Delegatic — User Stories

Canonical user-story catalog. Used for Playwright tests + Claude Design input for governance dashboard.

Scope: OS-006 agent governance shim — org trees, policies, signed authorization, audit. Unit-test surface covered: delegatic/test/** (42 tests — authorization kernel v0.1.0). Full org-tree + effective-policy layer still spec-only.

Story 1 · Create organization hierarchy and attach policies

• Persona: Enterprise governance officer setting up multi-team platform

• Goal: Define org tree (HQ → Engineering → ML Team) with monotonic policy inheritance

• Prerequisite: Delegatic running; auth setup

• Steps:

1. Navigate to dashboard; create root org "Acme Corp"

2. Create "Engineering" under Acme; set policy allow_agent_deploy: true, max_agents: 50

3. Create "ML Team" under Engineering; propose max_agents: 5 (tightening)

4. PolicyEngine validates: tighter than parent? YES. Accept.

5. ETS cache invalidated; PubSub broadcasts change

• Success: Org tree visible; policy inheritance resolved in <5μs; child policies enforce tighter constraints

• Covers: PolicyEngine.compute_effective, monotonic validation, ETS cache — ~15 unit tests

• UI status: planned (spec-complete; kernel ships today without UI)

• Claude Design hook: Org tree visualization with policy badges showing inheritance chain

Story 2 · Attach goal reference and enforce goal-aware policy

• Persona: Operator linking Graphonomous goals to governance decisions

• Goal: Attach a goal_id to org; enforce policy constraints keyed by goal context

• Prerequisite: Graphonomous goal exists; Delegatic org exists

• Steps:

1. POST /orgs/{org_id}/goals {goal_id, tags}

2. OrgGoal attachment created (reference-only; goal content NOT copied)

3. Task arrives with goal_id + action deploy_agent

4. PolicyEngine checks: goal attached? parent goal not suspended? action authorized under this goal?

5. Audit event recorded with goal_id

• Success: Goal-scoped audit trail; policies keyed by durable intent; revocation doesn't delete goal

• Covers: org_goal attachment, reference storage, audit creation, goal-aware routing — ~10 unit tests

• UI status: planned

• Claude Design hook: Audit log entry with goal_id + collapsed goal context panel

Story 3 · Enforce deny-by-default + detect violations

• Persona: Security team auditing access

• Goal: No agent can deploy unless explicitly authorized; detect violations in real-time

• Prerequisite: Root org allow_agent_deploy: false; child orgs inherit

• Steps:

1. Intern Team agent attempts deploy_agent

2. PolicyEngine walks ancestors: Intern Team → ML Team → Engineering → Acme

3. Effective: allow_agent_deploy: false (inherited from root)

4. Verdict: DENY

5. Audit event logged; dashboard highlights violation; PubSub alerts

• Success: Policy enforced; violation auditable; no privilege escalation

• Covers: monotonic AND logic, ancestor traversal, deny-by-default — ~12 unit tests

• UI status: planned

• Claude Design hook: Real-time violation ticker; drill-down to policy inheritance chain

Story 4 · Re-authorize destructive action at replay time

• Persona: Dark-factory operator replaying learned workflow on different machine

• Goal: Trace includes file deletions → each destructive step re-authorized via Delegatic

• Prerequisite: body-os trace recorded; Delegatic policy attached

• Steps:

1. Destination agent calls body.os.replay(Trace) with action: "file_delete@/ws/tmp/*"

2. Before executing: check_policy(org_id, action: "file_delete", goal_id)

3. Delegatic verifies: actor in org? scope attached? action ≤ policy limit?

4. If authorized: proceed; audit event with trace_id. If denied: fail-fast, emit SurpriseSignal

• Success: Cross-machine replay is governed; destructive actions re-authorized per environment

• Covers: policy check at commit time, audit emission with trace_id — ~8 unit tests (v0.1 kernel)

• UI status: planned

• Claude Design hook: Re-auth checkpoint UI — action + decision + actor approval log

Story 5 · Export compliance report with goal-linked audit trail

• Persona: Compliance officer proving to auditor that agent actions were authorized

• Goal: Report: "All deployments under approved goals; all destructive actions policy-checked"

• Prerequisite: 1000+ audit events collected

• Steps:

1. GET /orgs/{root}/compliance/export?period=q1&filter=goal_scoped

2. AuditWriter outputs append-only event stream grouped by goal_id with causal chain

3. Report: for each goal, actions taken, decisions, outcomes, surprises, links to Graphonomous/PRISM

4. Cryptographic commitment: report hash stored immutably

• Success: Compliance team proves governance to external auditor

• Covers: audit export, event grouping, report generation — future phase

• UI status: planned (Phase 3 enterprise)

• Claude Design hook: Timeline view — goal with linked audit events, expandable to policy decisions + outcomes

Tests to implement first (MCP-only until UI exists): Story 3 (deny-by-default) — use MCP inspector pattern to show policy decision + audit trail; strong investor story.