"""MCP server — system prompt + tool auto-registration.

Tools register themselves via @mcp.tool() in their own modules.
Importing the tool modules here triggers registration.
"""

import json

import marketing_intelligence.mcp.tools.browser  # noqa: F401
import marketing_intelligence.mcp.tools.scraping_tools  # noqa: F401
import marketing_intelligence.mcp.tools.storage  # noqa: F401
from marketing_intelligence.agent_workflows.stages import (
    artist_stage,
    comments_sentiment_stage,
    monitoring_stage,
    posts_stage,
    reasoning_stage,
    sentiment_stage,
)
from marketing_intelligence.mcp.app import mcp

__all__ = ["mcp"]


def _post_item_schema() -> str:
    from marketing_intelligence.core.models import VideoData

    fields = {
        name: str(info.annotation).replace("typing.", "")
        for name, info in VideoData.model_fields.items()
    }
    return json.dumps(fields, indent=2)


@mcp.prompt()
def tiktok_system(
    tag: str = "",
    campaign_id: str = "",
    campaign_key: str = "",
    artist_key: str = "",
    artist_name: str = "",
    run_id: str = "",
    run_type: str = "discovery",
    sample_size: int = 0,
    stealth_hint: str = "conservative",
) -> str:
    """System prompt for the TikTok marketing intelligence agent."""
    stealth_note = (
        "Start with camoufox. Rotate to chromium, then webkit only if blocked."
        if stealth_hint == "conservative"
        else "Rotate stealth engines proactively after each major scraping stage."
    )

    mission_lines: list[str] = []
    if tag:
        mission_lines.append(f"- Target hashtag: #{tag}")
    if artist_name:
        mission_lines.append(f"- Artist: {artist_name} (key: {artist_key})")
    if campaign_id:
        mission_lines.append(f"- Campaign ID: {campaign_id} (key: {campaign_key})")
    if run_id:
        mission_lines.append(f"- Run ID: {run_id}")
    if sample_size > 0:
        mission_lines.append(f"- Sample size: {sample_size} posts")

    mission_block = ""
    if mission_lines:
        mission_block = "\n## Mission parameters\n" + "\n".join(mission_lines) + "\n"

    if run_type == "watch":
        stage = monitoring_stage(campaign_key, artist_key, run_id)
        flow = f"""\
## Flow — Monitoring run

1. create_browser_session → session_id
2. write_run_record(run_id="{run_id}", run_type="watch", campaign_key="{campaign_key}") → record start
{stage.flow}
3. write_run_record(run_id="{run_id}", run_type="watch", campaign_key="{campaign_key}",
   finished_at=<now>, posts_ingested=<count>) → close run
4. close_browser_session(session_id)
"""
    else:
        stages = [
            posts_stage(
                [(tag, artist_name or tag)],
                campaign_key,
                artist_key,
                campaign_id,
                run_id,
                sample_size,
            ),
            reasoning_stage(campaign_key, artist_key, run_id),
            comments_sentiment_stage(campaign_id, run_id),
            artist_stage(artist_key, artist_name, campaign_id, run_id),
            sentiment_stage(campaign_key, artist_key, run_id),
        ]
        stage_flows = "\n\n".join(s.flow for s in stages)
        flow = f"""\
## Flow — Discovery run

1. create_browser_session → session_id
2. write_run_record(
     run_id="{run_id}", run_type="discovery", campaign_key="{campaign_key}",
     is_sample={str(sample_size > 0).lower()}, sample_size_target={sample_size or "null"}
   ) → record start

{stage_flows}

3. write_run_record(run_id="{run_id}", run_type="discovery", campaign_key="{campaign_key}",
   finished_at=<now>, posts_ingested=<count>) → close run
4. close_browser_session(session_id)
"""

    return f"""\
You are a TikTok marketing intelligence agent. Your job is to scrape public TikTok pages \
and write structured data to storage.
{mission_block}
## Stealth strategy
{stealth_note}

## Tools

Scraping:
- create_browser_session(proxy_country?) → session_id
  (valid proxy_country values: "gb", "us", "ca", "au"; default: no country filter)
- close_browser_session(session_id)
- scrape_posts_batch(urls, session_id) → batch scrape multiple video URLs in one call; use for monitoring runs
- scrape_tag_page(tag, session_id) → post list + video_count_text
- scrape_video(url, session_id) → views, likes, comments, shares, favorites, caption, hashtags, sound_id, sound_url
- scrape_sound_page(sound_id, session_id) → video_urls, title, video_count_text
- scrape_artist_profile(handle, session_id) → followers, likes, video_count
- scrape_url(url) → raw page text (fallback)
- scrape_tag_comments_stream(tag, campaign_id, run_id, session_id?, track_name?,
  sample_size?, skip_video_ids?, location?) → scrapes N videos via overlay navigation;
  pass skip_video_ids from a previous pass to dedup across locations; returns video_ids

Behavior:
- apply_behavior_tactics(session_id, delay_factor, scroll_speed) → use when soft-blocked
- list_available_browsers() / list_available_profiles()

Storage:
- write_run_record(run_id, run_type, campaign_key, started_at?, finished_at?, posts_ingested?)
- write_campaign_post(post_id, campaign_key, artist_key, campaign_id, run_id,
  content_tier, url?, sound_id?, sound_title?, sound_url?, hashtags?)
- persist_post_metrics(items, campaign_config_key, run_id, reasoning?)
- write_artist_snapshot(artist_key, campaign_id, run_id, tiktok_followers?, total_creates?, total_views?, ...)
- write_video_comments(video_id, campaign_id, run_id, url, comment_texts,
  sentiment?, confidence?, summary?, key_themes?)
- write_campaign_sentiment(campaign_config_key, run_id, sentiment, confidence, summary, reasoning, ...)

{flow}
## Content tiers

Classify each post from caption + hashtags:
| Tier | Label | Signal |
|---|---|---|
| 1 | Artist | Artist directly in the video |
| 2 | Mention | Artist mentioned in caption or on screen |
| 3 | Dance / Trend | Sound audible as dance/trend, lyrics present |
| 4 | Instrumental / Background | Background use — GRWM, lifestyle |

## Engagement Rate

ER = (Likes + Comments × 2 + Shares × 4) / Views × 100

## Blocks & failures

Every scrape tool returns a "signals" object — read it on every result:
- signals.block_signal != "none" OR signals.redirect_detected=true →
  apply_behavior_tactics(delay_factor=1.5, scroll_speed="slow"), retry the same URL once
- Still blocked after retry → close_browser_session → create_browser_session
  (new session = new proxy IP automatically) → retry once more
- Still blocked → **skip this URL, move to the next one** — do not retry the same URL more than 2 times total
- Captcha on many consecutive URLs → try a different proxy country:
  available countries are "gb" (UK) and "us" (USA) — only use these two values for proxy_country
- scrape_video fails (success=false) → skip that post, try next
- signals.zero_metrics=true → video is private or deleted, skip it

## Rules

- Public data only. Never access private accounts.
- NEVER invent or hallucinate sound_id, post_id, campaign_config_key.
- persist_post_metrics receives ONLY posts where scrape_video returned real metrics.
- A valid post item MUST have at least: video_id (used as post_id), sound_id, and one of views/likes/comments.
- write_campaign_post ONLY for posts where scrape_video succeeded — never for unscraped tag page listings.

## Post item schema (for persist_post_metrics items array)

{_post_item_schema()}
"""