Version: 1.0 | Date: April 2026 | Type: Full Implementation (Build Pipeline + MCP + Supabase)
You are building Agentelic — the engineering layer of the [&] Protocol dark factory pipeline. Agentelic takes SpecPrompt specifications as input and produces tested, deployable agent artifacts through a deterministic 4-stage pipeline with template-based code generation.
Read `docs/spec/README.md` fully before writing a single line. It is the authoritative spec.
Agentelic is the build stage of the dark factory pipeline:
SpecPrompt (spec in) → Agentelic (build) → OS-008 (enforce) → FleetPrompt (distribute) → RuneFort (observe)Language: Elixir 1.17+ / OTP 27
Pipeline: 4-stage build (PARSE → TEST INTAKE → GENERATE → COMPILE/TEST)
MCP: JSON-RPC over HTTP (agent_create, agent_build, agent_test, agent_deploy,
agent_status, template_list, template_pin, spec_validate, test_explain)
Database: PostgreSQL via shared Supabase (agentelic.* schema, migration range 080-089)
Auth: Supabase Auth (shared [&] ecosystem — amp.profiles, amp.workspaces)
Testing: ExUnit + deterministic test DSL from spec acceptance criteria
Deploy: Fly.io (MCP server + build pipeline worker)Create this structure inside agentelic.com/:
agentelic.com/
├── lib/
│ ├── agentelic/
│ │ ├── agents/
│ │ │ └── agent.ex # Agentelic.Agents.Agent Ecto schema (section 4.4.1)
│ │ ├── builds/
│ │ │ └── build.ex # Agentelic.Builds.Build Ecto schema (section 4.4.2)
│ │ ├── testing/
│ │ │ ├── test_run.ex # Agentelic.Testing.TestRun schema (section 4.4.3)
│ │ │ ├── test_result.ex # Agentelic.Testing.TestResult embedded schema
│ │ │ ├── dsl.ex # Agentelic.Test.DSL — test case generation from specs
│ │ │ ├── runner.ex # Agentelic.Testing.Runner — execute tests with mocks
│ │ │ └── compiled_intake.ex # Agentelic.Testing.CompiledIntake — SpecPrompt test bridge
│ │ ├── deploy/
│ │ │ └── deployment.ex # Agentelic.Deploy.Deployment schema (section 4.4.4)
│ │ ├── pipeline/
│ │ │ ├── orchestrator.ex # Agentelic.Pipeline.Orchestrator — 4-stage pipeline runner
│ │ │ ├── parser.ex # Stage 1: PARSE — SpecPrompt.Spec parsing
│ │ │ ├── test_intake.ex # Stage 1.5: TEST INTAKE — compiled test loading
│ │ │ ├── generator.ex # Stage 2: GENERATE — template-based code generation
│ │ │ ├── compiler.ex # Stage 3: COMPILE — build artifact
│ │ │ └── tester.ex # Stage 4: TEST — run deterministic tests
│ │ ├── templates/
│ │ │ ├── registry.ex # Agentelic.Templates.Registry — template version management
│ │ │ ├── manifest.ex # Agentelic.Templates.Manifest — template manifest schema
│ │ │ └── renderer.ex # Agentelic.Templates.Renderer — EEx template expansion
│ │ ├── triggers/
│ │ │ ├── supabase_listener.ex # Listen for spec.specs inserts via Supabase Realtime
│ │ │ ├── cloudevents.ex # Accept CloudEvents webhooks
│ │ │ └── github_webhook.ex # Accept GitHub push events on SPEC.md changes
│ │ ├── publisher.ex # Emit ConsolidationEvent to FleetPrompt on build success
│ │ └── mcp/
│ │ ├── server.ex # MCP JSON-RPC server (section 4.5)
│ │ └── tools.ex # Tool definitions and handlers
│ ├── agentelic.ex # Application entry point
│ └── agentelic_web/
│ ├── router.ex # Phoenix router (API + webhook endpoints)
│ └── controllers/
│ ├── pipeline_controller.ex # POST /api/pipeline/trigger, /api/pipeline/github
│ └── agent_controller.ex # REST API for agents
├── test/
│ ├── pipeline/
│ │ ├── parser_test.exs
│ │ ├── generator_test.exs
│ │ └── tester_test.exs
│ ├── testing/
│ │ ├── dsl_test.exs
│ │ └── compiled_intake_test.exs
│ ├── templates/
│ │ └── registry_test.exs
│ └── fixtures/
│ ├── customer_support_spec.md # Customer-support SPEC.md from SpecPrompt section 3.3
│ └── compiled_tests.json # Pre-compiled test assertions
├── priv/
│ └── templates/ # Bundled default templates (small set for MVP)
│ ├── elixir/
│ │ └── mcp-server/
│ │ ├── template.json
│ │ ├── mix.exs.eex
│ │ └── lib/agent.ex.eex
│ └── typescript/
│ └── mcp-server/
│ ├── template.json
│ ├── package.json.eex
│ └── src/agent.ts.eex
├── mix.exs
├── Dockerfile
└── fly.tomlSet up Phoenix project with Ecto, Supabase config
Implement Ecto schemas (sections 4.4.1-4.4.4)
Agentelic.Agents.Agent with workspace_id, user_id, product_type, status state machine
Agentelic.Builds.Build with template_version, template_hash, compiled_tests_hash
Agentelic.Testing.TestRun with workspace_id, compiled_tests_hash
Agentelic.Deploy.Deployment with workspace_id, governance_policy_hash, approved_by
Implement spec parser (pipeline/parser.ex)
Consume SpecPrompt SPEC.md files → SpecPrompt.Spec struct
Compute spec_hash (SHA-256)
Validate required sections
Implement CLI — agent_create, agent_build, agent_status commands
Implement template registry (templates/registry.ex)
List templates by {framework, product_type}
Version pinning: agent-level > workspace-level > global default
Template immutability enforcement
Implement template manifest schema and validation
Implement template renderer (templates/renderer.ex)
EEx-based template expansion
Input: SpecPrompt.Spec + ampersand.json + template files
Output: generated source files
Bundle starter templates in priv/templates/
Implement Test DSL (testing/dsl.ex)
from_spec/1 — generate test cases from SpecPrompt acceptance tests
Assertion types: :contains, :not_contains, :tool_called, :tool_not_called, :escalated, :constraint_respected
Mock specs: {tool_name, match_args, return}
Implement compiled test intake (testing/compiled_intake.ex)
Load SpecPrompt.CompiledTest[] from JSON
Only approved tests (approved == true) are used
Fallback: generate compilation prompt for LLM-assisted compilation
Cache by {spec_hash, test_index}
Implement test runner (testing/runner.ex)
Mock MCP tool calls based on test case mocks
Capture agent output + tool calls
Validate against assertions
Produce TestRun with individual TestResults
Write tests using customer-support fixture
Implement 4-stage pipeline (pipeline/orchestrator.ex)
Stage 1: PARSE — call parser, produce SpecPrompt.Spec
Stage 1.5: TEST INTAKE — load compiled tests, record compiled_tests_hash on Build
Stage 2: GENERATE — call template renderer with pinned template version
Stage 3: COMPILE — shell out to mix compile / npm run build / etc.
Stage 4: TEST — call test runner, evaluate pass/fail
Record timing for each stage on Build
Deterministic: same {spec_hash, template_hash} → same artifact_hash
Implement MCP server (mcp/server.ex)
JSON-RPC over HTTP, MCP protocol v2025-03-26
9 tools: agent_create, agent_build, agent_test, agent_deploy, agent_status, template_list, template_pin, spec_validate, test_explain
Implement pipeline triggers (triggers/)
Supabase Realtime listener on spec.specs inserts/updates
CloudEvents webhook at POST /api/pipeline/trigger
GitHub webhook at POST /api/pipeline/github
Implement publisher (publisher.ex)
On build success, emit ConsolidationEvent to FleetPrompt
CloudEvents v1 envelope format
Create migration in ampersand-supabase/migrations/ (range 080-089):
080_agentelic_schema.sql — create agentelic.agents, agentelic.builds, agentelic.test_runs, agentelic.deployments
081_agentelic_rls.sql — RLS policies scoped to workspace_id
All tables use workspace-based RLS matching kag.*, rune.* pattern
Implement deployment stages — staging → canary → production
Production approval gate — approved_by must be workspace admin
Rollback support — new deployment pointing to previous build
OpenSentience integration — manifest packaging for OpenSentience runtime
Builds are deterministic. Same {spec_hash, template_hash} → same artifact_hash. No floating "latest" template references.
Templates are immutable. Once published, a template version never changes. New versions require new records.
Only approved compiled tests are used. Unapproved tests are skipped. This is the trust boundary.
Production deployments require human approval. approved_by must be a workspace admin.
Agentelic does NOT make LLM calls for test compilation. The compiled test intake generates a prompt; the agent or user executes it. Agentelic is a tool, not an agent.
Pipeline triggers validate source_hash. A stale spec hash rejects the trigger.
workspace_id is required on all tables. RLS enforces multi-tenant isolation.
user_id on agents is audit-only. Created_by at agent creation; RLS uses workspace_id.
| System | Direction | Protocol | What |
|---|---|---|---|
| SpecPrompt | SpecPrompt → Agentelic | CloudEvents / Supabase Realtime / GitHub | ConsolidationEvent triggers build pipeline |
| FleetPrompt | Agentelic → FleetPrompt | CloudEvents / Supabase Realtime | ConsolidationEvent on build success (artifact shipped) |
| OS-008 Harness | Agentelic ↔ OS-008 | MCP | Harness orchestrates build agent sessions |
| OpenSentience | Agentelic → OpenSentience | Manifest packaging | Deploy agents to OpenSentience runtime |
| Graphonomous | Agentelic → Graphonomous | MCP | Store build history, learned heuristics |
| Delegatic | Agentelic → Delegatic | MCP | Deploy gates, compliance checks |
| PRISM | Agentelic → PRISM | OutcomeSignal | Deploy success/regression feedback |
[x] agent_create creates agent from customer-support SPEC.md
[x] agent_build runs the 4-stage pipeline and produces a build artifact
[x] Same {spec_hash, template_hash} always produces same artifact_hash (deterministic)
[x] agent_test runs deterministic tests with mocked tool calls and produces pass/fail results
[x] Compiled test intake loads approved SpecPrompt.CompiledTest[] and converts to Test.DSL format
[x] Template pinning works at agent-level, workspace-level, and global default
[x] MCP server discovers tools via tools/list and executes all 9 tools
[x] Pipeline triggers fire on SpecPrompt ConsolidationEvent (all 3 transports)
[x] Build success emits ConsolidationEvent to FleetPrompt
[x] Production deployment requires approved_by validation against workspace admin
[x] Supabase migration applies cleanly alongside existing amp.*, kag.*, rune.* schemas