# Step 1: the agent on the engine Built 2026-10-20. ## What was built - **The folder,** copied from `../gcp_solution_architecture_agent` without its engine: `agents/`, `providers/`, `skills/`, `docs/`, `CONTEXT.md`, `k8s/`, `scripts/` and `eval/datasets/`. Left behind: its `app/` (the engine is now `../orchestration_engine`), its tests and evaluation harness (written against the old `app.*` modules), its plans and scratch (its history stays with it), its CI workflows and requirements files. - **`app/solution_architecture/`:** the Workflow and its subject's modules, from the engine's example of them (`tests/reference_workflow`), imports renamed. `assembly.workflow_for` builds the Workflow for the settings and the provider bundle: the Terraform runner and local tools, the pillar advisors, whether published modules may be used, and the reviews before a submission. - **`app/main.py`:** `serve(WORKFLOW, ROOT, build=workflow_for)`. The engine's `serve` and `create_app` took only a fixed Workflow; they now take `build`, used for the assembly while the card describes the Workflow (`../orchestration_engine/execution_engine/main.py`, with a test). - **`scripts/`:** imports moved to the engine and to `app.solution_architecture.modules`. - **Packaging as the operating model agent's:** `pyproject.toml` (depends on `execution-engine`), a `Dockerfile` taking the engine from a named build context and installing Terraform, `.dockerignore`, `.gitignore`, `CLAUDE.md`, `README.md`, these plans. `git init`, no commit. - **Tests:** the agent assembled from its own folder, every role file present; a whole pre-approved Execution on a scripted runtime, from a small complete spec, delivering every file the operating model agent digests and the solution guide. - **On the local platform:** `solution-architecture` in the dev kit's personal compose files (port 10040), as the operating model agent is, with Google Cloud's remote MCP servers on and reached with the mounted credentials. The Orchestration engine routes `solution_architecture` to it, and the sequence `architecture_and_operating_model` to it and then to the operating model agent. ## What differed from the plan - **The sample scope is refused by the real scoping Skill.** The engine's tests skip scoping with a scripted shortcut; here the catalogue's essential fields must be settled, and a field that binds must be carried by a requirement. The test starts from a small complete spec and carries its four binding fields with the engine's own `carrying` script. - **The schema review is real.** With Terraform installed it reads the sample's placeholder resource types against the Google provider and refuses them once; the scripted target architect submits again in the same run. ## Verified - 2 tests pass; `ruff` and `mypy` clean. The engine: 552 tests pass. - The scripted Execution's files, named as the engine publishes them as A2A artifacts, with the Orchestration engine's completion record, were given to the operating model agent's real digest (`digest.architecture`). It read two Items from the components, the Customer case, and the architecture Execution as their source. - The container answers `/health` with no MCP discovery gaps. ## Not done - No run with real models yet (step 2). - The reference's evaluation harness is not ported (step 3); its datasets are copied. - `k8s/` is as the reference had it.