localaistack-cli

name	localaistack-cli
description	Use this skill when a task involves inspecting, explaining, or operating the LocalAIStack CLI at ./build/las or .\build\las.exe (command name localaistack), including model, module, provider, service, system, init, and failure subcommands.

LocalAIStack CLI

Use this skill when the user asks to operate ./build/las or .\build\las.exe, explain its flags, debug a failing LocalAIStack CLI workflow, or map a task onto localaistack subcommands.

Command identity

• The built binary is ./build/las on Linux/macOS and .\build\las.exe on Windows PowerShell.
• Its Cobra command name is localaistack.
• Global flags:
  • --config <path> overrides the config file. Default path is $HOME/.localaistack/config.yaml.
  • --verbose enables verbose output.
• localaistack init and localaistack system init are the same initializer.

Safe operating defaults

• Prefer ./build/las <command> --help on Linux/macOS, or .\build\las.exe <command> --help on Windows, before guessing flags.
• Prefer read-only commands first: model list, model search, module list, provider list, failure list.
• For model run, use --dry-run first unless the user explicitly wants the process started.
• Do not use model rm --force, module uninstall, or module purge unless the user explicitly asked for deletion.
• If a command depends on machine state and fails, inspect ./build/las failure list and ./build/las failure show <id> before retrying blindly.

Preconditions

• module config-plan and non-Ollama model run read base system info from ~/.localaistack/base_info.json.
• If that file is missing, run ./build/las system init or ./build/las init.
• system init writes config to ~/.localaistack/config.yaml and refreshes base_info.json.
• model run --smart-run also needs a valid LLM config in the user config file.

Top-level commands

init / system init

Use to bootstrap config and base system info.

Common non-interactive form:

./build/las system init \
  --api-key "$API_KEY" \
  --language en \
  --assistant-provider siliconflow \
  --assistant-model deepseek-ai/DeepSeek-V4-Flash

Important flags:

• --config-path
• --api-key
• --language
• --assistant-provider, --assistant-model, --assistant-base-url, --assistant-timeout-seconds
• --translation-provider, --translation-model, --translation-base-url, --translation-timeout-seconds

model

Use for model discovery, download, local inventory, runtime launch, repair, and smart-run cache inspection.

Decision rules:

• model search <query> searches remote sources. Default --source all.
• model download <model-id> [file] downloads a model. --file and positional [file] are mutually exclusive.
• model list lists locally downloaded models under ~/.localaistack/models.
• model repair <model-id> repairs missing config/tokenizer files for Hugging Face or ModelScope models.
• model rm <model-id> --force deletes a local model.
• model run <model-id> starts a runtime for a local model.

Source rules:

• Explicit --source accepts ollama, huggingface or hf, and modelscope.
• Without --source, the CLI tries to parse the source from the model id.

Runtime rules for model run:

• Ollama source: shells out to ollama run <model>.
• Local model directory with .safetensors: uses vllm.
• Local model directory with .gguf: uses llama.cpp.
• If both safetensors and gguf exist, safetensors/vLLM path is chosen first.

Practical patterns:

./build/las model search qwen3 --source all --limit 5
./build/las model download hf/Qwen/Qwen3-8B
./build/las model run hf/Qwen/Qwen3-8B --dry-run
./build/las model run hf/bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF Q4_K_M --dry-run
./build/las model repair hf/Qwen/Qwen3-8B
./build/las model smart-run-cache list --runtime vllm

Important model run flags:

• Common: --dry-run, --source, --file, --host, --port
• Smart planning: --smart-run, --smart-run-debug, --smart-run-refresh, --smart-run-strict
• llama.cpp: --threads, --ctx-size, --n-gpu-layers, --tensor-split, --batch-size, --ubatch-size, --auto-batch, --temperature, --top-p, --top-k, --min-p, --presence-penalty, --repeat-penalty, --chat-template-kwargs
• vLLM: --vllm-max-model-len, --vllm-gpu-memory-utilization, --vllm-trust-remote-code, --text-only

Failure handling for model run:

• Missing ollama in PATH: install the Ollama module first.
• Missing vllm in PATH: install the vllm module first.
• Missing base info: run ./build/las system init.
• For repeated smart-run issues, inspect and clear cache entries with model smart-run-cache.

module

Use for software module lifecycle and configuration planning.

Commands:

• module list
• module check <module>
• module install <module>
• module update <module>
• module uninstall <module>
• module purge <module>
• module setting <module> <args...>
• module config-plan <module>

module config-plan notes:

• Reads base info from ~/.localaistack/base_info.json.
• --model <id> adds model context.
• --dry-run prints the plan without saving.
• --apply saves the plan to ~/.localaistack/config-plans/<module>.json.
• --output json is best when another agent needs structured output.
• --planner-strict fails if planning cannot produce a valid plan.

Recommended flow:

./build/las module list
./build/las module check vllm
./build/las module config-plan vllm --output json --dry-run
./build/las module install vllm

provider

Use ./build/las provider list to list built-in LLM providers. This is read-only.

service

Use for service lifecycle:

• service start <name>
• service stop <name>
• service status <name>

Treat these as stateful operations. Check status before and after changes.

system

Current behavior:

• system init is real and writes config plus base info.
• system detect currently only prints Detecting hardware....
• system info currently only prints System information:.

Do not assume system detect or system info return structured hardware data unless the code changes.

failure

Use to inspect recorded failures after planner, install, or run problems.

Commands:

• failure list --limit 20
• failure list --phase model_run --output json
• failure show <event-id>

failure show returns JSON with both the event and suggested advice.

Agent workflow

1. Confirm whether the task is read-only, mutating, or destructive.
2. If the task touches planning or runtime, ensure config and base info exist.
3. Use --help or read the command source before assuming unsupported flags.
4. Prefer --dry-run or JSON/text inspection before launching long-running services.
5. After failures, inspect failure list / failure show and only then retry with adjusted flags.

Source files worth checking when behavior is unclear

• internal/cli/root.go
• internal/cli/commands/commands.go
• internal/cli/commands/init.go
• internal/cli/commands/failure_commands.go