FleetPrompt — User Stories

Canonical user-story catalog. Each story is a Playwright-testable journey covering unit-tested code paths. Used for (a) e2e test generation in e2e-playwright/tests/user-stories/fleetprompt/, and (b) as Claude Design input for UI generation.

Scope: Agent marketplace with trust scoring, crystallization, install + review. Unit-test surface covered: test/fleet_prompt/** (157 tests).

Story 1 · Search and discover an agent

• Persona: DevOps engineer looking for production-ready agents

• Goal: Find a trusted customer-support agent that meets org security requirements

• Prerequisite: Deployer logged in; at least 1 published agent in fleet.agents

• Steps:

1. Click ⌘K search bar on marketplace homepage

2. Type "customer support"; filter min_trust=70

3. View filtered results ranked by trust score

4. Click an agent card to open detail page

5. Review manifest, permissions, version history, test results

• Success: Agent manifest with trust badge visible; user can verify behavior before install

• Covers: FleetPrompt.Search.*, trust filter, tsvector ranking — ~20 unit tests

• UI status: exists-today

• Claude Design hook: Search result card (name · trust badge · description · category tags · test pass rate)

Story 2 · Install agent with permission review

• Persona: Infrastructure team installing an agent into production

• Goal: Deploy a vetted agent with explicit permission acceptance + org policy validation

• Prerequisite: Agent selected; deployer has org admin role

• Steps:

1. Click "Install" on agent detail page

2. Review declared permissions (orders:read, refunds:create, graphonomous:*)

3. Accept permission set

4. System runs Delegatic policy check + MCP dependency resolution

5. System deploys manifest to OpenSentience runtime with capability tokens

• Success: Agent live in production; permissions enforced at OS runtime; audit event recorded

• Covers: FleetPrompt.InstallEngine (all 7 steps), Permission.map_to_tokens, Delegatic.policy_check, AuditWriter.log — ~30 unit tests

• UI status: exists-today (/ backed by install_engine_test.exs)

• Claude Design hook: Permission review modal — per-capability toggle, policy conflict alerts, audit confirmation

Story 3 · Monitor agent trust score changes

• Persona: Marketplace admin tracking agent reputation

• Goal: Watch trust scores update in real-time as agents accumulate usage + audit data

• Prerequisite: TrustWorker running; at least 1 ReputationUpdate token in flight

• Steps:

1. Navigate to /trust dashboard showing all agents

2. View agents ranked by trust tier (Excellent / Good / Fair / Low)

3. New test result arrives; TrustWorker recomputes

4. PubSub broadcasts → LiveView updates without reload

• Success: Real-time trust deltas visible across browsers

• Covers: FleetPrompt.Trust.Engine, TrustWorker.recompute, Phoenix.PubSub — ~15 unit tests

• UI status: exists-today (backend live; route currently returns 500 in production — known regression)

• Claude Design hook: LiveView trust tier cards with animated score transitions + historical sparkline

Story 4 · Crystallize an interaction trace into a manifest

• Persona: Developer who has been coding with Claude

• Goal: Convert session interaction traces into a FleetPrompt manifest so another dev can install the skill

• Prerequisite: Trace seeded in Graphonomous (via act.store_trace); FleetPrompt workspace provisioned

• Steps:

1. Navigate to /publish

2. Click "Crystallize from Graphonomous"

3. Enter trace_id; submit

4. System polls Graphonomous (retrieve.replay) → runs Skills.Crystallizer

5. Assert draft manifest appears in /drafts with source_id = trace_id

• Success: Manifest row in fleet.manifests; crystallization row in fleet.skill_crystallizations; idempotent on re-run

• Covers: Skills.Crystallizer (23 tests) + PollWorker (12) + GraphonomousClient.HTTP (18) + InstallEngine step 4 (9) — 62 unit tests in one user story

• UI status: backend exists; UI planned

• Claude Design hook: Publish page with Graphonomous connector widget + manifest preview + idempotency badge

Story 5 · Publish an agent after build success (webhook flow)

• Persona: Agentelic build pipeline emitting ConsolidationEvent to FleetPrompt

• Goal: Auto-publish tested manifest once build passes

• Prerequisite: Agentelic emits CloudEvent; spec_hash + compiled_tests_hash present

• Steps:

1. Agentelic build succeeds

2. CloudEvent ConsolidationEvent → FleetPrompt webhook

3. FleetPrompt validates envelope + workspace_id

4. Verifies spec_hash via SpecPrompt registry

5. Atomic write: fleet.agents + fleet.manifests + search index

• Success: Agent searchable within seconds; initial trust score assigned

• Covers: FleetPrompt.PipelineIntake, CloudEvent.parse, Trust.compute_initial, atomic registry write — ~25 unit tests

• UI status: mcp-only (internal admin feature)

• Claude Design hook: Webhook config UI + event inspector (admin pane)

Story 6 · Fork and customize an agent

• Persona: Team customizing a published agent

• Goal: Create a private variant with provenance tracking

• Prerequisite: Agent viewable; user in a workspace

• Steps:

1. Click "Fork" on agent detail

2. Enter new name + description

3. System copies manifest with forked_from reference

4. Fork starts at trust_score=0 (independent chain)

• Success: Fork visible in user's workspace; lineage preserved

• Covers: Forks.create_fork, Registry.clone_with_provenance, Trust.reset_for_fork

• UI status: planned

• Claude Design hook: Fork dialog with parent-attribution banner

Tests to implement first (highest value): Story 1, Story 4 (crystallize — the dark-factory hero flow), Story 3.