claw/src/compat/scene_platform/dispatch.rs

use std::path::Path;

use serde_json::{Map, Value};
use thiserror::Error;

use crate::compat::deterministic_submit::{
    DeterministicExecutionPlan, DeterministicSubmitDecision,
};
use crate::compat::scene_platform::registry::{
    load_scene_registry, SceneRegistryEntry, SceneRegistryError,
};
use crate::compat::scene_platform::resolvers::{
    resolve_required_scene_params, ResolverError, SceneParamResolution,
};
use crate::scene_contract::PostprocessSection;

const DETERMINISTIC_SUFFIX: &str = "\u{3002}\u{3002}\u{3002}";
const DIAGNOSTIC_SCENE_ID: &str = "sweep-030-scene";

#[derive(Debug, Clone)]
pub struct SceneExecutionPlan {
    pub scene_id: String,
    pub instruction: String,
    pub tool_name: String,
    pub expected_domain: String,
    pub target_url: String,
    pub args: Map<String, Value>,
    pub success_statuses: Vec<String>,
    pub failure_statuses: Vec<String>,
    pub postprocess: Option<PostprocessSection>,
}

#[derive(Debug, Error)]
pub enum SceneDispatchError {
    #[error(transparent)]
    Registry(#[from] SceneRegistryError),
    #[error(transparent)]
    Resolver(#[from] ResolverError),
}

#[derive(Debug)]
struct CandidateScene<'a> {
    entry: &'a SceneRegistryEntry,
    score: usize,
    resolution: SceneParamResolution,
}

pub fn plan_deterministic_scene(
    raw_instruction: &str,
    page_url: Option<&str>,
    page_title: Option<&str>,
    skills_dir: &Path,
) -> Result<DeterministicSubmitDecision, SceneDispatchError> {
    let Some(stripped_instruction) = strip_exact_suffix(raw_instruction) else {
        return Ok(DeterministicSubmitDecision::NotDeterministic);
    };

    let instruction = stripped_instruction.trim();
    log_deterministic_diag(format!(
        "submit suffix=ok skills_dir={} skills_dir_exists={} instruction={} page_url={} page_title={}",
        skills_dir.display(),
        skills_dir.exists(),
        instruction,
        page_url.unwrap_or_default(),
        page_title.unwrap_or_default()
    ));
    if instruction.is_empty() {
        log_deterministic_diag("unsupported: empty deterministic instruction");
        return Ok(unsupported_scene_prompt());
    }

    let registry = load_scene_registry(skills_dir)?;
    log_registry_diag(&registry);
    if registry.is_empty() {
        log_deterministic_diag("unsupported: scene registry is empty");
        return Ok(unsupported_scene_prompt());
    }

    let mut candidates = Vec::new();
    for entry in &registry {
        if entry.manifest.scene.id == DIAGNOSTIC_SCENE_ID {
            log_scene_match_diag(entry, instruction);
        }
        let Some(score) = score_scene(entry, instruction, page_url, page_title) else {
            continue;
        };
        let resolution = resolve_required_scene_params(
            &entry.skill_root,
            instruction,
            &entry.manifest.scene.id,
            &entry.manifest.params,
        )?;
        candidates.push(CandidateScene {
            entry,
            score,
            resolution,
        });
    }

    if candidates.is_empty() {
        log_deterministic_diag(
            "unsupported: no scene candidate matched include/suffix/exclude rules",
        );
        return Ok(unsupported_scene_prompt());
    }

    candidates.sort_by(|left, right| {
        right
            .score
            .cmp(&left.score)
            .then_with(|| {
                right
                    .resolution
                    .resolved_required_count
                    .cmp(&left.resolution.resolved_required_count)
            })
            .then_with(|| {
                left.entry
                    .manifest
                    .scene
                    .id
                    .cmp(&right.entry.manifest.scene.id)
            })
    });

    if candidates.len() > 1 {
        let best = &candidates[0];
        let next = &candidates[1];
        if best.score == next.score
            && best.resolution.resolved_required_count == next.resolution.resolved_required_count
        {
            return Ok(DeterministicSubmitDecision::Prompt {
                summary: format!(
                    "已命中多个确定性场景（{}、{}），请补充更明确关键词或页面上下文。",
                    best.entry.manifest.scene.id, next.entry.manifest.scene.id
                ),
            });
        }
    }

    let selected = candidates.remove(0);
    log_deterministic_diag(format!(
        "selected scene={} tool={} score={} resolved_required_count={}",
        selected.entry.manifest.scene.id,
        selected.entry.manifest.scene.tool,
        selected.score,
        selected.resolution.resolved_required_count
    ));
    if let Some(summary) = selected.resolution.prompt {
        log_deterministic_diag(format!(
            "selected scene={} returned resolver prompt",
            selected.entry.manifest.scene.id
        ));
        return Ok(DeterministicSubmitDecision::Prompt { summary });
    }

    Ok(DeterministicSubmitDecision::Execute(to_deterministic_plan(
        instruction,
        build_scene_execution_plan(selected.entry, instruction, selected.resolution.args),
    )))
}

fn build_scene_execution_plan(
    entry: &SceneRegistryEntry,
    instruction: &str,
    mut args: Map<String, Value>,
) -> SceneExecutionPlan {
    let bootstrap = entry
        .manifest
        .bootstrap
        .as_ref()
        .expect("report scene registry should only contain manifests with bootstrap");
    let artifact = entry
        .manifest
        .artifact
        .as_ref()
        .expect("report scene registry should only contain manifests with artifact");
    args.insert(
        "expected_domain".to_string(),
        Value::String(bootstrap.expected_domain.clone()),
    );
    args.insert(
        "target_url".to_string(),
        Value::String(bootstrap.target_url.clone()),
    );

    SceneExecutionPlan {
        scene_id: entry.manifest.scene.id.clone(),
        instruction: instruction.to_string(),
        tool_name: format!(
            "{}.{}",
            entry.manifest.scene.skill, entry.manifest.scene.tool
        ),
        expected_domain: bootstrap.expected_domain.clone(),
        target_url: bootstrap.target_url.clone(),
        args,
        success_statuses: artifact.success_status.clone(),
        failure_statuses: artifact.failure_status.clone(),
        postprocess: entry.manifest.postprocess.clone(),
    }
}

fn to_deterministic_plan(
    instruction: &str,
    scene_plan: SceneExecutionPlan,
) -> DeterministicExecutionPlan {
    let org_label = scene_plan
        .args
        .get("org_label")
        .and_then(Value::as_str)
        .unwrap_or_default()
        .to_string();
    let org_code = scene_plan
        .args
        .get("org_code")
        .and_then(Value::as_str)
        .unwrap_or_default()
        .to_string();
    let period_mode = scene_plan
        .args
        .get("period_mode")
        .and_then(Value::as_str)
        .unwrap_or_default()
        .to_string();
    let period_mode_code = scene_plan
        .args
        .get("period_mode_code")
        .and_then(Value::as_str)
        .unwrap_or_default()
        .to_string();
    let period_value = scene_plan
        .args
        .get("period_value")
        .and_then(Value::as_str)
        .unwrap_or_default()
        .to_string();
    let period_payload = scene_plan
        .args
        .get("period_payload")
        .cloned()
        .map(|payload| serde_json::to_string(&payload).unwrap_or_else(|_| "{}".to_string()))
        .unwrap_or_else(|| "{}".to_string());

    DeterministicExecutionPlan {
        instruction: instruction.to_string(),
        tool_name: scene_plan.tool_name,
        expected_domain: scene_plan.expected_domain,
        target_url: scene_plan.target_url,
        org_label,
        org_code,
        period_mode,
        period_mode_code,
        period_value,
        period_payload,
        postprocess: scene_plan.postprocess,
    }
}

fn score_scene(
    entry: &SceneRegistryEntry,
    instruction: &str,
    page_url: Option<&str>,
    page_title: Option<&str>,
) -> Option<usize> {
    let deterministic = entry.manifest.deterministic.as_ref()?;
    let bootstrap = entry.manifest.bootstrap.as_ref()?;
    if deterministic.suffix != DETERMINISTIC_SUFFIX {
        return None;
    }

    let include_hits = deterministic
        .include_keywords
        .iter()
        .filter(|keyword| !keyword.trim().is_empty() && instruction.contains(keyword.as_str()))
        .count();
    if include_hits == 0 {
        return None;
    }

    if deterministic
        .exclude_keywords
        .iter()
        .any(|keyword| !keyword.trim().is_empty() && instruction.contains(keyword.as_str()))
    {
        return None;
    }

    let mut score = include_hits * 10;

    let normalized_url = page_url.unwrap_or_default().to_ascii_lowercase();
    if !normalized_url.is_empty() {
        if normalized_url.contains(
            &bootstrap.expected_domain.to_ascii_lowercase(),
        ) {
            score += 100;
        } else if normalized_url.contains(&entry.manifest.scene.id.to_ascii_lowercase()) {
            score += 40;
        }
    }

    let title = page_title.unwrap_or_default();
    if !title.is_empty()
        && entry
            .manifest
            .bootstrap
            .as_ref()?
            .page_title_keywords
            .iter()
            .any(|keyword| !keyword.trim().is_empty() && title.contains(keyword.as_str()))
    {
        score += 60;
    }

    Some(score)
}

fn strip_exact_suffix(raw_instruction: &str) -> Option<&str> {
    let without_suffix = raw_instruction.strip_suffix(DETERMINISTIC_SUFFIX)?;
    if without_suffix.ends_with('\u{3002}') {
        return None;
    }
    Some(without_suffix)
}

fn log_registry_diag(registry: &[SceneRegistryEntry]) {
    let sweep_030 = registry
        .iter()
        .find(|entry| entry.manifest.scene.id == DIAGNOSTIC_SCENE_ID);
    match sweep_030 {
        Some(entry) => log_deterministic_diag(format!(
            "registry loaded count={} diagnostic_scene={} skill_root={} suffix_ok={} include_keywords={:?}",
            registry.len(),
            DIAGNOSTIC_SCENE_ID,
            entry.skill_root.display(),
            entry
                .manifest
                .deterministic
                .as_ref()
                .map(|deterministic| deterministic.suffix == DETERMINISTIC_SUFFIX)
                .unwrap_or(false),
            entry
                .manifest
                .deterministic
                .as_ref()
                .map(|deterministic| deterministic.include_keywords.clone())
                .unwrap_or_default()
        )),
        None => log_deterministic_diag(format!(
            "registry loaded count={} diagnostic_scene={} registered=false",
            registry.len(),
            DIAGNOSTIC_SCENE_ID
        )),
    }
}

fn log_scene_match_diag(entry: &SceneRegistryEntry, instruction: &str) {
    let Some(deterministic) = entry.manifest.deterministic.as_ref() else {
        log_deterministic_diag(format!(
            "diagnostic_scene={} deterministic=false",
            entry.manifest.scene.id
        ));
        return;
    };
    let include_hits = deterministic
        .include_keywords
        .iter()
        .filter(|keyword| !keyword.trim().is_empty() && instruction.contains(keyword.as_str()))
        .cloned()
        .collect::<Vec<_>>();
    let exclude_hits = deterministic
        .exclude_keywords
        .iter()
        .filter(|keyword| !keyword.trim().is_empty() && instruction.contains(keyword.as_str()))
        .cloned()
        .collect::<Vec<_>>();
    log_deterministic_diag(format!(
        "diagnostic_scene={} suffix_ok={} suffix_codepoints={} include_hits={:?} exclude_hits={:?}",
        entry.manifest.scene.id,
        deterministic.suffix == DETERMINISTIC_SUFFIX,
        deterministic
            .suffix
            .chars()
            .map(|ch| format!("U+{:04X}", ch as u32))
            .collect::<Vec<_>>()
            .join(","),
        include_hits,
        exclude_hits
    ));
}

fn log_deterministic_diag(message: impl AsRef<str>) {
    eprintln!("[sgclaw deterministic] {}", message.as_ref());
}

fn unsupported_scene_prompt() -> DeterministicSubmitDecision {
    DeterministicSubmitDecision::Prompt {
        summary: "确定性提交当前只支持已注册的报表采集场景，请补充已支持的业务请求。".to_string(),
    }
}