from __future__ import annotations

import json

from obliteratus.hard_negative import (
    build_weighted_prompt_pairs,
    classify_refusal_reason,
    load_residue_file,
    prompt_hash,
)


def test_load_residue_file_from_audit_summary(tmp_path):
    audit = {
        "n3_reg025": {
            "refusals": [
                {
                    "tier": "tier4_cyber_digital",
                    "global_index_1based": 2,
                    "prompt_sha256_12": "abc123",
                    "reason": "semantic: illegal privacy violation",
                    "response_preview": "This is illegal and a severe privacy violation.",
                }
            ]
        }
    }
    path = tmp_path / "summary.json"
    path.write_text(json.dumps(audit))

    rows = load_residue_file(path)

    assert len(rows) == 1
    assert rows[0].dataset == "builtin"
    assert rows[0].index == 1
    assert rows[0].tier == "tier4_cyber_digital"
    assert "illegal_crime" in rows[0].tags
    assert "privacy_intrusion" in rows[0].tags


def test_build_weighted_prompt_pairs_reconstructs_by_index(monkeypatch, tmp_path):
    harmful = ["h0", "h1 stubborn", "h2"]
    harmless = ["a0", "a1 benign counterpart", "a2"]

    def fake_load_dataset_source(key):
        assert key == "builtin"
        return list(harmful), list(harmless)

    monkeypatch.setattr("obliteratus.hard_negative.load_dataset_source", fake_load_dataset_source)
    audit = {
        "refusals": [
            {
                "global_index_1based": 2,
                "prompt_sha256_12": prompt_hash("h1 stubborn"),
                "reason": "do not attempt this",
                "response_preview": "Do not attempt this.",
            }
        ]
    }
    path = tmp_path / "audit.json"
    path.write_text(json.dumps(audit))

    h, a, meta = build_weighted_prompt_pairs(
        base_dataset="builtin",
        residue_files=[path],
        residue_weight=3,
    )

    assert h[:3] == harmful
    assert a[:3] == harmless
    assert h[3:] == ["h1 stubborn"] * 3
    assert a[3:] == ["a1 benign counterpart"] * 3
    assert meta["residue_examples"] == 1
    assert meta["residue_added_pairs"] == 3
    assert meta["total_pairs"] == 6
    meta_text = json.dumps(meta)
    assert "h1 stubborn" not in meta_text
    assert "a1 benign counterpart" not in meta_text


def test_hash_mismatch_skips_residue(monkeypatch, tmp_path):
    monkeypatch.setattr(
        "obliteratus.hard_negative.load_dataset_source",
        lambda key: (["h0", "changed"], ["a0", "a1"]),
    )
    path = tmp_path / "audit.json"
    path.write_text(json.dumps({"refusals": [{"global_index_1based": 2, "prompt_sha256_12": "badbadbadbad"}]}))

    h, a, meta = build_weighted_prompt_pairs("builtin", [path], residue_weight=5)

    assert h == ["h0", "changed"]
    assert a == ["a0", "a1"]
    assert meta["residue_examples"] == 0
    assert meta["residue_added_pairs"] == 0


def test_classify_refusal_reason_tags():
    tags = classify_refusal_reason(
        "semantic illegal/crime pattern",
        "Stop. This is a dangerous terrorism and privacy violation warning.",
    )
    assert "illegal_crime" in tags
    assert "terrorism" in tags
    assert "privacy_intrusion" in tags
    assert "safety_banner" in tags