"""Self-reflection — the act → critique → revise loop.

This is what makes NaijaTaste AI an agent rather than a one-shot pipeline.
After a first-pass output, the agent critiques its own work against the
persona and, if the critique finds problems, revises.

Two public entry points:

  reflect_on_review(...)   — Task A: critique + refine a generated review
  reflect_on_recommendations(...) — Task B: critique + refine a top-N list

Each runs at most `max_iterations` revise cycles (default 2). The loop
stops early once the critique passes (no blocking issues). Every cycle is
logged so the paper can report how often refinement triggered and what it
changed.

Reference: Madaan et al. 2023, "Self-Refine: Iterative Refinement with
Self-Feedback"; Shinn et al. 2023, "Reflexion".
"""
from __future__ import annotations

import logging
from dataclasses import dataclass, field
from typing import Optional

from pydantic import BaseModel, Field

from core.llm import LLMClient
from core.persona import UserPersona

log = logging.getLogger(__name__)


# ──────────────────────────────────────────────────────────────────────────────
# Critique schemas
# ──────────────────────────────────────────────────────────────────────────────

class ReviewCritique(BaseModel):
    """The critique LLM's assessment of a generated review (Task A)."""
    rating_text_consistent: bool = Field(
        description="True if the review text matches the star rating "
                    "(e.g. a 4-star review doesn't read like a 2-star pan)"
    )
    voice_match: bool = Field(
        description="True if the review sounds like THIS user — their length, "
                    "register, vocabulary, and quirks"
    )
    on_topic: bool = Field(
        description="True if the review is about the actual item, not generic filler"
    )
    issues: str = Field(
        description="If any check failed, a specific 1-2 sentence description of what "
                    "to fix. If all passed, the string 'none'."
    )

    @property
    def passed(self) -> bool:
        return self.rating_text_consistent and self.voice_match and self.on_topic


class RecommendationCritique(BaseModel):
    """The critique LLM's assessment of a top-N recommendation list (Task B)."""
    titles_are_real: bool = Field(
        description="True if the recommended items look like real products, "
                    "not review-headline fragments"
    )
    well_matched: bool = Field(
        description="True if the picks genuinely fit the persona's tastes"
    )
    reasoning_grounded: bool = Field(
        description="True if each pick's reasoning cites specific persona signals, "
                    "not generic filler"
    )
    diverse_enough: bool = Field(
        description="True if the list isn't 10 near-identical items"
    )
    issues: str = Field(
        description="If any check failed, a specific 1-2 sentence description of what "
                    "to fix. If all passed, the string 'none'."
    )

    @property
    def passed(self) -> bool:
        return (self.titles_are_real and self.well_matched
                and self.reasoning_grounded and self.diverse_enough)


# ──────────────────────────────────────────────────────────────────────────────
# Reflection trace (for logging / paper reporting)
# ──────────────────────────────────────────────────────────────────────────────

@dataclass
class ReflectionTrace:
    """Record of what the reflection loop did — useful for the paper."""
    iterations_run: int = 0
    critiques: list[str] = field(default_factory=list)   # issues found each cycle
    passed_final: bool = False
    refined: bool = False  # True if at least one revision happened


# ──────────────────────────────────────────────────────────────────────────────
# Task A — review reflection
# ──────────────────────────────────────────────────────────────────────────────

def _critique_review(llm: LLMClient, persona: UserPersona,
                     item_title: str, item_domain: str,
                     rating: float, review: str) -> ReviewCritique:
    """One critique pass over a generated review."""
    prompt = (
        f"You are a strict editor checking whether an AI-generated review "
        f"faithfully imitates a specific user. Be critical — your job is to "
        f"catch problems, not to be nice.\n\n"
        f"{'=' * 55}\n"
        f"THE USER\n"
        f"{'=' * 55}\n"
        f"{persona.to_prompt_block()}\n\n"
        f"{'=' * 55}\n"
        f"ITEM REVIEWED\n"
        f"{'=' * 55}\n"
        f"Domain: {item_domain}\n"
        f"Title: {item_title}\n\n"
        f"{'=' * 55}\n"
        f"THE GENERATED REVIEW (check this)\n"
        f"{'=' * 55}\n"
        f"Rating: {rating}\u2605\n"
        f"Review: {review}\n\n"
        f"{'=' * 55}\n"
        f"YOUR CHECKS\n"
        f"{'=' * 55}\n"
        f"1. rating_text_consistent: Does the review TEXT match the {rating}-star "
        f"rating? A 4-5 star review should read positive; a 1-2 star review should "
        f"read negative; a 3 should read mixed.\n"
        f"2. voice_match: Does it sound like THIS user? Check their typical review "
        f"length ({persona.avg_review_length:.0f} words avg), tone ({persona.tone}), "
        f"and quirks. A terse user given a long essay = fail. A user who writes in "
        f"all-caps given lowercase = fail.\n"
        f"3. on_topic: Is the review about the actual item, or is it generic filler "
        f"that could apply to anything?\n\n"
        f"If any check fails, describe specifically what to fix in 'issues'. "
        f"If all pass, set 'issues' to 'none'."
    )
    return llm.structured(
        prompt, ReviewCritique, model="reasoning",
        system="You are a meticulous editor. Catch every inconsistency.",
    )


def _refine_review(llm: LLMClient, persona: UserPersona,
                   item_title: str, item_domain: str,
                   prev_rating: float, prev_review: str,
                   critique_issues: str) -> tuple[float, str]:
    """Regenerate a review given critique feedback. Returns (rating, review)."""

    class RefinedReview(BaseModel):
        rating: float = Field(description="Star rating 1.0-5.0")
        review: str = Field(description="The improved review in the user's voice")

    prompt = (
        f"You previously wrote a review imitating a specific user, but an editor "
        f"found problems. Rewrite the review to fix them.\n\n"
        f"{'=' * 55}\n"
        f"THE USER\n"
        f"{'=' * 55}\n"
        f"{persona.to_prompt_block()}\n\n"
        f"ITEM: [{item_domain}] {item_title}\n\n"
        f"YOUR PREVIOUS ATTEMPT:\n"
        f"  Rating: {prev_rating}\u2605\n"
        f"  Review: {prev_review}\n\n"
        f"EDITOR'S FEEDBACK — fix these specific issues:\n"
        f"  {critique_issues}\n\n"
        f"Rewrite the review addressing the feedback. Keep what worked; fix what "
        f"the editor flagged. Stay in the user's authentic voice."
    )
    result = llm.structured(
        prompt, RefinedReview, model="reasoning",
        system="You are an expert behavioral simulator revising your work based on feedback.",
    )
    return result.rating, result.review


def reflect_on_review(llm: LLMClient, persona: UserPersona,
                      item_title: str, item_domain: str,
                      rating: float, review: str,
                      max_iterations: int = 2) -> tuple[float, str, ReflectionTrace]:
    """Critique a generated review and refine it if needed.

    Returns: (final_rating, final_review, trace)

    The loop:
      1. Critique the current review.
      2. If it passes → stop, return as-is.
      3. If it fails → refine using the critique, then critique again.
      4. Stop after max_iterations even if still imperfect.
    """
    trace = ReflectionTrace()
    cur_rating, cur_review = rating, review

    for i in range(max_iterations):
        try:
            critique = _critique_review(llm, persona, item_title, item_domain,
                                        cur_rating, cur_review)
        except Exception as e:
            log.warning(f"Review critique failed ({type(e).__name__}); "
                        f"keeping current review")
            break

        trace.iterations_run = i + 1
        if critique.passed:
            trace.critiques.append("passed")
            trace.passed_final = True
            log.info(f"Review reflection: passed on iteration {i + 1}")
            break

        trace.critiques.append(critique.issues)
        log.info(f"Review reflection iter {i + 1}: issues = {critique.issues}")

        # Refine
        try:
            cur_rating, cur_review = _refine_review(
                llm, persona, item_title, item_domain,
                cur_rating, cur_review, critique.issues,
            )
            trace.refined = True
        except Exception as e:
            log.warning(f"Review refine failed ({type(e).__name__}); "
                        f"keeping pre-refine review")
            break

    return cur_rating, cur_review, trace


# ──────────────────────────────────────────────────────────────────────────────
# Task B — recommendation reflection
# ──────────────────────────────────────────────────────────────────────────────

def _critique_recommendations(llm: LLMClient, persona: UserPersona,
                              recommendations: list[dict],
                              mode: str) -> RecommendationCritique:
    """One critique pass over a recommendation list."""
    rec_block = "\n".join(
        f"  #{i+1} [{r['domain']}] {r['title']}\n     Why: {r['reasoning']}"
        for i, r in enumerate(recommendations)
    )
    prompt = (
        f"You are a strict reviewer checking the quality of a recommendation "
        f"list. Be critical — catch problems.\n\n"
        f"{'=' * 55}\n"
        f"THE USER\n"
        f"{'=' * 55}\n"
        f"{persona.to_prompt_block()}\n\n"
        f"{'=' * 55}\n"
        f"THE RECOMMENDATIONS (mode: {mode})\n"
        f"{'=' * 55}\n"
        f"{rec_block}\n\n"
        f"{'=' * 55}\n"
        f"YOUR CHECKS\n"
        f"{'=' * 55}\n"
        f"1. titles_are_real: Do these look like real product titles? FAIL if any "
        f"are review-headline fragments like 'Fast paced great read' or 'An "
        f"enjoyable read' or 'Loved it!'.\n"
        f"2. well_matched: Do the picks genuinely fit this user's tastes?\n"
        f"3. reasoning_grounded: Does each 'Why' cite specific persona signals, "
        f"or is it generic filler?\n"
        f"4. diverse_enough: Is there real variety, or are these 10 near-identical "
        f"items?\n\n"
        f"If any check fails, describe specifically what to fix in 'issues' "
        f"(e.g. 'items #4, #7, #9 have review-headline titles — replace them'). "
        f"If all pass, set 'issues' to 'none'."
    )
    return llm.structured(
        prompt, RecommendationCritique, model="reasoning",
        system="You are a meticulous recommendation-quality auditor.",
    )


def reflect_on_recommendations(llm: LLMClient, persona: UserPersona,
                               recommendations: list[dict], mode: str,
                               refine_fn,
                               max_iterations: int = 2,
                               ) -> tuple[list[dict], ReflectionTrace]:
    """Critique a recommendation list and refine if needed.

    Unlike review reflection, refinement here can't just rewrite text — it
    needs to re-run reranking with feedback. So the caller passes a
    `refine_fn(issues: str) -> list[dict]` that re-runs the rerank with the
    critique injected, and this function orchestrates the loop.

    Returns: (final_recommendations, trace)
    """
    trace = ReflectionTrace()
    cur_recs = recommendations

    for i in range(max_iterations):
        try:
            critique = _critique_recommendations(llm, persona, cur_recs, mode)
        except Exception as e:
            log.warning(f"Recommendation critique failed ({type(e).__name__}); "
                        f"keeping current list")
            break

        trace.iterations_run = i + 1
        if critique.passed:
            trace.critiques.append("passed")
            trace.passed_final = True
            log.info(f"Recommendation reflection: passed on iteration {i + 1}")
            break

        trace.critiques.append(critique.issues)
        log.info(f"Recommendation reflection iter {i + 1}: issues = {critique.issues}")

        # Refine via the caller-supplied function
        try:
            refined = refine_fn(critique.issues)
            if refined:
                cur_recs = refined
                trace.refined = True
        except Exception as e:
            log.warning(f"Recommendation refine failed ({type(e).__name__}); "
                        f"keeping pre-refine list")
            break

    return cur_recs, trace