runtime

Google evidence runtime

This runtime is the optional API behind the Google stack visual workbench. It turns an AP2 result plus AP2 / Verifiable Intent evidence into a verifier-gated decision:

verify AP2 packet -> run A2A handoff -> sign ADK Python allow decision -> sign ADK Python tool outcome -> sign ADK Python terminal error outcome

Run locally:

pnpm --filter @atrib/integration google-evidence-runtime
curl http://127.0.0.1:8080/v1/runtime-state
curl -X POST http://127.0.0.1:8080/api/runs \
  -H 'content-type: application/json' \
  -d '{"mode":"replay","prompt":"Continue only after AP2 verification."}'

The default replay uses the committed AP2 / VI reference fixture. Point the runtime at a captured official Google AP2 sample packet with:

ATRIB_AP2_INTEROP_RESULT_JSON=/tmp/google-ap2-live/atrib-packet/ap2-result.json \
ATRIB_AP2_INTEROP_EVIDENCE_JSON=/tmp/google-ap2-live/atrib-packet/ap2-vi-evidence.json \
ATRIB_AP2_INTEROP_TRANSACTION_RECORD_JSON=/tmp/google-ap2-live/atrib-packet/atrib-transaction-record.json \
pnpm --filter @atrib/integration google-evidence-runtime

Endpoints:

• GET /health: service liveness.
• GET /api/runs: recent active runtime runs, held in memory.
• POST /api/runs: creates one verifier-gated AP2 -> A2A -> ADK Python run.
• GET /api/runs/:runId: returns one active run.
• GET /v1/runtime-state: replay verifier state for the visual workbench.
• POST /v1/replay/google-ap2-sample: explicit replay trigger.
• POST /v1/verify-ap2: accepts inline AP2 result, evidence, and transactionRecord JSON, or { "mode": "replay" }.
• POST /v1/analytics/write: writes the current gate row to BigQuery only when BIGQUERY_WRITE_ENABLED=1.
• GET /v1/merchant-adapter: minimal packet contract for a merchant or payment participant.

Operator BigQuery write:

GOOGLE_CLOUD_PROJECT=<project> \
BIGQUERY_DATASET=<dataset> \
BIGQUERY_TABLE=<table> \
BIGQUERY_WRITE_ENABLED=1 \
pnpm --filter @atrib/integration google-evidence-runtime

curl -X POST http://127.0.0.1:8080/v1/analytics/write \
  -H 'content-type: application/json' \
  -d '{"mode":"replay"}'

Public deployments should leave BIGQUERY_WRITE_ENABLED unset. The runtime can show live verifier state and active run rows without exposing a public BigQuery write endpoint.

The active run uses the Python ADK decision-ledger proof. It signs the success path and a handler-error path where a real FunctionTool raises after an allow decision and on_tool_error_callback signs the terminal outcome. The container image installs Python, uv, google-adk==2.3.0, and cryptography in addition to the Node workspace.

Deploy:

gcloud artifacts repositories create <repo> \
  --project=<project> \
  --location=<region> \
  --repository-format=docker

gcloud builds submit . \
  --project=<project> \
  --config=packages/integration/examples/google-stack-chain/runtime/cloudbuild.yaml \
  --substitutions=_TAG=$(date +%Y%m%d%H%M%S),_IMAGE=<image>

gcloud run deploy atrib-google-evidence-runtime \
  --project=<project> \
  --region=<region> \
  --image=<region>-docker.pkg.dev/<project>/<repo>/<image>:latest \
  --allow-unauthenticated \
  --memory=1Gi \
  --cpu=1 \
  --set-env-vars=GOOGLE_CLOUD_PROJECT=<project>,BIGQUERY_DATASET=<dataset>,BIGQUERY_TABLE=<table>

The deploy command prints the runtime URL. Keep deployment-specific URLs out of committed public docs and pass them with ?runtime= or local config when needed.