Pilot Ml Training Pipeline Setup

Deploy an end-to-end ML training pipeline with 4 agents. Use this skill when: 1. User wants to set up a machine learning training pipeline 2. User is configu...

Security Scan

The skill's instructions, required binaries, and actions are coherent with setting up a 4-agent ML pipeline; it only relies on pilotctl and clawhub and does not ask for unrelated credentials or suspicious installs.

Purpose & Capability

Name/description match the runtime instructions: the SKILL.md walks through installing agent skills (via clawhub), setting hostnames (pilotctl), writing a JSON manifest, and performing handshakes — all expected for deploying a multi-agent ML pipeline.

Instruction Scope

Instructions are scoped to pipeline setup: ask role/prefix, run clawhub install for role-specific skills, call pilotctl to set hostname/handshake, and write a role manifest to ~/.pilot/setups/ml-training-pipeline.json. The only file/path the skill asks to write is a role-specific config in the user's Pilot config directory, which is appropriate for this purpose.

Install Mechanism

This is instruction-only (no install spec). It instructs the user to run clawhub to install other pilot-* skills; that is reasonable for a meta-setup skill. There are no downloads or archive extractions in this skill itself.

Credentials

The skill requests no environment variables or credentials and only requires the binaries pilotctl and clawhub. Those requirements align with the documented commands; no unrelated secrets are requested.

Persistence & Privilege

always:false and normal model invocation; the skill writes only its own manifest to ~/.pilot/setups/, and does not modify other skills or system-wide settings. No elevated persistence or cross-skill config changes are requested.

Assessment

This skill appears to do what it says: it sets up 4 agents by installing role-specific pilot-* skills and configuring hostnames/handshakes. Before running it, verify you trust the pilotctl and clawhub binaries (install sources and checksums), review the pilot-* skills that will be installed (they may request credentials or network access), and only perform handshakes with intended hostnames: handshakes create mutual trust and enable file/metric transfer between agents, so confirm network endpoints and confidentiality requirements for your datasets and models.

ML Training Pipeline Setup

Deploy 4 agents spanning data prep, training, evaluation, and serving.

Roles

Role	Hostname	Skills	Purpose
data-prep	<prefix>-data-prep	pilot-dataset, pilot-share, pilot-task-chain	Cleans and transforms datasets
trainer	<prefix>-trainer	pilot-dataset, pilot-model-share, pilot-metrics, pilot-task-chain	Trains models, tracks metrics
evaluator	<prefix>-evaluator	pilot-model-share, pilot-metrics, pilot-review, pilot-task-chain	Evaluates and gates promotion
serving	<prefix>-serving	pilot-model-share, pilot-health, pilot-webhook-bridge, pilot-load-balancer, pilot-metrics	Serves inference requests

Setup Procedure

Step 1: Ask the user which role this agent should play and what prefix to use.

Step 2: Install the skills for the chosen role:

# For data-prep:
clawhub install pilot-dataset pilot-share pilot-task-chain
# For trainer:
clawhub install pilot-dataset pilot-model-share pilot-metrics pilot-task-chain
# For evaluator:
clawhub install pilot-model-share pilot-metrics pilot-review pilot-task-chain
# For serving:
clawhub install pilot-model-share pilot-health pilot-webhook-bridge pilot-load-balancer pilot-metrics

Step 3: Set the hostname:

pilotctl --json set-hostname <prefix>-<role>

Step 4: Write the role-specific JSON manifest to ~/.pilot/setups/ml-training-pipeline.json.

Step 5: Tell the user to initiate handshakes with direct communication peers.

Manifest Templates Per Role

data-prep

{
  "setup": "ml-training-pipeline", "role": "data-prep", "role_name": "Data Preparation",
  "hostname": "<prefix>-data-prep",
  "description": "Cleans, validates, and transforms raw datasets. Shares processed data with the trainer.",
  "skills": {
    "pilot-dataset": "Exchange structured datasets with schema negotiation.",
    "pilot-share": "Send cleaned dataset files to <prefix>-trainer.",
    "pilot-task-chain": "Chain data prep steps into sequential pipeline."
  },
  "peers": [{ "role": "trainer", "hostname": "<prefix>-trainer", "description": "Receives prepared datasets" }],
  "data_flows": [{ "direction": "send", "peer": "<prefix>-trainer", "port": 1001, "topic": "dataset-ready", "description": "Cleaned datasets" }],
  "handshakes_needed": ["<prefix>-trainer"]
}

trainer

{
  "setup": "ml-training-pipeline", "role": "trainer", "role_name": "Model Trainer",
  "hostname": "<prefix>-trainer",
  "description": "Receives prepared datasets, runs training jobs, tracks metrics, and shares trained model artifacts.",
  "skills": {
    "pilot-dataset": "Receive prepared datasets from data-prep.",
    "pilot-model-share": "Send trained model checkpoints to evaluator.",
    "pilot-metrics": "Track and publish training loss, accuracy, epochs.",
    "pilot-task-chain": "Chain training steps sequentially."
  },
  "peers": [
    { "role": "data-prep", "hostname": "<prefix>-data-prep", "description": "Sends prepared datasets" },
    { "role": "evaluator", "hostname": "<prefix>-evaluator", "description": "Receives trained models" }
  ],
  "data_flows": [
    { "direction": "receive", "peer": "<prefix>-data-prep", "port": 1001, "topic": "dataset-ready", "description": "Cleaned datasets" },
    { "direction": "send", "peer": "<prefix>-evaluator", "port": 1001, "topic": "training-complete", "description": "Model checkpoints and metrics" }
  ],
  "handshakes_needed": ["<prefix>-data-prep", "<prefix>-evaluator"]
}

evaluator

{
  "setup": "ml-training-pipeline", "role": "evaluator", "role_name": "Model Evaluator",
  "hostname": "<prefix>-evaluator",
  "description": "Scores trained models against benchmarks and gates promotion to serving.",
  "skills": {
    "pilot-model-share": "Receive models from trainer, promote approved models to serving.",
    "pilot-metrics": "Compare benchmarks, detect drift.",
    "pilot-review": "Gate model promotion with approval workflow.",
    "pilot-task-chain": "Chain evaluation steps."
  },
  "peers": [
    { "role": "trainer", "hostname": "<prefix>-trainer", "description": "Sends trained models" },
    { "role": "serving", "hostname": "<prefix>-serving", "description": "Receives approved models" }
  ],
  "data_flows": [
    { "direction": "receive", "peer": "<prefix>-trainer", "port": 1001, "topic": "training-complete", "description": "Model checkpoints" },
    { "direction": "send", "peer": "<prefix>-serving", "port": 1001, "topic": "model-approved", "description": "Approved models" },
    { "direction": "receive", "peer": "<prefix>-serving", "port": 1002, "topic": "inference-metrics", "description": "Drift detection data" }
  ],
  "handshakes_needed": ["<prefix>-trainer", "<prefix>-serving"]
}

serving

{
  "setup": "ml-training-pipeline", "role": "serving", "role_name": "Model Server",
  "hostname": "<prefix>-serving",
  "description": "Loads approved models, serves inference, monitors health, and load-balances.",
  "skills": {
    "pilot-model-share": "Receive approved models from evaluator.",
    "pilot-health": "Monitor inference endpoint health and latency.",
    "pilot-webhook-bridge": "Trigger external alerts on serving failures.",
    "pilot-load-balancer": "Distribute inference requests across replicas.",
    "pilot-metrics": "Report QPS, latency, drift metrics to evaluator."
  },
  "peers": [{ "role": "evaluator", "hostname": "<prefix>-evaluator", "description": "Sends approved models, receives metrics" }],
  "data_flows": [
    { "direction": "receive", "peer": "<prefix>-evaluator", "port": 1001, "topic": "model-approved", "description": "Approved models" },
    { "direction": "send", "peer": "<prefix>-evaluator", "port": 1002, "topic": "inference-metrics", "description": "Inference metrics for drift" }
  ],
  "handshakes_needed": ["<prefix>-evaluator"]
}

Data Flows

• data-prep → trainer : cleaned datasets (port 1001)
• trainer → evaluator : model checkpoints and metrics (port 1001)
• evaluator → serving : approved models (port 1001)
• serving → evaluator : inference metrics for drift detection (port 1002)

Workflow Example

# On data-prep:
pilotctl --json send-file <prefix>-trainer ./datasets/training-v5.parquet
pilotctl --json publish <prefix>-trainer dataset-ready '{"name":"training-v5","rows":150000}'
# On trainer:
pilotctl --json send-file <prefix>-evaluator ./models/resnet-v5.pt
pilotctl --json publish <prefix>-evaluator training-complete '{"model":"resnet-v5","accuracy":0.967}'
# On evaluator:
pilotctl --json send-file <prefix>-serving ./models/resnet-v5.pt
pilotctl --json publish <prefix>-serving model-approved '{"model":"resnet-v5","benchmark":0.971}'

Dependencies

Requires pilot-protocol skill, pilotctl binary, clawhub binary, and a running daemon.