import json
import sys
from pathlib import Path


def read_json(path):
    return json.loads(Path(path).read_text(encoding="utf-8"))


def load_models(root):
    return [read_json(path) for path in sorted((root / "models").glob("*.model.json"))]


def load_selector_rules(root):
    return read_json(root / "selector" / "selector_rules.json")


def rules_by_model(selector_rules):
    return {item["model_id"]: item for item in selector_rules.get("models", [])}


def hit_any(user_input, signals):
    return [signal for signal in signals if signal in user_input]


def instruction_segment(user_input):
    segments = [user_input]
    for delimiter in ("：", ":", "\n"):
        if delimiter in user_input:
            segments.append(user_input.split(delimiter, 1)[0])
    return min(segments, key=len)


def has_analysis_override(user_input, global_rules):
    return bool(hit_any(user_input, global_rules.get("explicit_analysis_override_phrases", [])))


def hard_no_call_hits(user_input, global_rules):
    signals = list(global_rules.get("hard_no_call_signals", []))
    signals.extend(global_rules.get("direct_execution_no_call_signals", []))
    hits = hit_any(user_input, signals)
    instruction_hits = instruction_task_no_call_hits(
        user_input,
        global_rules.get("instruction_task_no_call_signals", []),
    )
    return hits + instruction_hits


def instruction_task_no_call_hits(user_input, signals):
    instruction = instruction_segment(user_input)
    if instruction != user_input:
        return hit_any(instruction, signals)
    stripped = user_input.strip()
    return [
        signal
        for signal in signals
        if 0 <= stripped.find(signal) <= 4
    ]


def depth_limited_qpi_override_hits(user_input, global_rules):
    depth_hits = hit_any(user_input, global_rules.get("depth_limiting_no_call_signals", []))
    override_hits = hit_any(user_input, global_rules.get("qpi_limited_analysis_override_phrases", []))
    return depth_hits, override_hits


def model_hard_exclusion_hits(user_input, model_rule):
    return hit_any(user_input, model_rule.get("hard_exclusion_triggers", []))


def score_model(model, model_rule, global_rules, user_input, task_type="", pipeline_position=""):
    model_id = model.get("model_id")
    weights = global_rules.get("weights", {})
    score = float(model_rule.get("base_score", 0)) / 100
    reasons = []
    penalties = []

    negative_hits = hit_any(user_input, model_rule.get("negative_triggers", [])) + hit_any(user_input, model.get("negative_triggers", []))
    depth_hits, qpi_limited_hits = depth_limited_qpi_override_hits(user_input, global_rules)
    suppress_qpi_depth_penalty = model_id == "qpi" and depth_hits and qpi_limited_hits
    if negative_hits and not has_analysis_override(user_input, global_rules) and not suppress_qpi_depth_penalty:
        penalty_key = "model_negative_penalty_qpi" if model_id == "qpi" else "model_negative_penalty_default"
        score -= weights.get(penalty_key, 0.5)
        penalties.append("negative trigger: " + "、".join(sorted(set(negative_hits))[:4]))

    trigger_hits = hit_any(user_input, model_rule.get("trigger_keywords", [])) + hit_any(user_input, model.get("trigger_keywords", []))
    trigger_hits = sorted(set(trigger_hits), key=trigger_hits.index)
    if trigger_hits:
        score += weights.get("trigger_keyword", 0.25)
        reasons.append("trigger keyword matched: " + "、".join(trigger_hits[:4]))

    input_type_matches = model_rule.get("input_type_matches", []) + model.get("input_types", [])
    input_hits = [input_type for input_type in input_type_matches if input_type in user_input or input_type == task_type]
    if input_hits:
        score += weights.get("input_type", 0.15)
        reasons.append("input type matched: " + "、".join(input_hits[:3]))

    if pipeline_position and model.get("pipeline_position") == pipeline_position:
        score += weights.get("pipeline_position", 0.2)
        reasons.append("pipeline_position matched")

    complexity_hits = hit_any(user_input, global_rules.get("complexity_signals", []))
    if complexity_hits:
        score += weights.get("complexity_signal", 0.15)
        reasons.append("complexity signal: " + "、".join(complexity_hits[:4]))

    priority = model.get("selection_priority")
    if isinstance(priority, int):
        score += max(0, min(priority, 10)) * weights.get("selection_priority_factor", 0.01)
        reasons.append(f"selection_priority={priority}")

    qpi_gate_hits = hit_any(user_input, global_rules.get("qpi_gate_signals", []))
    if model_id == "qpi" and qpi_gate_hits:
        score += weights.get("qpi_gate_bonus", 0.25)
        reasons.append("QPI problem-definition gate matched")
    if model_id == "qpi":
        has_positive_signal = bool(
            trigger_hits
            or input_hits
            or complexity_hits
            or qpi_gate_hits
            or has_analysis_override(user_input, global_rules)
            or task_type in {"problem_definition", "question_analysis"}
            or pipeline_position in {"pre_analysis", "problem_definition"}
        )
        if not has_positive_signal:
            score -= weights.get("qpi_missing_positive_signal_penalty", 0.25)
            penalties.append("QPI positive signal missing")
    if model_id == "qpi" and "QPI 已判断" in user_input:
        score -= weights.get("qpi_already_completed_penalty", 0.25)
        penalties.append("QPI already completed")

    if model_id == "intellectual_archaeology":
        heavy_hits = hit_any(user_input, global_rules.get("ia_heavy_signals", []))
        if heavy_hits:
            score += weights.get("ia_heavy_bonus", 0.2)
            reasons.append("IA heavy-depth signal: " + "、".join(heavy_hits[:3]))
        if "QPI 已判断" in user_input:
            score += weights.get("ia_qpi_completed_bonus", 0.25)
            reasons.append("QPI-before-IA gate satisfied")
        if global_rules.get("qpi_before_ia_gate", True) and qpi_gate_hits and "QPI 已判断" not in user_input:
            score -= weights.get("ia_qpi_gate_penalty", 0.45)
            penalties.append("QPI-before-IA gate: " + "、".join(qpi_gate_hits[:3]))
        if not heavy_hits and "QPI 已判断" not in user_input:
            score -= weights.get("ia_no_heavy_gate_penalty", 0.2)
            penalties.append("no IA heavy-depth gate")

    score = min(max(score, 0.0), 1.0)
    return {
        "model_id": model_id,
        "score": round(score, 2),
        "reasons": reasons,
        "penalties": penalties
    }


def hard_no_call_result(root, user_input, hits):
    rejected = []
    for model in load_models(root):
        rejected.append({
            "model_id": model.get("model_id"),
            "score": 0.0,
            "reasons": [],
            "penalties": ["hard no-call signal: " + "、".join(sorted(set(hits))[:4])],
            "decision": "rejected"
        })
    return {
        "input": user_input,
        "selected_models": [],
        "rejected_models": rejected,
        "no_call": True,
        "routing_notes": "hard no-call gate matched; explicit analysis override was not present."
    }


def recommend(root, user_input, task_type="", pipeline_position="", threshold=None):
    selector_rules = load_selector_rules(root)
    global_rules = selector_rules.get("global_rules", {})
    threshold = global_rules.get("no_call_threshold", 0.35) if threshold is None else threshold

    no_call_hits = hard_no_call_hits(user_input, global_rules)
    depth_hits, qpi_limited_hits = depth_limited_qpi_override_hits(user_input, global_rules)
    has_depth_limited_qpi_override = bool(depth_hits and qpi_limited_hits)
    if (
        global_rules.get("hard_no_call_first", True)
        and no_call_hits
        and not has_analysis_override(user_input, global_rules)
        and not has_depth_limited_qpi_override
    ):
        return hard_no_call_result(root, user_input, no_call_hits)

    model_rules = rules_by_model(selector_rules)
    scored = []
    for model in load_models(root):
        model_id = model.get("model_id")
        model_rule = model_rules.get(model_id, {})
        item = score_model(model, model_rule, global_rules, user_input, task_type, pipeline_position)
        exclusion_hits = model_hard_exclusion_hits(user_input, model_rule)
        if exclusion_hits:
            item["score"] = 0.0
            item["hard_excluded"] = True
            item["penalties"].append("model hard exclusion signal: " + "、".join(sorted(set(exclusion_hits))[:4]))
        scored.append(item)
    no_call = all(item["score"] < threshold for item in scored)
    selected = []
    rejected = []

    for item in scored:
        output_item = {
            "model_id": item["model_id"],
            "score": item["score"],
            "reasons": item["reasons"],
            "penalties": item["penalties"],
            "decision": "selected" if item["score"] >= threshold and not no_call and not item.get("hard_excluded") else "rejected"
        }
        if output_item["decision"] == "selected":
            selected.append(output_item)
        else:
            rejected.append(output_item)

    selected.sort(key=lambda item: item["score"], reverse=True)
    rejected.sort(key=lambda item: item["score"], reverse=True)

    return {
        "input": user_input,
        "selected_models": selected,
        "rejected_models": rejected,
        "no_call": no_call,
        "routing_notes": "selector is rule-based and driven by selector/selector_rules.json; no LLM, no vector search, no answer generation."
    }


def main():
    root = Path(__file__).resolve().parents[1]
    example = "团队每次都说要长期主义，但一到季度 KPI 就回到短期动作，这到底是什么问题？"
    result = recommend(root, example, task_type="question_analysis", pipeline_position="pre_analysis")
    print(json.dumps(result, ensure_ascii=False, indent=2))
    return 0


if __name__ == "__main__":
    sys.exit(main())