@raishin/vanguard-frontier-agentic 1.4.0 → 1.6.0

package/skills/gcp/gcp-firebase-developer/SKILL.md

@@ -0,0 +1,75 @@
+---
+name: gcp-firebase-developer
+description: "Build, configure, and operate Firebase-powered web and mobile applications — covering Firestore, Firebase Auth, Firebase Hosting, Cloud Functions for Firebase, Firebase Storage, App Check, Firebase Remote Config, and Firebase Analytics. Use when building mobile/web apps with Firebase, setting up authentication flows, designing Firestore data models, deploying Firebase Hosting, or configuring Firebase security rules."
+allowed-tools: Read Grep Glob
+metadata:
+  author: "github: Raishin"
+  version: "0.1.0"
+  updated: "2026-05-09"
+  category: platform
+---
+
+# GCP Firebase Developer
+
+## Overview
+
+Firebase is Google's app development platform for web and mobile. It provides a unified suite of backend services, SDKs, and tooling. Key products: Firestore (NoSQL real-time database), Firebase Auth (identity), Hosting (global CDN), Cloud Functions for Firebase (server logic), Firebase Storage (file storage), App Check (abuse prevention), Remote Config, Analytics.
+
+## Quick Start
+
+```bash
+# Install Firebase CLI
+npm install -g firebase-tools
+
+# Login
+firebase login
+
+# Initialize project
+firebase init
+
+# Deploy
+firebase deploy
+```
+
+## Reference Directory
+
+Load only when needed:
+
+| Scenario | Trigger Keywords | Reference |
+|---|---|---|
+| Firestore data modeling | collection, document, subcollection, NoSQL schema, query design | references/firestore.md |
+| Authentication | auth, sign-in, OAuth, email, phone, anonymous, custom claims | references/auth.md |
+| Security Rules | rules, allow, deny, Firestore rules, Storage rules | references/security-rules.md |
+| Cloud Functions | functions, trigger, callable, HTTP, Pub/Sub, Firestore trigger | references/functions.md |
+| Hosting | deploy, hosting, CDN, SPA, redirect, custom domain | references/hosting.md |
+| App Check | abuse, attestation, reCAPTCHA, DeviceCheck, Play Integrity | references/app-check.md |
+| Emulator Suite | local testing, emulator, integration test | references/emulators.md |
+| Firebase Extensions | extension, marketplace, prebuilt | references/extensions.md |
+
+## Core Operating Rules
+
+- Firebase SDKs come in Web (v9 modular), React Native, Flutter, iOS (Swift), and Android (Kotlin/Java) variants. Always confirm the platform before providing SDK code.
+- Firestore security rules are the primary access control layer for client-side apps — never assume the app logic is sufficient authorization.
+- Use Firebase Emulator Suite for local development and integration testing before deploying to production.
+- App Check is mandatory for production apps to prevent unauthorized API access and quota abuse.
+- Cloud Functions for Firebase run on Cloud Run (gen2) by default in new projects — prefer gen2 for better performance, concurrency, and VPC connectivity.
+- Firebase Auth supports custom claims for role-based access — use claims, not Firestore documents, as the authority for authorization decisions.
+- Never expose service account keys or private Firebase config in client-side code. The Firebase client SDK config (apiKey, projectId) is not a secret — it is safe to embed in client bundles.
+- Use Firebase Hosting rewrites to serve SPAs and proxy Cloud Functions securely.
+
+## Response Shape
+
+Platform/SDK confirmation, Firestore data model, security rules, auth flow, functions, hosting config, App Check setup, test plan with emulators.
+
+## Official Docs
+
+- https://firebase.google.com/docs
+- https://firebase.google.com/docs/firestore
+- https://firebase.google.com/docs/auth
+- https://firebase.google.com/docs/hosting
+- https://firebase.google.com/docs/functions
+- https://firebase.google.com/docs/app-check
+
+## Security Notes
+
+Read-only skill. Do not deploy to production, modify Firestore security rules, or change Firebase project settings without explicit approval. Client config (apiKey, projectId) is public — service account keys are private and must never be embedded in client code.

package/skills/gcp/gcp-firebase-developer/metadata.json

@@ -0,0 +1,22 @@
+{
+  "id": "gcp-firebase-developer",
+  "name": "GCP Firebase Developer",
+  "type": "skill",
+  "provider": "gcp",
+  "harnesses": ["codex", "claude-code", "cursor", "gemini", "kiro", "other"],
+  "summary": "Build, configure, and operate Firebase-powered web and mobile applications — covering Firestore, Firebase Auth, Firebase Hosting, Cloud Functions for Firebase, Firebase Storage, App Check, Remote Config, and Analytics.",
+  "source_type": "original",
+  "official_docs": [
+    "https://firebase.google.com/docs",
+    "https://firebase.google.com/docs/firestore",
+    "https://firebase.google.com/docs/auth",
+    "https://firebase.google.com/docs/hosting",
+    "https://firebase.google.com/docs/functions",
+    "https://firebase.google.com/docs/app-check"
+  ],
+  "security_notes": "Read-only skill. Do not deploy to production, modify Firestore security rules, or change Firebase project settings without explicit approval. Client config (apiKey, projectId) is public — service account keys are private and must never be embedded in client code.",
+  "last_verified": "2026-05-09",
+  "path": "skills/gcp/gcp-firebase-developer",
+  "author": "github: Raishin",
+  "version": "0.1.0"
+}

package/skills/gcp/gcp-gcs-data-perimeter-governor/SKILL.md

@@ -0,0 +1,60 @@
+---
+name: gcp-gcs-data-perimeter-governor
+description: Govern Google Cloud Storage data perimeters — uniform bucket-level access enforcement, public access prevention, VPC Service Controls perimeter coverage, IAM Conditions for time-bounded access, Object Lifecycle policies, and data residency compliance.
+allowed-tools: Read Grep Glob
+metadata:
+  author: "github: Raishin"
+  version: "0.1.0"
+  updated: "2026-05-09"
+  category: security
+---
+
+# GCP GCS Data Perimeter Governor
+
+## Purpose
+
+Act as the GCP GCS data perimeter reviewer who treats any allUsers/allAuthenticatedUsers binding as a CRITICAL finding requiring immediate remediation and refuses to clear a bucket's perimeter posture without verifying VPC-SC coverage and uniform bucket-level access.
+
+## When to use
+
+Use this skill for:
+
+- GCS bucket IAM audit (allUsers/allAuthenticatedUsers detection, least-privilege enforcement)
+- Uniform bucket-level access (UBL) enforcement review
+- VPC Service Controls perimeter coverage for storage.googleapis.com
+- Org policy constraints/storage.publicAccessPrevention verification for regulated environments
+- IAM Conditions review for time-bounded and attribute-based GCS access
+- Object Lifecycle policy safety review (irreversible delete rules, versioning prerequisite)
+- Data residency configuration review (dual-region, multi-region bucket location vs. requirements)
+- Bucket Lock and retention policy review for compliance workloads
+- HMAC key and signed URL token hygiene assessment
+
+## Lean operating rules
+
+- Prefer live GCP evidence from sanitized `gsutil` and `gcloud storage` output when available; otherwise use official Google Cloud documentation.
+- allUsers and allAuthenticatedUsers bindings are CRITICAL — GCS buckets with these bindings are public and indexed within minutes; remediation must be immediate.
+- Uniform bucket-level access must be enabled — legacy ACLs cannot be consistently audited and create conflicting access paths.
+- VPC Service Controls perimeters without storage.googleapis.com in scope are incomplete and do not prevent GCS data exfiltration.
+- Org policy constraints/storage.publicAccessPrevention at the org level provides a stronger control than bucket-level public access prevention — verify it is enforced for regulated environments.
+- Object Lifecycle delete rules are irreversible — always confirm versioning is enabled before reviewing or approving delete lifecycle rules.
+- Separate confirmed facts from inference. If state was not queried or shown, say so.
+- Challenge missing VPC-SC coverage, legacy ACL usage, absent public access prevention org policy, and delete lifecycle rules on unversioned buckets.
+- Keep the answer scoped, reversible, least-privilege, and explicit about blockers or unknowns.
+- Load references only when needed; do not pull all deep guidance into short answers.
+
+## References
+
+Load these only when needed:
+
+- [Workflow and output contract](references/workflow-and-output.md) — use when executing the full GCS perimeter review, IAM audit, or formatting the final answer.
+- [Official sources](references/official-sources.md) — use when grounding GCP GCS access control and VPC-SC service behavior or checking the detailed source list.
+
+## Response minimum
+
+Return, at minimum:
+
+- the scoped bucket target and evidence level,
+- the public access exposure status (allUsers/allAuthenticatedUsers check result),
+- the main perimeter risks (missing VPC-SC, disabled UBL, absent org policy),
+- the safest remediation actions with explicit priority order,
+- the assumptions or blockers that prevent stronger conclusions.

package/skills/gcp/gcp-gcs-data-perimeter-governor/metadata.json

@@ -0,0 +1,28 @@
+{
+  "id": "gcp-gcs-data-perimeter-governor",
+  "name": "GCP GCS Data Perimeter Governor",
+  "type": "skill",
+  "provider": "gcp",
+  "harnesses": [
+    "codex",
+    "claude-code",
+    "cursor",
+    "gemini",
+    "kiro",
+    "other"
+  ],
+  "summary": "Govern Google Cloud Storage data perimeters — uniform bucket-level access enforcement, public access prevention, VPC Service Controls perimeter coverage, IAM Conditions for time-bounded access, Object Lifecycle policies, and data residency compliance.",
+  "source_type": "original",
+  "official_docs": [
+    "https://cloud.google.com/storage/docs/access-control/uniform-bucket-level-access",
+    "https://cloud.google.com/storage/docs/public-access-prevention",
+    "https://cloud.google.com/vpc-service-controls/docs/supported-products",
+    "https://cloud.google.com/storage/docs/lifecycle",
+    "https://cloud.google.com/storage/docs/bucket-lock"
+  ],
+  "security_notes": "GCS buckets with allUsers binding are indexed by search engines and data scrapers within minutes of creation — remediation must be immediate. VPC-SC perimeter around GCS requires testing in dry-run mode first — enforcement mode can break legitimate GCS access from outside the perimeter instantly.",
+  "last_verified": "2026-05-09",
+  "path": "skills/gcp/gcp-gcs-data-perimeter-governor",
+  "author": "github: Raishin",
+  "version": "0.1.0"
+}

package/skills/gcp/gcp-gcs-data-perimeter-governor/references/official-sources.md

@@ -0,0 +1,19 @@
+# Official sources
+
+Use this reference only when you need source grounding for GCP GCS access control, VPC Service Controls, and data residency service behavior or the detailed source list.
+
+## Google Cloud documentation
+
+Use these as starting points, not as proof of the user's live GCP state:
+- https://cloud.google.com/storage/docs/access-control/uniform-bucket-level-access
+- https://cloud.google.com/storage/docs/public-access-prevention
+- https://cloud.google.com/vpc-service-controls/docs/supported-products
+- https://cloud.google.com/storage/docs/lifecycle
+- https://cloud.google.com/storage/docs/bucket-lock
+- https://cloud.google.com/storage/docs/access-control/iam-conditions
+- https://cloud.google.com/storage/docs/locations
+- https://cloud.google.com/resource-manager/docs/organization-policy/restricting-domains
+
+## Grounding rule
+
+Official documentation explains GCP GCS access control and VPC-SC behavior. It does not prove the user's current bucket IAM bindings, org policy enforcement, or lifecycle policy configuration. Prefer live GCP CLI/API evidence or sanitized user-provided evidence for current-state claims.

package/skills/gcp/gcp-gcs-data-perimeter-governor/references/workflow-and-output.md

@@ -0,0 +1,67 @@
+# Workflow and output contract
+
+Use this reference only when performing the full GCS perimeter review, IAM audit, VPC-SC coverage analysis, or data residency compliance pass.
+
+## Review domains
+
+Check these areas before giving a verdict:
+- Public access exposure: allUsers and allAuthenticatedUsers IAM bindings across all buckets in scope
+- Uniform bucket-level access: UBL enabled/disabled state, legacy ACL presence
+- VPC Service Controls: storage.googleapis.com included in perimeter, dry-run vs. enforcement mode
+- Org policy: constraints/storage.publicAccessPrevention enforced at org or folder level
+- IAM Conditions: time-bounded bindings, resource path conditions, IP-based conditions
+- Object Lifecycle: delete rules with versioning prerequisite check, transition rules review
+- Data residency: bucket location (single-region vs. dual-region vs. multi-region) vs. compliance requirement
+- Bucket Lock: retention policy existence, lock status for compliance workloads
+- HMAC keys: active HMAC keys associated with service accounts (exfiltration risk)
+
+## Safe workflow
+
+1. **Frame scope**
+   - Bucket list and project:
+   - Data classification (regulated/internal/public):
+   - Compliance requirements (GDPR, HIPAA, data residency):
+   - Required outcome:
+   - Explicit non-goals:
+2. **Collect evidence**
+   - Prefer live GCP CLI/API read-only evidence if available.
+   - Otherwise inspect repository IaC/config, sanitized user evidence, or official Google Cloud docs.
+   - Label each finding as `live evidence`, `repo evidence`, `user-provided evidence`, `documentation-based`, or `inference`.
+3. **Stress-test perimeter**
+   - Are any buckets bound to allUsers or allAuthenticatedUsers?
+   - Is UBL enabled on all buckets?
+   - Is storage.googleapis.com included in the VPC-SC perimeter?
+   - Is the org policy constraints/storage.publicAccessPrevention enforced?
+   - Are delete lifecycle rules on unversioned buckets?
+   - What evidence is missing?
+4. **Recommend the smallest safe action**
+   - Prefer narrow scope, staged rollout (VPC-SC dry-run first), validation, and rollback.
+   - Flag allUsers bindings as CRITICAL requiring immediate removal.
+   - If the safest action is to stop and gather evidence, say that plainly.
+
+## Output contract
+
+Return this structure:
+```markdown
+# GCP GCS Data Perimeter Governor: <scope>
+## Executive verdict
+- Status: SECURE / EXPOSED / PARTIAL / NEEDS EVIDENCE
+- Public access risk: <clean / CRITICAL — allUsers binding detected>
+- Biggest gap:
+- Evidence level:
+## Scope and assumptions
+- Confirmed:
+- Unknown:
+- Out of scope:
+## Findings
+| Severity | Finding | Evidence | Why it matters | Minimum safe action |
+|---|---|---|---|---|
+## Recommended remediation actions
+1. <action> — owner: <owner>, validation: <check>, rollback: <rollback>
+## Data residency and compliance
+- Bucket locations: <summary>
+- Org policy enforcement: <present / absent / unknown>
+- Bucket Lock: <configured / missing / unknown>
+## Residual risk
+- <risk or explicit none>
+```

package/skills/gcp/gcp-gemini-api-developer/SKILL.md

@@ -0,0 +1,97 @@
+---
+name: gcp-gemini-api-developer
+description: "Build, integrate, and debug Gemini API applications on Google Cloud Agent Platform (formerly Vertex AI) using the unified google-genai SDK. Covers text generation, multimodal inputs, function calling, structured output, embeddings, context caching, batch prediction, streaming, Live API (bidirectional voice/video), and model tuning across Python, TypeScript/JavaScript, Go, Java, and C#. Use when building Gemini-powered applications, migrating from deprecated Vertex AI or google-generativeai SDKs, or integrating Gemini capabilities into a GCP-hosted service."
+allowed-tools: Read Grep Glob
+metadata:
+  author: "github: Raishin"
+  version: "0.1.0"
+  updated: "2026-05-09"
+  category: ai
+---
+
+# GCP Gemini API Developer
+
+## IMPORTANT BRANDING NOTE
+
+"Agent Platform" (full name: Gemini Enterprise Agent Platform) was previously named "Vertex AI". Many web resources still use "Vertex AI" branding.
+
+## CRITICAL SDK NOTE
+
+The following SDKs are DEPRECATED and must NOT be used:
+- `google-cloud-aiplatform` (Python)
+- `@google-cloud/vertexai` (Node.js)
+- `google-generativeai` (Python — Gemini Developer API SDK)
+
+Use ONLY the unified Gen AI SDK:
+- Python: `google-genai` (`pip install google-genai`)
+- JavaScript/TypeScript: `@google/genai` (`npm install @google/genai`)
+- Go: `google.golang.org/genai` (`go get google.golang.org/genai`)
+- Java: `com.google.genai:google-genai`
+- C#/.NET: `Google.GenAI` (`dotnet add package Google.GenAI`)
+
+## Authentication
+
+```bash
+# Application Default Credentials (ADC) — preferred
+export GOOGLE_CLOUD_PROJECT='your-project-id'
+export GOOGLE_CLOUD_LOCATION='global'  # Use 'global' for automatic region routing
+export GOOGLE_GENAI_USE_VERTEXAI=true
+```
+
+## Model Selection
+
+- `gemini-3.1-pro-preview` — complex reasoning, coding, research (1M tokens)
+- `gemini-3-flash-preview` — fast, balanced, multimodal (1M tokens)
+- `gemini-3.1-flash-lite-preview` — high-frequency lightweight tasks
+- `gemini-3-pro-image-preview` — image generation and editing
+- `gemini-live-2.5-flash-native-audio` — Live Realtime API (bidirectional audio/video)
+- DEPRECATED (do not use): gemini-2.0-*, gemini-1.5-*, gemini-1.0-*, gemini-pro
+
+## Quick Start (Python)
+
+```python
+from google import genai
+client = genai.Client()  # picks up env vars automatically
+response = client.models.generate_content(
+    model="gemini-3-flash-preview",
+    contents="Explain transformer architecture"
+)
+print(response.text)
+```
+
+## Reference Directory
+
+Load only when needed:
+
+| Scenario | Trigger Keywords | Reference |
+|---|---|---|
+| Text + multimodal | chat, image, video, audio, streaming | references/text-multimodal.md |
+| Function calling + tools | tool use, function call, grounding, code execution | references/tools.md |
+| Structured output | JSON, schema, structured, typed response | references/structured-output.md |
+| Embeddings | embedding, semantic search, vector | references/embeddings.md |
+| Context caching | cache, large context, caching tokens | references/caching.md |
+| Batch prediction | batch, async, large dataset | references/batch.md |
+| Live API | live, realtime, voice, video streaming, bidirectional | references/live-api.md |
+| Model tuning | fine-tune, SFT, preference tuning | references/tuning.md |
+| Safety | safety filter, threshold, harm category | references/safety.md |
+| SDK migration | migrate, deprecated, upgrade | references/migration.md |
+
+## Core Rules
+
+- ALWAYS use the unified `google-genai` SDK family. If user code imports `google-cloud-aiplatform`, `@google-cloud/vertexai`, or `google-generativeai`, flag it as deprecated and provide migration guidance.
+- Use `location="global"` (global endpoint) by default for automatic capacity routing. Only use a specific region if the user explicitly requests it.
+- Initialize the client without parameters when environment variables are set — don't hardcode project/location in code.
+- `gemini-3.1-pro-preview` ≠ `gemini-3-pro-preview` — the latter does NOT exist; use the correct model IDs.
+- Context caching (`CachedContent`) reduces cost for repeated large contexts (system prompts, documents) — recommend it proactively for production workloads with stable large contexts.
+- For production, consult docs for stable model version aliases rather than using `-preview` models.
+- Batch prediction (`BatchJob`) is for async large-dataset inference — use it instead of looping `generate_content()` for bulk processing.
+
+## Official Docs
+
+- https://cloud.google.com/vertex-ai/generative-ai/docs/overview
+- https://cloud.google.com/vertex-ai/generative-ai/docs/sdks/overview
+- https://cloud.google.com/vertex-ai/generative-ai/docs/context-cache/context-cache-overview
+
+## Security Notes
+
+Read-only advisory. Never embed API keys or service account credentials in code examples. Use ADC and environment variables. Do not call batch jobs or fine-tuning jobs on production data without explicit user approval.

package/skills/gcp/gcp-gemini-api-developer/metadata.json

@@ -0,0 +1,19 @@
+{
+  "id": "gcp-gemini-api-developer",
+  "name": "GCP Gemini API Developer",
+  "type": "skill",
+  "provider": "gcp",
+  "harnesses": ["codex", "claude-code", "cursor", "gemini", "kiro", "other"],
+  "summary": "Build, integrate, and debug Gemini API applications on Google Cloud Agent Platform using the unified google-genai SDK — covering text generation, multimodal inputs, function calling, structured output, embeddings, context caching, batch prediction, Live API, and model tuning.",
+  "source_type": "original",
+  "official_docs": [
+    "https://cloud.google.com/vertex-ai/generative-ai/docs/overview",
+    "https://cloud.google.com/vertex-ai/generative-ai/docs/sdks/overview",
+    "https://cloud.google.com/vertex-ai/generative-ai/docs/context-cache/context-cache-overview"
+  ],
+  "security_notes": "Read-only advisory. Never embed API keys or service account credentials in code examples. Use ADC and environment variables. Do not call batch jobs or fine-tuning jobs on production data without explicit user approval.",
+  "last_verified": "2026-05-09",
+  "path": "skills/gcp/gcp-gemini-api-developer",
+  "author": "github: Raishin",
+  "version": "0.1.0"
+}

package/skills/gcp/gcp-gke-platform-operator/SKILL.md

@@ -0,0 +1,136 @@
+---
+name: gcp-gke-platform-operator
+description: Operate GKE clusters (Standard and Autopilot), manage node pools, configure Workload Identity, enforce Binary Authorization, plan node pool upgrades, and review cluster security posture.
+allowed-tools: Read Grep Glob
+metadata:
+  author: "github: Raishin"
+  version: "0.2.0"
+  updated: "2026-05-09"
+  category: platform
+---
+
+# GCP GKE Platform Operator
+
+## Purpose
+
+Act as a rigorous GKE platform operator. Keep GKE clusters secure, upgraded, and operating with zero-trust pod identity and image provenance enforcement.
+
+## Reference Directory
+
+Load the relevant reference based on the request. Prefer the most specific match.
+
+| Scenario | Trigger Keywords | Reference |
+|---|---|---|
+| Golden Path & Defaults | golden path, Day-0, production defaults, cluster creation | [Golden path and Day-0 checklist](#golden-path-configuration) |
+| Networking | private cluster, VPC, subnet, Gateway API, DNS, ingress, datapath | [Networking section](#networking) |
+| Security & IAM | Workload Identity, Secret Manager, RBAC, Binary Auth, hardening | [Security section](#security) |
+| Scaling | HPA, VPA, autoscaler, NAP, scale pods, scale nodes | [Scaling section](#scaling) |
+| Cost | Spot VMs, rightsizing, CUD, budget, OPTIMIZE_UTILIZATION | [Cost section](#cost) |
+| AI/ML Inference | LLM serving, GPU, TPU, vLLM, GIQ, model deployment | [AI/ML Inference section](#aiml-inference) |
+| Upgrades | maintenance window, release channel, patching, version | [Upgrades section](#upgrades) |
+| Observability | monitoring, Prometheus, Grafana, metrics, alerts | [Observability section](#observability) |
+| Multi-tenancy | namespace isolation, team access, RBAC planning | [Multi-tenancy section](#multi-tenancy) |
+| Batch & HPC | batch jobs, high performance, MPI, parallel workloads | [Batch & HPC section](#batch--hpc) |
+| Backup & DR | backup, restore, disaster recovery, CMEK | [Backup & DR section](#backup--dr) |
+| Storage | PVC, persistent volume, StorageClass, Filestore, GCS FUSE | [Storage section](#storage) |
+
+## Day-0 vs Day-1 Decisions
+
+**Day-0 decisions** are made at cluster creation and are hard or impossible to change afterwards. Always surface and confirm these before generating any cluster config:
+
+| Decision | Why It's Hard to Change |
+|---|---|
+| Autopilot vs Standard | Cannot convert after creation |
+| Private nodes (`enablePrivateNodes`) | Requires cluster recreation to change |
+| VPC and subnet | Network range cannot be shrunk post-creation |
+| IP allocation policy (pod/svc CIDRs) | Cannot be modified after creation |
+| Private endpoint enforcement | Changing opens public control plane |
+| Workload Identity pool | Requires workload reconfiguration |
+| Release channel | Changing channel may trigger immediate upgrade |
+
+**Day-1 decisions** can be changed after cluster creation (some require node pool recreation or short downtime):
+- Secret Manager integration
+- Monitoring/logging component list
+- Binary Authorization policy
+- RBAC bindings
+- Node pool machine type (via new pool + drain)
+- Maintenance exclusion windows
+
+## When to use
+
+Use this skill for:
+
+- GKE cluster type selection (Standard vs. Autopilot) and initial setup
+- Node pool design, sizing, and upgrade planning
+- Workload Identity configuration and audit
+- Binary Authorization policy setup and enforcement path
+- Release channel selection and upgrade strategy
+- Cluster security posture review (network policies, Pod Security Standards, RBAC)
+
+## Key GKE specifics
+
+- GKE Autopilot: Google manages nodes, you manage pods. Billing is per Pod CPU/memory. Cannot run privileged containers or DaemonSets. Best for most workloads.
+- GKE Standard: you manage nodes. More flexibility but more operational burden.
+- Workload Identity: maps Kubernetes ServiceAccounts to GCP Service Accounts via annotation — eliminates SA key files from pods. Always prefer over mounted key files.
+- Binary Authorization: enforces image signatures at admission. Must be set to WARN mode before ENFORCE mode — enforce mode will break deployments if images are unsigned.
+- Node pool upgrades: cluster must be on a release channel (Rapid/Regular/Stable) for automated upgrades. Manual upgrades for custom versioning.
+- Release channels: Rapid > Regular > Stable in terms of how quickly new Kubernetes versions arrive. Use Regular for production.
+
+## Lean operating rules
+
+- Prefer official GCP documentation and live evidence over memory or inference.
+- Separate confirmed facts from inference. If state was not queried or shown, say so.
+- Challenge missing Workload Identity, Binary Authorization in permissive mode, skipped node pool upgrades, and overbroad RBAC.
+- Keep the answer scoped, reversible, least-privilege, and explicit about blockers or unknowns.
+- Load references only when needed; do not pull all deep guidance into short answers.
+
+## References
+
+Load these only when needed:
+
+- [Workflow and output contract](references/workflow-and-output.md) — use when executing the full review or formatting the final answer.
+- [Official sources](references/official-sources.md) — use when grounding GKE behavior or checking the detailed source list.
+
+## AI/ML Inference
+
+GKE supports GPU/TPU workloads for LLM inference via the GKE Inference Quickstart (GIQ) — a `gcloud` workflow that generates optimized Kubernetes manifests for specific model + accelerator + serving framework combinations.
+
+### Discovery and manifest generation
+
+```bash
+# List all supported models
+gcloud container ai profiles models list
+
+# Find valid accelerator + server combinations for a model
+gcloud container ai profiles list --model=gemma-2-9b-it
+
+# Generate an optimized manifest
+gcloud container ai profiles manifests create \
+  --model=gemma-2-9b-it \
+  --model-server=vllm \
+  --accelerator-type=nvidia-l4 \
+  --target-ntpot-milliseconds=50 > inference.yaml
+
+# Deploy
+kubectl apply -f inference.yaml
+```
+
+Supported model-server values: `vllm`, `tgi`, `triton`, `tensorrt-llm`
+Common accelerator types: `nvidia-l4`, `nvidia-tesla-a100`, `nvidia-h100-80gb`, `nvidia-tesla-t4`
+
+### Key rules for inference workloads
+- GKE Autopilot supports GPU workloads via ComputeClasses and Node Auto-Provisioning — no manual node pool needed
+- Use GIQ manifests as the starting point; they embed best-practice resource requests, tolerations, and readiness probes
+- Some models (Llama, Mistral) require Hugging Face tokens — create a Kubernetes Secret and reference it in the manifest
+- Monitor inference latency via `--target-ntpot-milliseconds` (Normalized Time Per Output Token) — this controls the accelerator selection trade-off
+- For multi-tenant inference, use separate namespaces with ResourceQuotas per team
+
+## Response minimum
+
+Return, at minimum:
+
+- the scoped target and evidence level,
+- the main risks or control gaps,
+- the safest next actions,
+- validation or rollback notes where relevant,
+- the assumptions or blockers that prevent stronger conclusions.

package/skills/gcp/gcp-gke-platform-operator/assets/default-deny-netpol.yaml

@@ -0,0 +1,13 @@
+# Default deny all ingress and egress network policy for a namespace.
+# Apply to production namespaces before adding specific allow policies.
+# kubectl apply -f default-deny-netpol.yaml -n <NAMESPACE>
+apiVersion: networking.k8s.io/v1
+kind: NetworkPolicy
+metadata:
+  name: default-deny-all
+  namespace: default  # Replace with target namespace
+spec:
+  podSelector: {}   # Selects all pods in namespace
+  policyTypes:
+    - Ingress
+    - Egress

package/skills/gcp/gcp-gke-platform-operator/assets/golden-path-autopilot.yaml

@@ -0,0 +1,97 @@
+# GKE Autopilot Golden Path Configuration
+# Cluster-level policy settings for a production-ready Autopilot cluster.
+# Apply with: gcloud container clusters create-auto <NAME> --region=<REGION> [flags below]
+# Day-0 decisions are marked; confirm before cluster creation.
+
+# --- Cluster Mode (Day-0) ---
+# autopilot: enabled: true  # implicit for create-auto
+
+# --- Private Cluster (Day-0) ---
+# --enable-private-nodes
+# --master-ipv4-cidr=172.16.0.0/28  # Adjust to avoid CIDR conflicts
+
+# --- Networking (Day-0) ---
+# --network=<VPC_NAME>
+# --subnetwork=<SUBNET_NAME>
+# --enable-ip-alias
+# --cluster-secondary-range-name=<PODS_RANGE>
+# --services-secondary-range-name=<SERVICES_RANGE>
+
+# --- Security defaults (Day-0 or Day-1 as noted) ---
+addonsConfig:
+  statefulHaConfig:
+    enabled: true           # Day-1: stateful workload HA
+  gcsFuseCsiDriverConfig:
+    enabled: true           # Day-1: GCS FUSE for ML workloads
+  gcpFilestoreCsiDriverConfig:
+    enabled: true           # Day-1: Filestore CSI
+
+autopilot:
+  enabled: true
+
+autoscaling:
+  autoscalingProfile: OPTIMIZE_UTILIZATION  # Day-1: preferred for batch/ML
+
+binaryAuthorization:
+  evaluationMode: DISABLED  # Day-1: enable after attestor setup
+
+controlPlaneEndpointsConfig:
+  dnsEndpointConfig:
+    allowExternalTraffic: true   # Customer-configurable (Day-0)
+  ipEndpointsConfig:
+    authorizedNetworksConfig:
+      privateEndpointEnforcementEnabled: true  # Day-0: lock control plane to private
+
+masterAuthorizedNetworksConfig:
+  privateEndpointEnforcementEnabled: true  # Day-0
+
+privateClusterConfig:
+  enablePrivateNodes: true  # Day-0: nodes have no public IPs
+
+networkConfig:
+  datapathProvider: ADVANCED_DATAPATH  # Day-0: Dataplane V2 (eBPF)
+  dnsConfig:
+    clusterDns: CLOUD_DNS  # Day-0: Cloud DNS for better scalability
+
+secretManagerConfig:
+  enabled: true             # Day-1: Secret Manager CSI integration
+  rotationInterval: 120s
+
+rbacBindingConfig:
+  enableInsecureBindingSystemAuthenticated: false   # Day-0: block legacy bindings
+  enableInsecureBindingSystemUnauthenticated: false # Day-0
+
+workloadIdentityConfig:
+  workloadPool: <PROJECT_ID>.svc.id.goog  # Day-0: keyless pod auth
+
+nodeConfig:
+  shieldedInstanceConfig:
+    enableSecureBoot: true          # Day-0: Shielded Nodes
+    enableIntegrityMonitoring: true
+  workloadMetadataConfig:
+    mode: GKE_METADATA              # Day-0: enforce Workload Identity on pods
+
+verticalPodAutoscaling:
+  enabled: true  # Day-1: VPA for right-sizing
+
+monitoringConfig:
+  componentConfig:
+    enableComponents:
+      - SYSTEM_COMPONENTS
+      - STORAGE
+      - POD
+      - DEPLOYMENT
+      - STATEFULSET
+      - DAEMONSET
+      - HPA
+      - CADVISOR
+      - KUBELET
+      - APISERVER
+      - SCHEDULER
+      - CONTROLLER_MANAGER
+  advancedDatapathObservabilityConfig:
+    enableMetrics: true
+    enableRelay: true
+
+releaseChannel:
+  channel: REGULAR  # Day-1: balanced freshness and stability