Terradev CLI Command Reference (v6.2.7)

Usage: terradev agent agentic-serving [OPTIONS] COMMAND [ARGS]...

  Agentic inference serving  KV cache TTL, prefix caching, LMCache, priority
  scheduling.

Options:
  --help  Show this message and exit.

Commands:
  configure    Configure agentic inference serving settings.
  helm-values  Print Helm values for agentic inference deployment.
  k8s          Print K8s deployment manifests for agentic inference.
  launch-args  Print engine launch arguments for copy-paste.
  lmcache-env  Print LMCache environment variables.
  show-config  Show current agentic serving configuration.

Usage: terradev agent cost [OPTIONS]

  Show real-time cost breakdown for a fleet by tier.

  Example:
    terradev agent cost --fleet-id ag_abc123

Options:
  --fleet-id TEXT        Fleet ID  [required]
  --format [table|json]
  --help                 Show this message and exit.

Usage: terradev agent deploy [OPTIONS]

  Provision a heterogeneous agent fleet across all tiers simultaneously.

  Provisions reasoning (H100), decode (A100), and CPU tools tiers in parallel
  using the existing DAGExecutor wave-parallel orchestration.

  Examples:
    terradev agent deploy --agents 16 --model meta-llama/Llama-3.1-70B-Instruct
    terradev agent deploy --agents 32 --dry-run
    terradev agent deploy --topology ./agent-fleet.yaml
    terradev agent deploy --agents 8 --planner-gpu H100_SXM --worker-gpu A100_SXM_80

Options:
  -n, --agents INTEGER            Number of concurrent agent loops
  -m, --model TEXT                Model to serve
  --reasoning [instant|thinking]
  --topology PATH                 Path to agent-fleet.yaml spec file
  --planner-gpu TEXT              Reasoning tier GPU type
  --planner-count INTEGER         Reasoning tier instance count
  --worker-gpu TEXT               Decode tier GPU type
  --worker-count INTEGER          Decode tier instance count
  --cpu-cores INTEGER             vCPU count for CPU tools tier
  -p, --providers TEXT            Cloud providers to use (e.g. runpod vastai)
  --max-price FLOAT               Max price per GPU/hr in USD
  --dry-run                       Show allocation plan without provisioning
  --format [table|json]
  --help                          Show this message and exit.

Usage: terradev agent langchain [OPTIONS] COMMAND [ARGS]...

  LangChain integration with workflows, LangGraph, and SGLang.

Options:
  --help  Show this message and exit.

Commands:
  create-langgraph  Create a LangGraph workflow.
  create-pipeline   Create an SGLang pipeline.
  create-workflow   Create a LangChain workflow.
  test              Test connection to LangChain service.

Usage: terradev agent langgraph [OPTIONS] COMMAND [ARGS]...

  LangGraph workflow orchestration with monitoring.

Options:
  --help  Show this message and exit.

Commands:
  create-workflow  Create a LangGraph workflow.
  deploy           Deploy a workflow.
  status           Get workflow status.
  test             Test connection to LangGraph service.

Usage: terradev agent letta [OPTIONS] COMMAND [ARGS]...

  Letta stateful agents with long-horizon memory management.

  Build LLM agents that manage their own context, memory, and state across
  sessions using Letta's virtual memory system.

Options:
  --help  Show this message and exit.

Commands:
  chat      Send a message to a Letta agent.
  create    Create a new stateful Letta agent.
  delete    Delete a Letta agent.
  list      List Letta agents.
  remember  Teach a Letta agent a durable fact.
  status    Show the state of a Letta agent.

Usage: terradev agent list [OPTIONS]

  List all known agent fleets.

  Example:
    terradev agent list

Options:
  --format [table|json]
  --help                 Show this message and exit.

Usage: terradev agent mcp [OPTIONS] COMMAND [ARGS]...

  Universal, high-throughput MCP protocol bridge.

Options:
  --help  Show this message and exit.

Commands:
  call      Call a tool on a single MCP server.
  registry  Manage the MCP server registry.
  serve     Start the MCP bridge and expose stdio or HTTP endpoints.

Usage: terradev agent mesh [OPTIONS] COMMAND [ARGS]...

  Decentralized agent-to-agent communication and state sync.

Options:
  --help  Show this message and exit.

Commands:
  card   Manage A2A Agent Cards.
  node   Manage a mesh node.
  peers  List known peers in the mesh.
  route  Show the selected route for a set of skills.
  task   Manage mesh tasks.

Usage: terradev agent plan [OPTIONS]

  Plan a heterogeneous agent fleet without provisioning.

  Shows the recommended tier configuration, hardware selection rationale, KV
  cache budget, and cost estimate based on arXiv:2605.26297 research.

  Examples:
    terradev agent plan --agents 16 --model meta-llama/Llama-3.1-70B-Instruct
    terradev agent plan --agents 32 --model meta-llama/Llama-3.1-8B-Instruct --format json
    terradev agent plan --agents 8 --planner-gpu H100_SXM --worker-gpu A100_SXM_80

Options:
  -n, --agents INTEGER            Number of concurrent agent loops to provision
                                  for  [required]
  -m, --model TEXT                Model to serve across the fleet
  --reasoning [instant|thinking]  Reasoning mode: instant (faster) or thinking
                                  (extended CoT, 45-67% more output tokens)
  --planner-gpu TEXT              Override reasoning tier GPU type (e.g.
                                  H100_SXM)
  --planner-count INTEGER         Override reasoning tier instance count
  --worker-gpu TEXT               Override decode tier GPU type (e.g.
                                  A100_SXM_80)
  --worker-count INTEGER          Override decode tier instance count
  --cpu-cores INTEGER             vCPU count for CPU tools tier instances
  --format [table|json]           Output format
  --help                          Show this message and exit.

Usage: terradev agent sandbox [OPTIONS] COMMAND [ARGS]...

  Ephemeral, hardware-isolated execution for untrusted agent payloads.

Options:
  --help  Show this message and exit.

Commands:
  run       Run an untrusted payload inside a sandbox.
  runtimes  List registered sandbox runtimes and availability.

Usage: terradev agent scale [OPTIONS]

  Scale a single fleet tier up or down without affecting other tiers.

  KV cache state on existing instances is PRESERVED during scale operations. New
  instances are added to the pool; the router distributes new requests to them.

  Examples:
    terradev agent scale --fleet-id ag_abc123 --tier decode --count 8
    terradev agent scale --fleet-id ag_abc123 --tier reasoning --count 3
    terradev agent scale --fleet-id ag_abc123 --tier cpu_tools --count 4

Options:
  --fleet-id TEXT                 Fleet ID  [required]
  --tier [reasoning|decode|cpu_tools]
                                  Tier to scale  [required]
  --count INTEGER                 New instance count for this tier  [required]
  -p, --providers TEXT            Providers to use for scale-out instances
  --help                          Show this message and exit.

Usage: terradev agent skill [OPTIONS] COMMAND [ARGS]...

  Manage skill.md files and attach them to Letta agents.

Options:
  --help  Show this message and exit.

Commands:
  attach  Attach a skill.md to a Letta agent as a durable memory block.
  init    Create a skill.md template for an agent.

Usage: terradev agent status [OPTIONS]

  Show live status of a fleet — tier health, KV hit rate, queue depth, cost.

  Key metrics explained:   kv_hit_rate  — target >0.85. Below 0.80 = cache
  thrashing (expensive recompute).   ttft_p95_ms  — reasoning tier target
  <2000ms. Above = scale out reasoning.   queue_depth  — decode tier pending
  requests. Above 6 = scale out decode.

  (Metrics from arXiv:2605.26297 empirical benchmarking)

Options:
  --fleet-id TEXT        Fleet ID returned by 'terradev agent deploy'
                         [required]
  --format [table|json]
  --help                 Show this message and exit.

Usage: terradev agent teardown [OPTIONS]

  Terminate all fleet instances and remove fleet state.

  Example:
    terradev agent teardown --fleet-id ag_abc123
    terradev agent teardown --fleet-id ag_abc123 --yes

Options:
  --fleet-id TEXT  Fleet ID to destroy  [required]
  --yes            Skip confirmation prompt
  --help           Show this message and exit.

Usage: terradev agent vector-db [OPTIONS] COMMAND [ARGS]...

  Provision vector databases for agent memory and retrieval.

  Examples:   terradev agent vector-db up --name agent-memory --adapter qdrant
  terradev agent vector-db up --name docs-weaviate --adapter weaviate   terradev
  agent vector-db down --name agent-memory

Options:
  --help  Show this message and exit.

Commands:
  down  Teardown a vector database provisioned for an agent fleet.
  up    Provision a vector database for an agent fleet.

Usage: terradev canary report [OPTIONS]

  Show a summary of recent canary test results.

Options:
  -o, --output [text|json]  Output format
  -f, --file PATH           Path to canary results JSONL file
  -p, --provider TEXT       Filter results by provider
  -g, --gpu TEXT            Filter results by GPU type
  --help                    Show this message and exit.

Usage: terradev canary tail [OPTIONS]

  Show the most recent canary records.

Options:
  -f, --file PATH      Path to canary results JSONL file
  -n, --limit INTEGER  Number of recent records to show
  --help               Show this message and exit.

Usage: terradev canary drift [OPTIONS]

  Run a provider API drift check against live endpoints.

Options:
  --all                          Check all provider contracts.
  -p, --provider TEXT            Check a specific provider contract.
  -d, --contracts-dir DIRECTORY  Directory containing provider contract YAML
                                 files.
  --format [human|json|jsonl]    Output format for the drift report.
  --help                         Show this message and exit.

Usage: terradev database crud [OPTIONS]

  Run a CRUD operation on a database component.

Options:
  -m, --manifest PATH             Path to universal manifest
  -a, --adapter TEXT
  -n, --name TEXT
  -c, --config TEXT
  --operation [insert|select|update|delete]
                                  [required]
  --table TEXT                    [required]
  --data TEXT                     JSON data payload
  --filters TEXT                  JSON filter payload
  --help                          Show this message and exit.

Usage: terradev database down [OPTIONS]

  Teardown a database stack.

Options:
  -m, --manifest PATH  [required]
  --help               Show this message and exit.

Usage: terradev database qdrant [OPTIONS] COMMAND [ARGS]...

  Qdrant-specific vector operations.

Options:
  --help  Show this message and exit.

Commands:
  create-collection  Create a Qdrant collection.
  delete-collection  Delete a Qdrant collection.
  scroll             Scroll points in a Qdrant collection.
  search             Vector similarity search in a Qdrant collection.
  upsert             Upsert points into a Qdrant collection.

Usage: terradev database search [OPTIONS]

  Run vector similarity search on a vector store component.

Options:
  -m, --manifest PATH  Path to universal manifest
  -a, --adapter TEXT
  -n, --name TEXT
  -c, --config TEXT
  --table TEXT         [required]
  --vector TEXT        JSON array of floats  [required]
  --top-k INTEGER
  --filters TEXT
  --help               Show this message and exit.

Usage: terradev database sql [OPTIONS]

  Execute raw SQL against a database adapter that supports SQL.

Options:
  -m, --manifest PATH  Path to universal manifest
  -a, --adapter TEXT
  -n, --name TEXT
  -c, --config TEXT
  --query TEXT         Raw SQL query  [required]
  --table TEXT
  --params TEXT        JSON query parameters
  --help               Show this message and exit.

Usage: terradev database up [OPTIONS]

  Initialize a database or vector store component.

Options:
  -m, --manifest PATH  Path to universal manifest
  -a, --adapter TEXT   Database adapter name
  -n, --name TEXT      Component name
  -c, --config TEXT    JSON adapter config
  --help               Show this message and exit.

Usage: terradev database weaviate [OPTIONS] COMMAND [ARGS]...

  Weaviate vector database operations.

Options:
  --help  Show this message and exit.

Commands:
  create-collection  Create a Weaviate collection.
  delete-collection  Delete a Weaviate collection.
  hybrid-search      Hybrid vector + BM25 search in a Weaviate collection.
  insert             Insert objects into a Weaviate collection.
  list-collections   List Weaviate collections.
  query              Vector similarity search in a Weaviate collection.
  up                 Initialize a Weaviate connection.

Usage: terradev environments approve [OPTIONS] PROMOTION_ID

  Approve and execute promotion

Options:
  --user TEXT  User approving promotion
  --help       Show this message and exit.

Usage: terradev environments history [OPTIONS]

  Show promotion history

Options:
  --artifact TEXT  Filter by artifact name
  --help           Show this message and exit.

Usage: terradev environments list [OPTIONS]

  List artifacts by environment

Options:
  --env TEXT  Filter by environment
  --help      Show this message and exit.

Usage: terradev environments promote [OPTIONS] ARTIFACT_NAME

  Request environment promotion

Options:
  --from [dev|staging|prod]  [required]
  --to [dev|staging|prod]    [required]
  --user TEXT                User requesting promotion
  --help                     Show this message and exit.

Usage: terradev eval compare [OPTIONS] MODEL_A MODEL_B

  Compare two models side-by-side

Options:
  --dataset TEXT  Dataset for comparison  [required]
  --metrics TEXT  Metrics to compare
  --output TEXT   Output file for comparison results
  --help          Show this message and exit.

Usage: terradev eval evaluation [OPTIONS]

  Run model or endpoint evaluation

Options:
  --model TEXT           Model checkpoint path
  --endpoint TEXT        API endpoint URL
  --dataset TEXT         Dataset path for evaluation
  --metrics TEXT         Metrics to evaluate
  --baseline TEXT        Baseline result file for comparison
  --workload-type TEXT   Workload type classification
  --duration INTEGER     Duration for endpoint evaluation (seconds)
  --output TEXT          Output file for results
  --format [json|table]
  --help                 Show this message and exit.

Usage: terradev gateway baseten [OPTIONS] COMMAND [ARGS]...

  Baseten inference provider commands

Options:
  --help  Show this message and exit.

Commands:
  chat       Send a chat/prompt to this provider.
  configure  Configure provider credentials.
  delete     Delete/terminate a deployed endpoint.
  deploy     Deploy an inference endpoint with this provider.
  list       List deployed endpoints for this provider.
  models     List available models or deployed endpoints for this provider.
  status     Get status of a deployed endpoint.

Usage: terradev gateway huggingface [OPTIONS] COMMAND [ARGS]...

  Huggingface inference provider commands

Options:
  --help  Show this message and exit.

Commands:
  chat       Send a chat/prompt to this provider.
  configure  Configure provider credentials.
  delete     Delete/terminate a deployed endpoint.
  deploy     Deploy an inference endpoint with this provider.
  list       List deployed endpoints for this provider.
  models     List available models or deployed endpoints for this provider.
  status     Get status of a deployed endpoint.

Usage: terradev gateway inferx [OPTIONS] COMMAND [ARGS]...

  Inferx inference provider commands

Options:
  --help  Show this message and exit.

Commands:
  chat       Send a chat/prompt to this provider.
  configure  Configure provider credentials.
  delete     Delete/terminate a deployed endpoint.
  deploy     Deploy an inference endpoint with this provider.
  list       List deployed endpoints for this provider.
  models     List available models or deployed endpoints for this provider.
  status     Get status of a deployed endpoint.

Usage: terradev gateway serve [OPTIONS]

  Start the API gateway server explicitly.

Options:
  --help  Show this message and exit.

Usage: terradev gateway siliconflow [OPTIONS] COMMAND [ARGS]...

  Siliconflow inference provider commands

Options:
  --help  Show this message and exit.

Commands:
  chat       Send a chat/prompt to this provider.
  configure  Configure provider credentials.
  delete     Delete/terminate a deployed endpoint.
  deploy     Deploy an inference endpoint with this provider.
  list       List deployed endpoints for this provider.
  models     List available models or deployed endpoints for this provider.
  status     Get status of a deployed endpoint.

Usage: terradev gateway status [OPTIONS]

  Show the running gateway server status.

Options:
  -h, --host TEXT     Gateway host
  -p, --port INTEGER  Gateway port
  --help              Show this message and exit.

Usage: terradev huggingface compare [OPTIONS] MODEL_ID

  Compare all hardware options for a model

Options:
  --help  Show this message and exit.

Usage: terradev huggingface configure [OPTIONS]

  Configure HuggingFace token

Options:
  --help  Show this message and exit.

Usage: terradev huggingface list [OPTIONS]

  List deployed spaces

Options:
  --help  Show this message and exit.

Usage: terradev huggingface optimize [OPTIONS] MODEL_ID

  Show hardware optimization recommendations for a model

Options:
  --budget FLOAT  Budget constraint in $/hour
  --help          Show this message and exit.

Usage: terradev huggingface preview [OPTIONS] MODEL_ID

  Preview the generated template without deploying

Options:
  --template TEXT    Template type
  --space-name TEXT  Space name (auto-generated if not provided)
  --help             Show this message and exit.

Usage: terradev huggingface space [OPTIONS] SPACE_NAME

  Deploy a model to HuggingFace Spaces with smart optimization

Options:
  --model-id TEXT  HuggingFace model ID  [required]
  --template TEXT  Template type (auto, chat, embedding, image)
  --hardware TEXT  Override hardware tier
  --sdk TEXT       SDK (gradio, streamlit, docker)
  --private        Create private space
  --env TEXT       Environment variables (key=value)
  --budget FLOAT   Budget constraint in $/hour
  --dry-run        Show what would be deployed without creating
  --help           Show this message and exit.

Usage: terradev huggingface spaces [OPTIONS] COMMAND
                                                     [ARGS]...

  HuggingFace Spaces management — create, list, manage, delete.

Options:
  --help  Show this message and exit.

Commands:
  create    Create a new Space with auto-generated app.
  delete    Delete a Space.
  hardware  Show or change hardware tier.
  info      Get Space details.
  list      List HuggingFace Spaces.
  logs      Show Space build/run logs.
  pause     Pause a running Space (stops billing).
  restart   Restart a Space (factory reboot).
  resume    Resume a paused Space.

Usage: terradev infer deploy [OPTIONS]

  Compare and select the cheapest inference option across providers.

  Queries all configured inference providers (GPU-based and inference-only) and
  prints the best quote. To actually deploy, use `terradev infer endpoint`.

  Supported providers: runpod, vastai, lambda_labs, baseten, huggingface,
  siliconflow, inferx.

Options:
  -m, --model TEXT                Model name or path  [required]
  -t, --type [llm|embedding|vision]
                                  Model type
  -p, --provider [runpod|vastai|lambda_labs|baseten|huggingface|siliconflow|inferx]
                                  Provider (runpod|vastai|lambda_labs|baseten|hu
                                  ggingface|siliconflow|inferx)
  -g, --gpu-type TEXT             GPU type preference
  -r, --region TEXT               Region preference
  --max-latency FLOAT             Max latency in ms
  --max-cost FLOAT                Max cost per request
  --help                          Show this message and exit.

Usage: terradev infer endpoint [OPTIONS] MODEL_PATH

  Deploy an inference endpoint for MODEL_PATH.

  MODEL_PATH is passed to inference-only providers (huggingface, baseten,
  siliconflow, inferx) as their deployment model. For GPU-VM providers (runpod,
  vastai, lambda_labs) it is used as the endpoint label.

  Use --provider to pin a specific provider or omit it to pick the cheapest
  quote. Use --dry-run to preview the selected provider before provisioning.

  Examples:   terradev infer endpoint meta-llama/Llama-3.1-8B-Instruct -n my-ep
  terradev infer endpoint my-org/my-model -n my-ep --provider siliconflow -g
  A100

Options:
  -n, --name TEXT                 Endpoint name (required)  [required]
  -p, --provider [runpod|vastai|lambda_labs|baseten|huggingface|siliconflow|inferx]
                                  Provider (runpod|vastai|lambda_labs|baseten|hu
                                  ggingface|siliconflow|inferx)
  -g, --gpu-type TEXT             GPU type (A100|H100|RTX4090)
  --min-workers INTEGER           Minimum workers
  --max-workers INTEGER           Maximum workers
  --idle-timeout INTEGER          Idle timeout in seconds
  --cost-optimize                 Enable cost optimization
  --dry-run                       Show deployment plan without deploying
  --help                          Show this message and exit.

Usage: terradev infer failover [OPTIONS]

  Run health checks and auto-failover for inference endpoints.

  Probes all registered inference endpoints. If a primary endpoint is unhealthy
  and has a backup configured, traffic automatically shifts to the backup
  provider.

  Open source feature - available to all users.

Options:
  --dry-run  Show what would happen without executing failover
  --help     Show this message and exit.

Usage: terradev infer route [OPTIONS]

  Find the best inference endpoint using latency-aware routing.

  Selects the optimal healthy endpoint based on strategy:   - latency: lowest
  average response time (default)   - cost: cheapest price per hour   - score:
  weighted combination of latency + cost

  Use --measure to run fresh ping/TTFB probes before selecting.

  Integrates with WebPageTest TTFB probes for real-world latency data.
  Set WPT_API_KEY env var to enable WebPageTest integration.

Options:
  -m, --model TEXT                Filter by model name
  -s, --strategy [latency|cost|score]
                                  Routing strategy (default: latency)
  --measure                       Run fresh latency measurements before routing
  --help                          Show this message and exit.

Usage: terradev infer status [OPTIONS]

  Show inference endpoint health, latency, and failover status.

  Displays all registered inference endpoints with their health state, average
  latency, provider, and failover configuration.

  Use --check to run live health probes before displaying.

Options:
  --check  Run live health probes before showing status
  --help   Show this message and exit.

Usage: terradev inferx deploy [OPTIONS]

  Deploy model to InferX serverless platform

Options:
  --model TEXT                    Model ID or HuggingFace model name  [required]
  --image TEXT                    Docker image for model
  --gpu-type TEXT                 GPU type
  --gpu-memory INTEGER            GPU memory in GB
  --max-concurrency INTEGER       Maximum concurrent requests
  --framework TEXT                Model framework
  --openai-compatible / --no-openai-compatible
                                  OpenAI-compatible API
  --timeout INTEGER               Request timeout in seconds
  --help                          Show this message and exit.

Usage: terradev inferx inferx-configure [OPTIONS]

  Configure InferX provider credentials

Options:
  --api-key TEXT                  InferX API key  [required]
  --endpoint TEXT                 InferX API endpoint
  --region TEXT                   Region for deployment
  --snapshot / --no-snapshot      Enable snapshot technology
  --gpu-slicing / --no-gpu-slicing
                                  Enable GPU slicing
  --multi-tenant / --no-multi-tenant
                                  Enable multi-tenant isolation
  --help                          Show this message and exit.

Usage: terradev inferx inferx-delete [OPTIONS]

  Delete model deployment

Options:
  --model-id TEXT  Model deployment ID  [required]
  --help           Show this message and exit.

Usage: terradev inferx inferx-optimize [OPTIONS]

  Analyze and optimize InferX costs with AI-powered recommendations

Options:
  --cluster-config TEXT           Cluster configuration file
  --usage-metrics TEXT            Usage metrics file
  --tier [economy|balanced|performance]
                                  Cost optimization tier
  --output TEXT                   Output file for cost report
  --implement                     Implement cost optimizations automatically
  --help                          Show this message and exit.

Usage: terradev inferx inferx-quote [OPTIONS]

  Get pricing quotes for InferX

Options:
  --gpu-type TEXT  GPU type to quote
  --region TEXT    Region for quote
  --help           Show this message and exit.

Usage: terradev inferx inferx-status [OPTIONS]

  Get model deployment status

Options:
  --model-id TEXT  Model deployment ID  [required]
  --help           Show this message and exit.

Usage: terradev inferx list [OPTIONS]

  List all deployed models

Options:
  --help  Show this message and exit.

Usage: terradev inferx usage [OPTIONS]

  Get account usage statistics

Options:
  --help  Show this message and exit.

Usage: terradev k8s create [OPTIONS] CLUSTER_NAME

  Create multi-cloud Kubernetes GPU cluster

Options:
  -g, --gpu TEXT                  GPU type (H100, A100, L40)  [required]
  -n, --count INTEGER             Number of GPU nodes  [required]
  --max-price FLOAT               Maximum price per hour
  --multi-cloud                   Use multi-cloud provisioning
  --prefer-spot                   Prefer spot instances
  --aws-region TEXT               AWS region
  --gcp-region TEXT               GCP region
  --control-plane [eks|gke|self-hosted]
                                  Control plane type
  --help                          Show this message and exit.

Usage: terradev k8s destroy [OPTIONS] CLUSTER_NAME

  Destroy Kubernetes cluster

Options:
  --help  Show this message and exit.

Usage: terradev k8s info [OPTIONS] CLUSTER_NAME

  Get detailed cluster information

Options:
  --help  Show this message and exit.

Usage: terradev k8s list [OPTIONS]

  List all Kubernetes clusters

Options:
  --help  Show this message and exit.

Usage: terradev lineage add-input [OPTIONS] EXECUTION_ID
                                  {dataset|model|config|checkpoint} ARTIFACT_ID

  Add input artifact to execution (manual override)

Options:
  --help  Show this message and exit.

Usage: terradev lineage add-output [OPTIONS] EXECUTION_ID
                                   {model|checkpoint|metrics|evaluation}
                                   ARTIFACT_ID

  Add output artifact to execution (manual override)

Options:
  --help  Show this message and exit.

Usage: terradev lineage auto [OPTIONS]

  Start automatic lineage tracking for a pipeline execution

Options:
  --pipeline TEXT           Pipeline ID  [required]
  --env [dev|staging|prod]  Execution environment
  --triggered-by TEXT       Who triggered this execution
  --help                    Show this message and exit.

Usage: terradev lineage complete [OPTIONS] EXECUTION_ID

  Complete execution and finalize lineage record

Options:
  --status TEXT  Final status (completed, failed)
  --help         Show this message and exit.

Usage: terradev lineage diff [OPTIONS] VERSION1 VERSION2

  Compare two pipeline executions

Options:
  --help  Show this message and exit.

Usage: terradev lineage export [OPTIONS]

  Export lineage data for compliance reports

Options:
  --format [json|csv]       Export format
  --model TEXT              Filter by model name
  --env [dev|staging|prod]  Filter by environment
  -o, --output TEXT         Output file (default: stdout)
  --help                    Show this message and exit.

Usage: terradev lineage graph [OPTIONS] ARTIFACT_ID

  Show lineage graph for artifact

Options:
  --direction [up|down|both]
  --help                      Show this message and exit.

Usage: terradev lineage production [OPTIONS]

  Show artifacts in production environment

Options:
  --type TEXT  Filter by artifact type
  --help       Show this message and exit.

Usage: terradev lineage register [OPTIONS]
                                 {dataset|model|checkpoint|metrics|config} NAME
                                 URI

  Register a new artifact for lineage tracking

Options:
  --env [dev|staging|prod]
  --hash TEXT               Artifact hash
  --size INTEGER            Size in bytes
  --user TEXT               User registering artifact
  --parent TEXT             Parent artifact ID
  --help                    Show this message and exit.

Usage: terradev lineage show [OPTIONS] MODEL_IDENTIFIER

  Show complete provenance of a model (auto-generated)

Options:
  --env [dev|staging|prod]  Filter by environment
  --help                    Show this message and exit.

Usage: terradev lineage trace [OPTIONS]

  Trace complete lineage from checkpoint or execution

Options:
  --checkpoint TEXT  Checkpoint ID to trace backwards from
  --execution TEXT   Execution ID to trace
  --help             Show this message and exit.

Usage: terradev local pool [OPTIONS]

  View or manage your hybrid compute pool (local + cloud instances).

  Shows all registered local/remote GPU hosts alongside active cloud instances.

  Example:

      terradev local pool

      terradev local pool --format json

      terradev local pool --remove workstation-4090

Options:
  --format [table|json]  Output format
  --remove TEXT          Remove a pool entry by name
  --help                 Show this message and exit.

Usage: terradev local register [OPTIONS]

  Register a local or remote GPU host into your compute pool.

  Example:

      terradev local register --name workstation-4090

      terradev local register --name lab-node-01 --host 10.0.0.5 --user ubuntu

Options:
  --name TEXT  Name for this pool entry  [required]
  --host TEXT  Remote host (omit for localhost)
  --user TEXT  SSH username for remote host
  --key TEXT   SSH private key path
  --help       Show this message and exit.

Usage: terradev local scan [OPTIONS]

  Scan local machine or remote host for GPUs.

  Uses Rust NVML bindings (5-10x faster than nvidia-smi) with automatic fallback
  to nvidia-smi parsing if the Rust extension is unavailable.

  Examples:

      terradev local scan

      terradev local scan --detailed

      terradev local scan --host 192.168.1.50 --user ubuntu --key ~/.ssh/id_rsa

      terradev local scan --register --name workstation-4090

Options:
  --host TEXT  Remote host IP/hostname to scan via SSH
  --user TEXT  SSH username for remote scan
  --key TEXT   Path to SSH private key for remote scan
  --detailed   Show full topology, PCIe, NUMA, clock details
  --register   Auto-register discovered GPUs into pool
  --name TEXT  Name for registered pool entry (auto-generated if omitted)
  --help       Show this message and exit.

Usage: terradev lora activate [OPTIONS]

  Activate a specific version across all replicas.

  Examples:     terradev lora activate -n customer-a -v abc123...

Options:
  -n, --name TEXT     Adapter name  [required]
  -v, --version TEXT  Version ID to activate  [required]
  --help              Show this message and exit.

Usage: terradev lora add [OPTIONS]

  Hot-load a LoRA adapter onto a running vLLM server.

  Examples:     terradev lora add -e http://10.0.0.1:8000 -n customer-a --path
  /adapters/customer-a     terradev lora add -e http://10.0.0.1:8000 -n
  customer-a --path /adapters/customer-a --register --base-model meta-
  llama/Llama-2-7b-hf

Options:
  -e, --endpoint TEXT  vLLM endpoint  [required]
  -n, --name TEXT      Adapter name (becomes the model name in API requests)
                       [required]
  --path TEXT          Path to adapter weights  [required]
  --api-key TEXT       vLLM API key
  --register           Also register in central registry
  --base-model TEXT    Base model (required with --register)
  --rank INTEGER       LoRA rank (default: 64)
  --help               Show this message and exit.

Usage: terradev lora cost-report [OPTIONS]

  Generate cost attribution report for LoRA adapters.

  Examples:     terradev lora cost-report -d 7     terradev lora cost-report -a
  customer-a     terradev lora cost-report -t tenant-123

Options:
  -d, --days INTEGER  Number of days to report (default: 30)
  -a, --adapter TEXT  Specific adapter to report on
  -t, --tenant TEXT   Specific tenant to report on
  --help              Show this message and exit.

Usage: terradev lora drift-check [OPTIONS]

  Check for performance drift in an adapter.

  Examples:     terradev lora drift-check -n customer-a     terradev lora drift-
  check -n customer-a -t 0.15

Options:
  -n, --name TEXT        Adapter name to check  [required]
  -v, --version TEXT     Specific version to check (default: active)
  -t, --threshold FLOAT  Drift threshold (default: 0.1)
  --source TEXT          Data source for drift detection
  --help                 Show this message and exit.

Usage: terradev lora list [OPTIONS]

  List loaded LoRA adapters.

  Examples:     terradev lora list -e http://10.0.0.1:8000     terradev lora
  list -e http://10.0.0.1:8000 --registry

Options:
  -e, --endpoint TEXT  vLLM endpoint (e.g. http://10.0.0.1:8000)  [required]
  --api-key TEXT       vLLM API key
  --registry           Show registry state instead of live endpoint
  --help               Show this message and exit.

Usage: terradev lora lorax [OPTIONS] COMMAND [ARGS]...

  LoRAX (LoRA eXchange) multi-LoRA inference server from Predibase.

  Deploy and manage LoRAX servers for serving thousands of fine-tuned models on
  a single GPU with dynamic adapter loading.

Options:
  --help  Show this message and exit.

Commands:
  deploy          Deploy a LoRAX server.
  generate        Generate text using LoRAX server.
  list-adapters   List loaded adapters on LoRAX server.
  load-adapter    Load a LoRA adapter onto LoRAX server.
  sync-registry   Sync Terradev LoRA registry with LoRAX server state.
  test            Test LoRAX server connectivity.
  unload-adapter  Unload a LoRA adapter from LoRAX server.

Usage: terradev lora peft [OPTIONS] COMMAND [ARGS]...

  HuggingFace PEFT adapter import and management.

  Download, validate, and prepare LoRA adapters from HuggingFace for use with
  vLLM, LoRAX, or other inference servers.

Options:
  --help  Show this message and exit.

Commands:
  delete    Delete a locally imported adapter.
  import    Import a LoRA adapter from HuggingFace.
  list      List all locally imported PEFT adapters.
  validate  Validate a PEFT adapter structure.

Usage: terradev lora register [OPTIONS]

  Register a LoRA adapter in the central registry with version tracking.

  Examples:     terradev lora register -n customer-a --path /adapters/customer-a
  -b meta-llama/Llama-2-7b-hf     terradev lora register -n customer-b --path
  /adapters/customer-b -b meta-llama/Llama-2-7b-hf --tenant t-123

Options:
  -n, --name TEXT        Adapter name  [required]
  --path TEXT            Path to adapter weights  [required]
  -b, --base-model TEXT  Base model name (e.g., meta-llama/Llama-2-7b-hf)
                         [required]
  --rank INTEGER         LoRA rank (default: 64)
  --tenant TEXT          Associate with tenant ID
  --metadata TEXT        JSON metadata string
  --help                 Show this message and exit.

Usage: terradev lora remove [OPTIONS]

  Hot-unload a LoRA adapter.

  Examples:     terradev lora remove -e http://10.0.0.1:8000 -n customer-a

Options:
  -e, --endpoint TEXT  vLLM endpoint  [required]
  -n, --name TEXT      Adapter name to unload  [required]
  --api-key TEXT       vLLM API key
  --help               Show this message and exit.

Usage: terradev lora rollback [OPTIONS]

  Rollback adapter to previous stable version.

  Examples:     terradev lora rollback -n customer-a     terradev lora rollback
  -n customer-a -v abc123... --replicas 10.0.0.1:8000,10.0.0.2:8000

Options:
  -n, --name TEXT        Adapter name to rollback  [required]
  -v, --to-version TEXT  Target version ID (default: previous stable)
  --replicas TEXT        Comma-separated list of replica endpoints (host:port)
  --help                 Show this message and exit.

Usage: terradev lora sync [OPTIONS]

  Synchronize adapter state across all replicas in a deployment.

  Examples:     terradev lora sync -d prod -n customer-a --replicas
  10.0.0.1:8000,10.0.0.2:8000

Options:
  -d, --deployment TEXT  Deployment name  [required]
  -n, --name TEXT        Adapter name  [required]
  --replicas TEXT        Comma-separated list of replica endpoints (host:port)
  --help                 Show this message and exit.

Usage: terradev lora versions [OPTIONS]

  List all versions of an adapter.

  Examples:     terradev lora versions -n customer-a

Options:
  -n, --name TEXT  Adapter name  [required]
  --help           Show this message and exit.

Usage: terradev migrate list-workloads [OPTIONS]

  List available workloads for migration

Options:
  --provider TEXT        Filter by provider
  --format [table|json]
  --help                 Show this message and exit.

Usage: terradev migrate migration [OPTIONS]

  Migrate workload between providers with detailed cost analysis

Options:
  --from TEXT         Source provider  [required]
  --to TEXT           Target provider  [required]
  --instance-id TEXT  Source instance ID
  --workload TEXT     Workload ID from JobStateManager
  --dry-run           Show migration plan without executing
  --help              Show this message and exit.

Usage: terradev ml deepeval [OPTIONS] COMMAND [ARGS]...

  LLM evaluation with DeepEval.

Options:
  --help  Show this message and exit.

Commands:
  evaluate  Evaluate a single LLM output with a DeepEval metric.
  init      Generate a starter DeepEval test file.
  install   Install the DeepEval package.
  metrics   List available DeepEval metrics.
  run       Run DeepEval tests.

Usage: terradev ml dvc [OPTIONS] COMMAND [ARGS]...

  DVC (Data Version Control) management.

Options:
  --help  Show this message and exit.

Commands:
  add-data    Add data to tracking.
  add-remote  Add remote storage (name:url).
  init        Initialize DVC repository.
  pull        Pull data from remote.
  push        Push data to remote.
  status      Show repository status.
  test        Test connection to DVC service.

Usage: terradev ml guardrails [OPTIONS] COMMAND [ARGS]...

  NeMo Guardrails  LLM output safety, jailbreak detection, PII masking.

Options:
  --help  Show this message and exit.

Commands:
  chat             Send a message through guardrails and show the result.
  generate-config  Generate default Colang 2.x guardrails configuration.
  k8s              Print K8s deployment manifest for guardrails server.
  test             Test connection to guardrails server.

Usage: terradev ml kserve [OPTIONS] COMMAND [ARGS]...

  KServe model deployment and management.

Options:
  --help  Show this message and exit.

Commands:
  test  Test connection to KServe service.

Usage: terradev ml langfuse [OPTIONS] COMMAND [ARGS]...

  Langfuse LLM observability  traces, scores, datasets, prompts.

Options:
  --help  Show this message and exit.

Commands:
  configure             Configure Langfuse credentials.
  datasets              List datasets.
  export-training-data  Export traces as instruction/response pairs for...
  k8s                   Print K8s deployment manifest for Langfuse.
  otel-env              Print OTEL env vars for instrumenting LLM apps.
  quality               Get quality metrics for drift detection.
  score                 Create a score for a trace.
  scores                List scores.
  test                  Test Langfuse connectivity.
  trace                 Get a single trace with observations.
  traces                List recent traces.

Usage: terradev ml mlflow-legacy [OPTIONS] COMMAND [ARGS]...

  MLflow experiment tracking and model registry.

Options:
  --help  Show this message and exit.

Commands:
  create-experiment  Create a new MLflow experiment.
  export             Export experiment data.
  list-experiments   List all MLflow experiments.
  list-runs          List runs in experiment.
  test               Test connection to MLflow service.

Usage: terradev ml ollama [OPTIONS] COMMAND [ARGS]...

  Local Ollama model management and inference.

Options:
  --help  Show this message and exit.

Commands:
  chat      Chat with an Ollama model.
  generate  Generate text with an Ollama model.
  info      Show detailed information about an Ollama model.
  list      List models available on the Ollama server.
  ps        List currently running Ollama models.
  pull      Pull an Ollama model onto the local server.

Usage: terradev ml phoenix [OPTIONS] COMMAND [ARGS]...

  Arize Phoenix LLM trace observability  traces, spans, OTEL.

Options:
  --help  Show this message and exit.

Commands:
  k8s       Print K8s deployment manifest for Phoenix server.
  otel-env  Print OTEL env vars to inject into serving pods.
  projects  List Phoenix projects.
  snippet   Print Python instrumentation snippet.
  spans     List recent spans for a project.
  test      Test connection to Phoenix server.
  trace     View full execution tree for a trace.

Usage: terradev ml qdrant [OPTIONS] COMMAND [ARGS]...

  Qdrant vector database  collections, search, RAG infrastructure.

Options:
  --help  Show this message and exit.

Commands:
  collections        List all collections.
  count              Count points in a collection.
  create-collection  Create a vector collection (auto-configured for...
  info               Get collection info and stats.
  k8s                Print K8s StatefulSet manifest for Qdrant.
  test               Test connection to Qdrant server.

Usage: terradev ml ray [OPTIONS] COMMAND [ARGS]...

  Enhanced Ray distributed computing with monitoring and dashboards.

Options:
  --help  Show this message and exit.

Commands:
  dashboard   Get dashboard URL.
  install     Show installation instructions.
  list-nodes  List cluster nodes.
  start       Start Ray cluster.
  status      Show cluster status.
  stop        Stop Ray cluster.
  test        Test connection to Ray service.

Usage: terradev ml sglang [OPTIONS] COMMAND [ARGS]...

  SGLang optimization and management with workload-specific auto-tuning

Options:
  --help  Show this message and exit.

Commands:
  detect           Auto-detect workload type and show optimization...
  install          Install SGLang with optimization stack
  router           Generate cache-aware router command for multi-replica...
  sglang-optimize  Auto-optimize SGLang configuration for workload type and...
  start            Start optimized SGLang server
  test             Test SGLang installation and configuration

Usage: terradev ml vllm [OPTIONS] COMMAND [ARGS]...

  vLLM optimization and management commands.

Options:
  --help  Show this message and exit.

Commands:
  analyze         Analyze current vLLM server workload and provide...
  auto-optimize   Automatically optimize vLLM configuration based on...
  benchmark       Benchmark vLLM endpoint performance.
  import-adapter  Import a LoRA adapter from HuggingFace for vLLM.
  import-model    Import a base model from HuggingFace for vLLM serving.
  lora            LoRA adapter management for vLLM serving engines.
  optimize        Generate optimized vLLM configurations using the 6...

Usage: terradev ml wandb [OPTIONS] COMMAND [ARGS]...

  Weights & Biases experiment tracking with dashboards, reports, and alerts.

Options:
  --help  Show this message and exit.

Commands:
  create-dashboard  Create Terradev dashboard in W&B.
  create-project    Create a new W&B project.
  create-report     Generate infrastructure report in W&B.
  dashboard-status  Get comprehensive dashboard status.
  list-projects     List all W&B projects.
  list-runs         List recent W&B runs.
  setup-alerts      Set up Terradev alerts in W&B.
  test              Test connection to W&B service.

Usage: terradev model-router classify [OPTIONS] TEXT

  Classify a message's step type for routing.

Options:
  --help  Show this message and exit.

Usage: terradev model-router configure [OPTIONS]

  Configure model routing endpoints and strategy.

Options:
  --strong-url TEXT               Strong model endpoint
  --strong-model TEXT
  --strong-api-key TEXT
  --weak-url TEXT                 Weak model endpoint
  --weak-model TEXT
  --weak-api-key TEXT             Weak model API key (if needed)
  --strategy [step_type|threshold|cascade|strong_only|weak_only]
                                  Routing strategy
  --cost-threshold FLOAT          Complexity threshold for threshold strategy
  --help                          Show this message and exit.

Usage: terradev model-router llmd-config [OPTIONS]

  Generate llm-d KV-cache-aware routing config.

Options:
  -f, --format [json|yaml]
  --help                    Show this message and exit.

Usage: terradev model-router stats [OPTIONS]

  Show routing statistics (in-memory, current session).

Options:
  -f, --format [json|text]
  --help                    Show this message and exit.

Usage: terradev model-router test [OPTIONS]

  Test model routing with a sample prompt.

Options:
  -p, --prompt TEXT         Test prompt
  -f, --format [json|text]
  --help                    Show this message and exit.

Usage: terradev providers export-example [OPTIONS]

  Export an example provider profiles YAML file.

  Example:   terradev providers export-example   terradev providers export-
  example -o ~/.terradev/custom_providers.yaml

Options:
  -o, --output PATH  Output file path (default: stdout)
  --help             Show this message and exit.

Usage: terradev providers list-profiles [OPTIONS]

  List all registered provider profiles (built-in and custom).

  Example:   terradev providers list-profiles   terradev providers list-profiles
  --format json

Options:
  -f, --format [table|json|yaml]  Output format (default: table)
  --help                          Show this message and exit.

Usage: terradev providers load-profiles [OPTIONS]

  Load custom provider profiles from a YAML or JSON file.

  Example:   terradev providers load-profiles ~/.terradev/custom_providers.yaml
  terradev providers load-profiles profiles.json --override

  Profile file format (YAML):   profiles:     my_provider:       api_style: rest
  auth_type: bearer       egress_cost: 0.05       supports_spot: true

Options:
  -p, --path PATH  Path to YAML or JSON file containing provider profiles
  --override       Override existing profiles with same name (default: skip
                   existing)
  --help           Show this message and exit.

Usage: terradev providers remove-profile [OPTIONS] NAME

  Remove a custom provider profile from the registry.

  Example:   terradev providers remove-profile my_custom_provider   terradev
  providers remove-profile my_custom_provider --force

Options:
  -f, --force  Remove without confirmation
  --help       Show this message and exit.

Usage: terradev providers show-profile [OPTIONS] NAME

  Show details for a specific provider profile.

  Example:   terradev providers show-profile runpod   terradev providers show-
  profile my_custom_provider --format json

Options:
  -f, --format [table|json|yaml]  Output format (default: table)
  --help                          Show this message and exit.

Usage: terradev record start [OPTIONS]

  Start recording a live workflow

Options:
  -n, --name TEXT    Recording name  [required]
  --output-dir TEXT  Recording output directory
  --help             Show this message and exit.

Usage: terradev record stop [OPTIONS]

  Stop recording and optionally export as pipeline

Options:
  -n, --name TEXT    Recording name  [required]
  --export TEXT      Export as YAML pipeline file
  --output-dir TEXT  Recording directory
  --help             Show this message and exit.

Usage: terradev retrain deploy [OPTIONS]

  Manually deploy an adapter from a completed retrain cycle.

  Use this when --auto-swap was not set and eval passed.

  Examples:     terradev retrain deploy --cycle-id retrain-abc12345 -e
  http://10.0.0.1:8000

Options:
  --cycle-id TEXT           Retrain cycle ID to deploy  [required]
  -e, --vllm-endpoint TEXT  vLLM endpoint  [required]
  --vllm-api-key TEXT
  -f, --format [json|text]
  --help                    Show this message and exit.

Usage: terradev retrain detect [OPTIONS]

  Check for drift without triggering a retrain.

  Examples:     terradev retrain detect -m llama-70b-prod     terradev retrain
  detect -m llama-70b-prod --threshold 0.80

Options:
  -m, --model TEXT          Model identifier  [required]
  --phoenix-endpoint TEXT
  --phoenix-project TEXT
  --baseline FLOAT
  --threshold FLOAT
  --min-samples INTEGER
  -f, --format [json|text]
  --help                    Show this message and exit.

Usage: terradev retrain drift [OPTIONS]

  Run a drift-triggered retrain cycle.

  Monitors Phoenix traces, detects quality drift, retrains a LoRA adapter,
  evaluates it, and optionally hot-swaps it onto a running vLLM server.

  Examples:     terradev retrain drift -m llama-70b-prod --auto-swap
  terradev retrain drift -m llama-70b-prod -e http://10.0.0.1:8000     terradev
  retrain drift -m llama-70b-prod --eval-threshold 0.90

Options:
  -m, --model TEXT           Model identifier (e.g. llama-70b-prod)  [required]
  --source [phoenix-traces]  Data source for drift detection
  --method [lora]            Fine-tuning method
  --eval-threshold FLOAT     Minimum eval score to deploy (0.0-1.0)
  --deploy [canary|direct]   Deployment strategy
  --auto-swap                Auto-deploy if eval passes (no manual approval)
  --phoenix-endpoint TEXT    Phoenix collector endpoint
  --phoenix-project TEXT     Phoenix project name
  -e, --vllm-endpoint TEXT   vLLM endpoint for eval and deploy
  --vllm-api-key TEXT        vLLM API key
  --baseline FLOAT           Baseline quality score
  --threshold FLOAT          Drift trigger threshold
  --min-samples INTEGER      Min samples before triggering
  -f, --format [json|text]
  --help                     Show this message and exit.

Usage: terradev retrain history [OPTIONS]

  Show retrain cycle history.

  Examples:     terradev retrain history     terradev retrain history -n 5 -f
  json

Options:
  -n, --limit INTEGER       Number of cycles to show
  -f, --format [json|text]
  --help                    Show this message and exit.

Usage: terradev train dpo [OPTIONS] [EXTRA]...

  Run preference optimization (DPO / SimPO / KTO / ORPO).

Options:
  --base-checkpoint TEXT          SFT checkpoint to start from  [required]
  --data TEXT                     Preference pairs data path  [required]
  --model TEXT                    Optional model override (defaults to base-
                                  checkpoint)
  --algorithm [dpo|simpo|kto|orpo|cpo]
                                  Preference optimization algorithm
  --framework [unsloth|trl|axolotl|llama-factory|ms-swift|openrlhf]
                                  Training framework
  --provider TEXT                 Cloud provider, or 'auto'
  --checkpoint TEXT               Output checkpoint directory
  --gpu-type TEXT                 GPU type (A100, H100, etc.)
  --gpu-count INTEGER
  --node-count INTEGER
  --gpus-per-node INTEGER
  --spot / --no-spot
  --max-price FLOAT
  --num-train-epochs INTEGER
  --per-device-batch-size INTEGER
  --gradient-accumulation-steps INTEGER
  --learning-rate FLOAT
  --warmup-ratio FLOAT
  --beta FLOAT                    DPO beta / SimPO beta
  --max-seq-length INTEGER
  --lora-rank INTEGER
  --lora-alpha INTEGER
  --from-provision TEXT
  -n, --nodes TEXT
  --output-bucket TEXT
  --dry-run
  -f, --format [json|text]
  --help                          Show this message and exit.

Usage: terradev train grpo [OPTIONS] [EXTRA]...

  Run GRPO / RLVR stage.

Options:
  --base-checkpoint TEXT          DPO/SFT checkpoint to start from  [required]
  --data TEXT                     Prompt / rollout data path  [required]
  --model TEXT                    Optional model override
  --reward-fn TEXT                Reward function name
  --rollout-provider TEXT         Provider for rollout workers
  --trainer-provider TEXT         Provider for GRPO trainer
  --framework [unsloth|trl|axolotl|llama-factory|ms-swift|openrlhf]
                                  GRPO framework
  --provider TEXT                 Combined provider override
  --checkpoint TEXT               Output checkpoint directory
  --gpu-type TEXT                 GPU type for trainer
  --gpu-count INTEGER
  --node-count INTEGER
  --gpus-per-node INTEGER
  --num-generations INTEGER       GRPO group size
  --spot / --no-spot
  --max-price FLOAT
  --num-train-epochs INTEGER
  --per-device-batch-size INTEGER
  --gradient-accumulation-steps INTEGER
  --learning-rate FLOAT
  --warmup-ratio FLOAT
  --max-seq-length INTEGER
  --from-provision TEXT
  -n, --nodes TEXT
  --output-bucket TEXT
  --dry-run
  -f, --format [json|text]
  --help                          Show this message and exit.

Usage: terradev train pipeline [OPTIONS]

  Run a multi-stage training pipeline from a YAML file.

Options:
  -c, --config PATH         Pipeline YAML  [required]
  --dry-run                 Print the DAG plan without launching
  --teardown                Tear down provisioned nodes after each stage
  -f, --format [json|text]
  --help                    Show this message and exit.

Usage: terradev train sft [OPTIONS] [EXTRA]...

  Run supervised fine-tuning (SFT) stage.

Options:
  --model TEXT                    Base model ID or path  [required]
  --data TEXT                     Training data path (local dir or s3://)
                                  [required]
  --framework [unsloth|trl|axolotl|llama-factory|ms-swift|openrlhf]
                                  Training framework
  --provider TEXT                 Cloud provider, or 'auto' for cheapest quote
  --checkpoint TEXT               Output checkpoint directory
  --gpu-type TEXT                 GPU type (A100, H100, etc.)
  --gpu-count INTEGER             Total GPUs
  --node-count INTEGER            Number of nodes
  --gpus-per-node INTEGER         GPUs per node
  --spot / --no-spot              Use spot/preemptible instances
  --max-price FLOAT               Max $/hr per GPU
  --num-train-epochs INTEGER
  --per-device-batch-size INTEGER
  --gradient-accumulation-steps INTEGER
  --learning-rate FLOAT
  --warmup-ratio FLOAT
  --max-seq-length INTEGER
  --lora-rank INTEGER
  --lora-alpha INTEGER
  --from-provision TEXT           Use provision group, or "latest"
  -n, --nodes TEXT                Node IP addresses
  --output-bucket TEXT            s3:// bucket to sync checkpoint
  --dry-run                       Print the command but do not launch
  -f, --format [json|text]
  --help                          Show this message and exit.

Usage: terradev train resume [OPTIONS]

  Resume a training job from checkpoint.

  Rebuilds config from job state and resumes with topology validation.

  Examples:     terradev train-resume -j job-abc123     terradev train-resume -j
  job-abc123 --checkpoint-id ckpt-xyz

Options:
  -j, --job-id TEXT         Job ID to resume  [required]
  --checkpoint-id TEXT      Checkpoint to resume from (default: latest)
  -f, --format [json|text]
  --help                    Show this message and exit.

Usage: terradev train start [OPTIONS] [SCRIPT_ARGS]...

  Launch distributed training jobs across provisioned GPU nodes.

  Orchestrates distributed training with automatic topology optimization,
  FlashOptim integration, and checkpoint management. Supports torchrun,
  DeepSpeed, Accelerate, and Megatron frameworks.

  Examples:   terradev train -s train.py --framework torchrun --gpus-per-node 8
  terradev train -c job.yaml                                      # Use YAML
  config   terradev train -s train.py -n 10.0.0.1 -n 10.0.0.2 --tp 2 -- --lr
  1e-4   terradev train -s train.py --from-provision latest             # Auto-
  resolve nodes   terradev train -s train.py --from-provision
  pg_1709123456_abc12345   terradev train -s train.py --pool workstation-4090
  # Use local pool entry   terradev train -s train.py --pool workstation-4090
  --overflow-to-cloud  # Cloud fallback

  Workflow:   1. Provision nodes: terradev provision -g H100 -n 4   2. Validate:
  terradev preflight (optional, auto-run by default)   3. Train: terradev train
  -s train.py --from-provision latest   4. Monitor: terradev monitor --job <job-
  id>   5. Checkpoint: terradev checkpoint list --job <job-id>

  FlashOptim (auto-applied):   When training with bf16/fp16 and 40GB+ VRAM,
  FlashOptim is automatically   enabled for gradient compression and checkpoint
  optimization.

  Frameworks:   - torchrun: PyTorch native distributed training (default)   -
  deepspeed: Microsoft DeepSpeed for large models   - accelerate: HuggingFace
  Accelerate   - megatron: NVIDIA Megatron-LM for massive models

  Next Steps:   Monitor training: terradev monitor --job <job-id>   Check
  status: terradev train-status --job <job-id>   Stop training: terradev train-
  stop --job <job-id>   View checkpoints: terradev checkpoint list --job <job-
  id>

Options:
  -c, --config PATH               YAML config file for training job
  -s, --script TEXT               Training script path (Python file)
  --framework [torchrun|deepspeed|accelerate|megatron]
                                  Distributed framework: torchrun (default),
                                  deepspeed, accelerate, megatron
  --backend [native|ray]          Launch backend: native (default), ray
                                  (optional, requires Ray cluster)
  -n, --nodes TEXT                Node IP addresses (multiple allowed)
  --from-provision TEXT           Use nodes from provision group (pg_xxx or
                                  "latest" for most recent)
  --pool TEXT                     Use local pool entry by name (e.g.,
                                  workstation-4090)
  --overflow-to-cloud             Fall back to cloud providers if local pool
                                  unavailable or insufficient
  --gpus-per-node INTEGER         GPUs per node (default: 8)
  --tp INTEGER                    Tensor parallel size for model parallelism
  --pp INTEGER                    Pipeline parallel size for model parallelism
  --total-steps INTEGER           Total training steps (for ETA calculation)
  --skip-preflight                Skip preflight GPU/NCCL/RDMA validation checks
  -f, --format [json|text]        Output format: text (default) or json
  --help                          Show this message and exit.

Usage: terradev train status [OPTIONS]

  Show training job status, GPU-hours, cost, and ETA.

  Queries the local SQLite job database  no external services needed.

  Examples:     terradev train-status     terradev train-status -j job-abc123
  terradev train-status -f json

Options:
  -j, --job-id TEXT         Job ID (empty = all running)
  -f, --format [json|text]
  --help                    Show this message and exit.

Usage: terradev train stop [OPTIONS]

  Stop a running training job.

  Kills training processes on all nodes in parallel.

  Examples:     terradev train-stop -j job-abc123

Options:
  -j, --job-id TEXT         Job ID to stop  [required]
  -f, --format [json|text]
  --help                    Show this message and exit.

Usage: terradev train unsloth [OPTIONS] COMMAND [ARGS]...

  Unsloth optimized local LLM training, serving, and coding agents.

  Unsloth slashes VRAM usage by up to 70% and doubles training speeds using
  optimized Triton kernels. Use `run` to serve a local model and `start` to
  attach a coding agent to it.

Options:
  --help  Show this message and exit.

Commands:
  run    Run an Unsloth local model server.
  start  Start a coding agent backed by Unsloth's local model server.
  stop   Stop a running Unsloth server started with `unsloth run`.

Usage: terradev triggers create [OPTIONS] NAME PIPELINE

  Create a new trigger

Options:
  --type [event|schedule|condition]
                                  Trigger type
  --event TEXT                    Event type to trigger on (dataset_landed,
                                  model_drift_detected, etc.)
  --schedule TEXT                 Cron schedule (e.g., "0 0 * * 0" for Sunday
                                  midnight)
  --condition TEXT                Condition expression (e.g., "drift_score >
                                  0.1")
  --env [dev|staging|prod]        Target environment
  --help                          Show this message and exit.

Usage: terradev triggers disable [OPTIONS] NAME

  Disable a trigger

Options:
  --help  Show this message and exit.

Usage: terradev triggers enable [OPTIONS] NAME

  Enable a trigger

Options:
  --help  Show this message and exit.

Usage: terradev triggers fire [OPTIONS] EVENT_TYPE

  Manually fire an event for testing

Options:
  --data TEXT    JSON data for the event
  --source TEXT  Event source
  --help         Show this message and exit.

Usage: terradev triggers list [OPTIONS]

  List all triggers

Options:
  --help  Show this message and exit.

Usage: terradev vault env [OPTIONS] PROVIDER

  Print environment-style export lines for a provider.

Options:
  --raw   Print raw values as shell export statements
  --help  Show this message and exit.

Usage: terradev vault get [OPTIONS] PROVIDER KEY

  Retrieve a secret. By default the value is masked.

Options:
  --raw   Print the raw secret (disabled in non-TTY / CI by default)
  --help  Show this message and exit.

Usage: terradev vault list [OPTIONS]

  List stored provider and key names (values are never shown).

Options:
  --help  Show this message and exit.

Usage: terradev vault remove [OPTIONS] PROVIDER [KEY]

  Remove a provider or a single key from the vault.

Options:
  --help  Show this message and exit.

Usage: terradev vault run [OPTIONS] COMMAND...

  Run a shell command with vault secrets injected into the environment.

  Examples:   terradev vault run -- terradev up --job train   terradev vault run
  --provider runpod -- python train.py

Options:
  -p, --provider TEXT  Only inject secrets for this provider
  --no-exec            Build the env and print export lines without running
  --help               Show this message and exit.

Usage: terradev vault set [OPTIONS] PROVIDER KEY

  Store a secret for a provider.

  Examples:   terradev vault set runpod api_key --from-env RUNPOD_API_KEY
  terradev vault set aws secret_key --from-stdin   cat key.txt | terradev vault
  set vastai api_key --from-stdin   terradev vault set runpod api_key --value
  rpa_xxx --no-persist

Options:
  --value TEXT     Secret value (not recommended for shell history)
  --from-env TEXT  Read the secret from an environment variable
  --from-stdin     Read the secret from stdin (best for CI and scripts)
  --no-persist     Do not write to disk; keeps the secret in env/session only
  --help           Show this message and exit.

Usage: terradev vault sync [OPTIONS]

  Import TERRADEV_* environment variables for supported cloud providers.

  This is the recommended command for CI/CD pipelines using GitHub Secrets:

    env:     TERRADEV_RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY }}
    TERRADEV_AWS_SECRET_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}   run:
    terradev vault sync

  Use --dry-run in a workflow to verify mapping before a real run. Use --all to
  also import non-provider/custom TERRADEV_* variables.

Options:
  --dry-run     Show what would be imported without persisting
  --no-persist  Keep imported secrets in env only; do not write the vault file
  --all         Import every TERRADEV_* variable, not just supported cloud
                providers
  --help        Show this message and exit.

Usage: terradev vault verify [OPTIONS]

  Check which providers are fully configured and which keys are missing.

Options:
  --help  Show this message and exit.

Usage: terradev warm-pool register [OPTIONS] MODEL_ID

  Register a model with the warm pool manager

Options:
  --priority INTEGER  Model priority for warming
  --help              Show this message and exit.

Usage: terradev warm-pool start [OPTIONS]

  Start the warm pool manager for intelligent pre-warming

Options:
  --strategy [traffic_based|time_based|priority_based|cost_optimized|latency_optimized]
                                  Warm pool strategy
  --max-warm INTEGER              Maximum models to keep warm
  --min-warm INTEGER              Minimum models to keep warm
  --help                          Show this message and exit.

Usage: terradev warm-pool status [OPTIONS]

  Get warm pool manager status

Options:
  --help  Show this message and exit.