gitdataai/libs/agent/orao/reason.rs

//! Reason phase: structured reasoning and plan generation via LLM.
//!
//! Takes the current [`PerceptionSnapshot`] and the task goal, sends them to
//! the configured LLM, and produces a structured [`ReasoningOutput`] with an
//! executable plan.

use crate::client::AiClientConfig;
use crate::error::AgentError;
use rig::agent::AgentBuilder;
use rig::client::CompletionClient;
use rig::completion::Prompt;

use super::types::{OraoConfig, PerceptionSnapshot, ReasoningOutput};

/// Prompt template for the ORAO reasoning phase.
const REASON_SYSTEM_PROMPT: &str = r#"You are an expert software engineering agent using the ORAO (Observe-Reason-Act-Observe) framework.

## Your Role: REASON Phase

You are currently in the REASON phase. You will receive:
1. A **task goal** — what needs to be accomplished
2. An **observation snapshot** — the current state of the environment

Your job is to produce a structured analysis and a step-by-step action plan.

## Output Format

You MUST respond with a valid JSON object matching this schema:

```json
{
  "analysis": "<concise analysis of the current state, problems identified, constraints>",
  "plan": [
    {
      "step_id": 1,
      "description": "<what this step does>",
      "action_type": "shell_command | file_write | file_edit | git_operation | tool_invoke | user_dialog",
      "command_or_content": "<the exact command to run or content to write>",
      "expected_result": "<what success looks like>",
      "fallback_on_failure": "<what to do if this fails, or null>"
    }
  ]
}
```

## Rules

1. **Be specific**: commands must be exact and complete (include necessary flags, paths, etc.)
2. **One action per step**: each step should do one thing
3. **Validate assumptions**: don't assume dependencies are installed or files exist
4. **Prefer small, safe steps**: each change should be minimal and reversible
5. **Include verification**: build/test after changes to verify correctness
6. **Plan size**: 1-10 steps is typical. Don't plan more than 15 steps without asking the user.
7. **Fallbacks**: for risky steps, specify what to try if the step fails

## Safety

- Prefer read-only verification before making changes
- Use version control (git) for reversible operations
- Flag any step that requires network access or system-level privileges"#;

/// Run the reasoning phase: send observation + task to the LLM and get a plan.
pub async fn reason(
    config: &AiClientConfig,
    model_name: &str,
    _orao_config: &OraoConfig,
    task_goal: &str,
    snapshot: &PerceptionSnapshot,
    round: usize,
    history_rounds: &[super::types::RoundRecord],
) -> Result<ReasoningOutput, AgentError> {
    let user_prompt = build_reason_prompt(task_goal, snapshot, round, history_rounds);

    let client = config.build_rig_client();
    let model = client.completion_model(model_name);

    let agent = AgentBuilder::new(model)
        .preamble(REASON_SYSTEM_PROMPT)
        .build();

    let response = agent
        .prompt(&user_prompt)
        .extended_details()
        .await
        .map_err(|e: rig::completion::PromptError| AgentError::OpenAi(e.to_string()))?;

    let output = response.output.trim().to_string();

    // Extract JSON from the response (it may be wrapped in ```json fences)
    let json_str = extract_json(&output);

    serde_json::from_str::<ReasoningOutput>(json_str).map_err(|e| {
        AgentError::Internal(format!(
            "Failed to parse reasoning output as JSON: {}. Raw output: {}",
            e, output
        ))
    })
}

/// Build the user prompt for the reason phase.
fn build_reason_prompt(
    task_goal: &str,
    snapshot: &PerceptionSnapshot,
    round: usize,
    history_rounds: &[super::types::RoundRecord],
) -> String {
    let mut prompt = format!(
        "## Task Goal\n\n{}\n\n## Round Number\n\n{}\n\n",
        task_goal, round
    );

    // ── Observation snapshot ────────────────────────────────────────────
    prompt.push_str("## Current Observation\n\n");

    if let Some(ref git_status) = snapshot.git_status {
        prompt.push_str(&format!("### Git Status\n```\n{}\n```\n\n", git_status));
    }

    if let Some(ref project_structure) = snapshot.project_structure {
        prompt.push_str(&format!(
            "### Project Structure\n```\n{}\n```\n\n",
            project_structure
        ));
    }

    if !snapshot.files.is_empty() {
        prompt.push_str("### Relevant Files\n\n");
        for file in &snapshot.files {
            prompt.push_str(&format!("- `{}` ({} bytes)\n", file.path, file.size_bytes));
            if let Some(ref preview) = file.content_preview {
                let truncated = if preview.len() > 2000 {
                    format!("{}... [truncated]", &preview[..2000])
                } else {
                    preview.clone()
                };
                prompt.push_str(&format!("```\n{}\n```\n\n", truncated));
            }
        }
    }

    if !snapshot.errors.is_empty() {
        prompt.push_str("### Errors / Warnings\n\n");
        for err in &snapshot.errors {
            prompt.push_str(&format!("- {}\n", err));
        }
        prompt.push('\n');
    }

    if let Some(ref prev_result) = snapshot.previous_action_result {
        prompt.push_str(&format!(
            "### Previous Action Result\n- Verdict: {:?}\n- Exit code: {:?}\n- stdout: {}\n- stderr: {}\n\n",
            prev_result.verdict,
            prev_result.exit_code,
            truncate_str(&prev_result.stdout, 1000),
            truncate_str(&prev_result.stderr, 1000),
        ));
    }

    // ── History summary ─────────────────────────────────────────────────
    if !history_rounds.is_empty() {
        prompt.push_str("## Previous Rounds\n\n");
        for record in history_rounds.iter().rev().take(3) {
            prompt.push_str(&format!(
                "- Round {}: {} | {} tokens | {}ms\n",
                record.round,
                truncate_str(&record.reasoning_summary, 120),
                record.tokens_input + record.tokens_output,
                record.duration_ms,
            ));
        }
        prompt.push('\n');
    }

    prompt.push_str(
        "## Instructions\n\nProduce a structured analysis and action plan in JSON format as specified.",
    );

    prompt
}

/// Extract JSON content from a string that may be wrapped in markdown fences.
fn extract_json(s: &str) -> &str {
    let s = s.trim();

    if s.starts_with("```json") {
        let inner = &s[7..];
        if let Some(end) = inner.rfind("```") {
            return inner[..end].trim();
        }
        return inner.trim();
    }

    if s.starts_with("```") {
        let inner = &s[3..];
        if let Some(end) = inner.rfind("```") {
            return inner[..end].trim();
        }
        return inner.trim();
    }

    // Heuristic: find first { and last }
    if let (Some(start), Some(end)) = (s.find('{'), s.rfind('}')) {
        return &s[start..=end];
    }

    s
}

fn truncate_str(s: &str, max_len: usize) -> String {
    if s.len() <= max_len {
        s.to_string()
    } else {
        format!("{}...", &s[..max_len])
    }
}