Skip to content
Astromesh Astromesh

Quick Start

  1. Run it.

    Terminal window
    git clone https://github.com/monaccode/astromesh-prisma.git
    cd astromesh-prisma
    uv sync --extra dev
    uv run uvicorn astromesh_prisma.app:app --port 8100

    astromesh_prisma.app:app is the app with the GCP registry and the in-memory gateway, as the Dockerfile runs it. The GCP client libraries are an extra (--extra gcp) that the tests do not need.

  2. Translate an agent. /v1/translate shows what a manifest becomes without storing it:

    Terminal window
    curl -s localhost:8100/v1/translate -H 'content-type: application/json' \
      -d "$(jq -Rs '{yaml: .}' support-agent.agent.yaml)" | jq

    The answer has the native resources (here a vertex.agent_engine with its ADK config, and a modelarmor.template if the agent redacts PII) and the gaps:

    {
      "kind": "Agent", "name": "support-agent", "provider": "gcp",
      "resources": [
        {"resource_type": "vertex.agent_engine", "name": "support-agent",
         "source_section": "identity", "config": {"…": "…"}}
      ],
      "gaps": [
        {"section": "model.fallback", "coverage": "unsupported",
         "detail": "model candidate(s) llama3.1:8b are dropped: the native agent calls only the first one"}
      ]
    }

    For a model the native agent can call, use Gemini the way the core reaches it, through litellm: source: litellm, model: vertex_ai/gemini-2.5-flash.

  3. Translate its knowledge pipeline, if the agent has spec.knowledge. The agent refers to the corpus through a binding; the pipeline’s own manifest is what provisions it:

    Terminal window
    curl -s localhost:8100/v1/translate -H 'content-type: application/json' \
      -d "$(jq -Rs '{yaml: .}' help-center.rag.yaml)" | jq '.resources[].resource_type'
    # "vertex.rag_corpus"

  4. Submit and reconcile.

    Terminal window
    curl -s localhost:8100/v1/specs -H 'content-type: application/json' \
      -d "$(jq -Rs '{yaml: .}' support-agent.agent.yaml)"
    # {"kind": "Agent", "name": "support-agent"}
    

    curl -s -X POST localhost:8100/v1/specs/Agent/support-agent/reconcile
    # {"created": ["support-agent"], "updated": [], "deleted": [], "gaps": ["…"]}

    Reconcile again without changes and nothing is created or updated. Change the prompt or a tool, submit, reconcile, and the engine shows up in updated.

Run the tests

Section titled “Run the tests”
Terminal window
uv run pytest -v

The suite runs offline. It includes the contract test against the core’s agent schema and the test that runs a generated agent inside ADK.

Next steps

Section titled “Next steps”