hongming-ceo-delegated
456a52d473
Secret scan / Scan diff for credential-shaped strings (push) Successful in 6s
CI / Template validation (static) (push) Successful in 7s
CI / Adapter unit tests (push) Successful in 34s
CI / Template validation (runtime) (push) Successful in 3m2s
CI / T4 tier-4 conformance (live) (push) Successful in 3m0s
CI / validate (push) Successful in 2s
molecule-ai-workspace-template-google-adk
Workspace template for the google-adk runtime — Gemini agents powered by
Google's Agent Development Kit (ADK), governed by Molecule's A2A org.
Design (RFC internal#730)
- ADK as engine only:
LlmAgent+Runner+McpToolset. Installgoogle-adk[mcp]==2.1.0— never the[a2a]extra (it pinsa2a-sdk<0.4, incompatible with the platform'sa2a-sdk>=1.0). - Tools over MCP: ADK's native
McpToolsetconnects (stdio) to the platform'sa2a_mcp_server— the same 8-tool surface the CLI runtimes use (delegate / memory / peers / messaging). No LangChain. - Molecule-authored executor (
google_adk_executor.py) drives ADK'sRunnerand bridges its event stream to the platform's a2a-1.xEventQueue/TaskUpdater, with heartbeat task accounting. OWASP / OTEL / Temporal are deferred (off/dormant in prod) per the RFC. - BYOK: set
GOOGLE_API_KEY(AI Studio) — or Vertex viaGOOGLE_GENAI_USE_VERTEXAI=1+GOOGLE_CLOUD_PROJECT.
Files
| File | Role |
|---|---|
adapter.py |
GoogleADKAdapter — builds MCPToolset + LlmAgent + Runner |
google_adk_executor.py |
Runner → a2a-1.x bridge + heartbeat |
_routing.py |
provider:model → bare Gemini id; AI Studio / Vertex |
config.yaml |
default runtime config (Gemini 2.5 Pro) |
Dockerfile / entrypoint.sh |
lean image (no node/T4); uid-1000 drop |
Status
Unit-tested (tests/, routing + executor helpers). Stage A/B/C verification +
cross-repo registration (molecule-core manifest/knownRuntimes/canvas;
controlplane runtime_image_pins/providers; molecule-ci validator) +
providers-projection are tracked follow-ups per RFC internal#730.