replica-omnisciente/.aurelio/garden/agent-garden/google-cloud.yaml
Raphael Cautus (Maestro) 749432fefc feat(brain): garden, mirrors, vault-sync, provider registry consolidation
- .aurelio/garden/: model + agent garden (Google Cloud entries)
- .aurelio/mirrors/: sync-mirrors.yaml + state tracking
- .aurelio/skills/gcp/: Google Cloud skill
- Consolidation audit + execution plan (2026-07-30)
- vault-sync.py: Obsidian → GBrain MCP ingestion daemon
- brain-to-gbrain.py: brain → GBrain migration tool
- Provider registry + dist mirrors updated
- .gitignore: exclude .runner, .mimocode/.cron-lock, drift/target

Co-authored-by: Álvaro de Campos <campos@portugalfuturista.org>
2026-07-31 14:57:24 +01:00

185 lines
4.8 KiB
YAML

# PF Agent Garden — Google Cloud
#
# Agents and agentic services that run on GCP or call Google Cloud / Workspace APIs.
# Distinct from the generic `gemini` chat provider.
platform: google-cloud
realm_default: europe-west1
workload_identity: required
agents:
- id: gcp-deployer
name: GCP Deployer
description: Provisions and mutates GCP resources (GKE, Cloud Run, Cloud Functions, Cloud Storage, Pub/Sub, BigQuery, Cloud SQL, Secret Manager) from declarative specs.
kind: infrastructure
surfaces:
- dirac
- portal
triggers:
- user-command
- git-push
inputs:
- terraform
- gcloud-yaml
- openapi
outputs:
- deployed-url
- resource-manifest
required_apis:
- deploymentmanager.googleapis.com
- cloudbuild.googleapis.com
- run.googleapis.com
- container.googleapis.com
- id: vertex-coder
name: Vertex Coder
description: Coding agent backed by Vertex AI Gemini. Reads/writes code in a GKE/Cloud Run sandbox and streams results through Dirac / MCP.
kind: coding
surfaces:
- dirac
- mcp
model: gemini-2.5-pro-preview-06-05
runtime: cloud-run
triggers:
- dirac-task
- mcp-tool-call
inputs:
- repository
- prompt
outputs:
- diff
- test-results
- id: gcp-cost-guardian
name: GCP Cost Guardian
description: Monitors billing, budgets, and resource labels; alerts on anomalies and suggests rightsizing.
kind: observability
surfaces:
- portal
- mcp
triggers:
- schedule
- billing-alert
inputs:
- billing-export
- resource-labels
outputs:
- alert
- recommendation
required_roles:
- roles/billing.viewer
- roles/monitoring.viewer
- id: workspace-delegator
name: Workspace Delegator
description: Calls Google Workspace APIs (Directory, Gmail, Calendar, Drive) via domain-wide delegation. Useful for tenant automation and mail/calendar agents.
kind: integration
surfaces:
- mcp
- portal
triggers:
- user-command
- workflow
inputs:
- workspace-scope
- service-account
outputs:
- api-result
- event
requires_domain_wide_delegation: true
scopes:
- https://www.googleapis.com/auth/admin.directory.user.readonly
- https://www.googleapis.com/auth/calendar
- https://www.googleapis.com/auth/gmail.modify
- https://www.googleapis.com/auth/drive
- id: pubsub-agent-router
name: Pub/Sub Agent Router
description: Receives events from Pub/Sub and dispatches them to other PF agents running on Cloud Run, GKE, or via MCP.
kind: orchestration
surfaces:
- mcp
- portal
triggers:
- pubsub-message
inputs:
- pubsub-payload
- routing-key
outputs:
- agent-invocation
required_apis:
- pubsub.googleapis.com
- run.googleapis.com
- id: bq-memory-analyst
name: BigQuery Memory Analyst
description: Queries structured agent memory and telemetry stored in BigQuery; returns synthesized answers with citations.
kind: memory
surfaces:
- mcp
- portal
model: gemini-2.5-flash-preview-06-05
triggers:
- user-query
- schedule
inputs:
- bigquery-dataset
- sql-question
outputs:
- sql-query
- result-summary
required_roles:
- roles/bigquery.dataViewer
- roles/bigquery.jobUser
- id: gcs-artifact-librarian
name: GCS Artifact Librarian
description: Indexes and retrieves artifacts from Cloud Storage buckets; supports code artifacts, documents, and media.
kind: memory
surfaces:
- mcp
- portal
triggers:
- user-query
- event
inputs:
- bucket-prefix
- metadata
outputs:
- signed-url
- object-list
required_roles:
- roles/storage.objectViewer
- id: secret-bootstrapper
name: Secret Bootstrapper
description: Loads runtime secrets from Secret Manager into agent workloads; never logs values and rotates on schedule.
kind: security
surfaces:
- mcp
- dirac
triggers:
- deployment
- rotation-schedule
inputs:
- secret-name
- workload-identity
outputs:
- mounted-secret
required_apis:
- secretmanager.googleapis.com
runtimes:
- name: cloud-run
use_for: stateless-http
scaling: request-driven
- name: gke-autopilot
use_for: long-running-workloads
scaling: workload-driven
- name: cloud-functions
use_for: event-driven-functions
notes:
- All agents default to `europe-west1` unless EU residency or latency requires otherwise.
- Production agents must use Workload Identity or dedicated service accounts; no downloaded keys.
- Agents calling Workspace APIs require domain-wide delegation registered in the Google Admin console.