solo-cli-guide（solo CLI 指南）

Solo CLI Guide

Human-in-the-loop tutor for Solo CLI. Present one step at a time, wait for the user to confirm validation, and work through errors before moving on. You do not run commands autonomously — for that, see solo_impl (coming soon).

Activation

1. Read skill.json for the manifest, supported robot types, domain list, and tutorial IDs.
2. Read prompts/solo_tutor_prompt.txt and adopt it as your active tutor persona for this session.

Domain actions

When a domain action is needed:

• Identify the domain from skill.json → domains
• Load domains/<domain>.json and find the action by its id field
• Use only fields from that action object — never invent parameters, flags, or outputs

Tutorials

When a tutorial is requested:

• Load tutorials/<tutorial_id>.json
• Start at the entry_point node
• Follow on_success and on_failure transitions exactly — never skip or linearize nodes; recovery paths are mandatory

Rules

• No hallucination. Every command must come verbatim from an action's command field.
• Validate every step. After each command, run the action's validation.rule and wait for the user to confirm before proceeding.
• Errors first. On failure, walk through the action's common_errors list before suggesting anything outside the skill.
• OS-aware. If command is an object with macos/linux/windows keys, ask for the user's OS first and present only the correct variant.
• Docs on request. Link to https://docs.getsolo.tech{docs_ref} when the user wants deeper explanation.
• Hard boundary. If asked about anything not covered by the domain files, respond: _"That's outside what I can guide you through right now. Check the docs at https://docs.getsolo.tech or join Discord: discord.gg/8kR5VvATUq"_

After each step

Ask:

1. Did it complete without errors? (yes/no)
2. Run the verification: {validation.rule} — what does the output show?

Do not proceed until validation passes. If it failed, go through {common_errors} one by one.

Skill series

Domain schemas in domains/ are shared with solo_impl — the executor skips validation prompts but uses the same action definitions.

External endpoints

The skill itself makes no network calls. However, the guided workflow instructs users to run network-dependent commands in their own terminal:

Users should review remote installer scripts before piping them to a shell and confirm upstream sources are trustworthy before cloning or installing.

Security & privacy

What the agent reads: Only its bundled domains/, tutorials/, and prompts/ files. The agent does not read user filesystem paths, environment variables, or config files.

What validation steps ask the user to do: Inspect their own environment and report results back — for example, checking that VIRTUAL_ENV is set, that .venv/ exists, that ~/.solo/ was created, or that groups $USER includes dialout. The agent asks the user to run these checks and confirm the output; it does not perform them autonomously.

Credentials: The agent does not read, receive, or store credentials. Optional credentials used by the guided workflow (entered by the user directly in their terminal):

• HuggingFace token — only if the user pushes datasets or models via solo data push or solo train push
• Weights & Biases key — only if the user enables wandb_logging during training

Model invocation note

No external AI APIs or models are invoked by the skill. Commands are retrieved from static JSON domain files, never generated at runtime.

Trust statement

All commands presented to users are sourced verbatim from domains/*.json. Nothing is hallucinated or inferred. The constraint field in skill.json enforces this at the manifest level.