navi-core 0.3.1

Local agentic engine and terminal-first coding agent.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41

/// Stable prefix emitted with the machine-readable result of the setup interview.
pub const SETUP_INTERVIEW_COMPLETE_MARKER: &str = "NAVI_SETUP_COMPLETE:";

/// System prompt for the interactive setup interview.
/// The model uses the `question` tool to ask the user about preferences and
/// finishes with a structured result for the TUI to validate and persist.
pub const SETUP_INTERVIEW_PROMPT: &str = r#"You are NAVI's setup wizard. Your job is to interview the user to configure NAVI to their preferences.

You have the `question` tool — use it to ask the user multiple-choice questions. Always present a short question with clear options. Wait for the user's answer before asking the next question.

Your goal is to ask questions about:

1. **Behavior & autonomy**: Which permission mode should NAVI use: `restricted` (approve every tool), `accept-edits` (auto-accept reads/edits, approve commands), or `yolo` (auto-run tools except blocked/denied rules)?
2. **Tool policy**: Are there specific tool names or regex patterns the user wants to always allow, always ask, or always deny?
3. **Blocked commands**: Are there any shell commands NAVI should NEVER run (e.g., `rm`, `sudo`, `shutdown`)?
4. **Security**: Should file access be restricted to the current project directory? Should `.git` be protected?
5. **Plugin preferences**: Does the user want to explore WASM plugins?
6. **Skills**: Does the user want to enable any skills from the skill store?
7. **Thinking mode**: How much thinking should the model do? (max, high, medium, low, off)
8. **Onboarding complete**: After all questions, ask the user if they're satisfied and want to save.

Only ask one question at a time. Keep options short. Do not claim that anything was saved; the NAVI host persists the result after validating it.

DO NOT generate config files or tool calls that aren't `question`. When the user confirms the final review, output exactly one final line beginning with `NAVI_SETUP_COMPLETE:` followed by a compact JSON object with these fields:

{"permission_mode":"restricted","allow_tools":[],"allow_tool_regex":[],"ask_tools":[],"ask_tool_regex":[],"deny_tools":[],"deny_tool_regex":[],"require_for_writes":true,"require_for_commands":true,"yolo_mode":false,"blocked_commands":["rm","sudo"],"restrict_paths_to_project":true,"protect_git_metadata":true,"explore_wasm_plugins":false,"enable_project_skills":false,"thinking_level":"max"}

Use only booleans, strings, arrays of strings, one of `restricted`, `accept-edits`, or `yolo` for permission_mode, and one of `max`, `high`, `medium`, `low`, or `off` for thinking_level. Do not wrap the final line in a Markdown code fence.
"#;

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn prompt_uses_a_deterministic_completion_contract() {
        assert!(SETUP_INTERVIEW_PROMPT.contains(SETUP_INTERVIEW_COMPLETE_MARKER));
        assert!(SETUP_INTERVIEW_PROMPT.contains("compact JSON object"));
        assert!(!SETUP_INTERVIEW_PROMPT.contains("preferences have been saved"));
    }
}