight production batches against contracted vendor SLA deadlines, computing risk scores for each batch based on remaining time and historical vendor delay patterns. Proactively surfaces batches at risk of breaching their planned dispatch date.

Tools: read_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Ranked SLA risk report for in-flight production batches with breach probability scores and escalation recommendations. Does not reschedule or reassign batches.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory; vendor and center identifiers are redacted in decision logs via redactSensitiveText. No mutation tools registered unless noted. Plan/manifest/recommendation outputs require ops_admin approval before action.

CF Worker: ReprintExceptionAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Triages incoming reprint exception requests by validating the justification, checking available production capacity, and recommending whether to approve or reject the emergency reprint workflow. The agent surfaces a recommendation that ops_admin must confirm before the approve_emergency_reprint approval gate is triggered.

Tools: read_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Triage recommendation (approve/reject/escalate) for the reprint exception request, with supporting evidence and a confidence score, returned for ops_admin review. Duplicate reprint requests for the same batch are flagged and blocked.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory; vendor and center identifiers are redacted in decision logs via redactSensitiveText. No mutation tools registered unless noted. Plan/manifest/recommendation outputs require ops_admin approval before action.

CF Worker: LogisticsMinimizationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Recommends optimized dispatch routing and batching strategies to minimize logistics cost and transit time across exam centers. Consolidates shipments, selects cost-effective carriers, and flags sub-optimal dispatch sequences before the warehouse releases packets.

Tools: read_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Optimized dispatch routing plan with carrier recommendations, consolidation opportunities, and estimated cost savings — advisory payload for ops_admin confirmation. Cost and carrier data are estimates; final logistics decisions require ops_admin confirmation.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory; vendor and center identifiers are redacted in decision logs via redactSensitiveText. No mutation tools registered unless noted. Plan/manifest/recommendation outputs require ops_admin approval before action.

CF Worker: WarehouseDispatchAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Monitors the readiness of the warehouse dispatch queue, verifying that all packets scheduled for dispatch have correct manifest data, assigned carriers, and no open exceptions before release. Provides a pre-dispatch checklist view surfacing packets that are not ready to leave the warehouse.

Tools: read_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Dispatch readiness checklist with packets partitioned into ready and blocked states, including reason codes and resolution guidance for blocked items. Does not update omr_packets dispatch status — that is reserved for the deterministic confirm_shipment_delivery tool.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory; vendor and center identifiers are redacted in decision logs via redactSensitiveText. No mutation tools registered unless noted. Plan/manifest/recommendation outputs require ops_admin approval before action.

CF Worker: ProductionCapacityAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Estimates available print and packaging capacity across registered vendors for the upcoming exam cycle and flags potential shortfalls relative to the projected print order volume. Computes committed versus available capacity per vendor to flag shortfalls by region or subject.

Tools: read_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Vendor capacity utilization report with committed vs. available capacity per vendor, shortfall flags by region, and recommended reallocation or additional vendor engagement actions. Vendor capacity figures are estimates based on configured throughput.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory; vendor and center identifiers are redacted in decision logs via redactSensitiveText. No mutation tools registered unless noted. Plan/manifest/recommendation outputs require ops_admin approval before action.

CF Worker: DeliveryExceptionAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Detects and triages delivery exceptions — packets not received within their SLA window, packets received with incorrect counts, or packets flagged as damaged — and recommends resolution actions including escalation, reshipment, or reprint. Provides a single view of all delivery anomalies.

Tools: read_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Delivery exception triage report with all open anomalies classified by type (overdue, count_mismatch, damaged, or undelivered) and severity, including resolution recommendations and links to the appropriate deterministic approval workflows. Minimum confidence 0.75 before classifying a delivery exception as critical.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory; vendor and center identifiers are redacted in decision logs via redactSensitiveText. No mutation tools registered unless noted. Plan/manifest/recommendation outputs require ops_admin approval before action.

CF Worker: ExamMonitoringAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Monitors live exam session health on exam day by tracking attendance submission rates and OMR packet return rates across all registered schools. Computes an on_track/at_risk/critical status signal and surfaces priority flags for control-tower review.

Tools: read_exam_status  ·  render_table

Outcome: AgentDecision with status (on_track/at_risk/critical), attendanceSubmissionRate, omrReturnRate, missingSchools count, omrPacketsDispatched/Received, confidenceScore, recommendation text, forceHumanReview=true.

Guardrails: Hard-blocked: change_score, publish_results, post_ledger_entry, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; corrective action requires operator approval. Student identifiers and roll numbers are redacted from decision logs. forceHumanReview is always true.

CF Worker: AttendanceReconciliationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Reconciles expected vs. actual attendance and OMR sheet counts for a specific school on a given exam date. Computes the mismatch between present students and received OMR sheets and classifies severity as none/minor/significant/critical.

Tools: read_exam_status  ·  render_card

Outcome: AgentDecision with mismatchSeverity classification, omrMismatch and registrationMismatch counts, absenteeRate, confidenceScore (base 0.90), recommendation, forceHumanReview=true, and advisoryBoundary confirming physical re-count requires operator-approved workflow.

Guardrails: Hard-blocked: change_score, publish_results, post_ledger_entry, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; corrective action requires operator approval. Student identifiers and roll numbers are redacted from decision logs. forceHumanReview is always true.

CF Worker: ReverseLogisticsAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Tracks the return shipment of OMR answer-script packets from exam centers to evaluation hubs by monitoring courier status, dispatch dates, and evaluation deadlines. Classifies impact (low/medium/high/critical) and deadline risk (on_time/at_risk/missed_deadline/unknown).

Tools: read_exam_status  ·  show_timeline

Outcome: AgentDecision with shipmentId, courierStatus, impactLevel, deadlineStatus, all date timestamps, confidenceScore (base 0.87), recommendation, forceHumanReview=true, and advisoryBoundary confirming recovery actions require operator-approved workflow.

Guardrails: Hard-blocked: change_score, publish_results, post_ledger_entry, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; corrective action requires operator approval. Student identifiers and roll numbers are redacted from decision logs. forceHumanReview is always true.

CF Worker: OmrIntakeAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Validates OMR scan ingestion completeness by comparing received packet and sheet counts against expected manifest values, and measuring barcode readability. Classifies intake as pass/warning/fail and surfaces discrepancies for operator review before evaluation begins.

Tools: read_exam_status  ·  render_table

Outcome: AgentDecision with intakeSeverity (pass/warning/fail), packetShortfall, sheetShortfall, unreadableRate, confidenceScore (base 0.90), recommendation, forceHumanReview=true, and advisoryBoundary confirming corrective holds require operator-approved workflow.

Guardrails: Hard-blocked: change_score, publish_results, post_ledger_entry, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; corrective action requires operator approval. Student identifiers and roll numbers are redacted from decision logs. forceHumanReview is always true.

CF Worker: ScanQualityAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Assesses OMR scan image quality for a given batch by measuring defect rates across unreadable, folded, partial-bubbling, and barcode-failure categories. Computes a composite qualityScore and flags batches as pass/warning/critical to direct targeted re-scan operations before evaluation.

Tools: read_exam_status  ·  render_card

Outcome: AgentDecision with qualityFlag (pass/warning/critical), qualityScore, defectRate, partialRate, per-defect-type counts, confidenceScore (base 0.89), recommendation, forceHumanReview=true, and advisoryBoundary confirming batch holds and re-scans require operator-approved workflow.

Guardrails: Hard-blocked: change_score, publish_results, post_ledger_entry, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; corrective action requires operator approval. Student identifiers and roll numbers are redacted from decision logs. forceHumanReview is always true.

CF Worker: EvaluationCoordinationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Monitors evaluation batch progress across all evaluators by computing throughput rate, remaining sheet count, estimated hours to completion, and SLA risk classification (on_track/at_risk/breached). Identifies stalled batches and recommends staffing or schedule adjustments for operator review.

Tools: read_exam_status  ·  render_table

Outcome: AgentDecision with slaRisk classification, throughputRate (sheets/hr), remainingSheets, estimatedHoursLeft, completionPercent, evaluatorsCount, confidenceScore (base 0.87), recommendation, forceHumanReview=true, and advisoryBoundary confirming staffing adjustments require operator-approved workflow.

Guardrails: Hard-blocked: change_score, publish_results, post_ledger_entry, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; corrective action requires operator approval. Student identifiers and roll numbers are redacted from decision logs. forceHumanReview is always true.

CF Worker: ResultIntegrityAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Checks result dataset completeness and integrity before publication readiness is declared, by counting duplicate ranks, impossible scores, percentile anomalies, and suspected cheating flags. Computes an anomalyRate and integrityScore, and sets publicationBlocked=true when cheating flags are present.

Tools: read_exam_status  ·  render_table

Outcome: AgentDecision with integrityScore, anomalyRate, per-category anomaly counts, mandatoryHumanReview boolean, publicationBlocked boolean, confidenceScore (base 0.88), recommendation, forceHumanReview=true. Agent CANNOT block or publish results directly — only flags.

Guardrails: Hard-blocked: change_score, publish_results, post_ledger_entry, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; corrective action requires operator approval. Student identifiers and roll numbers are redacted from decision logs. forceHumanReview is always true.

CF Worker: ScoreOverrideAssistant extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Reviews score override requests against policy bounds, computing the score delta and required documentation checklist based on magnitude and direction of the change. Returns an advisory approval or rejection recommendation with full documentation requirements for human decision. change_score is unconditionally hard-blocked.

Tools: read_exam_status  ·  render_card

Outcome: AgentDecision with scoreDelta, riskLevel (low/medium/high), documentsRequired list, workflowRequired=score_override_approval_workflow, supportingEvidence references, confidenceScore (base 0.85), recommendation, forceHumanReview=true. Agent CANNOT apply score changes independently under any circumstances.

Guardrails: Hard-blocked: change_score, publish_results, post_ledger_entry, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; corrective action requires operator approval. Student identifiers and roll numbers are redacted from decision logs. forceHumanReview is always true.

CF Worker: ExamReevaluationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Triages reevaluation applications by eligibility criteria and recommends processing priority, SLA timeline, evidence checklist, and workflow steps including re-scan authorization or fraud investigation escalation. Handles parent_dispute, scan_error, marking_error, and suspected_fraud.

Tools: read_exam_status  ·  render_card

Outcome: AgentDecision with requestSLA (24h for scan_error, 72h otherwise), requiresRescan/requiresInvestigation flags, evidenceChecklist array, workflowSteps with per-step approval requirements, confidenceScore (base 0.86), recommendation, forceHumanReview=true.

Guardrails: Hard-blocked: change_score, publish_results, post_ledger_entry, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; corrective action requires operator approval. Student identifiers and roll numbers are redacted from decision logs. forceHumanReview is always true.

CF Worker: OmrEvaluationSubmissionAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Validates OMR evaluation CSV submission metadata — line counts, match rates, unmatched and duplicate counts, and impossible score counts — before advising on evaluation pipeline entry. Operates exclusively on aggregate validation metadata; raw scores and student identifiers are unconditionally excluded from its context by schema design.

Tools: read_exam_status  ·  render_card

Outcome: AgentDecision with validationStatus (passed/has_gaps/failed), matchRate, errorRate, unmatchedCount, duplicateCount, impossibleScoreCount, totalErrors, confidenceScore (base 0.88). NEVER accept, process, or include raw exam scores or student identifiers in agent context.

Guardrails: Hard-blocked: change_score, publish_results, post_ledger_entry, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; corrective action requires operator approval. Student identifiers and roll numbers are redacted from decision logs. forceHumanReview is always true.

CF Worker: ExamSessionPlanningAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Plans and validates exam session schedules across registered centers for a given exam cycle. Surfaces scheduling conflicts, seat capacity gaps, and center-readiness blockers before the session window opens. All output is advisory; session creation and lock-in requires operator approval.

Tools: read_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Advisory decision payload with session planning status, capacity gap summary, conflict list, confidence score (>=0.85), and recommended corrective actions for operator approval. forceHumanReview=true for all scheduling mutations.

Guardrails: Hard-blocked: change_score, publish_results, post_ledger_entry, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; corrective action requires operator approval. Student identifiers and roll numbers are redacted from decision logs. forceHumanReview is always true.

CF Worker: ExamCenterReadinessAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Assesses operational readiness of each exam center prior to session day, covering infrastructure, invigilator confirmation, stationery receipt, and communication connectivity. Surfaces center-specific readiness scores and flags centers requiring pre-emptive intervention.

Tools: read_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Readiness assessment payload per center including readiness tier (ready/at_risk/not_ready), dimension breakdown, overall readiness rate, confidence score, and escalation recommendations. Agent cannot modify center records, invigilator assignments, or stationery dispatch orders.

Guardrails: Hard-blocked: change_score, publish_results, post_ledger_entry, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; corrective action requires operator approval. Student identifiers and roll numbers are redacted from decision logs. forceHumanReview is always true.

CF Worker: OmrScanIntakeAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Validates the completeness and integrity of OMR scan batches at the point of intake into the evaluation pipeline. Checks scan counts, barcode readability, and allocation manifest alignment before evaluation begins.

Tools: read_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Advisory intake validation result with severity level (pass/warning/fail), shortfall counts, barcode failure rate, recommended disposition (proceed/rescan/hold), and confidence score. Agent cannot ingest raw answer data or student-OMR mappings.

Guardrails: Hard-blocked: change_score, publish_results, post_ledger_entry, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; corrective action requires operator approval. Student identifiers and roll numbers are redacted from decision logs. forceHumanReview is always true.

CF Worker: BarcodeMismatchAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Detects and classifies OMR barcode mismatches between scanned sheets and the allocation manifest, distinguishing between unreadable codes, phantom barcodes, and cross-batch contamination. Identifies sheets requiring manual re-scan or physical re-verification before evaluation proceeds.

Tools: read_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Mismatch classification report with per-type counts (unreadable, not-in-manifest, duplicate, cross-batch), mismatch rate, severity tier (acceptable/elevated/critical), affected sheet list references, and operator action recommendations.

Guardrails: Hard-blocked: change_score, publish_results, post_ledger_entry, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; corrective action requires operator approval. Student identifiers and roll numbers are redacted from decision logs. forceHumanReview is always true.

CF Worker: EvaluationQualityAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Monitors and assesses the quality of the evaluation process by analysing evaluator consistency, inter-rater agreement rates, and marking scheme adherence across batches. Detects systematic marking errors or outlier evaluators before result integrity is compromised.

Tools: read_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Evaluation quality report with per-evaluator consistency flags, batch-level quality risk tiers (acceptable/review_required/halt_recommended), anomaly counts, confidence score, and remediation recommendations. Individual student scores are never processed in LLM context.

Guardrails: Hard-blocked: change_score, publish_results, post_ledger_entry, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; corrective action requires operator approval. Student identifiers and roll numbers are redacted from decision logs. forceHumanReview is always true.

CF Worker: ScoreAnomalyAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Detects statistical anomalies in score distributions across subjects, centers, and cohorts to surface potential marking errors, data corruption, or integrity violations before result publication. Provides a pre-publication anomaly gate. Individual student identifiers and raw scores are excluded from LLM context.

Tools: read_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Score anomaly report with distribution statistics, anomaly type breakdown, severity level (none/minor/moderate/critical), publication hold advisory, and confidence score. Agent cannot modify scores or directly block publication — output is advisory only.

Guardrails: Hard-blocked: change_score, publish_results, post_ledger_entry, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; corrective action requires operator approval. Student identifiers and roll numbers are redacted from decision logs. forceHumanReview is always true.

CF Worker: ResultPublicationReadinessAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Performs a comprehensive pre-publication readiness check across all gating criteria — data completeness, integrity anomalies, mandatory human review clearances, and regulatory compliance flags — before the result publication workflow is permitted to proceed.

Tools: read_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Publication readiness verdict (ready/conditional/blocked), gating checklist with open items, compliance flag status, confidence score, and recommended workflow action. A "ready" verdict does not auto-publish; ops_admin must initiate the publication approval workflow independently.

Guardrails: Hard-blocked: change_score, publish_results, post_ledger_entry, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; corrective action requires operator approval. Student identifiers and roll numbers are redacted from decision logs. forceHumanReview is always true.

CF Worker: CertificateDistributionAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Monitors and validates the distribution pipeline for generated certificates, tracking delivery to schools and flagging undelivered, bounced, or duplicate dispatch events. Ensures every eligible candidate receives their certificate and surfaces distribution exceptions for resolution.

Tools: read_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Distribution status report with per-school delivery rates, exception type counts (undelivered/bounced/duplicate/expired-link), unresolved gap list, confidence score, and recommended follow-up actions. Re-dispatch decisions require operator approval via the designated certificate workflow.

Guardrails: Hard-blocked: change_score, publish_results, post_ledger_entry, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; corrective action requires operator approval. Student identifiers and roll numbers are redacted from decision logs. forceHumanReview is always true.

CF Worker: ReevaluationTriageAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Triages incoming reevaluation applications by eligibility, reason type, and SLA priority to produce a prioritised processing queue. Prioritises: fraud > scan_error > marking_error > parent_dispute. Manages high-volume reevaluation request intake efficiently.

Tools: read_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Triage report with a prioritised reevaluation queue, eligibility flags, SLA deadlines, recommended evaluator assignments, and confidence score. Agent cannot reject or approve reevaluation applications — recommendations only. Score corrections require a separate supervisor-approved workflow.

Guardrails: Hard-blocked: change_score, publish_results, post_ledger_entry, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; corrective action requires operator approval. Student identifiers and roll numbers are redacted from decision logs. forceHumanReview is always true.

CF Worker: ExamIncidentAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Captures, classifies, and prioritises exam-day incident reports from invigilators and center coordinators covering malpractice suspicions, centre disruptions, technical failures, and medical emergencies. Maintains a structured incident log and surfaces high-severity events requiring immediate escalation.

Tools: read_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Structured incident classification with severity tier, recommended escalation path (control_tower_alert/regional_coordinator_notify/exam_ops_review/log_and_monitor), pattern cross-reference summary, confidence score. Agent cannot cancel exam sessions, disqualify candidates, or trigger law enforcement escalations directly.

Guardrails: Hard-blocked: change_score, publish_results, post_ledger_entry, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; corrective action requires operator approval. Student identifiers and roll numbers are redacted from decision logs. forceHumanReview is always true.

CF Worker: ExamControlTowerAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Serves as the central aggregation and prioritisation layer for all live exam day signals — consolidating attendance rates, incident flags, OMR dispatch status, and centre connectivity into a unified operational picture. Real-time situational awareness hub for exam execution across all centers.

Tools: read_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Control tower advisory with overall execution health score, ranked intervention list, center-level signal summary, confidence score, and recommended priority actions. Agent cannot directly dispatch field coordinators, send official communications, or halt exam sessions. Only aggregate metrics are processed in LLM context.

Guardrails: Hard-blocked: change_score, publish_results, post_ledger_entry, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; corrective action requires operator approval. Student identifiers and roll numbers are redacted from decision logs. forceHumanReview is always true.

CF Worker: OfflineExamWorkerAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Monitors and validates offline exam session data synchronisation from centres operating without live connectivity, ensuring queued submissions, autosave payloads, and QR fallback bundles are correctly received and reconciled once connectivity is restored.

Tools: read_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Offline sync status report with per-session classification (fully_synced/partial_sync/sync_failed/awaiting_qr_fallback), payload gap counts, recovery recommendations, confidence score, and human-review flag for unresolved sessions. Agent cannot apply or merge synced payloads directly.

Guardrails: Hard-blocked: change_score, publish_results, post_ledger_entry, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; corrective action requires operator approval. Student identifiers and roll numbers are redacted from decision logs. forceHumanReview is always true.

CF Worker: ResultReportAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Generates structured result report advisories for school-level, district-level, and national-level summaries after result publication clearance, including rank distribution analytics, participation rates, and subject-wise performance breakdowns. Processes aggregate statistics only; individual student PII excluded.

Tools: read_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Structured result report advisory with rank distribution data, participation metrics, subject performance summaries, anomaly flags, and confidence score. Distribution to schools requires operator approval. Agent cannot access result records before publication has been cleared via the approved human workflow.

Guardrails: Hard-blocked: change_score, publish_results, post_ledger_entry, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; corrective action requires operator approval. Student identifiers and roll numbers are redacted from decision logs. forceHumanReview is always true.

CF Worker: ExamIntegrityAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Performs a holistic exam integrity audit by correlating malpractice incident flags, score anomaly signals, attendance irregularities, and proctoring event patterns to surface coordinated integrity violations. Final integrity clearance gate before result publication is permitted.

Tools: read_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Integrity audit advisory with per-center/session risk-level assignments (clear/elevated/high_risk/critical), correlated signal summary, recommended dispositions (clear_to_publish/hold_for_review/escalate_to_investigation), confidence score. Any "critical" integrity finding mandates mandatory human review by ops_admin before any downstream action. Agent does not gate publication directly — feeds result_publication_readiness_agent.

Guardrails: Hard-blocked: change_score, publish_results, post_ledger_entry, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; corrective action requires operator approval. Student identifiers and roll numbers are redacted from decision logs. forceHumanReview is always true.

CF Worker: ResultPublicationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Validates result dataset completeness by comparing processed results against total candidates and checking for pending verifications before recommending publication approval to an ops_admin human workflow gate. Computes a completion rate and assigns readiness states: ready_for_human_approval, pending_final_verification, or not_ready.

Tools: get_result_readiness  ·  get_ranking_conflicts  ·  get_reevaluation_queue  ·  get_certificate_status  ·  get_publication_checklist  ·  get_award_allocation  ·  approve_reevaluation  ·  reject_reevaluation  ·  publish_results  ·  revoke_certificate

Outcome: AgentDecision with examId, completionRate, pendingVerifications, readinessStatus, advisoryBoundary, confidenceScore (0.82-0.87), forceHumanReview=true. publish_results requires critical-level Approval gate and is blocked from direct agent invocation. approve/reject_reevaluation require high-level Approval gate.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. forceHumanReview is always true — output is advisory only. No tools registered for most agents (advisory classifier); all actions require human-approved workflows.

CF Worker: RankingIntegrityAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Checks ranking consistency across score bands, ties, and merit criteria by computing an anomaly rate from duplicate ranks, rank gaps, and percentile anomalies across a ranked candidate set. Classifies ranking integrity as clean, minor_issues, or integrity_risk.

Tools: (advisory classifier — no registered tools)

Outcome: AgentDecision with examId, totalRanked, duplicateRanks, rankGaps, percentileAnomalies, anomalyRate, integrityStatus (clean/minor_issues/integrity_risk), advisoryBoundary, confidenceScore (0.82-0.90). Ranking corrections require dedicated integrity workflow.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. forceHumanReview is always true — output is advisory only. No tools registered for most agents (advisory classifier); all actions require human-approved workflows.

CF Worker: CertificateGenerationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Validates certificate batch parameters including eligible candidate count, template version, and customizations, then produces a generation plan estimating batch count and generation time before recommending ops_admin authorization. No certificates are created during agent execution.

Tools: (advisory classifier — no registered tools)

Outcome: AgentDecision (decisionType: orchestration_intent) with batchId, eligibleCandidates, certificateTemplate, estimatedGenerationTimeMinutes, batchesRequired, batchSize, advisoryBoundary, confidenceScore 0.88, forceHumanReview=true. Certificate generation only proceeds after authorized workflow is triggered externally.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. forceHumanReview is always true — output is advisory only. No tools registered for most agents (advisory classifier); all actions require human-approved workflows.

CF Worker: CertificateRevocationAssistant extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Reviews certificate revocation requests by evaluating the revocation reason, determining required approver roles, assigning a risk tier (exam_fraud/court_order → critical; identity_mismatch → high; data_correction → medium), and building a mandatory evidence checklist.

Tools: (advisory classifier — no registered tools)

Outcome: AgentDecision (decisionType: exception_priority) with candidateId, certificateId, revocationReason, requiredApprovals, riskLevel, mandatoryEvidenceChecklist, advisoryBoundary, confidenceScore 0.91, forceHumanReview=true. revoke_certificates is hard-blocked for this agent — certificate revocation is irreversible.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. forceHumanReview is always true — output is advisory only. No tools registered for most agents (advisory classifier); all actions require human-approved workflows.

CF Worker: AwardAllocationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Allocates merit positions, subject toppers, and award categories based on result and ranking data by computing total awardees from a top percentage threshold and breaking down counts per award category rank range. Outputs require exam_ops review before any award records are created or announced.

Tools: (advisory classifier — no registered tools)

Outcome: AgentDecision (decisionType: school_health_score) with examId, totalEligible, topPercent, totalAwardees, awardBreakdown per category, advisoryBoundary, confidenceScore 0.87, forceHumanReview=true. No award records are created until ops_admin approves through the designated workflow.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. forceHumanReview is always true — output is advisory only. No tools registered for most agents (advisory classifier); all actions require human-approved workflows.

CF Worker: ParentCommunicationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Drafts result and certificate delivery communication intents for operator approval, selecting the appropriate template key and SLA priority for each communication type (result_ready, certificate_ready, reevaluation_update, award_notification) and delivery channel (email, SMS, WhatsApp). No external message is sent by the agent.

Tools: (advisory classifier — no registered tools)

Outcome: AgentDecision (decisionType: communication_draft) with candidateId, communicationType, templateKey, channel, slaPriority (same_day/one_business_day/two_business_days), advisoryBoundary, confidenceScore 0.85, forceHumanReview=true. No message is dispatched until ops_admin approves and triggers the send workflow.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. forceHumanReview is always true — output is advisory only. No tools registered for most agents (advisory classifier); all actions require human-approved workflows.

CF Worker: SchoolSummaryAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Generates per-school performance summaries including attendance rate, pass rate, average and top scores, subject breakdowns, and merit counts for coordinator dispatch. Classifies each school into a performance tier (outstanding/good/average/below_average) based on average score thresholds.

Tools: (advisory classifier — no registered tools)

Outcome: AgentDecision (decisionType: school_health_score) with schoolId, examId, attendanceRate, passRate, averageScore, topScore, passCount, awardCount, schoolPerformanceTier, advisoryBoundary, confidenceScore 0.86. Summary is advisory for coordinator review; official results published through designated workflow.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. forceHumanReview is always true — output is advisory only. No tools registered for most agents (advisory classifier); all actions require human-approved workflows.

CF Worker: ParentReevaluationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Triages parent reevaluation requests by validating eligibility windows, confirming the SLA (72 hours), and summarizing supporting evidence requirements for exam_ops processing. Produces a structured advisory packet including required evidence items and the next-step instruction for submission to the ops_admin workflow.

Tools: (advisory classifier — no registered tools)

Outcome: AgentDecision (decisionType: orchestration_intent) with requestId, candidateId, parentName, subject, currentScore, requestReason, isEligible=true, slaHours=72, evidenceRequired checklist (reevaluation_request_form, parent_identity_proof, original_answer_sheet_copy), nextStep, advisoryBoundary, confidenceScore 0.84. No score is changed.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. forceHumanReview is always true — output is advisory only. No tools registered for most agents (advisory classifier); all actions require human-approved workflows.

CF Worker: CandidateVerificationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Evaluates candidate identity signals — registration status, ID document submission, and face-match score — before online exam session admission. Produces verificationStatus (verified/face_match_failed/manual_review_required/id_missing/rejected) with a confidence score for the ops_admin approval workflow.

Tools: read_online_exam_status  ·  render_card

Outcome: Structured JSON with verificationStatus enum, faceMatchScore, idType, examId, candidateId, confidenceScore, recommendation, advisoryBoundary, forceHumanReview=true. No session state is mutated. Session admission requires ops_admin approval through the identity verification workflow.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory, schema-validated, and append-only logged. mutatesBusinessTruth=false enforced at runtime. Minimum confidence threshold 0.85 before routing to approval.

CF Worker: ExamSessionAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Monitors real-time online exam session health by evaluating timer elapsed time, question completion rate, connection drop counts, and autosave failure counts. Surfaces sessionHealth classification (healthy/degraded/critical) with anomaly flags for ops_admin control-tower review.

Tools: read_online_exam_status  ·  render_card

Outcome: Structured JSON with sessionId, candidateId, elapsedMinutes, completionRate, connectionDrops, autosaveFailures, sessionHealth (healthy/degraded/critical), confidenceScore (base 0.88, −0.02 per connection drop/autosave failure), forceHumanReview=true. Server-authoritative timers immutable to this layer.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory, schema-validated, and append-only logged. mutatesBusinessTruth=false enforced at runtime. Minimum confidence threshold 0.85 before routing to approval.

CF Worker: AntiCheatingAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Evaluates proctoring signals — tab-switch count, copy-paste events, face-detection failures, and multiple-person detection — to compute a risk score and suspicion level (low/medium/high/critical). Flags high and critical cases for super_admin review through the exam integrity workflow. Cannot invalidate attempts.

Tools: read_online_exam_status  ·  render_card

Outcome: Structured JSON with sessionId, candidateId, riskScore (weighted sum), suspicionLevel, flaggedForReview boolean, anomalousBehaviorSignals list, confidenceScore (0.7 + riskScore × 0.3), forceHumanReview=true. Agent CANNOT invalidate candidates or exam attempts — disqualification requires super_admin approval.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory, schema-validated, and append-only logged. mutatesBusinessTruth=false enforced at runtime. Minimum confidence threshold 0.85 before routing to approval.

CF Worker: ConnectivityRecoveryAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Detects the impact of network interruptions on candidate exam sessions by analysing disconnection duration, reconnection status, and autosave checkpoint timing. Produces recoveryStatus and timeExtensionRecommendationMinutes (0/5/min(ceil(duration/60), 30)) for ops_admin consideration.

Tools: read_online_exam_status  ·  render_card

Outcome: Structured JSON with sessionId, candidateId, disconnectionDuration, recoveryStatus (disconnected/auto_recovered/recovery_pending_review/critical_intervention_required), lostWorkRisk, timeExtensionRecommendationMinutes, confidenceScore, forceHumanReview=true. Agent cannot modify exam sessions, timers, or extend time.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory, schema-validated, and append-only logged. mutatesBusinessTruth=false enforced at runtime. Minimum confidence threshold 0.85 before routing to approval.

CF Worker: SubmissionIntegrityAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Validates the integrity of online exam submissions by checking checksum validity, duplicate submission flags, and client-to-server timestamp deltas (flagged if >30,000 ms). Produces integrityStatus and requiresInvestigation flag for ops_admin review.

Tools: read_online_exam_status  ·  render_card

Outcome: Structured JSON with submissionId, candidateId, examId, checksumValid, duplicateSubmission, timeDeltaMs, integrityStatus (valid/checksum_failed/duplicate_submission/timing_anomaly), requiresInvestigation boolean, confidenceScore, forceHumanReview=true. Agent CANNOT accept or reject submissions.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory, schema-validated, and append-only logged. mutatesBusinessTruth=false enforced at runtime. Minimum confidence threshold 0.85 before routing to approval.

CF Worker: OnlineEvaluationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Tracks the MCQ auto-grading pipeline progress for online submissions by computing evaluation completion rate and time remaining against the evaluation deadline. Classifies pipeline status as complete/on_track/at_risk/critical and recommends operator intervention when falling behind.

Tools: read_online_exam_status  ·  render_card

Outcome: Structured JSON with examId, totalSubmissions, evaluatedCount, autoGradableCount, manualGradingPending, completionRate, hoursRemaining, evaluationStatus, confidenceScore (base 0.87), forceHumanReview=true. Agent CANNOT modify or finalise any scores.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory, schema-validated, and append-only logged. mutatesBusinessTruth=false enforced at runtime. Minimum confidence threshold 0.85 before routing to approval.

CF Worker: OnlineSupportAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Classifies live exam support tickets by issue type (login_failure, timer_discrepancy, submission_failure, proctoring_tech_issue, question_display_error) and severity to derive SLA deadlines and recommended response templates. Flags critical and submission/timer issues for immediate ops_admin escalation.

Tools: read_online_exam_status  ·  render_card

Outcome: Structured JSON with ticketId, candidateId, issueType, severity, slaHours (critical=1/high=4/medium=12/low=24), escalationRequired boolean, responseTemplate name, confidenceScore, forceHumanReview=true. Technical interventions and session adjustments require ops_admin execution.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory, schema-validated, and append-only logged. mutatesBusinessTruth=false enforced at runtime. Minimum confidence threshold 0.85 before routing to approval.

CF Worker: OnlineCandidateVerificationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Verifies candidate identity and eligibility before granting access to the online exam portal. Receives face-match scores, ID document metadata, and registration status, then produces an advisory verification decision (verified/face_match_failed/manual_review_required/id_missing/rejected).

Tools: read_online_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Governed AgentDecision with verificationStatus, faceMatchScore, idType, confidenceScore, advisoryBoundary, humanReviewRequired=true, routed to the ops_admin identity verification approval workflow. NEVER admit or reject a candidate directly.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory, schema-validated, and append-only logged. mutatesBusinessTruth=false enforced at runtime. Minimum confidence threshold 0.85 before routing to approval.

CF Worker: OnlineSessionRecoveryAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Triages interrupted online exam sessions and recommends recovery actions when a candidate's session terminates abnormally (browser crash, forced logout, or platform error). Inspects session state, autosave timestamps, and elapsed time.

Tools: read_online_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Governed AgentDecision with recoveryRecommendation, dataLossRisk assessment, optional timeExtensionMinutes, humanReviewRequired=true, routed to the ops_admin session recovery approval workflow. NEVER modify exam session state or timers.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory, schema-validated, and append-only logged. mutatesBusinessTruth=false enforced at runtime. Minimum confidence threshold 0.85 before routing to approval.

CF Worker: OnlineProctoringTriageAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Triages live proctoring alerts generated by the adaptive proctoring layer. Classifies incoming events (face detection failures, tab switches, audio anomalies, multiple persons detected) by severity and priority, then routes to the appropriate review queue.

Tools: read_online_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Governed AgentDecision with alertSeverity, escalationPath (invigilator queue for low/medium; ops_admin review queue for high/critical), triagePriority, humanReviewRequired=true. All proctoring events are append-only and cannot automatically invalidate attempts.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory, schema-validated, and append-only logged. mutatesBusinessTruth=false enforced at runtime. Minimum confidence threshold 0.85 before routing to approval.

CF Worker: OnlineSubmissionIntegrityAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Verifies the integrity of online exam submissions by inspecting checksums, detecting duplicate submissions, and flagging timing anomalies between client submission time and server receipt time. Pre-evaluation gate in the Phase 6 online exam submission pipeline.

Tools: read_online_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Governed AgentDecision with integrityStatus (valid/checksum_failed/duplicate_submission/timing_anomaly), requiresInvestigation flag, timeDeltaMs, humanReviewRequired=true, routed to ops_admin submission integrity approval workflow. NEVER accept or reject submissions.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory, schema-validated, and append-only logged. mutatesBusinessTruth=false enforced at runtime. Minimum confidence threshold 0.85 before routing to approval.

CF Worker: OnlineAutosaveMonitorAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Monitors the 30-second autosave cycle of the Phase 6 offline-first online exam PWA, detecting autosave failures, stale save timestamps, and IndexedDB sync gaps. Surfaces autosave health degradation so operators can intervene before candidate data is lost.

Tools: read_online_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Governed AgentDecision with autosaveStatus (healthy/degraded/at_risk_of_data_loss), consecutiveFailureCount, estimatedDataAtRisk, recommendedAction, humanReviewRequired flag (when degraded). Online exam timers are server-authoritative; autosave writes are service-role gated. NEVER modify session state, exam timers, or autosave cadence.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory, schema-validated, and append-only logged. mutatesBusinessTruth=false enforced at runtime. Minimum confidence threshold 0.85 before routing to approval.

CF Worker: OnlineSupportTriageAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Triages incoming online exam support tickets raised by candidates during active exam windows, classifying them by issue type and severity to produce SLA-bound response and escalation recommendations. Routes critical and submission-related issues for immediate ops_admin escalation.

Tools: read_online_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Governed AgentDecision with ticketId, issueType, severity, slaHours, escalationRequired, responseTemplate, humanReviewRequired=true, surfaced in the support_user triage dashboard. NEVER perform technical interventions or session adjustments.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory, schema-validated, and append-only logged. mutatesBusinessTruth=false enforced at runtime. Minimum confidence threshold 0.85 before routing to approval.

CF Worker: OnlineAccessibilityAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Advises on accessibility accommodations for candidates with special needs during Phase 6 online exam sessions, including extended time eligibility, assistive technology compatibility, and screen reader guidance. Surfaces accommodation gaps and WCAG compliance flags.

Tools: read_online_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Governed AgentDecision with accommodationStatus, detectedGaps, complianceFlags, recommendedActions, humanReviewRequired=true, routed to the ops_admin accessibility review workflow. NEVER grant or modify exam accommodations.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory, schema-validated, and append-only logged. mutatesBusinessTruth=false enforced at runtime. Minimum confidence threshold 0.85 before routing to approval.

CF Worker: OnlineNetworkHealthAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Monitors network quality signals for active online exam sessions, analyzing bandwidth, latency, packet loss, and WebSocket stability to detect degradation before it causes session disruptions or data loss.

Tools: read_online_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Governed AgentDecision with networkHealthStatus (healthy/degraded/critical), thresholdBreaches, recommendedActions (reconnect/switch to mobile data/reduce background traffic), humanReviewRequired flag. NEVER modify exam session state or connectivity settings.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory, schema-validated, and append-only logged. mutatesBusinessTruth=false enforced at runtime. Minimum confidence threshold 0.85 before routing to approval.

CF Worker: OnlineDeviceRiskAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Assesses device-level risk signals for candidates attempting to join Phase 6 online exams, evaluating browser fingerprint anomalies, virtual machine indicators, emulator flags, unsupported OS/browser combinations, and rooted/jailbroken device signals.

Tools: read_online_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Governed AgentDecision with deviceRiskScore, riskClassification (low/medium/high/critical), riskFactors, recommendedAction (allow/flag-for-review/block-pending-review), humanReviewRequired=true. NEVER block or admit a candidate based on device signals alone.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory, schema-validated, and append-only logged. mutatesBusinessTruth=false enforced at runtime. Minimum confidence threshold 0.85 before routing to approval.

CF Worker: OnlineExamTimerAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Monitors server-authoritative exam timer state for Phase 6 online exam sessions, detecting clock skew between the client timer and the server-side countdown, flagging timer manipulation attempts when client remaining time exceeds server remaining time beyond threshold.

Tools: read_online_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Governed AgentDecision with skewSeconds, skewClassification (acceptable <5s/warning 5-30s/anomalous >30s), manipulationSuspected flag, serverRemainingSeconds, humanReviewRequired=true. NEVER modify exam timers or extend/reduce candidate time. Online exam timers are server-authoritative.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory, schema-validated, and append-only logged. mutatesBusinessTruth=false enforced at runtime. Minimum confidence threshold 0.85 before routing to approval.

CF Worker: OnlineAiCheatingSignalAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Analyses AI-assisted cheating signals captured during Phase 6 online exam sessions, including suspicious response patterns consistent with AI-generated answers, answer-timing anomalies, and cross-candidate answer similarity spikes. Produces an advisory AI cheating risk score.

Tools: read_online_exam_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Governed AgentDecision with aiCheatingRiskScore, detectedPatterns, crossCandidateSimilarity, flaggedForReview flag, humanReviewRequired=true, routed to the super_admin exam integrity approval workflow. NEVER invalidate a candidate session or exam attempt based on AI signal alone.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory, schema-validated, and append-only logged. mutatesBusinessTruth=false enforced at runtime. Minimum confidence threshold 0.85 before routing to approval.

CF Worker: SyllabusMappingAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Maps curriculum content items to syllabus taxonomy nodes for tracking coverage across grades, subjects, and boards (CBSE, ICSE, Cambridge, State, IB). Computes coverage rate and determines mapping status (complete/partial/incomplete).

Tools: read_academic_summary  ·  render_table

Outcome: AgentDecision (decisionType: column_mapping_suggestion) with coverageRate, mappingStatus (complete/partial/incomplete), unmapped topic count, confidenceScore (0.82-0.92), and advisoryBoundary confirming curriculum changes require academic committee approval.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only: agents cannot mutate official exam result truth, student records, or official academic standing. No external communications. All output is advisory, logged, and schema-validated.

CF Worker: CourseBlueprintAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Generates course structure blueprints aligned to learning objectives and syllabus coverage targets for human authoring review. Computes total hours, module count, and a skeletal weekly schedule from course duration and learning objective inputs.

Tools: read_academic_summary  ·  render_card

Outcome: AgentDecision (decisionType: orchestration_intent) with courseId, totalHours, moduleCount, learningObjectiveCount, weeklySchedule anchor points (introduction + revision_and_assessment), advisoryBoundary. confidenceScore 0.78. Official course content requires academic committee review and approval before implementation.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only: agents cannot mutate official exam result truth, student records, or official academic standing. No external communications. All output is advisory, logged, and schema-validated.

CF Worker: LessonGenerationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Recommends lesson content structures, learning objective sequences, and difficulty progressions for human content authors. Proposes a section count, estimated activities, and an ordered lesson structure based on topic, subject, grade, duration, and learning style.

Tools: read_academic_summary  ·  render_card

Outcome: AgentDecision (decisionType: orchestration_intent) with sectionCount, estimatedActivities, lessonStructure array (introduction/core_concept/worked_examples/practice_exercise if ≥60min/summary), advisoryBoundary requiring teacher review before delivery to students. confidenceScore 0.76.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only: agents cannot mutate official exam result truth, student records, or official academic standing. No external communications. All output is advisory, logged, and schema-validated.

CF Worker: PracticeGenerationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Recommends practice question set parameters aligned to diagnosed skill gaps and Bloom's taxonomy difficulty levels. Classifies practice sessions as standard or mistake-remediation based on recent error topics, and estimates session duration.

Tools: read_academic_summary  ·  render_card

Outcome: AgentDecision (decisionType: orchestration_intent) with practiceType (standard_practice/mistake_remediation), estimatedDurationMinutes, focusedOnMistakes flag, recentMistakeTopics, advisoryBoundary requiring teacher approval for actual question delivery.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only: agents cannot mutate official exam result truth, student records, or official academic standing. No external communications. All output is advisory, logged, and schema-validated.

CF Worker: ContentReviewAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Reviews learning content quality across accuracy, clarity, age-appropriateness, and DPDP policy compliance, flagging concerns for editorial review. Accepts structured signals (factual errors, grammar errors, difficulty suitability) and produces a reviewStatus and priority classification.

Tools: read_academic_summary  ·  render_card

Outcome: AgentDecision (decisionType: anomaly_detection) with issueCount, reviewStatus (approved/minor_revision_needed/major_revision_needed), reviewPriority (high if factual error, else medium), individual error flags, advisoryBoundary. Content publication is blocked until human SME review.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only: agents cannot mutate official exam result truth, student records, or official academic standing. No external communications. All output is advisory, logged, and schema-validated.

CF Worker: StudyPathAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Generates personalized study path recommendations from performance data, skill gaps, and exam proximity. Computes days remaining to exam, selects study intensity (intensive/focused/steady), and prioritizes weak topics (top 5 weak topics).

Tools: read_academic_summary  ·  render_card

Outcome: AgentDecision (decisionType: orchestration_intent) with daysRemaining, studyIntensity (intensive <14 days/focused <30/steady ≥30), weakTopicsCount, strongTopicsCount, priorityTopics list (up to 5), advisoryBoundary requiring teacher review before implementation.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only: agents cannot mutate official exam result truth, student records, or official academic standing. No external communications. All output is advisory, logged, and schema-validated.

CF Worker: RevisionAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Recommends spaced-repetition revision schedules based on retention decay signals, topic counts, and upcoming exam timelines. Computes a daily revision pace (revisionsPerDay = min(ceil(topics/max(daysToExam,1)), 10)) and priority level (high/medium/low).

Tools: read_academic_summary  ·  render_table

Outcome: AgentDecision (decisionType: orchestration_intent) with daysToExam, topicCount, revisionsPerDay, revisionPriority (high ≤3 days/medium ≤7/low >7), suggestedSchedule text, advisoryBoundary. Actual sessions are guided by student and teacher.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only: agents cannot mutate official exam result truth, student records, or official academic standing. No external communications. All output is advisory, logged, and schema-validated.

CF Worker: MockTestAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Recommends mock test configurations including topic mix, difficulty distribution, and time allocation for adaptive practice sessions. Validates that difficulty ratios sum to 1 (tolerance 0.01) and computes exact question counts per tier and an estimated passing score (questionCount × 0.6).

Tools: read_academic_summary  ·  render_card

Outcome: AgentDecision (decisionType: orchestration_intent) with easyCount, mediumCount, hardCount, timeAllotedMinutes, estimatedPassingScore, advisoryBoundary requiring teacher or platform workflow authorization for delivery. Difficulty distribution must sum to 1.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only: agents cannot mutate official exam result truth, student records, or official academic standing. No external communications. All output is advisory, logged, and schema-validated.

CF Worker: MistakeAnalysisAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Identifies recurring mistake patterns across topics and question types and classifies student performance into tiers (strong <20% error rate/developing <40%/needs_support ≥40%). Surfaces remediation priorities for teacher and tutor review by ranking topics by error count.

Tools: read_academic_summary  ·  render_table

Outcome: AgentDecision (decisionType: anomaly_detection) with errorRate, performanceTier (strong/developing/needs_support), topMistakeTopics list (up to 5), incorrectAnswers, attemptedQuestions, advisoryBoundary. Performance intervention requires teacher review and action.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only: agents cannot mutate official exam result truth, student records, or official academic standing. No external communications. All output is advisory, logged, and schema-validated.

CF Worker: ReadinessAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Scores exam readiness from practice performance, gap coverage, and mastery scores across multiple topics. Classifies students as ready/borderline/not_ready relative to a configurable target score (default 60). Identifies weakest topics needing urgent attention.

Tools: read_academic_summary  ·  render_card

Outcome: AgentDecision (decisionType: school_health_score) with avgMastery, readinessStatus (ready/borderline/not_ready), weakestTopics list (up to 5, below targetScore), topicsChecked, advisoryBoundary. Official exam entry decisions are governed by registration rules and administrator oversight.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only: agents cannot mutate official exam result truth, student records, or official academic standing. No external communications. All output is advisory, logged, and schema-validated.

CF Worker: AiTutorAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Provides hint-level tutoring guidance within teacher-configured policy bounds across Olympiad subjects (Mathematics, Science, English, GK). Maintains per-session conversation history (up to 40 messages) in SQLite. Supports hintLevel: none/subtle/moderate/full — full_explanation triggers mandatory teacher review.

Tools: read_academic_summary  ·  render_card

Outcome: AgentDecision (decisionType: communication_draft) with responseType (concept_pointer/leading_question/step_breakdown/full_explanation), requiresTeacherReview flag (true if hintLevel=full), hintLevel, advisoryBoundary blocking official answer evaluation, exam record modification, or replacement of teacher-led instruction. confidenceScore 0.75.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only: agents cannot mutate official exam result truth, student records, or official academic standing. No external communications. All output is advisory, logged, and schema-validated.

CF Worker: MotivationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Generates motivational nudges and progress acknowledgment messages calibrated to recent test scores, study streaks, and days remaining to exam. Selects from three templates (encouragement_excellence/encouragement_progress/encouragement_restart) based on motivationLevel classification (high/medium/low).

Tools: read_academic_summary  ·  render_card

Outcome: AgentDecision (decisionType: communication_draft) with motivationLevel (high/medium/low), messageTemplate, daysToExam, advisoryBoundary requiring teacher or parent portal review before any delivery to students. All student communications require teacher or parent portal review before delivery — no direct dispatch. confidenceScore 0.73.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only: agents cannot mutate official exam result truth, student records, or official academic standing. No external communications. All output is advisory, logged, and schema-validated.

CF Worker: VersaConceptAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Provides concise (2-3 sentence) concept-level explanations calibrated to a student's IRT ability estimate (theta). Accepts message format CONCEPT:{uuid}|THETA:{float}|Q:{question}. Maps theta to level (beginner <-1/intermediate -1 to 1/advanced >1) and prepends ability directive before forwarding to LLM. Never reveals exam answers.

Tools: read_academic_summary  ·  render_card

Outcome: AgentDecision (decisionType: communication_draft) with conceptId, theta, level (beginner/intermediate/advanced), question text, and advisoryBoundary — never reveals exam answers; all content is advisory and calibrated to student ability. confidenceScore 0.80.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only: agents cannot mutate official exam result truth, student records, or official academic standing. No external communications. All output is advisory, logged, and schema-validated.

CF Worker: LearningGapAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Identifies advisory learning gaps by comparing a student's per-topic scores against a configurable mastery threshold. Used in the Academic Intelligence portal (phase7) by teachers, exam-ops coordinators, and the student portal to surface which topics fall below mastery.

Tools: read_academic_summary  ·  render_card  ·  render_table  ·  show_timeline

Outcome: AgentDecision with gapTopics array (ranked by ascending score with per-topic gap size), masteredTopics, overallReadiness (ready/partially_ready/not_ready), confidenceScore (0.82 + mastered count factor), advisory boundary. Cannot mutate official grades or progress records.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only: agents cannot mutate official exam result truth, student records, or official academic standing. No external communications. All output is advisory, logged, and schema-validated.

CF Worker: PracticeRecommendationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Recommends targeted practice sessions for individual students based on diagnosed skill gaps, recent mistake patterns, and Bloom's taxonomy difficulty levels. Actual question generation and delivery require separate teacher-authorized workflow steps.

Tools: read_academic_summary  ·  render_card  ·  render_table  ·  show_timeline

Outcome: AgentDecision with recommended practiceType, topic focus list, difficulty distribution, estimated session duration, and confidenceScore. Output is advisory and cannot directly trigger question delivery or mutate student progress records.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only: agents cannot mutate official exam result truth, student records, or official academic standing. No external communications. All output is advisory, logged, and schema-validated.

CF Worker: AdaptiveMocktestAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Recommends adaptive mock test configurations for students by calibrating topic mix, difficulty distribution, and time allocation based on current ability estimates and performance history. All configuration outputs require teacher or platform workflow authorization before delivery.

Tools: read_academic_summary  ·  render_card  ·  render_table  ·  show_timeline

Outcome: AgentDecision with recommended mock test parameters: question count, time allocation, per-tier difficulty counts, estimated passing score, and confidenceScore. Output is advisory only and cannot directly deliver or score any assessment.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only: agents cannot mutate official exam result truth, student records, or official academic standing. No external communications. All output is advisory, logged, and schema-validated.

CF Worker: QuestionQualityAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Reviews individual question items for quality signals including accuracy, clarity, age-appropriateness, distractor effectiveness, and DPDP policy compliance. Used in the content governance workflow (phase7) before items are published to the item bank.

Tools: read_academic_summary  ·  render_card  ·  render_table  ·  show_timeline

Outcome: AgentDecision with reviewStatus (approved/minor_revision_needed/major_revision_needed), reviewPriority (high if factual errors/medium otherwise), issue count, per-criterion flags, and confidenceScore. Output is advisory; item edits, retirement, and publication decisions require content governance team approval.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only: agents cannot mutate official exam result truth, student records, or official academic standing. No external communications. All output is advisory, logged, and schema-validated.

CF Worker: StudentTutorGuardrailAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Enforces teacher-configured hint policy boundaries and content safety constraints on AI tutor responses before they are delivered to students. Verifies responses do not exceed permitted hint level, contain unsafe content, or bypass subject and grade-level restrictions. All overrides are logged for teacher review.

Tools: read_academic_summary  ·  render_card  ·  render_table  ·  show_timeline

Outcome: AgentDecision with guardrail verdict (permit/downgrade/block), violation signals, hint-policy compliance status, and confidenceScore. Output governs tutor content delivery but cannot itself send messages to students — delivery requires downstream workflow authorization.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only: agents cannot mutate official exam result truth, student records, or official academic standing. No external communications. All output is advisory, logged, and schema-validated.

CF Worker: PsychometricCalibrationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Provides advisory IRT calibration signals for exam items, including difficulty parameter estimates, discrimination index validation, and item fit statistics. All calibration outputs require content governance team approval before item labels or bank records are changed.

Tools: read_academic_summary  ·  render_card  ·  render_table  ·  show_timeline

Outcome: AgentDecision with calibrationAction (recalibrate/confirm/flag_for_review), empirical vs current label comparison, discriminationQuality rating (good/poor), and confidenceScore. Item label changes require content governance approval.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only: agents cannot mutate official exam result truth, student records, or official academic standing. No external communications. All output is advisory, logged, and schema-validated.

CF Worker: AcademicContentSafetyAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Reviews learning content, tutor responses, and practice materials for safety violations including age-inappropriate content, factual misinformation, DPDP policy non-compliance, and harmful language before delivery to students. Pre-publication safety gate operated by editorial administrators.

Tools: read_academic_summary  ·  render_card  ·  render_table  ·  show_timeline

Outcome: AgentDecision with safety verdict (safe/advisory_flag/block), violation category list, humanReviewRequired flag, and confidenceScore. Output gates content delivery but cannot itself publish or delete content — publication requires editorial workflow authorization.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only: agents cannot mutate official exam result truth, student records, or official academic standing. No external communications. All output is advisory, logged, and schema-validated.

CF Worker: LearningProgressAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Aggregates and summarises a student's learning progress across topics, practice sessions, and revision completion to produce an advisory progress snapshot for teachers, parents, and the student portal. Surfaces mastery trajectory, study streak, and readiness delta over time.

Tools: read_academic_summary  ·  render_card  ·  render_table  ·  show_timeline

Outcome: AgentDecision with progress tier (improving/stable/declining), mastery delta, study streak, revision coverage, practice frequency summary, and confidenceScore. Official progress reports, grade changes, and parent communications require teacher and academic coordinator authorization.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only: agents cannot mutate official exam result truth, student records, or official academic standing. No external communications. All output is advisory, logged, and schema-validated.

CF Worker: PortalWorkspaceRouterAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Routes authenticated users to the correct portal workspace based on their assigned role (student/parent/teacher/school_coordinator/vendor/finance/support). Evaluates session role context and emits a redirect or workspace-load instruction so the portal shell renders the appropriate dashboard.

Tools: read_portal_context  ·  render_card

Outcome: Structured routing decision with resolved portal workspace path, active session flag, and render_card payload for the portal shell. Actual navigation is enforced by the Next.js proxy (src/proxy.ts). NEVER redirect users to a workspace they do not have an RLS-verified role for.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: ParentPortalSummaryAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Generates a consolidated summary dashboard for parents aggregating their linked children's registration status, upcoming exam dates, recent results, and outstanding fee dues. Serves the /parent dashboard route in Phase 8. Distinct from parent_summary_agent — operates at portal-session level across potentially multiple children.

Tools: read_portal_context  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Structured portal summary with per-child status cards, timeline of upcoming exam dates and deadlines, outstanding fee alerts, and advisory recommendation — all scoped to the authenticated parent's linked children. NEVER surface data for children not linked to the authenticated parent via RLS.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: StudentPortalCompanionAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Acts as a conversational assistant embedded in the student portal, helping students navigate their exam journey — answering questions about admit cards, exam dates, result status, and study recommendations. Synthesises the student's current phase context and surfaces the most relevant next action.

Tools: read_portal_context  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Phase-aware summary card with the student's most relevant next action, plus answers to direct queries about registration, exam, results, or certificate — all scoped strictly to their own student record. NEVER expose another student's data. NEVER confirm result publication before qualification_status is set.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: TeacherPortalAssistantAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Provides teachers with a unified portal assistant covering class performance overviews, student at-risk signals, grading backlog status, and invigilator duty reminders. Combines the advisory insight capability of teacher_insight_agent with portal-session awareness. Serves all /teacher/* routes in Phase 8.

Tools: read_portal_context  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Class health dashboard card, prioritised list of at-risk students requiring attention, grading backlog status, and exam-day duty reminders — all advisory and scoped to teacher's own assigned classes. NEVER modify official grade records.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: SchoolPortalOpsAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Provides school coordinators with a comprehensive operational status view of their school's Olympiad journey — synthesising registration health, exam kit readiness, invoice dues, and certificate pipeline into a single ops-level advisory. Primary agent for the school dashboard (/school) home panel.

Tools: read_portal_context  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Single-screen ops dashboard summarising registration health, exam readiness, finance status, and certificate pipeline with a prioritised action list and overall school Olympiad health score. NEVER expose data from other schools. NEVER trigger any operational workflow.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: VendorPortalStatusAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Surfaces consolidated task and order status for vendors accessing the vendor portal, showing open purchase orders, pending deliveries, overdue shipments, and outstanding coordination items. Portal-facing counterpart to vendor_coordination_portal_agent for the vendor's own authenticated session on /vendor.

Tools: read_portal_context  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Vendor dashboard card with open order count, pending delivery count, overdue items, and prioritised action items — all scoped to the authenticated vendor account. NEVER confirm dispatch or delivery — logistics mutations require ops_admin workflow.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: FinancePortalAssistantAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Assists finance staff and school coordinators on the finance portal by answering questions about invoices, payment reconciliation, outstanding dues, and fee collection analytics. Targets finance-role users who need cross-school summaries, reconciliation status, and collection-rate analytics.

Tools: read_portal_context  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Collection analytics summary with outstanding balance, overdue count, reconciliation status, and anomaly flags — advisory and read-only, scoped to the authenticated user's access level. NEVER process or confirm payments; reconciliation performed by the deterministic finance workflow.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: SupportPortalAssistantAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Provides support staff with a conversational assistant on the support portal for triaging inbound tickets, looking up school and student context, and drafting scripted responses. Extends the routing logic of support_routing_agent within an authenticated support-staff portal session.

Tools: read_portal_context  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Ticket classification, contextual school/student data, draft scripted response, and escalation recommendation — advisory only; ticket resolution and communications require authorised support workflow execution. NEVER send communications to schools or parents directly.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: NotificationPreferenceAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Helps portal users (students, parents, school coordinators, teachers) review and understand their current notification preferences — channels enabled (email, SMS, in-app), notification categories opted in or out, and digest frequency. Serves the notification preferences panel within any portal role. Does not mutate preferences directly.

Tools: read_portal_context  ·  render_card  ·  render_table

Outcome: Plain-language summary of notification preferences, list of active channels and opted-in categories, and advisory recommendations for any critical categories that appear disabled — read-only advisory output directing the user to the settings page. NEVER mutate notification preferences.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: MultiPortalPermissionAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Advises on cross-portal permission and access-scoping questions — explaining what data categories and portal routes a given role is permitted to access, and flagging any permission anomalies in the authenticated session. Translates technical permission boundaries into plain-language explanations.

Tools: read_portal_context  ·  render_card  ·  render_table

Outcome: Plain-language permission summary for the queried role, list of permitted portal routes and data categories, any detected permission anomalies, and escalation recommendation if anomalies are present — all advisory. Actual permission changes require super_admin action through the RLS configuration workflow.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: StudentStudyAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Surfaces personalised study recommendations on the student portal. Given today's study hours, weekly totals, and upcoming exam schedule, classifies study status as on_track/light/minimal and recommends an adjusted daily target. Adds 2h if urgent exam (≤7 days), else 1h, capped at 8h.

Tools: read_portal_context  ·  render_card

Outcome: AgentDecision (decisionType: communication_draft) with studyStatus classification, urgentExamName and daysUntilUrgentExam (null if none), recommendedDailyHours, advisory boundary. confidenceScore fixed at 0.72.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: ParentSummaryAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Generates parent-facing progress summaries and key highlights for the parent portal. Given a student's recent test scores and attendance rate, produces a performanceSummary (excellent/good/needs_attention) and attendanceSummary (excellent/satisfactory/low).

Tools: read_portal_context  ·  render_card

Outcome: AgentDecision (decisionType: school_health_score) with avgScore, performanceSummary, attendanceSummary, attendanceRate, upcomingExamDays, advisory boundary. Confidence score ranges 0.70-0.85. Performance interventions require teacher communication through official channels.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: TeacherInsightAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Surfaces class-level performance insights for teacher review on the teacher portal. Given class composition (total students, above/below average counts) and grading backlog, computes a class health score and trend classification.

Tools: read_portal_context  ·  render_table

Outcome: AgentDecision (decisionType: school_health_score) with classHealthScore (aboveAverage/totalStudents × 100), classTrend (performing_well/mixed/needs_attention), gradingBacklog (high/medium/low), pendingAssignmentGrading, upcomingClassDays. confidenceScore fixed at 0.74.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: TeacherInterventionAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Recommends targeted interventions for at-risk students and surfaces the appropriate communication template for teacher use. Triggered by interventionTrigger (low_test_score/attendance_drop/missing_assignments/parent_request).

Tools: read_portal_context  ·  render_card

Outcome: AgentDecision (decisionType: communication_draft) with interventionTrigger, urgency level (high/medium), and selected interventionTemplate name. confidenceScore fixed at 0.71. Teacher communications must be sent through official school communication channels.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: VendorCoordinationPortalAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Surfaces vendor task status and action items on the vendor portal. Given a vendor's open orders, pending deliveries, and overdue items, classifies coordination status (escalation_needed/attention_required/on_track) and contact priority.

Tools: read_portal_context  ·  render_table

Outcome: AgentDecision (decisionType: vendor_risk_score) with coordinationStatus, contactPriority, openOrders, pendingDeliveries, overdueItems, advisory boundary. Confidence score ranges ~0.23-0.73 depending on overdue count. Purchase orders and vendor communications require ops_admin action.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: SupportRoutingAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Classifies and routes support tickets to the correct internal team (tech_support/finance_ops/exam_ops/registration_ops/certificate_ops). Computes SLA hours based on urgency (critical=2h/high=8h/medium=24h/low=48h) and sets escalateToManager flag for critical urgency.

Tools: read_portal_context  ·  render_card

Outcome: AgentDecision (decisionType: reply_classification) with assignedTeam, slaHours, escalateToManager flag, issueCategory, urgency, reportedByRole. confidenceScore fixed at 0.78.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: TenantAdminAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Assists tenant admins with configuration and onboarding monitoring on the admin portal. Given a tenant's active user count, storage used, subscription tier, and enabled feature flags, assesses storage health and user density and surfaces actionable admin recommendations.

Tools: read_portal_context  ·  render_card

Outcome: AgentDecision (decisionType: anomaly_detection) with storageHealthStatus (critical/warning/healthy), userDensity (activeUsers/userCapacity), featureFlagsCount, adminActions array. Confidence score ranges 0.70-0.75 based on storage health. Tenant configuration changes require super_admin action.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: SchoolCoordinatorPortalAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Helps school coordinators understand their registration status, batch uploads, team members, and school journey through the Olympiad process on the /school/* portal routes. Reads live Supabase data via registered tools.

Tools: get_portal_registration_status  ·  get_portal_batch_summary  ·  get_portal_team_members  ·  read_portal_context  ·  render_card

Outcome: AgentDecision (decisionType: school_health_score) with registrationRate, registeredStudents/totalStudents counts, activeBatches, pendingExceptions, and recommendations array. Live data fetched from registration_batches, students, and school_sub_users tables.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: InvoiceAssistantPortalAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Helps school coordinators understand their invoices, outstanding dues, payment history, and reconciliation status on the /school/invoices portal page. Reads live Supabase invoice and payment data and surfaces total outstanding balances, overdue counts, and payment instructions.

Tools: get_portal_invoice_list  ·  get_portal_outstanding_dues  ·  get_portal_payment_history  ·  read_portal_context  ·  render_card  ·  render_table

Outcome: AgentDecision (decisionType: anomaly_detection) with invoiceCount, totalOutstanding (INR), overdueCount, lastPaymentDate, and advisory boundary. Live invoice and payment records from invoices and payments tables. NEVER mark invoices paid or process payments.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: ExamMonitorPortalAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Helps school coordinators monitor exam readiness — admit card availability, exam schedule, invigilator assignments, and OMR upload guidance on the /school/exam-kit portal route. Surfaces a readiness checklist.

Tools: get_portal_exam_kit_status  ·  get_portal_exam_schedule  ·  get_portal_invigilator_assignments  ·  read_portal_context  ·  render_card  ·  show_timeline

Outcome: AgentDecision (decisionType: school_health_score) with examReady boolean, readinessFlags array, invigilatorsAssigned count, admitCardsReady and omrUploaded booleans. Live session and invigilator data from exam_sessions and school_sub_users tables. NEVER schedule exams or assign invigilators.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: SchoolSummaryPortalAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Helps school coordinators review student results, school performance analytics, and understand how their school performed in the Olympiad. Serves /school/results and /school dashboard routes. Uses two-step join (students.school_id → result_records.student_id) to scope results per school.

Tools: get_portal_student_results  ·  get_portal_school_performance  ·  read_portal_context  ·  render_card  ·  render_table

Outcome: AgentDecision (decisionType: school_health_score) with performanceRating (excellent/good/average/below_average), averageScore, passRate, topRank, totalStudentsExamined. Live results from result_records scoped via students.school_id join. NEVER modify or recalculate published results.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: CertificateGenerationPortalAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Helps school coordinators check certificate status, track which students have received certificates, and understand the certificate download process on the /school/certificates portal route. Uses two-step join (students.school_id → certificates.student_id) to scope data.

Tools: get_portal_certificate_status  ·  get_portal_certificate_list  ·  read_portal_context  ·  render_card  ·  render_table

Outcome: AgentDecision (decisionType: school_health_score) with totalEligible, certificatesGenerated, certificatesDistributed, pendingDistribution, pendingGeneration, distributionRate. Live data from certificates table scoped via students.school_id join. NEVER generate, revoke, or modify certificates.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: ExamReadinessAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Interprets client-reported device and network check results to tell a student whether their device is ready for an online exam. Evaluates 5 checks: browser compatibility, camera access, audio access, network speed (minimum 2 Mbps), and fullscreen support.

Tools: read_portal_context  ·  render_card

Outcome: AgentDecision (decisionType: school_health_score) with ready boolean, issues string array, checksPassed count (0-5), checksTotal (5). Confidence 0.95 if ready, 0.75 otherwise. Agent cannot grant or deny exam access — exam entry is controlled by the exam platform.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: RankingDisplayAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Displays a student's Olympiad ranking, percentile, and score with a motivational narrative on the student portal results page. Produces a human-readable performance tier and encouragement message calibrated to the result.

Tools: read_portal_context  ·  render_card

Outcome: AgentDecision (decisionType: school_health_score) with score, percentile, rank (all nullable), and narrative string. Official ranking data lives in result_records — this agent displays it for informational purposes only. Confidence 0.95 if rank is available, 0.70 otherwise.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: ProgressHistoryAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Narrates a student's multi-year Olympiad performance trajectory by analysing score history across exam cycles. Classifies the trend as improving/declining/stable/insufficient_data and provides a personalised narrative.

Tools: read_portal_context  ·  render_table

Outcome: AgentDecision (decisionType: school_health_score) with cycles array, trend enum (improving/declining/stable/insufficient_data), and trendNarrative string. Confidence 0.90 if ≥2 cycles, 0.60 for first-time student.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: ParentConsentAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Provides parents with a transparent advisory summary of what personal data is held about their children and how to exercise their rights under DPDP Act 2023. Displays consent version, data categories, and directs withdrawal requests to the support team. Serves /parent/* routes.

Tools: read_portal_context  ·  render_card

Outcome: AgentDecision (decisionType: compliance_status_check) with hasChildren boolean, studentCount, dataCategories string, and consentVersion. Consent mutation requires human workflow approval and audit trail. NEVER auto-process consent withdrawals or delete personal data. Confidence 0.90 with confirmed children, 0.75 without.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: ParentPaymentAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Provides parents with an advisory payment status summary — outstanding invoice count, total amount due, and next steps. Classifies urgency (high >₹5000/medium >0/none) and directs parents to the Invoices page for payment via school payment link.

Tools: read_portal_context  ·  render_table

Outcome: AgentDecision (decisionType: finance_payment_advisory) with allPaid boolean, outstandingCount, totalOutstandingAmount, currency, and urgency level. Confidence score fixed at 0.90. Actual payment processing is external to this system. NEVER process payments or mark invoices paid.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. NEVER expose data outside the authenticated user's RLS-verified scope. All output is advisory and read-only; mutations require authorised human workflows.

CF Worker: ProvenanceAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Traces the authorship and revision history of content items to assess chain-of-custody integrity. Scores provenance based on chain length and flags items with unclear, contested, or undocumented origin for IP review.

Tools: read_compliance_controls  ·  render_card

Outcome: AgentDecision with provenanceScore (0-1), provenanceStatus (clean ≥0.95/review_needed ≥0.80/suspect <0.80), chain length, source system, content type, prioritised recommendation, and advisoryBoundary statement. forceHumanReview=true always.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. forceHumanReview is always true — output is advisory only. Agent cannot block, remove, modify, publish, or freeze any content item autonomously. All consequential actions require super_admin or ops_admin approval through the content governance workflow.

CF Worker: RightsCheckAssistant extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Evaluates content usage rights by cross-referencing license type against the intended usage context. Flags potential rights violations (e.g., proprietary content distributed outside internal systems) and recommends legal review before publication or distribution.

Tools: read_compliance_controls  ·  render_card

Outcome: AgentDecision with rightsStatus (cleared/unverified/potential_violation), requiresLegalReview boolean, license type, usage context, confidence score, and recommendation. All outputs are advisory; licensing decisions require legal team approval. forceHumanReview=true always.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. forceHumanReview is always true — output is advisory only. Agent cannot block, remove, modify, publish, or freeze any content item autonomously. All consequential actions require super_admin or ops_admin approval through the content governance workflow.

CF Worker: SimilarityReviewAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Scores content similarity against the existing exam corpus using n-gram and structural matching signals. Classifies plagiarism risk as low/medium/high and flags items exceeding configurable thresholds for editorial review.

Tools: read_compliance_controls  ·  render_table

Outcome: AgentDecision with maxSimilarity (rounded to 2dp), highSimilarityCount, plagiarismRisk (low/medium/high), referenceCount, confidence score, and recommendation. maxSimilarity >0.9 → high; >0.7 → medium; else low. Plagiarism determination requires academic committee and legal review.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. forceHumanReview is always true — output is advisory only. Agent cannot block, remove, modify, publish, or freeze any content item autonomously. All consequential actions require super_admin or ops_admin approval through the content governance workflow.

CF Worker: AcademicReviewAssistant extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Reviews academic content for factual accuracy, difficulty calibration, and curriculum alignment against Bloom's taxonomy and syllabus standards. Produces a review verdict (approved/revision_required/rejected) and a priority classification for flagged items.

Tools: read_compliance_controls  ·  render_card

Outcome: AgentDecision with reviewVerdict (approved/revision_required/rejected), reviewPriority (critical if factual error/medium otherwise), factualAccuracy, difficultyCalibration, curriculumAlignment ratings, issueCount, confidence score. Content approval requires authorized reviewer sign-off.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. forceHumanReview is always true — output is advisory only. Agent cannot block, remove, modify, publish, or freeze any content item autonomously. All consequential actions require super_admin or ops_admin approval through the content governance workflow.

CF Worker: AnswerKeyReviewAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Validates answer key completeness and flags disputed or unreviewed answers ahead of candidate evaluation. Computes a verification rate and surfaces answer key status (verified/disputed/pending_review) with a risk classification.

Tools: read_compliance_controls  ·  render_table

Outcome: AgentDecision with verificationRate, keyStatus (verified/disputed/pending_review), riskLevel (high >5 disputed/medium >0/low none), counts of verified/disputed/unreviewed answers, confidence score, recommendation. Answer key changes require exam controller approval.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. forceHumanReview is always true — output is advisory only. Agent cannot block, remove, modify, publish, or freeze any content item autonomously. All consequential actions require super_admin or ops_admin approval through the content governance workflow.

CF Worker: TranslationQaAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Assesses translated content quality against the source using word-count deviation, terminology issues, grammatical errors, and cultural adaptation signals. Outputs a QA score and status (pass/revision_required/failed) to guide professional translator review.

Tools: read_compliance_controls  ·  render_card

Outcome: AgentDecision with qaScore (0-1), qaStatus (pass ≥0.9/revision_required ≥0.7/failed <0.7), wordCountDeviation, terminologyIssues, grammaticalErrors, culturalAdaptationNeeded flag, totalIssues, recommendation. Translation revision and approval require professional translator and academic review.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. forceHumanReview is always true — output is advisory only. Agent cannot block, remove, modify, publish, or freeze any content item autonomously. All consequential actions require super_admin or ops_admin approval through the content governance workflow.

CF Worker: ContentQualityAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Scores content items across readability, completeness, media assets, broken links, and accessibility to determine publication readiness. Flags items below quality thresholds with publishable/needs_revision/blocked status. The only Group 10 agent with live database tools.

Tools: get_review_queue  ·  get_ip_flags  ·  get_answer_key_conflicts  ·  get_translation_queue  ·  get_leakage_risks  ·  get_governance_log  ·  approve_content_item  ·  reject_content_item  ·  flag_for_takedown  ·  approve_translation  ·  read_compliance_controls  ·  render_table

Outcome: AgentDecision with qualityScore (0-1), qualityStatus (publishable ≥0.85/needs_revision ≥0.70/blocked <0.70), readabilityScore, completenessPercent, brokenLinksCount, accessibilityIssues, recommendation. approve_content_item, reject_content_item, approve_translation require explicit human confirmation. flag_for_takedown requires approval gate.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. forceHumanReview is always true — output is advisory only. Agent cannot block, remove, modify, publish, or freeze any content item autonomously. All consequential actions require super_admin or ops_admin approval through the content governance workflow.

CF Worker: TakedownTriageAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Triages incoming content takedown requests by classifying urgency, routing to the appropriate review authority (super_admin_and_legal or ops_admin_and_legal), and recommending whether a content freeze is warranted. Handles reports from automated scans, school admins, legal notices, and regulatory bodies.

Tools: read_compliance_controls  ·  render_card

Outcome: AgentDecision with urgency (critical/high/medium), reviewRoute (super_admin_and_legal for critical; ops_admin_and_legal for high/medium), contentFreeze recommendation (boolean, true only for critical urgency), reportReason, reportedBy, confidence score, recommendation. Content removal and regulatory responses require super_admin and legal team approval.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. forceHumanReview is always true — output is advisory only. Agent cannot block, remove, modify, publish, or freeze any content item autonomously. All consequential actions require super_admin or ops_admin approval through the content governance workflow.

CF Worker: ExamLeakageRiskAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Evaluates signals of potential exam content leakage including search-engine exposure, social media references, unauthorised downloads, and suspicious IP access patterns. Produces a composite riskScore and leakage risk level, and recommends containment actions.

Tools: read_runtime_status  ·  render_table

Outcome: AgentDecision with riskScore (0-1), leakageRisk (critical >0.50/high >0.20/medium >0.05/low ≤0.05), leakageSignals, socialMediaMentions, unauthorizedDownloads, suspiciousIPCount, recommendedActions array, confidence score. Exam postponement, legal action, and security escalation require exam controller and legal team approval.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. forceHumanReview is always true — output is advisory only. Agent cannot block, remove, modify, publish, or freeze any content item autonomously. All consequential actions require super_admin or ops_admin approval through the content governance workflow.

CF Worker: ContentPublishApprovalAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Aggregates the results of all upstream review gates (academic, rights, quality, answer key) and tracks collected versus required human sign-offs to determine publish readiness. Surfaces a publish checklist for authorized human reviewers when all gates are passed, or blocks publication when hard gates are unresolved.

Tools: read_compliance_controls  ·  render_card

Outcome: AgentDecision with publishApprovalStatus (ready_for_human_publish_approval/not_ready/blocked), blockingGates count, missingSignoffs list, all upstream review statuses, confidence score. Actual content publication requires authorized human approval and a separate audited publish action. Approved roles: content_reviewer, academic_reviewer, super_admin.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. forceHumanReview is always true — output is advisory only. Agent cannot block, remove, modify, publish, or freeze any content item autonomously. All consequential actions require super_admin or ops_admin approval through the content governance workflow.

CF Worker: AiGuardrailAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Evaluates incoming AI prompts and completions against registered guardrail policies, detecting jailbreak attempts, blocked keywords, and sensitive-data exposure. Classifies the overall threat level as low/medium/high/critical and recommends an action (allow_with_logging/flag_for_monitoring/block_for_review/block_and_escalate).

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with requestId, agentName, jailbreakAttempted flag, triggeredPolicies list, blockedKeywords list, threatLevel, actionRecommended, advisoryBoundary note, confidenceScore (0.96 − triggeredPolicies.length × 0.01 − jailbreakAttempted ? 0.05 : 0). All enforcement requires super_admin approval.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. All output is advisory only — actual enforcement, blocking, policy updates, and model changes are implemented exclusively in the deterministic middleware layer or require super_admin approval through the AI governance workflow. humanReviewRequired is unconditionally true.

CF Worker: PromptGovernanceAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Classifies prompt templates for safety, bias, and compliance alignment before deployment. Detects PII-extraction requests, system-prompt-leak attempts, and instruction injection risks, then assigns promptStatus (approved/review_required/rejected).

Tools: read_compliance_controls  ·  render_card

Outcome: AgentDecision with promptId, agentName, promptVersion, governanceIssues array, promptStatus (approved/review_required/rejected), advisoryBoundary note, confidenceScore (0.95 − governanceIssues.length × 0.05). Agent cannot deploy, modify, or roll back prompt versions. Prompt rejection requires super_admin approval.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. All output is advisory only — actual enforcement, blocking, policy updates, and model changes are implemented exclusively in the deterministic middleware layer or require super_admin approval through the AI governance workflow. humanReviewRequired is unconditionally true.

CF Worker: AiDecisionAuditAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Audits AI decision log entries for scope compliance (modelDecisionScope must be advisory), confidence threshold adherence (decisions below 0.85 must have humanReviewRequired flagged), and required input data category documentation. Enforces the advisory-only constraint across all 288 agent implementations.

Tools: read_runtime_status  ·  render_table

Outcome: AgentDecision with decisionId, agentName, decisionType, modelDecisionScope, auditStatus (compliant/non_compliant_scope/review_threshold_not_met/missing_data_categories), complianceScore, humanReviewRequired, inputDataCategoriesCount, advisoryBoundary. Agent CANNOT modify or delete logged AI decisions.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. All output is advisory only — actual enforcement, blocking, policy updates, and model changes are implemented exclusively in the deterministic middleware layer or require super_admin approval through the AI governance workflow. humanReviewRequired is unconditionally true.

CF Worker: ComplianceAssistantAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Evaluates AI governance posture against active compliance controls across frameworks including DPDP, CERT-In, RBI, ISO 27001, and OWASP Top 10. Scores each control's risk level and assigns remediationUrgency (immediate/within_30_days/routine).

Tools: read_compliance_controls  ·  render_table

Outcome: AgentDecision with controlId, controlName, framework, currentStatus (compliant/partial/non_compliant/not_assessed), riskScore, remediationUrgency (immediate/within_30_days/routine), findingCount, advisoryBoundary. Remediation actions, compliance certifications, and regulatory submissions require compliance team and management approval.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. All output is advisory only — actual enforcement, blocking, policy updates, and model changes are implemented exclusively in the deterministic middleware layer or require super_admin approval through the AI governance workflow. humanReviewRequired is unconditionally true.

CF Worker: McpToolSecurityAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Validates MCP tool invocations against four security checks: policy registration, schema validation, permission scope, and rate limits. Derives securityStatus (approved/conditional/blocked) and sets blockRecommended when permission checks fail or two or more checks fail.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with toolName, requestingAgentName, individual check results (toolPolicyRegistered/schemaValidationPassed/permissionCheckPassed/rateWithinLimits), failedChecks count, securityStatus, blockRecommended flag, advisoryBoundary, confidenceScore (0.97 − failedChecks × 0.05). Tool blocking and permission changes require super_admin approval.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. All output is advisory only — actual enforcement, blocking, policy updates, and model changes are implemented exclusively in the deterministic middleware layer or require super_admin approval through the AI governance workflow. humanReviewRequired is unconditionally true.

CF Worker: EvaluationReviewAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Monitors model evaluation metrics against historical baseline thresholds to detect performance regressions. Computes delta between current and baseline scores, classifies result as regression_detected/acceptable_degradation/improvement, and assigns alertLevel (critical/warning/none).

Tools: read_runtime_status  ·  render_table

Outcome: AgentDecision with evaluationId, modelName, taskType, metricName, baseline, current, delta, sampleSize, isRegression flag, evaluationStatus (regression_detected/improvement/acceptable_degradation), alertLevel (critical if |delta| ≥0.1/warning if |delta| <0.1/none), advisoryBoundary. Model rollbacks require ML engineering team and ops_admin approval.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. All output is advisory only — actual enforcement, blocking, policy updates, and model changes are implemented exclusively in the deterministic middleware layer or require super_admin approval through the AI governance workflow. humanReviewRequired is unconditionally true.

CF Worker: OpsControlTowerAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Provides a unified cross-domain operational status summary spanning finance, production, exam, and academic domains for ops_admin users. Surfaces live exception queues, SLA breach counts, pending approvals, and overall system health, and derives a priority action list.

Tools: get_registration_upload_summary  ·  get_schools_operations_summary  ·  get_exam_sessions_status  ·  get_production_batch_status  ·  get_results_publication_status  ·  get_online_exam_monitor_status  ·  get_academic_intelligence_summary  ·  get_operations_finance_summary  ·  assign_exception  ·  resolve_exception  ·  escalate_exception  ·  hold_batch  ·  approve_batch  ·  read_runtime_status  ·  render_table

Outcome: AgentDecision with opsStatus (critical/degraded/at_risk/healthy), criticalExceptions, slaBreaches, openExceptions, pendingApprovals, systemHealthScore, priorityActions array, confidenceScore, advisoryBoundary. resolve_exception and approve_batch require Approval gate (high-risk). assign_exception, escalate_exception, hold_batch require Confirm step.

Guardrails: HARD-BLOCKED: post_ledger_entry, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls. Advisory only — all remediation actions must flow through official ops_admin workflows. forceHumanReview is unconditionally true. All output is schema-validated and logged. RLS is never bypassed.

CF Worker: ExceptionPrioritizationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Ranks individual operational exceptions by composite impact score derived from financial impact (40%), schools affected (30%), and exam days impacted (30%). Outputs a priority tier (P1-P4) and escalation route (super_admin_immediate/ops_admin_same_day/standard_queue).

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with exceptionId, exceptionType, affectedDomain (10 permitted enum values), financialImpact, schoolsAffected, examDaysImpacted, impactScore (0-1), priority (P1 ≥0.7/P2 ≥0.4/P3 ≥0.2/P4 otherwise), escalationRoute, confidenceScore, advisoryBoundary.

Guardrails: HARD-BLOCKED: post_ledger_entry, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls. Advisory only — all remediation actions must flow through official ops_admin workflows. forceHumanReview is unconditionally true. All output is schema-validated and logged. RLS is never bypassed.

CF Worker: SlaRiskPredictionAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Predicts SLA breach probability for a named operational process by comparing current throughput against required throughput and time remaining until deadline. Surfaces breach risk as breached/at_risk/on_track to allow early intervention.

Tools: read_runtime_status  ·  render_table

Outcome: AgentDecision with processName, currentThroughput, requiredThroughput, throughputGap, hoursRemaining, estimatedHoursToComplete, backlogSize, slaRisk (breached/at_risk/on_track), confidenceScore, advisoryBoundary. SLA escalation and resource allocation require ops_admin approval.

Guardrails: HARD-BLOCKED: post_ledger_entry, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls. Advisory only — all remediation actions must flow through official ops_admin workflows. forceHumanReview is unconditionally true. All output is schema-validated and logged. RLS is never bypassed.

CF Worker: OperationalHealthAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Scores the operational health of a named system using uptime percentage (40%), error rate (40%), average response time penalty (10%), alert penalty (1% per alert), and incident penalty (2% per incident). Classifies system status as healthy/degraded/critical.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with systemName, uptimePercent, errorRate, avgResponseMs, activeAlerts, recentIncidents, healthScore (0-1), healthStatus (healthy ≥0.9/degraded ≥0.7/critical otherwise), confidenceScore (= healthScore), advisoryBoundary. Incident response and system restarts require ops_admin approval.

Guardrails: HARD-BLOCKED: post_ledger_entry, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls. Advisory only — all remediation actions must flow through official ops_admin workflows. forceHumanReview is unconditionally true. All output is schema-validated and logged. RLS is never bypassed.

CF Worker: DigitalTwinSimulationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Runs what-if operational simulations — including exam-day load surges, vendor failures, and registration spikes — against a baseline metrics snapshot without touching live data. Produces projected impact percentages with confidence bands and reliability classifications.

Tools: read_runtime_status  ·  render_table

Outcome: AgentDecision with simulationId, scenarioName, projectedImpactPercent, confidenceBand, simulationRunCount, metricsSimulated, impactDirection (positive/negative/neutral), projectionReliability (high ≥0.8/medium ≥0.5/low), simulationDisclaimer, advisoryBoundary. Simulations NEVER overwrite or modify operational truth. Management sign-off required before any action on simulation results.

Guardrails: HARD-BLOCKED: post_ledger_entry, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls. Advisory only — all remediation actions must flow through official ops_admin workflows. forceHumanReview is unconditionally true. All output is schema-validated and logged. RLS is never bypassed.

CF Worker: SchoolOpsSummaryAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Generates per-exam-cycle operational health summaries tracking school progress through 5 pipeline stages: registration, fee payment, materials dispatch, exam conduct, and results publication. Derives stage completion rates and cycleStatus (on_track/partial/behind).

Tools: read_runtime_status  ·  render_table

Outcome: AgentDecision with examCycleId, totalSchools, stageRates object (5 stage completion ratios), overallCompletion (0-1), cycleStatus (on_track ≥0.9/partial ≥0.6/behind otherwise), confidenceScore, advisoryBoundary.

Guardrails: HARD-BLOCKED: post_ledger_entry, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls. Advisory only — all remediation actions must flow through official ops_admin workflows. forceHumanReview is unconditionally true. All output is schema-validated and logged. RLS is never bypassed.

CF Worker: FinancialAnomalyAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Detects unusual payment patterns, revenue deviations, and collection anomalies by analysing transaction-level signals. Produces an anomalyScore and risk tier (high/medium/low) with an investigation priority recommendation. Identifies periods requiring audit and compliance escalation.

Tools: read_runtime_status  ·  render_table

Outcome: AgentDecision with reportingPeriod, totalTransactions, suspiciousTransactions, duplicatePayments, roundAmountFlags, unusualTimingFlags, anomalyRate, anomalyScore (0-1, clamp(anomalyRate×5 + roundAmountFlags×0.001 + unusualTimingFlags×0.001)), anomalyRisk (high ≥0.5/medium ≥0.2/low otherwise), advisoryBoundary. Agent CANNOT modify financial records.

Guardrails: HARD-BLOCKED: post_ledger_entry, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls. Advisory only — all remediation actions must flow through official ops_admin workflows. forceHumanReview is unconditionally true. All output is schema-validated and logged. RLS is never bypassed.

CF Worker: VendorRiskAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Scores vendor risk from delivery reliability score, quality score, on-time delivery rate, payment dispute count, and contract value. Produces a composite riskScore and tier (low/medium/high/critical). Recommends targeted actions such as escalate_to_procurement or schedule_vendor_review.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with vendorId, vendorName, contractValue, deliveryReliabilityScore, qualityScore, paymentDisputeCount, onTimeDeliveryRate, riskScore (0-1), vendorRiskTier (low <0.2/medium <0.4/high <0.6/critical ≥0.6), vendorActions array, advisoryBoundary. Contract actions, payments, and vendor termination require procurement and finance team approval.

Guardrails: HARD-BLOCKED: post_ledger_entry, change_score, publish_results, revoke_certificates, delete_audit_logs, bypass_rls. Advisory only — all remediation actions must flow through official ops_admin workflows. forceHumanReview is unconditionally true. All output is schema-validated and logged. RLS is never bypassed.

CF Worker: ItemDifficultyAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Classifies empirical item difficulty from candidate response data using facility index (pCorrect), distractor engagement rates, and average time spent. Assigns a difficulty level (easy/medium/hard) and an item quality verdict (good/review_needed/retire) and flags items where response patterns fall outside normal ranges.

Tools: read_academic_data  ·  render_table

Outcome: AgentDecision with itemId, totalAttempts, pCorrect, difficultyLevel (easy ≥0.70/medium 0.40-0.69/hard <0.40), distractorEngagement, avgTimeSpentSeconds, itemQuality (good/review_needed/retire), confidenceScore (base 0.88 scaled by attempt count up to 500), forceHumanReview=true.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Advisory only — agents cannot mutate official exam result truth, student records, item labels, or official academic standing. Item retirement, difficulty label changes, and content edits require content team approval through the item governance workflow.

CF Worker: AbilityEstimationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Estimates a student's latent ability level from their item response pattern using a simplified IRT signal. Takes per-item correctness and difficulty, applies a prior ability estimate, and outputs an ability score on a -4 to +4 scale with a proficiency band. For formative insight only — not linked to official results.

Tools: read_academic_data  ·  render_card

Outcome: AgentDecision with candidateId, itemCount, correctCount, avgDifficulty, abilityEstimate (float -4 to +4), proficiencyBand (below_basic/basic/proficient/advanced), confidenceScore (base 0.75 scaled by item count up to 40), advisoryBoundary confirming estimates do not affect official standing.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Advisory only — agents cannot mutate official exam result truth, student records, item labels, or official academic standing. Item retirement, difficulty label changes, and content edits require content team approval through the item governance workflow.

CF Worker: LearningGapAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Identifies learning gaps by comparing a student's per-topic scores against a configurable mastery threshold. Ranks gap topics by deficit magnitude, lists mastered topics, and assigns an overall readiness verdict (ready/partially_ready/not_ready).

Tools: read_academic_data  ·  render_table

Outcome: AgentDecision with studentId, gapTopics array (topic, score, gap delta), masteredTopics array, overallReadiness (ready if 0 gaps/partially_ready if ≤30%/not_ready >30%), masteryThreshold, confidenceScore (0.82 + masteredTopics.length × 0.01), recommendation with top-3 focus topics. Curriculum changes require teacher approval.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Advisory only — agents cannot mutate official exam result truth, student records, item labels, or official academic standing. Item retirement, difficulty label changes, and content edits require content team approval through the item governance workflow.

CF Worker: MisconceptionTrackingAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Tracks recurring misconceptions at the item level by analyzing distractor selection rates and their associated misconception labels. Identifies the dominant misconception, classifies its prevalence (high ≥0.40/moderate 0.20-0.39/low <0.20), and raises a curriculum alert flag when high-prevalence misconception coincides with low correct-option uptake.

Tools: read_academic_data  ·  render_card

Outcome: AgentDecision with itemId, topMisconception label, topMisconceptionRate, misconceptionPrevalence, correctOptionRate, needsCurriculumAlert boolean, confidenceScore (0.84 − (1 − correctOptionRate) × 0.10), forceHumanReview = needsCurriculumAlert. Curriculum corrections require content team and academic coordinator approval.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Advisory only — agents cannot mutate official exam result truth, student records, item labels, or official academic standing. Item retirement, difficulty label changes, and content edits require content team approval through the item governance workflow.

CF Worker: AdaptiveRecommendationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Recommends an adaptive practice plan for a student by computing a target difficulty level aligned to their current IRT ability estimate and Bloom's taxonomy progression. Selects up to 5 priority weak topics, determines an appropriate item count, and flags readiness for advanced content.

Tools: read_academic_data  ·  render_card

Outcome: AgentDecision with studentId, targetDifficulty (float 0.10-0.90, set just above current ability), priorityTopics (top 5 weak topics), recommendedItemCount (clamp(10 - masteredTopics.length, 3, 20)), readinessForAdvanced boolean (true when abilityEstimate ≥1.5), confidenceScore, forceHumanReview=false (informational). Mandatory practice schedules require teacher approval.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Advisory only — agents cannot mutate official exam result truth, student records, item labels, or official academic standing. Item retirement, difficulty label changes, and content edits require content team approval through the item governance workflow.

CF Worker: PerformanceForecastingAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Forecasts a student's exam performance trajectory by projecting score improvement based on current score, practice session history, average improvement per session, and days remaining before the exam. Assigns a forecast band and urgency level and flags at-risk learners for early intervention.

Tools: read_academic_data  ·  render_card

Outcome: AgentDecision with studentId, currentScore, projectedScore, forecastBand (below_passing <33/marginal 33-49/passing 50-74/distinction ≥75), daysUntilExam, urgency (high when below_passing AND <30 days/medium for marginal/low otherwise), confidenceScore (0.70 + min(sessions, 30) × 0.005), forceHumanReview = (urgency === high). Forecasts are probabilistic and do not affect official exam standing.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Advisory only — agents cannot mutate official exam result truth, student records, item labels, or official academic standing. Item retirement, difficulty label changes, and content edits require content team approval through the item governance workflow.

CF Worker: CohortAnalyticsAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Generates cohort-level academic performance analytics including score distribution shape, mean, standard deviation, coefficient of variation, performance classification, and intervention flags. The only Group 13 agent with active read tools (6 analytics tools) providing access to cycle KPIs, revenue trends, school participation, subject performance, agent throughput, and CSV exports.

Tools: get_cycle_kpis  ·  get_revenue_trends  ·  get_school_participation  ·  get_subject_performance  ·  get_agent_throughput  ·  export_aggregated_csv

Outcome: AgentDecision with cohortId, cohortSize, meanScore, stdDeviation, coefficientOfVariation, distributionShape (tight <0.15/normal 0.15-0.34/dispersed ≥0.35), cohortPerformance (below_average/average/above_average), interventionNeeded boolean, confidenceScore. All 6 tools strip PII — no individual student names or financial amounts returned; only aggregated metrics. Academic interventions require academic coordinator and institution approval.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Advisory only — agents cannot mutate official exam result truth, student records, item labels, or official academic standing. Item retirement, difficulty label changes, and content edits require content team approval through the item governance workflow.

CF Worker: DifficultyCalibrationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Recommends item difficulty label recalibration when empirical response data diverges from the human-assigned label. Computes an empirical difficulty classification from the facility index, evaluates the discrimination index for item quality, and recommends one of three actions: recalibrate/confirm/flag_for_review.

Tools: read_academic_data  ·  render_card

Outcome: AgentDecision with itemId, currentDifficultyLabel, empiricalDifficulty, empiricalLabel, discriminationIndex, discriminationQuality (good ≥0.20/poor <0.20), labelMismatch boolean, calibrationAction (recalibrate if mismatch + good discrimination/flag_for_review if poor discrimination/confirm otherwise), confidenceScore, forceHumanReview = (calibrationAction !== confirm). Item difficulty label changes require content governance team approval through the item calibration workflow.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Advisory only — agents cannot mutate official exam result truth, student records, item labels, or official academic standing. Item retirement, difficulty label changes, and content edits require content team approval through the item governance workflow.

CF Worker: WebhookIntelligenceAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Analyses inbound webhook delivery telemetry to detect failure rate spikes, latency anomalies, and degraded endpoint health. Classifies each webhook as healthy/degraded/failing and surfaces the last known failure reason.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with webhookId, failureRate (failedDeliveries/totalDeliveries), webhookHealth (healthy <5%/degraded 5-19%/failing ≥20%), avgLatencyMs, latencyStatus (normal <500ms/slow 500-1999ms/critical ≥2000ms), lastFailureReason, confidenceScore (0.88 − failureRate × 0.2), forceHumanReview when failing.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only — endpoint configuration changes, pipeline restarts, data remediation, and ETL reprocessing require ops_admin approval. Agent cannot halt, restart, or reconfigure any pipeline. All outputs include confidence score, human review flag, and correlation ID.

CF Worker: EtlMonitoringAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Monitors ETL pipeline job health by evaluating processing completion rates, error rates, and run-duration ratios against expected baselines. Classifies each pipeline run as on_track/slow/failing and flags stalled or anomalously slow jobs.

Tools: read_runtime_status  ·  render_table

Outcome: AgentDecision with pipelineId, processingRate, errorRate, durationRatio, pipelineStatus (failing if errorRate ≥5% or processingRate <50%/slow if durationRatio >1.5x/on_track otherwise), confidenceScore (0.87 − errorRate × 0.5), forceHumanReview when failing.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only — endpoint configuration changes, pipeline restarts, data remediation, and ETL reprocessing require ops_admin approval. Agent cannot halt, restart, or reconfigure any pipeline. All outputs include confidence score, human review flag, and correlation ID.

CF Worker: DataQualityAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Scores dataset quality across completeness (null rates per column), consistency (schema violation rate), and deduplication (duplicate rate) dimensions. Assigns qualityGrade (good ≥0.85/acceptable 0.60-0.84/poor <0.60) and identifies high-null columns (>10%) requiring remediation.

Tools: read_data_quality_summary  ·  read_runtime_status  ·  render_table

Outcome: AgentDecision with datasetId, totalRows, qualityScore (1 − maxNullRate×0.4 − duplicateRate×0.3 − schemaViolationRate×0.3), qualityGrade, highNullColumns list, duplicateRate, schemaViolationRate, confidenceScore, forceHumanReview when qualityGrade is poor.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only — endpoint configuration changes, pipeline restarts, data remediation, and ETL reprocessing require ops_admin approval. Agent cannot halt, restart, or reconfigure any pipeline. All outputs include confidence score, human review flag, and correlation ID.

CF Worker: FeatureStoreAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Validates feature store freshness by comparing the age of the stalest feature against defined SLA thresholds and computing a staleness ratio across the feature set. Classifies health as fresh/stale/critical and advises when feature refresh is required to prevent ML model accuracy degradation.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with featureSetId, stalestFeatureAgeHours, maxAllowedStalenessHours, staleRate, stalenessRatio, featureStoreHealth (critical ≥2.0/stale 1.0-1.99/fresh <1.0), confidenceScore (0.87 − stalenessRatio × 0.1), forceHumanReview when critical.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only — endpoint configuration changes, pipeline restarts, data remediation, and ETL reprocessing require ops_admin approval. Agent cannot halt, restart, or reconfigure any pipeline. All outputs include confidence score, human review flag, and correlation ID.

CF Worker: BiInsightAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Summarises business intelligence dashboard signals by comparing current metric snapshots against historical benchmarks and computing relative deviations. Classifies insight priority as routine/noteworthy/critical based on deviation magnitude (>50% → critical; >20% → noteworthy).

Tools: read_runtime_status  ·  render_table

Outcome: AgentDecision with reportName, periodLabel, metricCount, significantDeviationCount, insightPriority (critical/noteworthy/routine), confidenceScore (0.82 + metricCount × 0.01), forceHumanReview when critical. Business decisions based on BI data require management approval.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only — endpoint configuration changes, pipeline restarts, data remediation, and ETL reprocessing require ops_admin approval. Agent cannot halt, restart, or reconfigure any pipeline. All outputs include confidence score, human review flag, and correlation ID.

CF Worker: SyncConflictAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Detects and triages data synchronisation conflicts between operational Supabase stores and downstream warehouse systems by analysing conflict type, records affected, and auto-resolvability. Recommends resolution strategy (auto_resolve/manual_review/escalate) and classifies conflict severity.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with conflictId, conflictType, sourceSystem, targetSystem, recordsAffected, conflictSeverity (high >10k/medium 1k-10k/low ≤1k), resolutionStrategy (auto_resolve if autoResolvable and <100 records/escalate if >10k/manual_review otherwise), confidenceScore (0.85 − min(recordsAffected,10000)/100000), forceHumanReview for non-auto_resolve. Even auto_resolve recommendations require explicit ops_admin approval.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only — endpoint configuration changes, pipeline restarts, data remediation, and ETL reprocessing require ops_admin approval. Agent cannot halt, restart, or reconfigure any pipeline. All outputs include confidence score, human review flag, and correlation ID.

CF Worker: BiDashboardAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Surfaces business intelligence dashboard signals from data warehouse query results, detects metric anomalies against historical baselines, and recommends investigation priorities for the ops portal. Monitors KPIs and flags critical deviations that require human review.

Tools: read_data_quality_summary  ·  read_runtime_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Advisory AgentDecision with insight priority classification (routine/noteworthy/critical), count of significant metric deviations, rendered BI dashboard card or table, and a recommendation string. forceHumanReview set true when any metric deviates more than 50% from its benchmark. Minimum confidence threshold 0.75.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only — endpoint configuration changes, pipeline restarts, data remediation, and ETL reprocessing require ops_admin approval. Agent cannot halt, restart, or reconfigure any pipeline. All outputs include confidence score, human review flag, and correlation ID.

CF Worker: McpToolGatewayAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Acts as the governed gateway for MCP tool invocations within the Finverse Agentic OS, routing and mediating tool calls from other agents through a permission-checked, schema-validated, and audited boundary. Enforces allowedTools constraints and blocks any action outside the safe tool catalog.

Tools: read_data_quality_summary  ·  read_runtime_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Advisory AgentDecision containing the proxied tool output, tool name invoked, confidenceScore, humanReviewRequired flag, and audit-ready evidence reference. Any tool call outside the governed catalog is blocked and surfaced as a policy violation with forceHumanReview set true. Cannot initiate external MCP calls outside the safe tool catalog or bypass the Cloudflare AI Gateway DPA requirement.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. Advisory only — endpoint configuration changes, pipeline restarts, data remediation, and ETL reprocessing require ops_admin approval. Agent cannot halt, restart, or reconfigure any pipeline. All outputs include confidence score, human review flag, and correlation ID.

CF Worker: OfflineAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Monitors offline data capture status across mobile and PWA clients by evaluating device-level sync lag, pending operation count, and storage utilisation. Surfaces risk indicators (syncLag/storageRisk/syncPriority) to the mobile operations dashboard for operator review.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with deviceId, offlineDurationMinutes, pendingOperations, storageUsageRate, syncLag (minimal/moderate/critical), storageRisk (safe/warning/critical), syncPriority (immediate/scheduled/low), confidenceScore (base 0.84 − storage pressure penalty), forceHumanReview=true when syncPriority is immediate.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Agent CANNOT initiate sync, perform data merges, apply rollbacks, or switch capture methods directly — all such actions require ops_admin approval. No mutation of business truth under any condition. All output is advisory, logged, and schema-validated.

CF Worker: SyncResolutionAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Triages synchronisation conflicts generated when offline mobile or PWA clients attempt to push records back to the server. Recommends merge strategies (auto-merge vs. manual review) based on conflict count, auto-mergeability, server/client version authority, and data-loss risk.

Tools: read_runtime_status  ·  render_table

Outcome: AgentDecision with deviceId, totalConflicts, autoMergeableCount, manualReviewRequired, serverVersionWins, clientVersionWins, dataLossRisk flag, resolutionComplexity (simple ≤10/moderate 11-100/complex >100), confidenceScore (base 0.86 − manual-conflict ratio penalty), recommendation, advisoryBoundary. No data merge is performed by the agent.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Agent CANNOT initiate sync, perform data merges, apply rollbacks, or switch capture methods directly — all such actions require ops_admin approval. No mutation of business truth under any condition. All output is advisory, logged, and schema-validated.

CF Worker: MobileSupportAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Classifies mobile application support queries from students and coordinators by platform (Android/iOS/web), app version, and error code, then recommends the correct resolution path or escalation routing. Distinguishes between self-service, Tier 1 support, and engineering escalation scenarios.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with userId, devicePlatform, appVersion, errorCode (or null), supportCategory (self_service/tier1_support/engineering_escalation), confidenceScore (0.87 for known errors/0.72 for no-error), forceHumanReview=true for engineering_escalation. Agent CANNOT perform account fixes, app rollbacks, or forced app updates.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Agent CANNOT initiate sync, perform data merges, apply rollbacks, or switch capture methods directly — all such actions require ops_admin approval. No mutation of business truth under any condition. All output is advisory, logged, and schema-validated.

CF Worker: AttendanceCaptureAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Validates attendance records captured from mobile and offline sources (biometric/QR code/barcode/manual) for completeness, failure rate, and time-window integrity before sync confirmation. Flags exam centres where capture rate is critically low or failure rate exceeds threshold.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with examCentreId, captureMethod, totalCandidates, capturedCount, failedCaptureCount, captureRate, failureRate, captureStatus (complete ≥95%/in_progress 60-94%/at_risk <60%), needsManualIntervention boolean, confidenceScore (base 0.88 − failure rate penalty), forceHumanReview=true when needsManualIntervention. Agent CANNOT write, correct, or delete attendance records.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Agent CANNOT initiate sync, perform data merges, apply rollbacks, or switch capture methods directly — all such actions require ops_admin approval. No mutation of business truth under any condition. All output is advisory, logged, and schema-validated.

CF Worker: TenantOnboardingAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Validates tenant onboarding checklist completion covering schema provisioning, RLS policy setup, admin user creation, and feature flag configuration. Computes a completion rate and classifies onboarding status as complete/in_progress/stalled, then derives churn risk level.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with tenantId, tenantName, plan (starter/growth/enterprise), completionRate, onboardingStatus (complete ≥1.0/stalled if >14 days + <50%/in_progress otherwise), churnRisk (high if stalled >21 days/medium if stalled/low otherwise), advisoryBoundary, recommendation (HIGH CHURN RISK alert when applicable), confidenceScore (0.82 + completionRate × 0.1), forceHumanReview when churnRisk is high.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Advisory only — account interventions, plan changes, quota enforcement, and billing adjustments require authorised ops_admin or SaaS admin approval. No mutation of tenant entitlement records, quota limits, or billing configurations. Tenant isolation enforced at RLS layer; outputs scoped to authenticated tenant only.

CF Worker: TenantAdminSaasAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Summarises tenant-level usage health, quota utilisation, and subscription renewal urgency for SaaS admin review. Computes storage and API usage rates, then classifies quota status (healthy/warning/critical) and renewal urgency (none/upcoming/immediate).

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with tenantId, storageUsageRate, apiUsageRate, quotaStatus (critical if either ≥0.9/warning if either ≥0.75/healthy otherwise), subscriptionRenewalDays, renewalUrgency (immediate ≤7 days/upcoming ≤30/none otherwise), advisoryBoundary, recommendation (CRITICAL QUOTA or RENEWAL URGENT alerts), confidenceScore (0.88 − max(storageUsageRate, apiUsageRate) × 0.1), forceHumanReview when critical or immediate.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Advisory only — account interventions, plan changes, quota enforcement, and billing adjustments require authorised ops_admin or SaaS admin approval. No mutation of tenant entitlement records, quota limits, or billing configurations. Tenant isolation enforced at RLS layer; outputs scoped to authenticated tenant only.

CF Worker: EntitlementEnforcementAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Evaluates feature access requests against tenant entitlement plans and returns an advisory decision of allow/upsell/block with supporting plan evidence. Determines whether the current plan permits the requested feature and, if not, identifies the appropriate upgrade target.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with tenantId, featureName, plan, planAllowsFeature, decision (allow if planAllowsFeature/upsell if not enterprise and not allowed/block if enterprise and not allowed), upgradeTarget (growth if starter/enterprise if growth/null), advisoryBoundary, confidenceScore=0.99 (deterministic), forceHumanReview=false. Actual access enforcement handled by deterministic entitlement service.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Advisory only — account interventions, plan changes, quota enforcement, and billing adjustments require authorised ops_admin or SaaS admin approval. No mutation of tenant entitlement records, quota limits, or billing configurations. Tenant isolation enforced at RLS layer; outputs scoped to authenticated tenant only.

CF Worker: UsageMeteringAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Monitors tenant usage metrics including API calls, storage, active users, and exam seat consumption against included plan allowances for a given billing period. Computes per-dimension overage amounts and an aggregate overageScore, then classifies the billing flag.

Tools: read_runtime_status  ·  render_table

Outcome: AgentDecision with tenantId, billingPeriod, per-dimension overage breakdown (candidates/apiCalls/storageGb/computeUnits), hasOverage flag, billingFlag (clean/overage_advisory/overage_critical where critical ≥0.3), advisoryBoundary, recommendation (CRITICAL OVERAGE alert), confidenceScore (0.94 − overageScore × 0.05), forceHumanReview when overage_critical. Billing invoices handled by deterministic billing service.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Advisory only — account interventions, plan changes, quota enforcement, and billing adjustments require authorised ops_admin or SaaS admin approval. No mutation of tenant entitlement records, quota limits, or billing configurations. Tenant isolation enforced at RLS layer; outputs scoped to authenticated tenant only.

CF Worker: WhiteLabelConfigurationAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Validates white-label configuration completeness including branding assets (logo, brand colour), custom domain settings, and overall completion rate for a tenant. Classifies configuration status as complete/partial/minimal and enumerates missing elements.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with tenantId, brandName, configCompletionRate, configStatus (complete ≥0.9/partial ≥0.5/minimal otherwise), missingElements list (customDomain/primaryColor/logoUrl if absent), advisoryBoundary, recommendation (lists missing elements or advises DNS verification), confidenceScore (0.86 + configCompletionRate × 0.1), forceHumanReview=false. Domain setup, SSL provisioning, and branding publish actions require ops_admin approval.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Advisory only — account interventions, plan changes, quota enforcement, and billing adjustments require authorised ops_admin or SaaS admin approval. No mutation of tenant entitlement records, quota limits, or billing configurations. Tenant isolation enforced at RLS layer; outputs scoped to authenticated tenant only.

CF Worker: ModelGatewayRoutingAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Recommends model routing decisions for incoming AI requests based on latency, cost, capability, and context-length signals. Evaluates primary model availability and acceptable latency thresholds (acceptable if <3000ms) to decide between primary and fallback model paths.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with selectedModel, routingDecision (primary or fallback), primaryModelLatencyMs, estimatedTokens, confidence score (0.95 if primary available/0.78 if routing to fallback), advisoryBoundary. Gateway configuration changes require separate super_admin approval workflow.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Never mutates model gateway configuration, provider credentials, billing settings, rate-limit policy, or agent guardrails directly. All routing changes, model rollbacks, sandbox terminations, and circuit state transitions require super_admin or AI ops approval. All outputs are advisory only and unconditionally require human review. No direct write access to any business truth table.

CF Worker: ModelFallbackAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Detects AI model provider degradation signals and recommends fallback routing to alternative providers or local stubs. Evaluates the failed model, failure reason, task type, and ordered list of fallback candidates to produce a ranked recommendation.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with failedModel, recommendedFallback (first candidate in fallbackCandidates list), fallbackAvailable boolean, taskType (classification/generation/tool_use/embedding), confidence score (0.85 if available/0.50 if no fallback), advisoryBoundary. forceHumanReview forced to true when no fallback is available.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Never mutates model gateway configuration, provider credentials, billing settings, rate-limit policy, or agent guardrails directly. All routing changes, model rollbacks, sandbox terminations, and circuit state transitions require super_admin or AI ops approval. All outputs are advisory only and unconditionally require human review. No direct write access to any business truth table.

CF Worker: RateLimitGuardAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Monitors API rate utilization across providers per tenant and recommends throttling, queuing, or request-shedding interventions. Computes request and token usage rates against per-minute limits and classifies tenant status as allow/throttle/block.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with tenantId, requestUsageRate, tokenUsageRate, throttleStatus (block if either ≥1.0/throttle if ≥0.9/allow otherwise), confidence score (1 − max(usageRate) × 0.05), advisoryBoundary. Actual rate-limit enforcement and policy changes require AI ops approval.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Never mutates model gateway configuration, provider credentials, billing settings, rate-limit policy, or agent guardrails directly. All routing changes, model rollbacks, sandbox terminations, and circuit state transitions require super_admin or AI ops approval. All outputs are advisory only and unconditionally require human review. No direct write access to any business truth table.

CF Worker: BudgetEnforcementAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Tracks AI inference spend against per-tenant and aggregate budget limits and recommends cutover or throttling actions when thresholds are approached or exceeded. Evaluates current spend, budget cap, and forecasted spend to classify budget health.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with tenantId, billingPeriod, spentUsdCents, budgetUsdCents, forecastedSpendUsdCents, spentRate, forecastedRate, budgetStatus (exceeded if spentRate ≥1/critical if forecastedRate ≥1/warning if either ≥0.8/healthy otherwise), confidence score, advisoryBoundary. AI feature suspension requires super_admin approval. forceHumanReview forced to true for exceeded and critical states.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Never mutates model gateway configuration, provider credentials, billing settings, rate-limit policy, or agent guardrails directly. All routing changes, model rollbacks, sandbox terminations, and circuit state transitions require super_admin or AI ops approval. All outputs are advisory only and unconditionally require human review. No direct write access to any business truth table.

CF Worker: DlpAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Scans AI inputs and outputs for data loss prevention policy violations including PII exposure, sensitive schema leakage, and DPDP-restricted data categories. Evaluates detected pattern confidence scores to classify each request as pass/redact-advisory/block-advisory.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with requestId, detectedPatterns, maxConfidence, hasSensitiveContent, dlpAction (block_advisory if maxConfidence ≥0.9 + sensitive/redact_advisory if sensitive below threshold/pass otherwise), sourcePortal, confidence score, advisoryBoundary. Content blocking enforcement is executed by the deterministic middleware. forceHumanReview forced to true for block_advisory outcomes.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Never mutates model gateway configuration, provider credentials, billing settings, rate-limit policy, or agent guardrails directly. All routing changes, model rollbacks, sandbox terminations, and circuit state transitions require super_admin or AI ops approval. All outputs are advisory only and unconditionally require human review. No direct write access to any business truth table.

CF Worker: MemoryGovernanceAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Audits agent memory stores for PII retention, scope boundary violations, and data retention policy compliance. Evaluates memory usage rate, stale entry fraction, and the presence of sensitive entries to classify memory health and recommend pruning or immediate audit actions.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with agentId, memoryEntries, staleEntries, sensitiveEntries, usageRate (memoryEntries/maxAllowedEntries), staleRate, memoryHealth (critical if ≥0.9 or any sensitiveEntries/needs_pruning if ≥0.7 or staleRate ≥0.3/healthy otherwise), confidence score, advisoryBoundary. Never prunes, deletes, or modifies memory entries directly. forceHumanReview forced to true when sensitiveEntries > 0 or usageRate ≥ 0.9.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Never mutates model gateway configuration, provider credentials, billing settings, rate-limit policy, or agent guardrails directly. All routing changes, model rollbacks, sandbox terminations, and circuit state transitions require super_admin or AI ops approval. All outputs are advisory only and unconditionally require human review. No direct write access to any business truth table.

CF Worker: ArtifactGovernanceAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Validates AI-generated artifact provenance, retention lifecycle, and access control policy compliance. Checks whether an artifact's actual retention period violates policy limits and whether it contains PII, then classifies governance action as retain/flag-for-deletion/immediate-review.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with artifactId, artifactType (generated_report/model_output/evidence_bundle/audit_export), retentionDays, policyRetentionDays, retentionCompliant, piiRisk (high if containsPii/low otherwise), governanceAction (immediate_review if PII + non-compliant/flag_for_deletion if non-compliant without PII/retain otherwise), confidence score, advisoryBoundary. Never deletes artifacts or modifies retention policies directly. forceHumanReview forced to true for immediate_review outcomes.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Never mutates model gateway configuration, provider credentials, billing settings, rate-limit policy, or agent guardrails directly. All routing changes, model rollbacks, sandbox terminations, and circuit state transitions require super_admin or AI ops approval. All outputs are advisory only and unconditionally require human review. No direct write access to any business truth table.

CF Worker: SandboxMonitorAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Monitors agent sandbox activity for unexpected external resource access patterns, file-system anomalies, and policy boundary violations. Evaluates blocked tool call rates, external network attempts, CPU, and memory usage to classify sandbox status.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with sandboxId, blockRate (blockedToolCalls/toolCallsExecuted), externalNetworkAttempts, resourcePressure (low/medium/high), sandboxStatus (breach_attempt if any externalNetworkAttempts or blockRate ≥0.3/suspicious if ≥0.1/normal otherwise), confidence score, advisoryBoundary. Sandbox termination requires security team approval. forceHumanReview forced to true for breach_attempt outcomes.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Never mutates model gateway configuration, provider credentials, billing settings, rate-limit policy, or agent guardrails directly. All routing changes, model rollbacks, sandbox terminations, and circuit state transitions require super_admin or AI ops approval. All outputs are advisory only and unconditionally require human review. No direct write access to any business truth table.

CF Worker: CircuitBreakerAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Detects cascading failure patterns across agent invocations and recommends circuit-break intervention to prevent runaway retry storms. Computes failure rate from success and failure counts and recommends the appropriate circuit state (closed/half_open/open).

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with serviceId, currentState (closed/open/half_open), recommendedState, failureRate, stateChange boolean, confidence score, advisoryBoundary. Never transitions circuit state directly — ops_admin approval required. forceHumanReview forced to true when recommending transition to open state. Service isolation actions require ops_admin approval.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Never mutates model gateway configuration, provider credentials, billing settings, rate-limit policy, or agent guardrails directly. All routing changes, model rollbacks, sandbox terminations, and circuit state transitions require super_admin or AI ops approval. All outputs are advisory only and unconditionally require human review. No direct write access to any business truth table.

CF Worker: AgentHealthMonitorAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Monitors agent session health metrics including error rates, latency distributions, guardrail violation counts, and Durable Object eviction signals, then recommends restart or escalation actions. Classifies each agent as healthy/degraded/critical based on 24-hour run history.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with agentSlug, totalRunsLast24h, errorRunsLast24h, errorRate, guardrailViolations, guardrailViolationRate, avgLatencyMs, agentHealth (critical if errorRate ≥0.2 or guardrailViolationRate ≥0.05/degraded if ≥0.05 or any violations/healthy otherwise), confidence score, advisoryBoundary. Agent disabling and rollback require AI ops approval. forceHumanReview forced to true when agentHealth is critical.

Guardrails: Hard-blocked: post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, delete_audit_logs, bypass_rls. Never mutates model gateway configuration, provider credentials, billing settings, rate-limit policy, or agent guardrails directly. All routing changes, model rollbacks, sandbox terminations, and circuit state transitions require super_admin or AI ops approval. All outputs are advisory only and unconditionally require human review. No direct write access to any business truth table.

CF Worker: SecurityIncidentTriageAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Classifies security incidents by severity (critical/high/medium/low) and attack vector, evaluating containment status and affected system scope to determine escalation path. Recommends initial containment advisory and escalation priority (CISO immediate/security team urgent/standard) for the security team.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with incidentId, incidentType, affectedSystems list, containmentStatus, severity level, escalation path recommendation, advisoryBoundary, confidence score (base 0.88 − 0.05 if >5 affected systems). Agent CANNOT isolate systems, suspend accounts, or file CERT-In reports.

Guardrails: Hard-blocked: delete_audit_logs, post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; no operational mutation. forceHumanReview=true unconditionally. Requires super_admin or ops_admin approval before any downstream action. Minimum confidence threshold: 0.95.

CF Worker: AccessGovernanceAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Reviews access grant history, unused permission accumulation, privilege escalation patterns, and dormant account signals for individual users. Computes an access risk score (low/medium/high) based on last login recency, unused permission count, and privileged access count.

Tools: read_runtime_status  ·  render_table

Outcome: AgentDecision with userId, currentRoles, lastLoginDays, dormantAccount flag (if >90 days), unusedPermissions list, privilegedAccessCount, accessRisk level, advisory boundary, confidence score (base 0.87 − privilegedAccessCount × 0.01). Agent CANNOT revoke roles, suspend accounts, or modify access control policies.

Guardrails: Hard-blocked: delete_audit_logs, post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; no operational mutation. forceHumanReview=true unconditionally. Requires super_admin or ops_admin approval before any downstream action. Minimum confidence threshold: 0.95.

CF Worker: AuditEvidenceAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Collects and summarizes audit evidence artifacts from decision logs, access logs, and event store to determine audit readiness for internal, regulatory, customer, or penetration test audits. Computes an evidence collection rate and classifies readiness as ready/partial/not_ready.

Tools: read_runtime_status  ·  render_table

Outcome: AgentDecision with auditRequestId, auditType, evidenceCollected, evidenceRequired, collectionRate, auditReadiness (ready ≥0.95/partial ≥0.70/not_ready otherwise), gapAreas list, confidence score (0.84 + collectionRate × 0.10). Agent CANNOT submit evidence to auditors or issue compliance declarations.

Guardrails: Hard-blocked: delete_audit_logs, post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; no operational mutation. forceHumanReview=true unconditionally. Requires super_admin or ops_admin approval before any downstream action. Minimum confidence threshold: 0.95.

CF Worker: CertInReportingAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Structures CERT-In cybersecurity incident reports within the mandated 6-hour initial reporting window, computing hours remaining to deadline and classifying reporting urgency as submitted/scheduled/urgent/overdue. Validates whether CERT-In reporting is required based on incident severity.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with incidentId, incidentSeverity, hoursToDeadline, reportSubmitted flag, certInRequired flag (true if critical or high severity), reportingUrgency (submitted/overdue/urgent ≤6h/scheduled otherwise), confidence score (0.95 − 0.05 if certInRequired and not submitted). Agent CANNOT submit CERT-In reports.

Guardrails: Hard-blocked: delete_audit_logs, post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; no operational mutation. forceHumanReview=true unconditionally. Requires super_admin or ops_admin approval before any downstream action. Minimum confidence threshold: 0.95.

CF Worker: DpdpComplianceAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Evaluates data processing activities for a given dataset against DPDP Act 2023 obligations including consent coverage, data minimization, purpose limitation, cross-border transfer review, Data Protection Manager appointment, and children's data protections.

Tools: read_runtime_status  ·  render_table

Outcome: AgentDecision with datasetId, consentCoverageRate, complianceGaps list (incomplete_consent_coverage/cross_border_transfer_review/dpm_not_appointed), dpdpStatus (non_compliant/partial/compliant), crossBorderTransfer flag, dpmAppointed flag, confidence score (0.87 + consentCoverageRate × 0.05). Agent CANNOT execute remediation actions, appoint DPM, or file regulatory submissions.

Guardrails: Hard-blocked: delete_audit_logs, post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; no operational mutation. forceHumanReview=true unconditionally. Requires super_admin or ops_admin approval before any downstream action. Minimum confidence threshold: 0.95.

CF Worker: ConsentManagementAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Reviews consent records for completeness, currency, and withdrawal status across a consent batch, computing active consent rate and expired consent rate to classify consent health as healthy/renewal_needed/critical. Recommends renewal campaign initiation.

Tools: read_runtime_status  ·  render_table

Outcome: AgentDecision with consentBatchId, consentVersion, totalSubjects, activeConsentRate ((consentedCount − withdrawnCount) / totalSubjects), withdrawnCount, expiredRate, consentHealth (critical if activeConsentRate <0.70 or expiredRate >0.30/renewal_needed if expiredRate >0.10 or activeConsentRate <0.85/healthy otherwise), confidence score (0.88 + activeConsentRate × 0.05). Agent CANNOT halt data processing, modify consent records, or trigger renewal campaigns.

Guardrails: Hard-blocked: delete_audit_logs, post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; no operational mutation. forceHumanReview=true unconditionally. Requires super_admin or ops_admin approval before any downstream action. Minimum confidence threshold: 0.95.

CF Worker: DataRetentionAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Evaluates data retention compliance for a given data category by comparing actual record ages against defined retention policy bounds, computing an over-retention rate and classifying status as compliant/overdue/critical. Flags records exceeding retention limits for DPO-approved purge review without triggering any deletion.

Tools: read_runtime_status  ·  render_table

Outcome: AgentDecision with dataCategory, retentionPolicyDays, oldestRecordAgeDays, recordsOverRetention, overRetentionRate, retentionStatus (critical if ≥10% over or oldest >2x policy/overdue if any exceed policy/compliant otherwise), confidence score (0.90 − overRetentionRate × 0.20). Agent CANNOT delete records or modify retention policies.

Guardrails: Hard-blocked: delete_audit_logs, post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; no operational mutation. forceHumanReview=true unconditionally. Requires super_admin or ops_admin approval before any downstream action. Minimum confidence threshold: 0.95.

CF Worker: BreachDetectionAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Analyzes access logs, audit events, and anomaly signals to compute exfiltration likelihood and classify breach likelihood as low/medium/high/confirmed. Assesses impact scale based on estimated affected record count and data categories involved.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with alertId, anomalySignals list, affectedRecordsEstimate, dataCategories, exfiltrationLikelihood, breachLikelihood (confirmed ≥0.80/high ≥0.50/medium ≥0.20/low otherwise), impactScale (severe ≥100k/significant ≥1k/minimal otherwise), confidence score (0.85 + exfiltrationLikelihood × 0.10). Agent CANNOT declare incidents, file CERT-In reports, or notify data subjects.

Guardrails: Hard-blocked: delete_audit_logs, post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; no operational mutation. forceHumanReview=true unconditionally. Requires super_admin or ops_admin approval before any downstream action. Minimum confidence threshold: 0.95.

CF Worker: VendorRiskAssessmentAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Scores third-party vendor data processing risk against DPA obligations by evaluating data access level, Data Processing Agreement status, security certifications (ISO27001/SOC2/DPDP_DPA), and time since last assessment. Classifies vendor risk as low/medium/high/critical.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with vendorId, vendorName, dataAccessLevel, certifications list, lastAssessmentDays, contractualDpaInPlace, dpaRequired flag (true if personal or sensitive data access), vendorRisk (critical if sensitive data without DPA/high if personal without DPA/medium if >365 days or missing certs/low otherwise), confidence score (0.87 − 0.05 for critical risk). Agent CANNOT suspend data sharing, execute DPAs, or disqualify vendors.

Guardrails: Hard-blocked: delete_audit_logs, post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; no operational mutation. forceHumanReview=true unconditionally. Requires super_admin or ops_admin approval before any downstream action. Minimum confidence threshold: 0.95.

CF Worker: AiRedTeamAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Systematically probes AI system components for adversarial vulnerabilities by analyzing results from prompt injection attempts, guardrail bypass attempts, and jailbreak attempts against a target agent. Computes a bypass rate and classifies security posture as robust/vulnerable/critical.

Tools: read_runtime_status  ·  render_table

Outcome: AgentDecision with testRunId, targetAgentSlug, totalAttempts, successfulBypasses, bypassRate, securityPosture (critical if ≥0.05/vulnerable if ≥0.01/robust otherwise), confidence score (0.86 + (1 − bypassRate) × 0.10). Agent CANNOT disable agents, modify guardrails, or deploy emergency patches.

Guardrails: Hard-blocked: delete_audit_logs, post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; no operational mutation. forceHumanReview=true unconditionally. Requires super_admin or ops_admin approval before any downstream action. Minimum confidence threshold: 0.95.

CF Worker: RegulatoryChangeAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Monitors regulatory change signals across DPDP, RBI, CERT-In, and SEBI frameworks by tracking compliance gaps and days until a regulation's effective date. Classifies preparedness urgency as compliant/planned/urgent/overdue.

Tools: read_runtime_status  ·  render_card

Outcome: AgentDecision with regulationId, regulationName, jurisdiction, daysUntilEffective, complianceGapCount, impactAreas list, preparednessUrgency (compliant if 0 gaps/overdue if past effective date with gaps/urgent if ≤30 days/planned otherwise), confidence score (0.86 − complianceGapCount × 0.01). Agent CANNOT issue compliance declarations, update policies, or file regulatory submissions.

Guardrails: Hard-blocked: delete_audit_logs, post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; no operational mutation. forceHumanReview=true unconditionally. Requires super_admin or ops_admin approval before any downstream action. Minimum confidence threshold: 0.95.

CF Worker: ParentConsentAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Helps parents understand what data Finverse holds about their children and how to exercise their rights under DPDP Act 2023 (access, correction, erasure, grievance). Explains data categories held (registration, results, AI tutor logs), guides consent status interpretation, and directs withdrawal requests to the support ticket workflow. Serves /parent/* routes.

Tools: get_parent_consent_status

Outcome: Advisory text response explaining consent status, data categories, or DPDP rights. When get_parent_consent_status is invoked, structured data includes active children count, consent status, and data category list. All consent mutations are deferred to the support ticket workflow. Agent CANNOT mutate consent records. requiresHumanApproval=false; all output is advisory and logged.

Guardrails: Hard-blocked: delete_audit_logs, post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; no operational mutation. forceHumanReview=true unconditionally. Requires super_admin or ops_admin approval before any downstream action. Minimum confidence threshold: 0.95.

CF Worker: ComplianceGovernanceAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Oversees the overall compliance governance posture of the Finverse Olympiad OS, cross-referencing active regulatory obligations (DPDP Act 2023, RBI, CERT-In, SEBI) against implemented controls. Used by ops_admin and super_admin in the Trust Control Tower portal to assess governance gaps and track remediation.

Tools: read_compliance_controls  ·  read_runtime_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Prioritised compliance gap report with obligation-to-control mapping, severity scores, timeline of upcoming obligation deadlines, and remediation recommendations — all flagged for super_admin or ops_admin review. All output is advisory; no compliance control policy is created or modified. Minimum confidence threshold: 0.95.

Guardrails: Hard-blocked: delete_audit_logs, post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; no operational mutation. forceHumanReview=true unconditionally. Requires super_admin or ops_admin approval before any downstream action. Minimum confidence threshold: 0.95.

CF Worker: ItComplianceAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Evaluates IT infrastructure and software controls against relevant compliance frameworks including ISO 27001 controls, RBI IT framework requirements, and CERT-In guidelines applicable to Finverse Academy operations. Surfaces control gaps, misconfigured policies, and missing evidence items.

Tools: read_compliance_controls  ·  read_runtime_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: IT compliance gap assessment showing control status, evidence gaps, severity classification (critical/high/medium), and recommended remediation steps — requiring ops_admin or super_admin sign-off. No system configuration changes are triggered.

Guardrails: Hard-blocked: delete_audit_logs, post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; no operational mutation. forceHumanReview=true unconditionally. Requires super_admin or ops_admin approval before any downstream action. Minimum confidence threshold: 0.95.

CF Worker: OwaspAuditAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Audits the Finverse Olympiad OS application and API surface against the OWASP Top 10 and relevant OWASP ASVS controls. Identifies web application security risks including injection flaws, broken authentication, security misconfigurations, and vulnerable dependencies.

Tools: read_compliance_controls  ·  read_runtime_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: OWASP audit report with per-category risk ratings (critical/high/medium/low), affected surface areas (portal routes/API endpoints), mitigation status, and prioritised remediation recommendations — gated behind super_admin approval before any action. Findings are advisory only; no code changes applied automatically.

Guardrails: Hard-blocked: delete_audit_logs, post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; no operational mutation. forceHumanReview=true unconditionally. Requires super_admin or ops_admin approval before any downstream action. Minimum confidence threshold: 0.95.

CF Worker: DpdpConsentAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Reviews consent records for data principals (students and parents) against DPDP Act 2023 requirements, checking for valid, informed, and freely given consent at each data processing stage. Flags missing, expired, or improperly recorded consent items for compliance remediation.

Tools: read_compliance_controls  ·  read_runtime_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: DPDP consent compliance report identifying gaps by data principal, processing purpose, and regulatory obligation, with recommended remediation steps requiring human review. Consent records are never mutated directly; outputs contain aggregate gap summaries only, not raw PII.

Guardrails: Hard-blocked: delete_audit_logs, post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; no operational mutation. forceHumanReview=true unconditionally. Requires super_admin or ops_admin approval before any downstream action. Minimum confidence threshold: 0.95.

CF Worker: DpdpChildrenGuardianAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Enforces heightened DPDP Act 2023 protections for data belonging to children (minors under 18) enrolled in the Finverse Academy Olympiad. Reviews processing activities involving minor data principals for age verification, verifiable parental consent, and children-specific data minimisation obligations.

Tools: read_compliance_controls  ·  read_runtime_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Children data protection compliance report detailing violations of DPDP Act 2023 children-specific obligations, with severity scores and remediation steps gated behind super_admin review. Minor PII is never exposed in outputs; reports reference only aggregate counts and anonymised gap identifiers.

Guardrails: Hard-blocked: delete_audit_logs, post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; no operational mutation. forceHumanReview=true unconditionally. Requires super_admin or ops_admin approval before any downstream action. Minimum confidence threshold: 0.95.

CF Worker: RbiPaymentComplianceAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Evaluates payment processing activities within Finverse Olympiad OS against RBI payment regulations including PPI guidelines, digital payment security standards, and payment aggregator norms. Reviews invoice, reconciliation, and payment proof workflows for regulatory compliance gaps.

Tools: read_compliance_controls  ·  read_runtime_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: RBI payment compliance gap assessment with control mapping, regulatory exposure scores, and prioritised remediation recommendations — requiring human approval before any payment workflow change. No payment records, ledger entries, or reconciliation data are mutated.

Guardrails: Hard-blocked: delete_audit_logs, post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; no operational mutation. forceHumanReview=true unconditionally. Requires super_admin or ops_admin approval before any downstream action. Minimum confidence threshold: 0.95.

CF Worker: CertInIncidentAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Structures and validates cybersecurity incident reports required under CERT-In directions within the mandated 6-hour initial reporting window. Distinct from cert_in_reporting_agent — specifically triages whether an incident crosses the CERT-In reportable threshold and maps incident attributes to CERT-In mandatory fields.

Tools: read_compliance_controls  ·  read_runtime_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: CERT-In incident reportability assessment with draft mandatory fields pre-populated, 6-hour window status, and filing recommendations — gated behind authorized personnel review before any submission. Incident details are not exposed outside the Trust Control Tower security workflow.

Guardrails: Hard-blocked: delete_audit_logs, post_ledger_entry, change_score, publish_results, revoke_certificates, approve_refunds, bypass_rls, access_service_role_secret, write_business_truth_directly, export_raw_pii, call_external_ai_without_dpa, store_primary_truth_in_durable_object. All output is advisory only; no operational mutation. forceHumanReview=true unconditionally. Requires super_admin or ops_admin approval before any downstream action. Minimum confidence threshold: 0.95.

CF Worker: PrivacyRightsAgent extends BaseFinverseAgent · Durable Object · SQLite · WebSocket hibernation

Purpose: Manages and tracks data principal rights requests under DPDP Act 2023 including rights of access, correction, erasure, and grievance redressal for students and parents. Reviews open rights requests for completeness, SLA compliance, and appropriate handling, surfacing overdue or mishandled requests.

Tools: read_compliance_controls  ·  read_runtime_status  ·  render_card  ·  render_table  ·  show_timeline

Outcome: Data principal rights requests compliance report showing SLA status, handling gaps, overdue items, and remediation recommendations — requiring ops_admin or super_admin review before action. No PII is exported; outputs reference request identifiers and aggregate SLA metrics only.