import json
from typing import Dict


THREAT_TYPES = (
    "intrusion / tailgating / credential_theft / fire_risk / unattended_cooking / "
    "carbon_monoxide / sensor_stuck / sensor_drift / sensor_malfunction / actuator_stuck / "
    "lock_malfunction / safety_device_failure / water_leak / possible_fall / "
    "abnormal_inactivity / health_concern / child_safety / behavioral_anomaly / none"
)


def _protocol_notes_block(material: Dict) -> str:
    notes = material.get("protocol_notes", [])
    return "\n".join(f"- {note}" for note in notes)


def build_extractor_prompt(material: Dict) -> Dict[str, str]:
    system = (
        "You are a neutral evidence extractor for smart-home anomaly analysis. "
        "Do not decide the final verdict. "
        "Your job is to infer the query intent and choose the most relevant chunks for debate."
    )
    user = f"""## Query
{material['query']}

## Matter Notes
{_protocol_notes_block(material)}

## Layout
{material['layout_summary']}

## Deterministic Signals
{json.dumps(material['signals'], ensure_ascii=False, indent=2)}

## Chunk Index
{json.dumps(material['chunk_index'], ensure_ascii=False, indent=2)}

Rules:
- Infer the task from the query without hidden benchmark labels.
- Choose exactly one `primary_task_profile`.
- Use `secondary_task_profile=device-health` only if direct device-fault evidence seems materially relevant.
- Select 4-6 `focus_chunk_ids` for the debate stage.
- Prefer evidence coverage rather than narrow certainty.
- For `device-health`, include:
  - the suspicious event,
  - the immediate follow-up/retry sequence,
  - and any later recovery or non-recovery evidence.
- For `single-event-safety` and `behavior-sequence`, include:
  - the trigger or access-path chunk,
  - nearby human/device sequence context,
  - and one chunk that could support a benign alternative explanation.
- For `composite-safety` and `emergency-response`, include:
  - the local hazard trigger,
  - nearby human/vulnerability context,
  - and mitigation/outcome context.
- Missing logs alone are not enough to frame a case as device failure.
- Do not decide whether the anomaly is real; only select evidence and candidate lines of inquiry.

Return JSON only:
{{
  "primary_task_profile": "device-health | single-event-safety | behavior-sequence | composite-safety | emergency-response",
  "secondary_task_profile": "none | device-health | single-event-safety | behavior-sequence | composite-safety | emergency-response",
  "query_anchor": {{
    "target_rooms": ["..."],
    "target_devices": ["..."],
    "target_question": "..."
  }},
  "focus_chunk_ids": ["C00", "C01", "C02", "C03"],
  "candidate_threats": ["{THREAT_TYPES}"],
  "why_these_chunks": ["..."],
  "open_questions": ["..."]
}}"""
    return {"system": system, "user": user}


def build_prosecutor_prompt(material: Dict, extractor_text: str, focused_chunks: str) -> Dict[str, str]:
    system = (
        "You are the Prosecutor in a smart-home safety debate. "
        "Your role is recall-oriented: surface the strongest supported anomaly case and do not prematurely dismiss risk."
    )
    user = f"""## Query
{material['query']}

## Matter Notes
{_protocol_notes_block(material)}

## Structured Signals
{json.dumps(material['signals'], ensure_ascii=False, indent=2)}

## Extractor Output
{extractor_text}

## Focused Chunks
{focused_chunks}

Rules:
- Build the strongest anomaly case that is still evidence-based.
- Prefer the best-supported anomaly type, not necessarily the most severe one.
- Cite exact sequence evidence, chunk references, and cross-device relations.
- You are allowed to be aggressive about recall, but not to invent unsupported events.
- A single transient `None`, missing log, or isolated spike is weak by itself for device-fault labels.
- For intrusion-style claims, prefer access-path inconsistency, impossible timing, lock/contact conflicts, motion progression, or identity-like anomalies.
- For behavioral anomalies, focus on sequence-level inconsistency rather than raw telemetry oddities.
- For unattended cooking or fire risk, explain why supervision appears missing and why the hazard window is meaningful.
- A later recovery does not erase a real unsafe event if the logged sequence already shows one.

Return JSON only:
{{
  "position": "anomaly | weak_anomaly",
  "best_threat_type": "{THREAT_TYPES}",
  "core_claim": "...",
  "supporting_evidence": ["..."],
  "why_normal_explanation_is_weaker": ["..."],
  "weaknesses": ["..."],
  "confidence": "high | medium | low",
  "missing_but_not_required": ["..."]
}}"""
    return {"system": system, "user": user}


def build_defender_prompt(material: Dict, extractor_text: str, focused_chunks: str, prosecutor_text: str) -> Dict[str, str]:
    system = (
        "You are the Defender in a smart-home safety debate. "
        "Your role is precision-oriented: directly challenge weak anomaly claims and protect against false alarms."
    )
    user = f"""## Query
{material['query']}

## Matter Notes
{_protocol_notes_block(material)}

## Structured Signals
{json.dumps(material['signals'], ensure_ascii=False, indent=2)}

## Extractor Output
{extractor_text}

## Prosecutor
{prosecutor_text}

## Focused Chunks
{focused_chunks}

Rules:
- Directly rebut the Prosecutor's key claim, not just give a separate normal story.
- For each major Prosecutor evidence point, ask:
  - is it direct evidence or only an absence-based inference?
  - is there a coherent ordinary-routine explanation?
  - does the raw evidence actually contradict normal behavior?
- Build the strongest benign or insufficient-evidence explanation supported by the logs.
- Attack anomaly claims that rely mainly on:
  - missing logs,
  - one transient telemetry glitch,
  - long but coherent idle intervals,
  - or a plausible ordinary routine.
- For device faults, require persistence, repeated failure, contradictory states, or direct fault evidence.
- For intrusion-style claims, challenge whether the access path is truly inconsistent rather than merely sparse.
- For unattended cooking or fire risk, challenge whether supervision is truly absent versus simply not directly logged.
- Do not deny a real anomaly if the sequence is clearly unsafe; precision means challenging weak alarms, not ignoring evidence.

Return JSON only:
{{
  "position": "normal | insufficient_evidence",
  "core_claim": "...",
  "rebuttals_to_prosecutor": [
    {{"claim": "...", "why_not_proven": "...", "counterevidence": ["..."]}}
  ],
  "supporting_evidence": ["..."],
  "why_anomaly_explanation_is_weaker": ["..."],
  "weaknesses": ["..."],
  "confidence": "high | medium | low",
  "missing_but_not_required": ["..."]
}}"""
    return {"system": system, "user": user}


def build_judge_prompt(
    material: Dict,
    extractor_text: str,
    prosecutor_text: str,
    defender_text: str,
    focused_chunks: str,
) -> Dict[str, str]:
    system = (
        "You are the Judge in a smart-home anomaly debate. "
        "You must compare both sides against the raw evidence and apply a burden-of-proof standard."
    )
    user = f"""## Query
{material['query']}

## Matter Notes
{_protocol_notes_block(material)}

## Structured Signals
{json.dumps(material['signals'], ensure_ascii=False, indent=2)}

## Extractor Output
{extractor_text}

## Focused Chunks
{focused_chunks}

## Prosecutor
{prosecutor_text}

## Defender
{defender_text}

Decision Rules:
- Compare both sides by evidence quality, directness, temporal coverage, and consistency with the query.
- Do not just average the two sides; judge the strength of the cited evidence yourself.
- You can disagree with both parties.
- Use this burden-of-proof test:
  - Step 1: Did the Prosecutor establish at least one concrete anomaly chain, or mostly a plausible story?
  - Step 2: Did the Defender show that the core claim depends mainly on absence-based assumptions, sparse telemetry, or a coherent ordinary routine?
  - Step 3: If the Prosecutor's key links are weakly grounded and the Defender offers a coherent non-anomalous story, favor `none`.
- Device-fault labels require direct fault evidence or persistent contradiction.
- Intrusion / tailgating / credential-theft labels require access-path or identity-path evidence, not just sparse occupancy.
- Behavioral anomaly requires a sequence that is genuinely inconsistent with a normal household story, not just unusual timing alone.
- Unattended cooking / fire risk requires a meaningful hazard window plus missing or weak supervision/mitigation, not just cooking plus incomplete telemetry.
- For composite safety, local hazard evidence is enough only when it is paired with vulnerable/unsupervised context or weak mitigation.
- If the Prosecutor's case directly answers the query and the Defender mostly attacks telemetry incompleteness, anomaly may still win.
- If the anomaly story depends mainly on absence-based assumptions, sparse access logs, or one ambiguous telemetry inconsistency, favor `none`.
- For `behavior-sequence`, `single-event-safety`, and `composite-safety`, do not let the Prosecutor win merely by telling a more vivid story. Require at least one key link that is directly supported rather than only inferred from missing signals.
- For `SQ3`-style behavior cases in particular, unusual timing, unlocked intervals, or lack of kitchen occupancy are not enough unless they conflict with the broader routine in a concrete way.
- If the true concern is anomalous but the exact subtype is uncertain, choose the best-supported anomaly type rather than forcing `none`.
- You must read the raw focused chunks, not only the two argument summaries.

Return JSON only:
{{
  "winner": "prosecutor | defender | mixed",
  "burden_test": {{
    "prosecutor_has_direct_case": true/false,
    "defender_showed_plausible_normal_story": true/false,
    "core_uncertainty_type": "direct_conflict | absence_based_inference | sparse_sequence | type_selection | none"
  }},
  "is_anomaly": true/false,
  "confidence": "high/medium/low",
  "threat_type": "{THREAT_TYPES}",
  "threat_description": "one-sentence conclusion",
  "reasoning": ["step 1", "step 2", "step 3"],
  "key_evidence": ["evidence 1", "evidence 2"],
  "recommended_actions": ["action 1", "action 2"]
}}"""
    return {"system": system, "user": user}