菜单
Deploy trained models across the AI fleet — federated artifact fusion, quality gate checks, version promotion, multi-cloud orchestration, and rollback. Use when deploying models, promoting artifacts, checking deployment health, or managing the fleet release cycle.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Build, run, screenshot, and drive the SCBE AI browser (python/scbe/ai_browser.py + browser_{controller,grid,map,camera,autodrive}.py). Use when asked to run/launch/drive/screenshot the AI browser, browse a site headlessly, observe a page as data, or autonomously fill+submit+navigate a website.
Audit, sort, route, and back up large Windows workspaces across local disks and cloud-sync folders without recursive mirror failures. Use when the user wants the filesystem cleaned up, cloud storage organized, backup targets chosen by free space and route quality, or hardware/storage bottlenecks identified before moving data.
Manages AI safety governance using 21D hyperbolic embeddings and phase tunnel gates. Invoke when analyzing agent drift, applying SCBE-AETHERMOORE protocols, packaging governance products for sale, or running the monetization pipeline.
Construct and evolve the 9-dimensional state vector xi = [c(t), tau(t), eta(t), q(t)] that drives SCBE-AETHERMOORE governance decisions. Use when building context vectors, evolving time/entropy/quantum dimensions, or preparing state for the governance gate.
Generate visual assets for the SCBE/Aethermoore project — manhwa panels, character sheets, website hero images, social media cards, product diagrams, and marketing banners. Use when creating any visual content for the project.
Publish articles and content across all platforms — GitHub Discussions, Dev.to, X/Twitter, Buffer, YouTube, and more. Use when posting, scheduling, or managing content distribution for SCBE/Aethermoore.
| name | scbe-fleet-deploy |
| description | Deploy trained models across the AI fleet — federated artifact fusion, quality gate checks, version promotion, multi-cloud orchestration, and rollback. Use when deploying models, promoting artifacts, checking deployment health, or managing the fleet release cycle. |
Manage the deployment lifecycle from trained artifacts to running fleet agents.
Trained Artifacts (HF, GCP, AWS)
|
v
Federated Fusion (training/federated_orchestrator.py)
|
v
Quality Gates (safety >= 0.95, quality >= 0.80)
|
v
Version Promotion (staging -> production)
|
v
Fleet Rollout (rolling update across agents)
|
v
Health Monitoring (coherence, latency, safety scores)
Gather training outputs from all providers:
# Each provider produces a manifest JSON
# HF: textgen models (LoRA adapters, full weights)
# GCP: embedding models (Vertex AI)
# AWS: runtime models (SageMaker endpoints)
ls training/manifests/
# hf_manifest.json gcp_manifest.json aws_manifest.json
python training/federated_orchestrator.py \
--manifests training/manifests/*.json \
--output training/fused_release.json \
--min-quality 0.80 \
--min-safety 0.95
Every artifact must pass all gates before promotion:
| Gate | Threshold | Metric |
|---|---|---|
| Quality | >= 0.80 | Task accuracy / BLEU / F1 |
| Safety | >= 0.95 | Governance compliance rate |
| Latency | <= 200ms p95 | Inference latency |
| Cost | <= $1.00/1K tokens | Compute cost |
Failed artifacts are rejected with detailed failure reports.
staging -> canary (10% traffic) -> production (100%)
Each stage requires:
Rolling update across the flock:
# Immediate rollback to last known good
python training/federated_orchestrator.py \
--rollback \
--target-version v2.1.0
Track these metrics after deployment:
# Per-agent metrics
agent.coherence # Should stay > 0.7
agent.error_rate # Should stay < 0.05
agent.response_time # Should stay < target
# Fleet-wide metrics
fleet.consensus_rate # BFT agreement percentage
fleet.safety_score # Governance compliance
fleet.task_throughput # Tasks completed per minute
| Provider | Role | Artifact Type | Region |
|---|---|---|---|
| HuggingFace | textgen | LoRA adapters, full weights | Global CDN |
| GCP Vertex | embed | Embedding models | us-central1 |
| AWS SageMaker | runtime | Inference endpoints | us-east-1 |
| File | Purpose |
|---|---|
training/federated_orchestrator.py | Multi-cloud artifact fusion + gates |
training/train_node_fleet_three_specialty.py | 3-head specialty training |
hydra/spine.py | Fleet coordination backbone |
hydra/swarm_governance.py | BFT consensus for promotions |
agents/browser/fleet_coordinator.py | Browser fleet management |
| Stage | Tongue | Meaning |
|---|---|---|
| Build | CA (Cascade) | Breaking down, compiling |
| Test | RU (Runethic) | Binding, validating rules |
| Stage | UM (Umbroth) | Hidden, not yet revealed |
| Deploy | KO (Kor'aelin) | Asserting into production |
| Monitor | AV (Avali) | Listening, watching |
| Rollback | DR (Draumric) | Structured retreat |