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_triage_prompt(material: Dict) -> Dict[str, str]:
    system = (
        "You are the triage coordinator for smart-home log analysis. "
        "Infer the task from the query without hidden benchmark labels. "
        "Anchor to the query target first, then choose a small number of chunks for inspection."
    )
    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:
- Choose exactly one `primary_task_profile`.
- Use `secondary_task_profile` only if it materially helps.
- If the query asks whether a specific device/room works normally, prefer `device-health`.
- If the query asks for an action or response plan, prefer `emergency-response`.
- If the query asks whether a recent event is a security threat, prefer `single-event-safety`, not `device-health`.
- If the query asks whether there is an abnormal behavior pattern over many hours, prefer `behavior-sequence`, not `device-health`.
- If the query asks for the home's current safety status or potential risks, prefer `composite-safety`, not `device-health`.
- Select only 1-4 `focus_chunk_ids` in round 1.
- If `primary_task_profile=device-health`, prioritize chunks that show:
  - the suspected abnormal event,
  - the immediate follow-up/retry sequence,
  - and any later recovery or stabilization evidence.
- For `device-health`, avoid a single-chunk verdict when the case involves locks, contacts, sensors, or a claimed transient malfunction. Prefer at least:
  - one chunk before or at the suspicious event,
  - and one chunk after it that can confirm recovery, failed recovery, or missing follow-through.
- For locks and access devices, include both the command sequence and the later state outcome. Do not conclude health from one locally coherent unlock/lock snippet alone.
- For `composite-safety` or `emergency-response`, do not narrow the case to a single suspicious device too early. Prefer chunk selection that covers:
  - the possible hazard trigger,
  - nearby human/activity context,
  - and any recovery or consequence evidence.
- For `single-event-safety` or `behavior-sequence`, prioritize chunks that show:
  - the access path or trigger event,
  - nearby human/device sequence context,
  - and whether the sequence is explained by an ordinary routine.
- Missing logs for a device are not themselves evidence of failure.
- For non-`device-health` queries, do not let a transient `None`, temperature spike, or sparse telemetry become the main task framing unless the query itself is about device failure.
- For `composite-safety`, `behavior-sequence`, or `emergency-response`, use `secondary_task_profile=device-health` only when the case explicitly contains direct device-fault evidence such as fault events, repeated command retries, persistent no-response, or safety-device failure indicators.

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_rooms": ["..."],
  "focus_devices": ["..."],
  "focus_chunk_ids": ["C00", "C01"],
  "suspected_patterns": ["..."],
  "why_these_chunks": ["..."]
}}"""
    return {"system": system, "user": user}


def build_investigator_prompt(
    material: Dict,
    triage_text: str,
    focused_chunks: str,
    supervisor_text: str = "",
    round_index: int = 1,
) -> Dict[str, str]:
    system = (
        "You are the investigator. Work only from the query, Matter notes, structured signals, and focused raw chunks. "
        "Construct competing normal and anomaly hypotheses with explicit evidence. "
        "Distinguish device faults from behavior or safety anomalies carefully."
    )
    supervisor_block = ""
    if supervisor_text:
        supervisor_block = f"""
## Supervisor Feedback
{supervisor_text}

Apply the supervisor feedback explicitly in this round.
"""
    user = f"""## Query
{material['query']}

## Matter Notes
{_protocol_notes_block(material)}

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

## Triage Output
{triage_text}
{supervisor_block}
## Focused Chunks
{focused_chunks}

Round: {round_index}

Rules:
- A raw value like `2466` for `TemperatureMeasurement.MeasuredValue` means `24.66 C`, not `2466 C`.
- Do not infer `sensor_malfunction` or `sensor_drift` from scaled temperature values alone.
- Device-fault hypotheses require direct evidence such as explicit alarm/fault events, repeated non-recovery, stuck values, impossible state transitions, or actuator commands failing to take effect.
- Behavior, intrusion, safety, and emergency hypotheses may be supported by coherent temporal patterns, cross-device inconsistencies, absence where an event should appear, or risky multi-step sequences even when no explicit fault code exists.
- If the query asks for abnormal behavior or safety risk, do not discard the anomaly path just because the system eventually recovered.
- If `primary_task_profile` is not `device-health`, treat telemetry issues (`None`, short dropouts, one isolated spike) as side evidence unless they directly explain the queried behavior/safety event.
- Keep the normal hypothesis competitive, but also keep at least one anomaly hypothesis whenever the logs contain a plausible risk pattern that still needs explanation.
- For `device-health`, eventual recovery does not erase an anomaly if the logs show repeated retries, alarm/fault events, contradictory state transitions, failed first attempts, or unstable actuator behavior.
- For `device-health`, an isolated suspicious reading that immediately returns to baseline is not enough for `sensor_malfunction` unless there is repetition, corroboration, failed recovery, or direct fault evidence.
- For lock/contact health checks, require explicit failed lock commands, contradictory lock/contact states, repeated retries, or persistent insecure state. Do not infer `lock_malfunction` only from delayed auto-lock, a long unlocked interval, or sparse mid-day access logs.
- A single transient `None`, one brief telemetry dropout, or the mere absence of logs for a device is not enough to claim `sensor_malfunction`, `safety_device_failure`, or a monitoring blind spot.
- For `sensor_malfunction` / `sensor_stuck` / `sensor_drift`, require persistence, repetition, failed recovery, or direct contradiction with other signals.
- For `unattended_cooking` / `fire_risk`, require not only heat/cook activity but also evidence of missing supervision, dangerous duration, failed mitigation, or hazardous escalation.
- When `primary_task_profile` is `composite-safety` or `emergency-response`, do not rely only on missing kitchen occupancy, the occupant being logged in another room, or the absence of a smoke alarm for `unattended_cooking`. Prefer a concrete unsupervised sequence such as:
  - people leaving the hazard area with no return for a meaningful interval,
  - vulnerable subject context with no mitigating check-back,
  - or a hazardous device left active through a clear outcome window.
- When `primary_task_profile` is `composite-safety` or `emergency-response`, temperature rise or one telemetry gap is insufficient by itself for `fire_risk`. Prefer corroboration such as sustained heat growth, alarm-related evidence, failed mitigation, or a clearly unsafe appliance sequence.
- For `intrusion` / `tailgating` / `credential_theft`, occupancy alone is insufficient. Prefer corroboration from access-control inconsistencies, motion progression, lock/contact conflicts, or impossible entry timing.
- Do not explain away a suspected intrusion / credential-theft / child-safety / behavior anomaly as "normal cooking" or "telemetry noise" unless that explanation directly accounts for the queried abnormal sequence.
- For `child_safety`, `possible_fall`, and `abnormal_inactivity`, require subject-specific risky context, not just generic household quietness or sparse logs.
- For `composite-safety` or `emergency-response`, explicitly check whether your story is overly dependent on a single device or a single chunk. If yes, weaken the anomaly claim.
- For `composite-safety`, a local hazard can still justify an anomaly if all three are present:
  - a clear hazard source or risky sequence,
  - human vulnerability / missing supervision / dangerous context,
  - and weak or delayed mitigation.
- For `composite-safety`, do not collapse to `none` merely because not every device category was observed, if a concrete hazardous sequence is already evidenced.
- For `composite-safety`, if a cook/heat/leak/fall hazard is suspected, make sure your evidence covers:
  - the trigger chunk,
  - one adjacent context chunk,
  - and one mitigation/outcome chunk before concluding normal.

Return JSON only:
{{
  "normal_hypotheses": [
    {{"id": "N1", "description": "...", "evidence": ["..."], "weaknesses": ["..."]}}
  ],
  "anomaly_hypotheses": [
    {{"id": "A1", "description": "...", "threat_type": "{THREAT_TYPES}", "evidence": ["..."], "weaknesses": ["..."]}}
  ],
  "most_discriminative_evidence": ["..."],
  "missing_information": ["..."]
}}"""
    return {"system": system, "user": user}


def build_supervisor_prompt(
    material: Dict,
    triage_text: str,
    investigator_text: str,
    focused_chunks: str,
    round_index: int = 1,
) -> Dict[str, str]:
    system = (
        "You are the supervisor. Check whether the current evidence collection is on-topic and sufficient. "
        "Flag protocol-format misunderstandings, false-alarm risk, and missing checks."
    )
    user = f"""## Query
{material['query']}

## Matter Notes
{_protocol_notes_block(material)}

## Triage Output
{triage_text}

## Investigator Output
{investigator_text}

## Focused Chunks
{focused_chunks}

## Available Chunk IDs
{json.dumps([chunk['chunk_id'] for chunk in material['chunk_index']], ensure_ascii=False)}

Round: {round_index}

Rules:
- If the investigator seems to misread Matter-scaled values as literal impossible temperatures, mark false-alarm risk high.
- If the analysis drifted away from the query target, mark `on_topic=false`.
- Distinguish two standards of evidence:
  - device-fault labels need direct fault evidence;
  - behavior/safety/emergency labels may rely on coherent temporal and cross-device evidence.
- If the query is about recent security events, abnormal behavior, or whole-home safety, and the investigation mostly debates telemetry quality, treat that as partial task drift.
- If the evidence is still ambiguous after this round, do not automatically convert that into a normal verdict.
- Use `recommended_action=refine_investigation` when another round could realistically help.
- Use `recommended_action=abstain` only when the current anomaly story is weak, underspecified, or mostly speculative after review.
- For `device-health`, eventual recovery does not automatically clear the case if the observed sequence already contains direct malfunction evidence such as repeated retries, contradictory states, or fault/alarm events.
- If a `device-health` conclusion is being made from one local chunk or one short nominal sequence only, prefer `refine_investigation` and request adjacent pre/post chunks.
- Mark false-alarm risk `high` if the anomaly story depends mainly on:
  - one transient `None` or brief data dropout,
  - missing logs from a device,
  - one isolated suspicious reading without consequence,
  - or a broad safety conclusion built from a single local device issue.
- Mark false-alarm risk `high` for `sensor_malfunction` if the story rests on one transient spike that immediately returns to baseline without repetition, alarms, or downstream consequences.
- Mark false-alarm risk `high` for `lock_malfunction` if the story rests mainly on delayed auto-lock, a long unlocked interval, or unobserved mid-gap access without an explicit failed lock attempt or contradictory final state.
- For `composite-safety` or `emergency-response`, mark false-alarm risk `high` for `unattended_cooking` / `fire_risk` if the story rests mainly on missing kitchen occupancy, another room showing occupancy, missing OFF logs near the truncation boundary, or one transient telemetry gap without hazardous escalation.
- For `composite-safety` and `emergency-response`, only mark `evidence_sufficient=true` if the analysis covers both the local trigger and the wider human/safety context.
- For `composite-safety`, if the current focused chunks miss the trigger chunk, the adjacent context chunk, or the mitigation/outcome chunk, request those chunks explicitly in `needs_more_chunks`.
- For `behavior-sequence` and `single-event-safety`, if the current story does not explain the access path / sequence logic directly, request the surrounding sequence chunks instead of debating device telemetry.
- If the conclusion is effectively "a device was not logged, therefore the home is unsafe", treat that as weak evidence unless there is corroboration.
- For `composite-safety`, if a clear hazardous sequence already includes a local trigger, vulnerable/unsupervised context, and missing or delayed mitigation, that local sequence may be sufficient even if some peripheral devices are not discussed.

Return JSON only:
{{
  "on_topic": true/false,
  "evidence_sufficient": true/false,
  "risk_of_false_alarm": "low | medium | high",
  "recommended_action": "allow_final_verdict | refine_investigation | abstain",
  "needs_more_chunks": ["C03", "C05"],
  "missing_checks": ["..."],
  "supervisor_notes": ["..."]
}}"""
    return {"system": system, "user": user}


def build_verifier_prompt(
    material: Dict,
    triage_text: str,
    investigator_text: str,
    supervisor_text: str,
    focused_chunks: str,
) -> Dict[str, str]:
    system = (
        "You are the final verifier. Make a precise final decision from the query, Matter notes, evidence, and competing hypotheses. "
        "You must obey the supervisor gate."
    )
    user = f"""## Query
{material['query']}

## Matter Notes
{_protocol_notes_block(material)}

## Triage
{triage_text}

## Investigator
{investigator_text}

## Supervisor
{supervisor_text}

## Focused Chunks
{focused_chunks}

Rules:
- If the supervisor recommends `abstain`, you must return `is_anomaly=false`, `threat_type=none`, and `confidence=low`.
- Do not equate `evidence_sufficient=false` with `no anomaly`. It means the case may still require a cautious low/medium-confidence decision.
- Device-fault labels require direct device-fault evidence; scaled Matter temperature values alone are insufficient.
- Behavior, intrusion, safety, and emergency anomalies may be concluded from coherent temporal, causal, or cross-device evidence even when no explicit fault code exists.
- If the anomaly pattern is strong but type selection is uncertain, prefer the best-supported non-device-fault label over collapsing to `none`.
- Do not dismiss an anomaly just because the system later recovered, if the logged sequence itself reflects a real unsafe or abnormal event.
- If the investigator presents a plausible anomaly hypothesis and the supervisor does not abstain, you must explicitly refute that anomaly in your reasoning before returning `none`.
- If the query is non-`device-health` and the main anomaly story directly answers the query, prefer a low/medium-confidence anomaly over `none` when the remaining uncertainty is about telemetry quality rather than about the event sequence itself.
- For `composite-safety` or `emergency-response`, do not use that rule when the anomaly story itself is weakly supported or depends mainly on absence-based supervision assumptions, sparse lock activity, or one ambiguous telemetry inconsistency.
- Use `none` when the evidence supports a normal explanation better than an anomaly explanation.
- For `device-health`, repeated retries, contradictory lock/contact transitions, explicit alarm events, or unstable actuator sequences can justify anomaly even if the device later recovers.
- Never escalate to `sensor_malfunction`, `safety_device_failure`, or `sensor_stuck` from a single transient dropout or from missing logs alone.
- Never escalate to `sensor_malfunction` from one isolated spike that immediately returns to baseline unless you also have repetition, corroborating device conflict, failed recovery, or explicit fault evidence.
- Never escalate to `lock_malfunction` from delayed auto-lock, a long unlocked interval, or sparse access activity alone. Require failed lock commands, repeated retries, contradictory lock/contact states, or persistent insecure outcome.
- For `device-health`, if the inspected evidence does not cover both the suspicious event and a later confirming outcome, lower confidence and prefer additional review over a strong final verdict.
- For `single-event-safety`, `behavior-sequence`, and `composite-safety`, do not return `none` solely because some telemetry is incomplete if the human/device sequence already supports a concrete anomaly better than a normal routine.
- For `composite-safety` or `emergency-response`, do return `none` when the anomaly depends mainly on missing supervision assumptions, sparse access logs, or a plausible ordinary routine that explains the same sequence without contradiction.
- For `composite-safety` and `emergency-response`, check scope before finalizing:
  - Did you assess the broader human/safety context?
  - Or did you overfit to one device/chunk?
  If you overfit to one local signal, lower confidence or reject the anomaly claim.
- For `unattended_cooking` / `fire_risk`, require a hazardous sequence, not merely cooking plus incomplete telemetry.
- For `intrusion` / `tailgating`, require access-path evidence, not just unusual occupancy or periodic motion.
- For `composite-safety`, if the logs show a concrete hazardous sequence with a trigger, vulnerable or unsupervised context, and delayed/weak mitigation, you may conclude anomaly even if the full-home picture is incomplete.

Return JSON only:
{{
  "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}