Project Mapping & Drupal Stack (for AI Assistants)

Moved from CLAUDE.md to reduce context size. Consult when working with specific projects.

Local to GitLab Project Mapping

Non-Git Directories: NAS-hosted apps — Mac: /Volumes/AgentPlatform/applications/; NAS SSH: /volume1/AgentPlatform/applications/

Working Directories

• __BARE_REPOS/ - Bare git repos; 1:1 match to GitLab. No new projects.
• TESTING_DEMOS/ - Drupal demo sites for testing
• WORKING_DEMOs/ - Active working demo sites
• worktrees/ - Git worktrees created from bare repos
• platform-agents/.agents-workspace/ - Agent registry and workspace config

__BARE_REPOS/ matches GitLab. No new projects. Only submodules and extensions within existing repos (~70 repos).

Bare repos under [object Object]

• drupal/ — Drupal modules/themes (one bare repo per module). Examples: drupal/ai_agents_client.git, drupal/recipe_onboarding.git, drupal/mcp_registry.git, drupal/cedar_policy.git, etc.
• agentdashboard — AgentDash demo repo
• agentmarketplace — Agent Marketplace demo repo
• Other dirs: apps/, infra/, models/, services/, tools/, etc.

Fleet Drupal sites

ai_agents_client installs on each target site (scale to many sites). AgentDash builds flows and deploys agents that connect to those sites to run operations.

Drupal stack: what has never been done (community impact)

1. One control plane, thousands of sites. AgentDash builds flows (FlowDrop + Orchestration API). Deploy OSSA agents via ai_agents_client. Agents run Drupal Tools on each site.
2. Same capability, three surfaces. Every capability exposed as Tool (AI/agents/MCP), Orchestration service (ActivePieces, n8n, Make), and FlowDrop node (visual workflows).
3. OpenAPI to entities to Canvas to agents. API Normalization imports OpenAPI → ECK entities → Canvas pages → agents update at scale.
4. Recipes + AI + fleet deployment. AI-generated recipes deployed to thousands of sites by agents.
5. OSSA + Tool + Cedar + MCP in one stack. OSSA manifests → Drupal AI Agent plugins. Tools power function_calling and MCP. Cedar enforces policy. MCP Registry for FlowDrop/Orchestration. Canvas for dashboards.
6. Contrib and custom submodules. ai_agents_ossa, api_normalization, mcp_registry, code_executor, cedar_policy, recipe_onboarding, ai_agents_client, external_migration, layout_system_converter — all with Orchestration, FlowDrop, Canvas submodules.

OSSA Fleet Change Control Plane (canonical product direction)

One sentence: A manifest-driven (OSSA) control plane where you describe a Drupal change, the platform generates a recipe, runs it through Cedar + SoD, executes rollout via ai_agents_client, and shows everything in a pulse view backed by A2A/MCP/tracer.

How projects plug in:

• OSSA: Drupal Fleet Operator agent manifest with capabilities (generate recipe, plan rollout, apply recipe, verify, rollback).
• recipe_onboarding: "Describe it" → AI recipe generation. Four Tools: discover, validate, generate, deploy.
• Cedar + SoD: tier_1_read (Analyzer), tier_2_write_limited (Reviewer), tier_3_full_access (Executor), tier_4_policy (Approver).
• Platform services: mesh, tracer, compliance-engine, workflow-engine, agent-brain/dragonfly, GKG.
• GitLab CI/CD: Git-traceable; promotion gates map to pipeline stages.

Implementation order: (1) Event envelope + correlation IDs. (2) Recipe generation + risk classification. (3) Cedar + SoD at single tool gateway. (4) Rollout engine (canary + waves). (5) Pulse view.

GitLab Epic: In project blueflyio/agent-platform/technical-docs; label ossa-fleet-control-plane.

OaAS (Outcome as Agentic Solution): platform alignment

OaAS is the enterprise model where vendors contract for outcomes delivered by AI agents, with pricing tied to performance.

Tools to use: compliance-engine (SoD + audit evidence), cedar_policy (pre-tool enforcement), agent-tracer + A2A + pulse (correlation/observability), recipe_onboarding + fleet rollout (unit of change).

Tools to build: (1) Cedar at tool gateway. (2) Outcome evidence export. (3) Pulse + Grafana dashboard. (4) OSSA agent manifests for OaAS.