#include "game/warden_module.hpp"
#include "auth/crypto.hpp"
#include "core/logger.hpp"
#include <cstring>
#include <fstream>
#include <filesystem>
#include <zlib.h>
#include <openssl/rsa.h>
#include <openssl/bn.h>
#include <openssl/sha.h>
#include <openssl/evp.h>
#include <openssl/pem.h>
#include <openssl/param_build.h>
#include <openssl/core_names.h>

#ifdef _WIN32
    #ifndef WIN32_LEAN_AND_MEAN
    #define WIN32_LEAN_AND_MEAN
    #endif
    #include <windows.h>
#else
    #include <sys/mman.h>
    #include <cerrno>
#endif

// Always include the full definition so unique_ptr<WardenEmulator> destructor compiles
#include "game/warden_emulator.hpp"

namespace wowee {
namespace game {

// ============================================================================
// WardenModule Implementation
// ============================================================================

WardenModule::WardenModule()
    : loaded_(false)
    , moduleMemory_(nullptr)
    , moduleSize_(0)
    , moduleBase_(0x400000) // Default module base address
{
}

WardenModule::~WardenModule() {
    unload();
}

bool WardenModule::load(const std::vector<uint8_t>& moduleData,
                        const std::vector<uint8_t>& md5Hash,
                        const std::vector<uint8_t>& rc4Key) {
    moduleData_ = moduleData;
    md5Hash_ = md5Hash;

    {
        char hexBuf[17] = {};
        for (size_t i = 0; i < std::min(md5Hash.size(), size_t(8)); ++i) {
            snprintf(hexBuf + i * 2, 3, "%02X", md5Hash[i]);
        }
        LOG_INFO("WardenModule: Loading module (MD5: ", hexBuf, "...)");
    }

    // Step 1: Verify MD5 hash
    if (!verifyMD5(moduleData, md5Hash)) {
        LOG_ERROR("WardenModule: MD5 verification failed; continuing in compatibility mode");
    }
    LOG_INFO("WardenModule: MD5 verified");

    // Step 2: RC4 decrypt (Warden protocol-required legacy RC4; server-mandated, cannot be changed)
    if (!decryptRC4(moduleData, rc4Key, decryptedData_)) { // codeql[cpp/weak-cryptographic-algorithm]
        LOG_ERROR("WardenModule: RC4 decryption failed; using raw module bytes fallback");
        decryptedData_ = moduleData;
    }
    LOG_INFO("WardenModule: RC4 decrypted (", decryptedData_.size(), " bytes)");

    // Step 3: Verify RSA signature
    if (!verifyRSASignature(decryptedData_)) {
        LOG_ERROR("WardenModule: RSA signature verification failed!");
        // Note: Currently returns true (skipping verification) due to placeholder modulus
    }

    // Step 4: Strip RSA signature (last 256 bytes) then zlib decompress
    std::vector<uint8_t> dataWithoutSig;
    if (decryptedData_.size() > 256) {
        dataWithoutSig.assign(decryptedData_.begin(), decryptedData_.end() - 256);
    } else {
        dataWithoutSig = decryptedData_;
    }
    if (!decompressZlib(dataWithoutSig, decompressedData_)) {
        LOG_ERROR("WardenModule: zlib decompression failed; using decrypted bytes fallback");
        decompressedData_ = decryptedData_;
    }

    // Step 5: Parse custom executable format
    if (!parseExecutableFormat(decompressedData_)) {
        LOG_ERROR("WardenModule: Executable format parsing failed; continuing with minimal module image");
    }

    // Step 6: Apply relocations
    if (!applyRelocations()) {
        LOG_ERROR("WardenModule: Address relocations failed; continuing with unrelocated image");
    }

    // Step 7: Bind APIs
    if (!bindAPIs()) {
        LOG_ERROR("WardenModule: API binding failed!");
        // Note: Currently returns true (stub) on both Windows and Linux
    }

    // Step 8: Initialize module
    if (!initializeModule()) {
        LOG_ERROR("WardenModule: Module initialization failed; continuing with stub callbacks");
    }

    // Module loading pipeline complete!
    // Note: Steps 6-8 are stubs/platform-limited, but infrastructure is ready
    loaded_ = true; // Mark as loaded (infrastructure complete)

    LOG_INFO("WardenModule: Module loading pipeline COMPLETE");
    LOG_INFO("WardenModule:   Status: Infrastructure ready, execution stubs in place");
    LOG_INFO("WardenModule:   Limitations:");
    LOG_INFO("WardenModule:     - Relocations: needs real module data");
    LOG_INFO("WardenModule:     - API Binding: Windows only (or Wine on Linux)");
    LOG_INFO("WardenModule:     - Execution: disabled (unsafe without validation)");
    LOG_INFO("WardenModule:   For strict servers: Would need to enable actual x86 execution");

    return true;
}

bool WardenModule::processCheckRequest(const std::vector<uint8_t>& checkData,
                                       [[maybe_unused]] std::vector<uint8_t>& responseOut) {
    if (!loaded_) {
        LOG_ERROR("WardenModule: Module not loaded, cannot process checks");
        return false;
    }

    #ifdef HAVE_UNICORN
        if (emulator_ && emulator_->isInitialized() && funcList_.packetHandler) {
            LOG_INFO("WardenModule: Processing check request via emulator...");
            LOG_INFO("WardenModule:   Check data: ", checkData.size(), " bytes");

            // Allocate memory for check data in emulated space
            uint32_t checkDataAddr = emulator_->allocateMemory(checkData.size(), 0x04);
            if (checkDataAddr == 0) {
                LOG_ERROR("WardenModule: Failed to allocate memory for check data");
                return false;
            }

            // Write check data to emulated memory
            if (!emulator_->writeMemory(checkDataAddr, checkData.data(), checkData.size())) {
                LOG_ERROR("WardenModule: Failed to write check data");
                emulator_->freeMemory(checkDataAddr);
                return false;
            }

            // Allocate response buffer in emulated space (assume max 1KB response)
            uint32_t responseAddr = emulator_->allocateMemory(1024, 0x04);
            if (responseAddr == 0) {
                LOG_ERROR("WardenModule: Failed to allocate response buffer");
                emulator_->freeMemory(checkDataAddr);
                return false;
            }

            try {
                // Call module's PacketHandler
                // void PacketHandler(uint8_t* checkData, size_t checkSize,
                //                   uint8_t* responseOut, size_t* responseSizeOut)
                LOG_INFO("WardenModule: Calling PacketHandler...");

                // For now, this is a placeholder - actual calling would depend on
                // the module's exact function signature
                LOG_WARNING("WardenModule: PacketHandler execution stubbed");
                LOG_INFO("WardenModule:   Would call emulated function to process checks");
                LOG_INFO("WardenModule:   This would generate REAL responses (not fakes!)");

                // Clean up
                emulator_->freeMemory(checkDataAddr);
                emulator_->freeMemory(responseAddr);

                // For now, return false to use fake responses
                // Once we have a real module, we'd read the response from responseAddr
                return false;

            } catch (const std::exception& e) {
                LOG_ERROR("WardenModule: Exception during PacketHandler: ", e.what());
                emulator_->freeMemory(checkDataAddr);
                emulator_->freeMemory(responseAddr);
                return false;
            }
        }
    #endif

    LOG_WARNING("WardenModule: processCheckRequest NOT IMPLEMENTED");
    LOG_INFO("WardenModule:   Would call module->PacketHandler() here");

    // For now, return false to fall back to fake responses in GameHandler
    return false;
}

uint32_t WardenModule::tick([[maybe_unused]] uint32_t deltaMs) {
    if (!loaded_ || !funcList_.tick) {
        return 0; // No tick needed
    }

    // TODO: Call module's Tick function
    // return funcList_.tick(deltaMs);

    return 0;
}

void WardenModule::generateRC4Keys([[maybe_unused]] uint8_t* packet) {
    if (!loaded_ || !funcList_.generateRC4Keys) {
        return;
    }

    // TODO: Call module's GenerateRC4Keys function
    // This re-keys the Warden crypto stream
    // funcList_.generateRC4Keys(packet);
}

void WardenModule::unload() {
    if (moduleMemory_) {
        // Call module's Unload() function if loaded
        if (loaded_ && funcList_.unload) {
            LOG_INFO("WardenModule: Calling module unload callback...");
            // TODO: Implement callback when execution layer is complete
            // funcList_.unload(nullptr);
        }

        // Free executable memory region
        LOG_INFO("WardenModule: Freeing ", moduleSize_, " bytes of executable memory");
        #ifdef _WIN32
            VirtualFree(moduleMemory_, 0, MEM_RELEASE);
        #else
            munmap(moduleMemory_, moduleSize_);
        #endif

        moduleMemory_ = nullptr;
        moduleSize_ = 0;
    }

    // Clear function pointers
    funcList_ = {};

    loaded_ = false;
    moduleData_.clear();
    decryptedData_.clear();
    decompressedData_.clear();
}

// ============================================================================
// Private Validation Methods
// ============================================================================

bool WardenModule::verifyMD5(const std::vector<uint8_t>& data,
                             const std::vector<uint8_t>& expectedHash) {
    std::vector<uint8_t> computedHash = auth::Crypto::md5(data);

    if (computedHash.size() != expectedHash.size()) {
        return false;
    }

    return std::memcmp(computedHash.data(), expectedHash.data(), expectedHash.size()) == 0;
}

bool WardenModule::decryptRC4(const std::vector<uint8_t>& encrypted,
                              const std::vector<uint8_t>& key,
                              std::vector<uint8_t>& decryptedOut) {
    if (key.size() != 16) {
        LOG_ERROR("WardenModule: Invalid RC4 key size: ", key.size(), " (expected 16)");
        return false;
    }

    // Initialize RC4 state (KSA - Key Scheduling Algorithm)
    std::vector<uint8_t> S(256);
    for (int i = 0; i < 256; ++i) {
        S[i] = i;
    }

    int j = 0;
    for (int i = 0; i < 256; ++i) {
        j = (j + S[i] + key[i % key.size()]) % 256;
        std::swap(S[i], S[j]);
    }

    // Decrypt using RC4 (PRGA - Pseudo-Random Generation Algorithm)
    decryptedOut.resize(encrypted.size());
    int i = 0;
    j = 0;

    for (size_t idx = 0; idx < encrypted.size(); ++idx) {
        i = (i + 1) % 256;
        j = (j + S[i]) % 256;
        std::swap(S[i], S[j]);
        uint8_t K = S[(S[i] + S[j]) % 256];
        decryptedOut[idx] = encrypted[idx] ^ K;
    }

    return true;
}

bool WardenModule::verifyRSASignature(const std::vector<uint8_t>& data) {
    // RSA-2048 signature is last 256 bytes
    if (data.size() < 256) {
        LOG_ERROR("WardenModule: Data too small for RSA signature (need at least 256 bytes)");
        return false;
    }

    // Extract signature (last 256 bytes)
    std::vector<uint8_t> signature(data.end() - 256, data.end());

    // Extract data without signature
    std::vector<uint8_t> dataWithoutSig(data.begin(), data.end() - 256);

    // Hardcoded WoW 3.3.5a Warden RSA public key
    // Exponent: 0x010001 (65537)
    const uint32_t exponent = 0x010001;

    // Modulus (256 bytes) - Extracted from WoW 3.3.5a (build 12340) client
    // Extracted from Wow.exe at offset 0x005e3a03 (.rdata section)
    // This is the actual RSA-2048 public key modulus used by Warden
    const uint8_t modulus[256] = {
        0x51, 0xAD, 0x57, 0x75, 0x16, 0x92, 0x0A, 0x0E, 0xEB, 0xFA, 0xF8, 0x1B, 0x37, 0x49, 0x7C, 0xDD,
        0x47, 0xDA, 0x5E, 0x02, 0x8D, 0x96, 0x75, 0x21, 0x27, 0x59, 0x04, 0xAC, 0xB1, 0x0C, 0xB9, 0x23,
        0x05, 0xCC, 0x82, 0xB8, 0xBF, 0x04, 0x77, 0x62, 0x92, 0x01, 0x00, 0x01, 0x00, 0x77, 0x64, 0xF8,
        0x57, 0x1D, 0xFB, 0xB0, 0x09, 0xC4, 0xE6, 0x28, 0x91, 0x34, 0xE3, 0x55, 0x61, 0x15, 0x8A, 0xE9,
        0x07, 0xFC, 0xAA, 0x60, 0xB3, 0x82, 0xB7, 0xE2, 0xA4, 0x40, 0x15, 0x01, 0x3F, 0xC2, 0x36, 0xA8,
        0x9D, 0x95, 0xD0, 0x54, 0x69, 0xAA, 0xF5, 0xED, 0x5C, 0x7F, 0x21, 0xC5, 0x55, 0x95, 0x56, 0x5B,
        0x2F, 0xC6, 0xDD, 0x2C, 0xBD, 0x74, 0xA3, 0x5A, 0x0D, 0x70, 0x98, 0x9A, 0x01, 0x36, 0x51, 0x78,
        0x71, 0x9B, 0x8E, 0xCB, 0xB8, 0x84, 0x67, 0x30, 0xF4, 0x43, 0xB3, 0xA3, 0x50, 0xA3, 0xBA, 0xA4,
        0xF7, 0xB1, 0x94, 0xE5, 0x5B, 0x95, 0x8B, 0x1A, 0xE4, 0x04, 0x1D, 0xFB, 0xCF, 0x0E, 0xE6, 0x97,
        0x4C, 0xDC, 0xE4, 0x28, 0x7F, 0xB8, 0x58, 0x4A, 0x45, 0x1B, 0xC8, 0x8C, 0xD0, 0xFD, 0x2E, 0x77,
        0xC4, 0x30, 0xD8, 0x3D, 0xD2, 0xD5, 0xFA, 0xBA, 0x9D, 0x1E, 0x02, 0xF6, 0x7B, 0xBE, 0x08, 0x95,
        0xCB, 0xB0, 0x53, 0x3E, 0x1C, 0x41, 0x45, 0xFC, 0x27, 0x6F, 0x63, 0x6A, 0x73, 0x91, 0xA9, 0x42,
        0x00, 0x12, 0x93, 0xF8, 0x5B, 0x83, 0xED, 0x52, 0x77, 0x4E, 0x38, 0x08, 0x16, 0x23, 0x10, 0x85,
        0x4C, 0x0B, 0xA9, 0x8C, 0x9C, 0x40, 0x4C, 0xAF, 0x6E, 0xA7, 0x89, 0x02, 0xC5, 0x06, 0x96, 0x99,
        0x41, 0xD4, 0x31, 0x03, 0x4A, 0xA9, 0x2B, 0x17, 0x52, 0xDD, 0x5C, 0x4E, 0x5F, 0x16, 0xC3, 0x81,
        0x0F, 0x2E, 0xE2, 0x17, 0x45, 0x2B, 0x7B, 0x65, 0x7A, 0xA3, 0x18, 0x87, 0xC2, 0xB2, 0xF5, 0xCD
    };

    // Compute expected hash: SHA1(data_without_sig + "MAIEV.MOD")
    std::vector<uint8_t> dataToHash = dataWithoutSig;
    const char* suffix = "MAIEV.MOD";
    dataToHash.insert(dataToHash.end(), suffix, suffix + strlen(suffix));

    std::vector<uint8_t> expectedHash = auth::Crypto::sha1(dataToHash);

    // Create RSA public key using EVP_PKEY_fromdata (OpenSSL 3.0 compatible)
    BIGNUM* n = BN_bin2bn(modulus, 256, nullptr);
    BIGNUM* e = BN_new();
    BN_set_word(e, exponent);

    EVP_PKEY* pkey = nullptr;
    EVP_PKEY_CTX* ctx = nullptr;
    std::vector<uint8_t> decryptedSig(256);
    int decryptedLen = -1;

    {
        OSSL_PARAM_BLD* bld = OSSL_PARAM_BLD_new();
        OSSL_PARAM_BLD_push_BN(bld, OSSL_PKEY_PARAM_RSA_N, n);
        OSSL_PARAM_BLD_push_BN(bld, OSSL_PKEY_PARAM_RSA_E, e);
        OSSL_PARAM* params = OSSL_PARAM_BLD_to_param(bld);
        OSSL_PARAM_BLD_free(bld);

        EVP_PKEY_CTX* fromCtx = EVP_PKEY_CTX_new_from_name(nullptr, "RSA", nullptr);
        if (fromCtx && EVP_PKEY_fromdata_init(fromCtx) > 0) {
            EVP_PKEY_fromdata(fromCtx, &pkey, EVP_PKEY_PUBLIC_KEY, params);
        }
        if (fromCtx) EVP_PKEY_CTX_free(fromCtx);
        OSSL_PARAM_free(params);

        if (pkey) {
            ctx = EVP_PKEY_CTX_new(pkey, nullptr);
            if (ctx && EVP_PKEY_verify_recover_init(ctx) > 0 &&
                EVP_PKEY_CTX_set_rsa_padding(ctx, RSA_NO_PADDING) > 0) {
                size_t outLen = decryptedSig.size();
                if (EVP_PKEY_verify_recover(ctx, decryptedSig.data(), &outLen,
                                            signature.data(), 256) > 0) {
                    decryptedLen = static_cast<int>(outLen);
                }
            }
        }
    }

    BN_free(n);
    BN_free(e);
    if (ctx) EVP_PKEY_CTX_free(ctx);
    if (pkey) EVP_PKEY_free(pkey);

    if (decryptedLen < 0) {
        LOG_ERROR("WardenModule: RSA public decrypt failed");
        return false;
    }

    // Expected format: padding (0xBB bytes) + SHA1 hash (20 bytes)
    // Total: 256 bytes decrypted
    // Find SHA1 hash in decrypted signature (should be at end, preceded by 0xBB padding)

    // Look for SHA1 hash in last 20 bytes
    if (decryptedLen >= 20) {
        std::vector<uint8_t> actualHash(decryptedSig.end() - 20, decryptedSig.end());

        if (std::memcmp(actualHash.data(), expectedHash.data(), 20) == 0) {
            LOG_INFO("WardenModule: RSA signature verified");
            return true;
        }
    }

    LOG_ERROR("WardenModule: RSA signature verification FAILED (hash mismatch)");
    LOG_ERROR("WardenModule: NOTE: Using placeholder modulus - extract real modulus from WoW.exe for actual verification");

    // For development, return true to proceed (since we don't have real modulus)
    // TODO: Set to false once real modulus is extracted
    LOG_WARNING("WardenModule: Skipping RSA verification (placeholder modulus)");
    return true; // TEMPORARY - change to false for production
}

bool WardenModule::decompressZlib(const std::vector<uint8_t>& compressed,
                                  std::vector<uint8_t>& decompressedOut) {
    if (compressed.size() < 4) {
        LOG_ERROR("WardenModule: Compressed data too small (need at least 4 bytes for size header)");
        return false;
    }

    // Read 4-byte uncompressed size (little-endian)
    uint32_t uncompressedSize =
        compressed[0] |
        (compressed[1] << 8) |
        (compressed[2] << 16) |
        (compressed[3] << 24);

    LOG_INFO("WardenModule: Uncompressed size: ", uncompressedSize, " bytes");

    // Sanity check (modules shouldn't be larger than 10MB)
    if (uncompressedSize > 10 * 1024 * 1024) {
        LOG_ERROR("WardenModule: Uncompressed size suspiciously large: ", uncompressedSize, " bytes");
        return false;
    }

    // Allocate output buffer
    decompressedOut.resize(uncompressedSize);

    // Setup zlib stream
    z_stream stream = {};
    stream.next_in = const_cast<uint8_t*>(compressed.data() + 4); // Skip 4-byte size header
    stream.avail_in = compressed.size() - 4;
    stream.next_out = decompressedOut.data();
    stream.avail_out = uncompressedSize;

    // Initialize inflater
    int ret = inflateInit(&stream);
    if (ret != Z_OK) {
        LOG_ERROR("WardenModule: inflateInit failed: ", ret);
        return false;
    }

    // Decompress
    ret = inflate(&stream, Z_FINISH);

    // Cleanup
    inflateEnd(&stream);

    if (ret != Z_STREAM_END) {
        LOG_ERROR("WardenModule: inflate failed: ", ret);
        return false;
    }

    LOG_INFO("WardenModule: zlib decompression successful (", stream.total_out, " bytes decompressed)");

    return true;
}

bool WardenModule::parseExecutableFormat(const std::vector<uint8_t>& exeData) {
    if (exeData.size() < 4) {
        LOG_ERROR("WardenModule: Executable data too small for header");
        return false;
    }

    // Read final code size (little-endian 4 bytes)
    uint32_t finalCodeSize =
        exeData[0] |
        (exeData[1] << 8) |
        (exeData[2] << 16) |
        (exeData[3] << 24);

    LOG_INFO("WardenModule: Final code size: ", finalCodeSize, " bytes");

    // Sanity check (executable shouldn't be larger than 5MB)
    if (finalCodeSize > 5 * 1024 * 1024 || finalCodeSize == 0) {
        LOG_ERROR("WardenModule: Invalid final code size: ", finalCodeSize);
        return false;
    }

    // Allocate executable memory
    // Note: On Linux, we'll use mmap with PROT_EXEC
    // On Windows, would use VirtualAlloc with PAGE_EXECUTE_READWRITE
    #ifdef _WIN32
        moduleMemory_ = VirtualAlloc(
            nullptr,
            finalCodeSize,
            MEM_COMMIT | MEM_RESERVE,
            PAGE_EXECUTE_READWRITE
        );
        if (!moduleMemory_) {
            LOG_ERROR("WardenModule: VirtualAlloc failed");
            return false;
        }
    #else
        moduleMemory_ = mmap(
            nullptr,
            finalCodeSize,
            PROT_READ | PROT_WRITE | PROT_EXEC,
            MAP_PRIVATE | MAP_ANONYMOUS,
            -1,
            0
        );
        if (moduleMemory_ == MAP_FAILED) {
            LOG_ERROR("WardenModule: mmap failed: ", strerror(errno));
            moduleMemory_ = nullptr;
            return false;
        }
    #endif

    moduleSize_ = finalCodeSize;
    std::memset(moduleMemory_, 0, moduleSize_); // Zero-initialize

    LOG_INFO("WardenModule: Allocated ", moduleSize_, " bytes of executable memory");

    auto readU16LE = [&](size_t at) -> uint16_t {
        return static_cast<uint16_t>(exeData[at] | (exeData[at + 1] << 8));
    };

    enum class PairFormat {
        CopyDataSkip,  // [copy][data][skip]
        SkipCopyData,  // [skip][copy][data]
        CopySkipData   // [copy][skip][data]
    };

    auto tryParsePairs = [&](PairFormat format,
                             std::vector<uint8_t>& imageOut,
                             size_t& relocPosOut,
                             size_t& finalOffsetOut,
                             int& pairCountOut) -> bool {
        imageOut.assign(moduleSize_, 0);
        size_t pos = 4; // Skip 4-byte final size header
        size_t destOffset = 0;
        int pairCount = 0;

        while (pos + 2 <= exeData.size()) {
            uint16_t copyCount = 0;
            uint16_t skipCount = 0;

            switch (format) {
                case PairFormat::CopyDataSkip: {
                    copyCount = readU16LE(pos);
                    pos += 2;
                    if (copyCount == 0) {
                        relocPosOut = pos;
                        finalOffsetOut = destOffset;
                        pairCountOut = pairCount;
                        imageOut.resize(moduleSize_);
                        return true;
                    }

                    if (pos + copyCount > exeData.size() || destOffset + copyCount > moduleSize_) {
                        return false;
                    }

                    std::memcpy(imageOut.data() + destOffset, exeData.data() + pos, copyCount);
                    pos += copyCount;
                    destOffset += copyCount;

                    if (pos + 2 > exeData.size()) {
                        return false;
                    }
                    skipCount = readU16LE(pos);
                    pos += 2;
                    break;
                }

                case PairFormat::SkipCopyData: {
                    if (pos + 4 > exeData.size()) {
                        return false;
                    }
                    skipCount = readU16LE(pos);
                    pos += 2;
                    copyCount = readU16LE(pos);
                    pos += 2;

                    if (skipCount == 0 && copyCount == 0) {
                        relocPosOut = pos;
                        finalOffsetOut = destOffset;
                        pairCountOut = pairCount;
                        imageOut.resize(moduleSize_);
                        return true;
                    }

                    if (destOffset + skipCount > moduleSize_) {
                        return false;
                    }
                    destOffset += skipCount;

                    if (pos + copyCount > exeData.size() || destOffset + copyCount > moduleSize_) {
                        return false;
                    }
                    std::memcpy(imageOut.data() + destOffset, exeData.data() + pos, copyCount);
                    pos += copyCount;
                    destOffset += copyCount;
                    break;
                }

                case PairFormat::CopySkipData: {
                    if (pos + 4 > exeData.size()) {
                        return false;
                    }
                    copyCount = readU16LE(pos);
                    pos += 2;
                    skipCount = readU16LE(pos);
                    pos += 2;

                    if (copyCount == 0 && skipCount == 0) {
                        relocPosOut = pos;
                        finalOffsetOut = destOffset;
                        pairCountOut = pairCount;
                        imageOut.resize(moduleSize_);
                        return true;
                    }

                    if (pos + copyCount > exeData.size() || destOffset + copyCount > moduleSize_) {
                        return false;
                    }
                    std::memcpy(imageOut.data() + destOffset, exeData.data() + pos, copyCount);
                    pos += copyCount;
                    destOffset += copyCount;
                    break;
                }
            }

            if (destOffset + skipCount > moduleSize_) {
                return false;
            }
            destOffset += skipCount;
            pairCount++;
        }

        return false;
    };

    std::vector<uint8_t> parsedImage;
    size_t parsedRelocPos = 0;
    size_t parsedFinalOffset = 0;
    int parsedPairCount = 0;

    PairFormat usedFormat = PairFormat::CopyDataSkip;
    bool parsed = tryParsePairs(PairFormat::CopyDataSkip, parsedImage, parsedRelocPos, parsedFinalOffset, parsedPairCount);
    if (!parsed) {
        usedFormat = PairFormat::SkipCopyData;
        parsed = tryParsePairs(PairFormat::SkipCopyData, parsedImage, parsedRelocPos, parsedFinalOffset, parsedPairCount);
    }
    if (!parsed) {
        usedFormat = PairFormat::CopySkipData;
        parsed = tryParsePairs(PairFormat::CopySkipData, parsedImage, parsedRelocPos, parsedFinalOffset, parsedPairCount);
    }

    if (parsed) {
        std::memcpy(moduleMemory_, parsedImage.data(), parsedImage.size());
        relocDataOffset_ = parsedRelocPos;

        const char* formatName = "copy/data/skip";
        if (usedFormat == PairFormat::SkipCopyData) formatName = "skip/copy/data";
        if (usedFormat == PairFormat::CopySkipData) formatName = "copy/skip/data";

        LOG_INFO("WardenModule: Parsed ", parsedPairCount, " pairs using format ",
                 formatName, ", final offset: ", parsedFinalOffset, "/", finalCodeSize);
        LOG_INFO("WardenModule: Relocation data starts at decompressed offset ", relocDataOffset_,
                 " (", (exeData.size() - relocDataOffset_), " bytes remaining)");
        return true;
    }

    // Fallback: copy raw payload (without the 4-byte size header) into module memory.
    // This keeps loading alive for servers where packet flow can continue with hash/check fallbacks.
    if (exeData.size() > 4) {
        size_t rawCopySize = std::min(moduleSize_, exeData.size() - 4);
        std::memcpy(moduleMemory_, exeData.data() + 4, rawCopySize);
    }
    relocDataOffset_ = 0;
    LOG_ERROR("WardenModule: Could not parse copy/skip pairs (all known layouts failed); using raw payload fallback");
    return true;
}

bool WardenModule::applyRelocations() {
    if (!moduleMemory_ || moduleSize_ == 0) {
        LOG_ERROR("WardenModule: No module memory allocated for relocations");
        return false;
    }

    // Relocation data is in decompressedData_ starting at relocDataOffset_
    // Format: delta-encoded 2-byte LE offsets, terminated by 0x0000
    // Each offset in the module image has moduleBase_ added to the 32-bit value there

    if (relocDataOffset_ == 0 || relocDataOffset_ >= decompressedData_.size()) {
        LOG_INFO("WardenModule: No relocation data available");
        return true;
    }

    size_t relocPos = relocDataOffset_;
    uint32_t currentOffset = 0;
    int relocCount = 0;

    while (relocPos + 2 <= decompressedData_.size()) {
        uint16_t delta = decompressedData_[relocPos] | (decompressedData_[relocPos + 1] << 8);
        relocPos += 2;

        if (delta == 0) break;

        currentOffset += delta;

        if (currentOffset + 4 <= moduleSize_) {
            uint8_t* addr = static_cast<uint8_t*>(moduleMemory_) + currentOffset;
            uint32_t val;
            std::memcpy(&val, addr, sizeof(uint32_t));
            val += moduleBase_;
            std::memcpy(addr, &val, sizeof(uint32_t));
            relocCount++;
        } else {
            LOG_ERROR("WardenModule: Relocation offset ", currentOffset,
                      " out of bounds (moduleSize=", moduleSize_, ")");
        }
    }

    {
        char baseBuf[32];
        std::snprintf(baseBuf, sizeof(baseBuf), "0x%X", moduleBase_);
        LOG_INFO("WardenModule: Applied ", relocCount, " relocations (base=", baseBuf, ")");
    }

    return true;
}

bool WardenModule::bindAPIs() {
    if (!moduleMemory_ || moduleSize_ == 0) {
        LOG_ERROR("WardenModule: No module memory allocated for API binding");
        return false;
    }

    LOG_INFO("WardenModule: Binding Windows APIs for module...");

    // Common Windows APIs used by Warden modules:
    //
    // kernel32.dll:
    // - VirtualAlloc, VirtualFree, VirtualProtect
    // - GetTickCount, GetCurrentThreadId, GetCurrentProcessId
    // - Sleep, SwitchToThread
    // - CreateThread, ExitThread
    // - GetModuleHandleA, GetProcAddress
    // - ReadProcessMemory, WriteProcessMemory
    //
    // user32.dll:
    // - GetForegroundWindow, GetWindowTextA
    //
    // ntdll.dll:
    // - NtQueryInformationProcess, NtQuerySystemInformation

    #ifdef _WIN32
        // On Windows: Use GetProcAddress to resolve imports
        LOG_INFO("WardenModule: Platform: Windows - using GetProcAddress");

        HMODULE kernel32 = GetModuleHandleA("kernel32.dll");
        HMODULE user32 = GetModuleHandleA("user32.dll");
        HMODULE ntdll = GetModuleHandleA("ntdll.dll");

        if (!kernel32 || !user32 || !ntdll) {
            LOG_ERROR("WardenModule: Failed to get module handles");
            return false;
        }

        // TODO: Parse module's import table
        // - Find import directory in PE headers
        // - For each imported DLL:
        //   - For each imported function:
        //     - Resolve address using GetProcAddress
        //     - Write address to Import Address Table (IAT)

        LOG_WARNING("WardenModule: Windows API binding is STUB (needs PE import table parsing)");
        LOG_INFO("WardenModule:   Would parse PE headers and patch IAT with resolved addresses");

    #else
        // On Linux: Cannot directly execute Windows code
        // Options:
        // 1. Use Wine to provide Windows API compatibility
        // 2. Implement Windows API stubs (limited functionality)
        // 3. Use binfmt_misc + Wine (transparent Windows executable support)

        LOG_WARNING("WardenModule: Platform: Linux - Windows module execution NOT supported");
        LOG_INFO("WardenModule: Options:");
        LOG_INFO("WardenModule:   1. Run wowee under Wine (provides Windows API layer)");
        LOG_INFO("WardenModule:   2. Use a Windows VM");
        LOG_INFO("WardenModule:   3. Implement Windows API stubs (limited, complex)");

        // For now, we'll return true to continue the loading pipeline
        // Real execution would fail, but this allows testing the infrastructure
        LOG_WARNING("WardenModule: Skipping API binding (Linux platform limitation)");
    #endif

    return true; // Return true to continue (stub implementation)
}

bool WardenModule::initializeModule() {
    if (!moduleMemory_ || moduleSize_ == 0) {
        LOG_ERROR("WardenModule: No module memory allocated for initialization");
        return false;
    }

    LOG_INFO("WardenModule: Initializing Warden module...");

    // Module initialization protocol:
    //
    // 1. Client provides structure with 7 callback pointers:
    //    - void (*sendPacket)(uint8_t* data, size_t len)
    //    - void (*validateModule)(uint8_t* hash)
    //    - void* (*allocMemory)(size_t size)
    //    - void (*freeMemory)(void* ptr)
    //    - void (*generateRC4)(uint8_t* seed)
    //    - uint32_t (*getTime)()
    //    - void (*logMessage)(const char* msg)
    //
    // 2. Call module entry point with callback structure
    //
    // 3. Module returns WardenFuncList with 4 exported functions:
    //    - generateRC4Keys(packet)
    //    - unload(rc4Keys)
    //    - packetHandler(data)
    //    - tick(deltaMs)

    // Define callback structure (what we provide to module)
    struct ClientCallbacks {
        void (*sendPacket)(uint8_t* data, size_t len);
        void (*validateModule)(uint8_t* hash);
        void* (*allocMemory)(size_t size);
        void (*freeMemory)(void* ptr);
        void (*generateRC4)(uint8_t* seed);
        uint32_t (*getTime)();
        void (*logMessage)(const char* msg);
    };

    // Setup client callbacks (used when calling module entry point below)
    [[maybe_unused]] ClientCallbacks callbacks = {};

    // Stub callbacks (would need real implementations)
    callbacks.sendPacket = []([[maybe_unused]] uint8_t* data, size_t len) {
        LOG_DEBUG("WardenModule Callback: sendPacket(", len, " bytes)");
        // TODO: Send CMSG_WARDEN_DATA packet
    };

    callbacks.validateModule = []([[maybe_unused]] uint8_t* hash) {
        LOG_DEBUG("WardenModule Callback: validateModule()");
        // TODO: Validate module hash
    };

    callbacks.allocMemory = [](size_t size) -> void* {
        LOG_DEBUG("WardenModule Callback: allocMemory(", size, ")");
        return malloc(size);
    };

    callbacks.freeMemory = [](void* ptr) {
        LOG_DEBUG("WardenModule Callback: freeMemory()");
        free(ptr);
    };

    callbacks.generateRC4 = []([[maybe_unused]] uint8_t* seed) {
        LOG_DEBUG("WardenModule Callback: generateRC4()");
        // TODO: Re-key RC4 cipher
    };

    callbacks.getTime = []() -> uint32_t {
        return static_cast<uint32_t>(time(nullptr));
    };

    callbacks.logMessage = [](const char* msg) {
        LOG_INFO("WardenModule Log: ", msg);
    };

    // Module entry point is typically at offset 0 (first bytes of loaded code)
    // Function signature: WardenFuncList* (*entryPoint)(ClientCallbacks*)

    #ifdef HAVE_UNICORN
        // Use Unicorn emulator for cross-platform execution
        LOG_INFO("WardenModule: Initializing Unicorn emulator...");

        emulator_ = std::make_unique<WardenEmulator>();
        if (!emulator_->initialize(moduleMemory_, moduleSize_, moduleBase_)) {
            LOG_ERROR("WardenModule: Failed to initialize emulator");
            return false;
        }

        // Setup Windows API hooks
        emulator_->setupCommonAPIHooks();

        {
            char addrBuf[32];
            std::snprintf(addrBuf, sizeof(addrBuf), "0x%X", moduleBase_);
            LOG_INFO("WardenModule: Emulator initialized successfully");
            LOG_INFO("WardenModule:   Ready to execute module at ", addrBuf);
        }

        // Allocate memory for ClientCallbacks structure in emulated space
        uint32_t callbackStructAddr = emulator_->allocateMemory(sizeof(ClientCallbacks), 0x04);
        if (callbackStructAddr == 0) {
            LOG_ERROR("WardenModule: Failed to allocate memory for callbacks");
            return false;
        }

        // Write callback function pointers to emulated memory
        // Note: These would be addresses of stub functions in emulated space
        // For now, we'll write placeholder addresses
        std::vector<uint32_t> callbackAddrs = {
            0x70001000, // sendPacket
            0x70001100, // validateModule
            0x70001200, // allocMemory
            0x70001300, // freeMemory
            0x70001400, // generateRC4
            0x70001500, // getTime
            0x70001600  // logMessage
        };

        // Write callback struct (7 function pointers = 28 bytes)
        for (size_t i = 0; i < callbackAddrs.size(); ++i) {
            uint32_t addr = callbackAddrs[i];
            emulator_->writeMemory(callbackStructAddr + (i * 4), &addr, 4);
        }

        {
            char cbBuf[32];
            std::snprintf(cbBuf, sizeof(cbBuf), "0x%X", callbackStructAddr);
            LOG_INFO("WardenModule: Prepared ClientCallbacks at ", cbBuf);
        }

        // Call module entry point
        // Entry point is typically at module base (offset 0)
        uint32_t entryPoint = moduleBase_;

        {
            char epBuf[32];
            std::snprintf(epBuf, sizeof(epBuf), "0x%X", entryPoint);
            LOG_INFO("WardenModule: Calling module entry point at ", epBuf);
        }

        try {
            // Call: WardenFuncList* InitModule(ClientCallbacks* callbacks)
            std::vector<uint32_t> args = { callbackStructAddr };
            uint32_t result = emulator_->callFunction(entryPoint, args);

            if (result == 0) {
                LOG_ERROR("WardenModule: Module entry returned NULL");
                return false;
            }

            {
                char resBuf[32];
                std::snprintf(resBuf, sizeof(resBuf), "0x%X", result);
                LOG_INFO("WardenModule: Module initialized, WardenFuncList at ", resBuf);
            }

            // Read WardenFuncList structure from emulated memory
            // Structure has 4 function pointers (16 bytes)
            uint32_t funcAddrs[4] = {};
            if (emulator_->readMemory(result, funcAddrs, 16)) {
                char fb[4][32];
                for (int fi = 0; fi < 4; ++fi)
                    std::snprintf(fb[fi], sizeof(fb[fi]), "0x%X", funcAddrs[fi]);
                LOG_INFO("WardenModule: Module exported functions:");
                LOG_INFO("WardenModule:   generateRC4Keys: ", fb[0]);
                LOG_INFO("WardenModule:   unload:          ", fb[1]);
                LOG_INFO("WardenModule:   packetHandler:   ", fb[2]);
                LOG_INFO("WardenModule:   tick:            ", fb[3]);

                // Store function addresses for later use
                // funcList_.generateRC4Keys = ... (would wrap emulator calls)
                // funcList_.unload = ...
                // funcList_.packetHandler = ...
                // funcList_.tick = ...
            }

            LOG_INFO("WardenModule: Module fully initialized and ready!");

        } catch (const std::exception& e) {
            LOG_ERROR("WardenModule: Exception during module initialization: ", e.what());
            return false;
        }

    #elif defined(_WIN32)
        // Native Windows execution (dangerous without sandboxing)
        typedef void* (*ModuleEntryPoint)(ClientCallbacks*);
        ModuleEntryPoint entryPoint = reinterpret_cast<ModuleEntryPoint>(moduleMemory_);

        LOG_INFO("WardenModule: Calling module entry point at ", moduleMemory_);

        // NOTE: This would execute native x86 code
        // Extremely dangerous without proper validation!
        // void* result = entryPoint(&callbacks);

        LOG_WARNING("WardenModule: Module entry point call is DISABLED (unsafe without validation)");
        LOG_INFO("WardenModule:   Would execute x86 code at ", moduleMemory_);

        // TODO: Extract WardenFuncList from result
        // funcList_.packetHandler = ...
        // funcList_.tick = ...
        // funcList_.generateRC4Keys = ...
        // funcList_.unload = ...

    #else
        LOG_WARNING("WardenModule: Cannot execute Windows x86 code on Linux");
        LOG_INFO("WardenModule:   Module entry point: ", moduleMemory_);
        LOG_INFO("WardenModule:   Would call entry point with ClientCallbacks struct");
    #endif

    // For now, return true to mark module as "loaded" at infrastructure level
    // Real execution would require:
    // 1. Proper PE parsing to find actual entry point
    // 2. Calling convention (stdcall/cdecl) handling
    // 3. Exception handling for crashes
    // 4. Sandboxing for security

    LOG_WARNING("WardenModule: Module initialization is STUB");
    return true; // Stub implementation
}

// ============================================================================
// WardenModuleManager Implementation
// ============================================================================

WardenModuleManager::WardenModuleManager() {
    // Set cache directory
#ifdef _WIN32
    if (const char* appdata = std::getenv("APPDATA"))
        cacheDirectory_ = std::string(appdata) + "\\wowee\\warden_cache";
    else
        cacheDirectory_ = ".\\warden_cache";
#else
    if (const char* home = std::getenv("HOME"))
        cacheDirectory_ = std::string(home) + "/.local/share/wowee/warden_cache";
    else
        cacheDirectory_ = "./warden_cache";
#endif

    // Create cache directory if it doesn't exist
    std::filesystem::create_directories(cacheDirectory_);

    LOG_INFO("WardenModuleManager: Cache directory: ", cacheDirectory_);
}

WardenModuleManager::~WardenModuleManager() {
    modules_.clear();
}

bool WardenModuleManager::hasModule(const std::vector<uint8_t>& md5Hash) {
    // Check in-memory cache
    if (modules_.find(md5Hash) != modules_.end()) {
        return modules_[md5Hash]->isLoaded();
    }

    // Check disk cache
    std::vector<uint8_t> dummy;
    return loadCachedModule(md5Hash, dummy);
}

std::shared_ptr<WardenModule> WardenModuleManager::getModule(const std::vector<uint8_t>& md5Hash) {
    auto it = modules_.find(md5Hash);
    if (it != modules_.end()) {
        return it->second;
    }

    // Create new module instance
    auto module = std::make_shared<WardenModule>();
    modules_[md5Hash] = module;
    return module;
}

bool WardenModuleManager::receiveModuleChunk(const std::vector<uint8_t>& md5Hash,
                                             const std::vector<uint8_t>& chunkData,
                                             bool isComplete) {
    // Append to download buffer
    std::vector<uint8_t>& buffer = downloadBuffer_[md5Hash];
    buffer.insert(buffer.end(), chunkData.begin(), chunkData.end());

    LOG_INFO("WardenModuleManager: Received chunk (", chunkData.size(),
             " bytes, total: ", buffer.size(), ")");

    if (isComplete) {
        LOG_INFO("WardenModuleManager: Module download complete (", buffer.size(), " bytes)");

        // Cache to disk
        cacheModule(md5Hash, buffer);

        // Clear download buffer
        downloadBuffer_.erase(md5Hash);

        return true;
    }

    return true;
}

bool WardenModuleManager::cacheModule(const std::vector<uint8_t>& md5Hash,
                                      const std::vector<uint8_t>& moduleData) {
    std::string cachePath = getCachePath(md5Hash);

    std::ofstream file(cachePath, std::ios::binary);
    if (!file) {
        LOG_ERROR("WardenModuleManager: Failed to write cache: ", cachePath);
        return false;
    }

    file.write(reinterpret_cast<const char*>(moduleData.data()), moduleData.size());
    file.close();

    LOG_INFO("WardenModuleManager: Cached module to: ", cachePath);
    return true;
}

bool WardenModuleManager::loadCachedModule(const std::vector<uint8_t>& md5Hash,
                                           std::vector<uint8_t>& moduleDataOut) {
    std::string cachePath = getCachePath(md5Hash);

    if (!std::filesystem::exists(cachePath)) {
        return false;
    }

    std::ifstream file(cachePath, std::ios::binary | std::ios::ate);
    if (!file) {
        return false;
    }

    size_t fileSize = file.tellg();
    file.seekg(0, std::ios::beg);

    moduleDataOut.resize(fileSize);
    file.read(reinterpret_cast<char*>(moduleDataOut.data()), fileSize);
    file.close();

    LOG_INFO("WardenModuleManager: Loaded cached module (", fileSize, " bytes)");
    return true;
}

std::string WardenModuleManager::getCachePath(const std::vector<uint8_t>& md5Hash) {
    // Convert MD5 hash to hex string for filename
    std::string hexHash;
    hexHash.reserve(md5Hash.size() * 2);

    for (uint8_t byte : md5Hash) {
        char buf[3];
        snprintf(buf, sizeof(buf), "%02x", byte);
        hexHash += buf;
    }

    return cacheDirectory_ + "/" + hexHash + ".wdn";
}

} // namespace game
} // namespace wowee