The Golang4U LLM Bible: Internals of the Intelligent Llama Manager

Preface: Why Native llama.cpp is Painful

Running Large Language Models (LLMs) locally using llama.cpp offers incredible privacy and performance. However, deploying it effectively is notoriously difficult. The native experience is raw: compiling binaries manually, dealing with unoptimized package manager downloads, juggling 50GB file transfers, calculating mental VRAM boundaries, and memorizing arcane CLI flags. Get a single context metric wrong, and your GPU violently crashes with an Out-Of-Memory (OOM) panic.

We built an all-in-one Go application—llama_manager.go—to solve this entirely. It is a highly fault-tolerant daemon that dynamically profiles physical hardware, reverse-engineers a model’s memory footprint on the fly, and calculates the exact parameters needed to run engines at absolute peak efficiency.

Here is the architectural Masterclass on the 1000+ lines of Go logic driving the system.

A backend engine needs an impenetrable foreground. llama_manager.go provides a seamless, dynamic, terminal UI using purely native OS calls and ANSI escapes.

1.1 Color-Coded Terminal UI

We use ANSI escape codes (\033[...]) to render clean, semantic output uniformly across Windows, macOS, and Linux.

1
2
3
4
5
6
7
8
const (
	colorReset  = "\033[0m"
	colorCyan   = "\033[36m"
	colorGreen  = "\033[32m"
	colorYellow = "\033[33m"
	colorRed    = "\033[31m"
	appDirName  = ".llama_manager"
)

1.2 Interactive Menus & Graceful Screen Clearing

We use bufio.NewReader(os.Stdin) for blocking terminal execution. This strictly enforces navigation, halting the UI at critical junctions. We pair this with graceful OS-specific terminal clearing to eliminate output clutter:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
func clearScreen() {
	cmd := exec.Command("clear")
	if runtime.GOOS == "windows" {
		cmd = exec.Command("cmd", "/c", "cls")
	}
	cmd.Stdout = os.Stdout
	cmd.Run()
}

func pause() {
	fmt.Print(colorYellow + "\nPress Enter to continue..." + colorReset)
	bufio.NewReader(os.Stdin).ReadBytes('\n')
}

func promptInput(prompt string) string {
	fmt.Print(prompt)
	reader := bufio.NewReader(os.Stdin)
	text, _ := reader.ReadString('\n')
	return strings.TrimSpace(text)
}

1.3 Live Update Notifier

When the script launches, an isolated Goroutine silently checks the official GitHub API without blocking the host menu. If an update exists, it visually injects a [Update Available] yellow badge on the main loop.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
var latestGitHubTag = ""

func fetchLatestReleaseAsync() {
	req, _ := http.NewRequest("GET", "https://api.github.com/repos/ggml-org/llama.cpp/releases/latest", nil)
	client := &http.Client{Timeout: 5 * time.Second}
	resp, err := client.Do(req)
	if err == nil {
		defer resp.Body.Close()
		var release struct { TagName string `json:"tag_name"` }
		if json.NewDecoder(resp.Body).Decode(&release) == nil {
			latestGitHubTag = release.TagName
		}
	}
}

Chapter 2: Hardware Profiling & Compiling

Packages like Homebrew and Winget ship slow, CPU-bound builds. To achieve peak efficiency, we aggressively profile the hardware and fetch the exact required variant.

2.1 Deep Hardware Detection

We detect CPU ecosystems and verify GPU endpoints recursively to bind to the optimal backend:

1
2
3
4
5
6
func detectHardware() (backend string, name string) {
	if runtime.GOOS == "darwin" { return "metal", "Apple Silicon/Metal" }
	if err := exec.Command("nvidia-smi").Run(); err == nil { return "cuda", "NVIDIA GPU (CUDA)" }
	if err := exec.Command("rocm-smi").Run(); err == nil { return "rocm", "AMD GPU (ROCm)" }
	return "vulkan", "Standard/APU (Vulkan)"
}

2.2 Suboptimal Install Detection

If the script notices a package manager binary in the host’s .PATH, it intercepts and violently stops using it. It advises the user to strip it using native uninstallation tools, warning that it compromises inference latency:

1
2
3
4
5
6
func checkSystemInstall(exePath, hwName, targetBackend string) {
    // A snippet showing how we handle bad PATH binaries
    if strings.Contains(strings.ToLower(exePath), "homebrew") {
        fmt.Println("To remove it, open your terminal and run:  " + colorRed + "brew uninstall llama.cpp" + colorReset)
    }
}

2.3 Automated Source Compilation & Dependency Advising

If a pre-built binary tailored for the system target is wholly missing from Hugging Face’s API wrapper, we perform automated source compilation (using cmake). If that fails because cmake or git aren’t installed, the daemon intervenes with literal shell copy-paste install commands specific to the OS flavor (dnf, apt, pacman, winget).

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
func adviseDependencies(missingTool string) {
	fmt.Println(colorRed + "\n[Error] Missing Compilation Dependency: " + missingTool + colorReset)
	switch runtime.GOOS {
	case "windows":
		if missingTool == "git" { fmt.Println("  winget install --id Git.Git -e --source winget") }
	case "darwin":
		fmt.Println("  brew install " + missingTool)
	case "linux":
		fmt.Println("  Debian/Ubuntu:  sudo apt update && sudo apt install " + missingTool + " build-essential")
	}
}

Chapter 3: Fault-Tolerant Model Engine (The Network)

Large GGUF files (7B to 70B parameter space) measure between 5 to 50 gigabytes. Standard HTTP streams will inevitably fail over unstable home internet lines.

3.1 Fetching from the Hugging Face API

A user simply types “qwen 7b q4”. We seamlessly query the Hugging Face HTTP endpoints, restrict to GGUF mappings, sort by historic downloads (indicating trust), and build an interactive file explorer.

3.2 The Stream Hijacker (Live Progress Bar)

A custom io.Writer interface violently hijacks the data stream output to provide live feedback by rewriting the active console line via carriage returns (\r\033[K) without gobbling RAM allocations.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
type progressWriter struct {
	io.Writer
	TotalDownloaded int64
	TotalSize       int64
	LastPercent     int64
}

func (pw *progressWriter) Write(p []byte) (int, error) {
	n, err := pw.Writer.Write(p)
	pw.TotalDownloaded += int64(n)
	if pw.TotalSize > 0 {
		percent := (pw.TotalDownloaded * 100) / pw.TotalSize
		if percent > pw.LastPercent {
			fmt.Printf("\r\033[KDownloading... %d%% (%d MB / %d MB) - Press Ctrl+C to Pause", percent, pw.TotalDownloaded/(1024*1024), pw.TotalSize/(1024*1024))
			pw.LastPercent = percent
		}
	}
	return n, err
}

3.3 HTTP 206 Resumable Block Chunking

If an interruption occurs, we write the data to a .part chunk. Upon restarting the request, we modify the TCP/IP Headers (HTTP Range) to bypass the previously collected bits and stream directly into os.O_APPEND.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
func downloadWithResume(ctx context.Context, dlURL, finalDest string) error {
	partDest := finalDest + ".part"
	var startBytes int64 = 0

	if info, err := os.Stat(partDest); err == nil {
		startBytes = info.Size() // Existing chunk!
	}
	req, _ := http.NewRequestWithContext(ctx, "GET", dlURL, nil)
	if startBytes > 0 {
		req.Header.Set("Range", fmt.Sprintf("bytes=%d-", startBytes)) // Resume!
	}
	
	resp, _ := dlClient.Do(req)
	
	flags := os.O_CREATE | os.O_WRONLY
	if resp.StatusCode == http.StatusPartialContent {
		flags |= os.O_APPEND // Append if Supported
	} else {
		flags |= os.O_TRUNC // Truncate and Restart 
	}
	
	outFile, _ := os.OpenFile(partDest, flags, 0644)
    // ... Copy stream via pw
}

Chapter 4: Intelligent Memory Mathematics

The core of our platform reverse engineers neural footprints to mathematically prevent systemic memory failures. If the Model Weights + the Conversational Application (Key-Value) Cache breaches VRAM limits, the OS forces panic.

4.1 Native VRAM Hardware Scanning

Extracting active physical capacity bypasses generic runtime estimates:

1
2
3
4
5
6
7
8
9
func getMemoryCapacity() (float64, string) {
	out, err := exec.Command("nvidia-smi", "--query-gpu=memory.total", "--format=csv,noheader,nounits").Output()
	if err == nil {
		lines := strings.Split(strings.TrimSpace(string(out)), "\n")
		vramMB, _ := strconv.ParseFloat(strings.TrimSpace(lines[0]), 64)
		if vramMB > 0 { return vramMB / 1024.0, "Dedicated NVIDIA VRAM" }
	}
    // Check AMD / macOS sysctl / system powershell fallbacks ...
}

4.2 Reverse-Engineer Neural Parameter Footprints

We inspect file names and sizes locally. For a Q4_K model, we assign a quantization density limit of 4.5 bits. Then we triangulate total parameter mass mathematically without reading external index files:

1
2
Bits_Per_Weight ≈ 4.5
Parameter_Count = (Gigabytes * 8.0) / Bits_Per_Weight

4.3 Safe Context Limit Constraints

Modern models cost ~15 Megabytes per 1 Billion internal parameters per every 1,024 context tokens. We strip OS Overhead (4.0GB Windows / 3.0GB Linux) proactively from the physical constraints. The derivative formula provides boundary parameters that completely guarantee safe limits:

1
2
Available_VRAM = Hardware_Ram - Weight_GB - Base_OS_Overhead
Max_Context = (Available_VRAM_in_MB / Model_Cost_Per_1024_Token) * 1024

Chapter 5: Deployment & Daemon Automation Pipeline

When firing the engine daemon, we construct massive OS execution command arrays containing specific tuning metrics natively hidden to typical users.

5.1 KV Cache Formating & Sub-Slot Slicing

KV Cache Quantization (-ctk, -ctv): We format memory blocks optionally using 4-Bit arrays (q4_0), squeezing cache volumes down by over 75% for massive prompt windowing.
Anti-Hyperthreading Optimization (-t): By halving CPU core threads mathematically (Total Cores / 2), we force llama.cpp to bind exclusively to distinct physical logic gates, bypassing hyper-threaded cache thrashing entirely.
Context Slicing Prevention (-np 1): Default daemon topologies eagerly aggressively chop memory context slots for multiple parallel tenant users. We enforce single tenant (-kvu) locks.
Batching & Defragmentation: llama.cpp evaluates tokens optimally at 512 chunks. The user can ramp batch (-b, -ub) sizing up to 2048 to significantly reduce Initial Time-to-First-Token processing. We inject --defrag-thold 0.1 to heal degraded memory fragmentation pools automatically.

5.2 Network & Security Layer

LLM daemon servers blindly bind to localhosts natively. We optionally mesh nodes using Tailscale for secure global execution:

1
2
3
4
5
6
7
func getTailscaleIP() string {
	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Second)
	defer cancel()
	out, err := exec.CommandContext(ctx, "tailscale", "ip", "-4").Output()
	if err == nil { return strings.TrimSpace(string(out)) }
	return ""
}

If exposed to WAN/LAN bounds (0.0.0.0), a purely native hex key payload guarantees a zero-trust model API layer instantly:

1
2
3
4
5
func generateAPIKey() string {
	bytes := make([]byte, 16)
	rand.Read(bytes)
	return hex.EncodeToString(bytes)
}

5.3 Bulletproof Daemons: Ghost Process Auto-Healing

If the daemon was pushed physically out of bounds previously, a stale .pid file lock usually destroys the startup flow. We heal the app state dynamically:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
func isProcessAlive(pid int) bool {
	if runtime.GOOS == "windows" {
		cmd := exec.Command("tasklist", "/FI", fmt.Sprintf("PID eq %d", pid))
		out, _ := cmd.Output()
		return strings.Contains(string(out), strconv.Itoa(pid))
	}
	process, err := os.FindProcess(pid)
	if err != nil { return false }
	return process.Signal(syscall.Signal(0)) == nil // Ping kernel payload zero check
}

5.4 Graceful Zero-Leak VRAM Signal Channels

If a user violently slams Ctrl+C into the foreground terminal, the Go layer will inherently drop execution silently. However, C++ binaries hold GPU architectures captive. A dropped pointer implies absolute VRAM lock, requiring PC reboots. Instead, we orchestrate multiplexed Channel signals:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
sigChan := make(chan os.Signal, 1)
signal.Notify(sigChan, os.Interrupt, syscall.SIGTERM) // Trap native OS interrupt

go func() {
	<-sigChan // Lock Goroutine on keyboard event
	fmt.Println("\nShutting down cleanly, freeing VRAM...")
	if cmd.Process != nil {
		cmd.Process.Signal(os.Interrupt)  // Flush Native process first!
	}
}()

cmd.Run() // Wait for subprocess death to gracefully exit Go
signal.Stop(sigChan)

The Code

   1
   2
   3
   4
   5
   6
   7
   8
   9
  10
  11
  12
  13
  14
  15
  16
  17
  18
  19
  20
  21
  22
  23
  24
  25
  26
  27
  28
  29
  30
  31
  32
  33
  34
  35
  36
  37
  38
  39
  40
  41
  42
  43
  44
  45
  46
  47
  48
  49
  50
  51
  52
  53
  54
  55
  56
  57
  58
  59
  60
  61
  62
  63
  64
  65
  66
  67
  68
  69
  70
  71
  72
  73
  74
  75
  76
  77
  78
  79
  80
  81
  82
  83
  84
  85
  86
  87
  88
  89
  90
  91
  92
  93
  94
  95
  96
  97
  98
  99
 100
 101
 102
 103
 104
 105
 106
 107
 108
 109
 110
 111
 112
 113
 114
 115
 116
 117
 118
 119
 120
 121
 122
 123
 124
 125
 126
 127
 128
 129
 130
 131
 132
 133
 134
 135
 136
 137
 138
 139
 140
 141
 142
 143
 144
 145
 146
 147
 148
 149
 150
 151
 152
 153
 154
 155
 156
 157
 158
 159
 160
 161
 162
 163
 164
 165
 166
 167
 168
 169
 170
 171
 172
 173
 174
 175
 176
 177
 178
 179
 180
 181
 182
 183
 184
 185
 186
 187
 188
 189
 190
 191
 192
 193
 194
 195
 196
 197
 198
 199
 200
 201
 202
 203
 204
 205
 206
 207
 208
 209
 210
 211
 212
 213
 214
 215
 216
 217
 218
 219
 220
 221
 222
 223
 224
 225
 226
 227
 228
 229
 230
 231
 232
 233
 234
 235
 236
 237
 238
 239
 240
 241
 242
 243
 244
 245
 246
 247
 248
 249
 250
 251
 252
 253
 254
 255
 256
 257
 258
 259
 260
 261
 262
 263
 264
 265
 266
 267
 268
 269
 270
 271
 272
 273
 274
 275
 276
 277
 278
 279
 280
 281
 282
 283
 284
 285
 286
 287
 288
 289
 290
 291
 292
 293
 294
 295
 296
 297
 298
 299
 300
 301
 302
 303
 304
 305
 306
 307
 308
 309
 310
 311
 312
 313
 314
 315
 316
 317
 318
 319
 320
 321
 322
 323
 324
 325
 326
 327
 328
 329
 330
 331
 332
 333
 334
 335
 336
 337
 338
 339
 340
 341
 342
 343
 344
 345
 346
 347
 348
 349
 350
 351
 352
 353
 354
 355
 356
 357
 358
 359
 360
 361
 362
 363
 364
 365
 366
 367
 368
 369
 370
 371
 372
 373
 374
 375
 376
 377
 378
 379
 380
 381
 382
 383
 384
 385
 386
 387
 388
 389
 390
 391
 392
 393
 394
 395
 396
 397
 398
 399
 400
 401
 402
 403
 404
 405
 406
 407
 408
 409
 410
 411
 412
 413
 414
 415
 416
 417
 418
 419
 420
 421
 422
 423
 424
 425
 426
 427
 428
 429
 430
 431
 432
 433
 434
 435
 436
 437
 438
 439
 440
 441
 442
 443
 444
 445
 446
 447
 448
 449
 450
 451
 452
 453
 454
 455
 456
 457
 458
 459
 460
 461
 462
 463
 464
 465
 466
 467
 468
 469
 470
 471
 472
 473
 474
 475
 476
 477
 478
 479
 480
 481
 482
 483
 484
 485
 486
 487
 488
 489
 490
 491
 492
 493
 494
 495
 496
 497
 498
 499
 500
 501
 502
 503
 504
 505
 506
 507
 508
 509
 510
 511
 512
 513
 514
 515
 516
 517
 518
 519
 520
 521
 522
 523
 524
 525
 526
 527
 528
 529
 530
 531
 532
 533
 534
 535
 536
 537
 538
 539
 540
 541
 542
 543
 544
 545
 546
 547
 548
 549
 550
 551
 552
 553
 554
 555
 556
 557
 558
 559
 560
 561
 562
 563
 564
 565
 566
 567
 568
 569
 570
 571
 572
 573
 574
 575
 576
 577
 578
 579
 580
 581
 582
 583
 584
 585
 586
 587
 588
 589
 590
 591
 592
 593
 594
 595
 596
 597
 598
 599
 600
 601
 602
 603
 604
 605
 606
 607
 608
 609
 610
 611
 612
 613
 614
 615
 616
 617
 618
 619
 620
 621
 622
 623
 624
 625
 626
 627
 628
 629
 630
 631
 632
 633
 634
 635
 636
 637
 638
 639
 640
 641
 642
 643
 644
 645
 646
 647
 648
 649
 650
 651
 652
 653
 654
 655
 656
 657
 658
 659
 660
 661
 662
 663
 664
 665
 666
 667
 668
 669
 670
 671
 672
 673
 674
 675
 676
 677
 678
 679
 680
 681
 682
 683
 684
 685
 686
 687
 688
 689
 690
 691
 692
 693
 694
 695
 696
 697
 698
 699
 700
 701
 702
 703
 704
 705
 706
 707
 708
 709
 710
 711
 712
 713
 714
 715
 716
 717
 718
 719
 720
 721
 722
 723
 724
 725
 726
 727
 728
 729
 730
 731
 732
 733
 734
 735
 736
 737
 738
 739
 740
 741
 742
 743
 744
 745
 746
 747
 748
 749
 750
 751
 752
 753
 754
 755
 756
 757
 758
 759
 760
 761
 762
 763
 764
 765
 766
 767
 768
 769
 770
 771
 772
 773
 774
 775
 776
 777
 778
 779
 780
 781
 782
 783
 784
 785
 786
 787
 788
 789
 790
 791
 792
 793
 794
 795
 796
 797
 798
 799
 800
 801
 802
 803
 804
 805
 806
 807
 808
 809
 810
 811
 812
 813
 814
 815
 816
 817
 818
 819
 820
 821
 822
 823
 824
 825
 826
 827
 828
 829
 830
 831
 832
 833
 834
 835
 836
 837
 838
 839
 840
 841
 842
 843
 844
 845
 846
 847
 848
 849
 850
 851
 852
 853
 854
 855
 856
 857
 858
 859
 860
 861
 862
 863
 864
 865
 866
 867
 868
 869
 870
 871
 872
 873
 874
 875
 876
 877
 878
 879
 880
 881
 882
 883
 884
 885
 886
 887
 888
 889
 890
 891
 892
 893
 894
 895
 896
 897
 898
 899
 900
 901
 902
 903
 904
 905
 906
 907
 908
 909
 910
 911
 912
 913
 914
 915
 916
 917
 918
 919
 920
 921
 922
 923
 924
 925
 926
 927
 928
 929
 930
 931
 932
 933
 934
 935
 936
 937
 938
 939
 940
 941
 942
 943
 944
 945
 946
 947
 948
 949
 950
 951
 952
 953
 954
 955
 956
 957
 958
 959
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
package main

import (
	"archive/tar"
	"archive/zip"
	"bufio"
	"compress/gzip"
	"context"
	"crypto/rand"
	"encoding/hex"
	"encoding/json"
	"fmt"
	"io"
	"net"
	"net/http"
	"net/url"
	"os"
	"os/exec"
	"os/signal"
	"path/filepath"
	"runtime"
	"strconv"
	"strings"
	"syscall"
	"time"
)

// --- Constants & Config ---
const (
	colorReset  = "\033[0m"
	colorCyan   = "\033[36m"
	colorGreen  = "\033[32m"
	colorYellow = "\033[33m"
	colorRed    = "\033[31m"
	appDirName  = ".llama_manager"
)

var (
	homeDir, _      = os.UserHomeDir()
	baseDir         = filepath.Join(homeDir, appDirName)
	binDir          = filepath.Join(baseDir, "bin")
	modelsDir       = filepath.Join(baseDir, "models")
	pidFile         = filepath.Join(baseDir, "server.pid")
	logFile         = filepath.Join(baseDir, "server.log")
	httpClient      = &http.Client{Timeout: 30 * time.Second}
	latestGitHubTag = ""
)

func init() {
	os.MkdirAll(binDir, 0755)
	os.MkdirAll(modelsDir, 0755)
	go fetchLatestReleaseAsync()
}

// --- Live Update Notifier ---
func fetchLatestReleaseAsync() {
	req, _ := http.NewRequest("GET", "https://api.github.com/repos/ggml-org/llama.cpp/releases/latest", nil)
	client := &http.Client{Timeout: 5 * time.Second}
	resp, err := client.Do(req)
	if err == nil {
		defer resp.Body.Close()
		var release struct {
			TagName string `json:"tag_name"`
		}
		if json.NewDecoder(resp.Body).Decode(&release) == nil {
			latestGitHubTag = release.TagName
		}
	}
}

// --- Process Management (Ghost Process Fix) ---
func isProcessAlive(pid int) bool {
	if runtime.GOOS == "windows" {
		cmd := exec.Command("tasklist", "/FI", fmt.Sprintf("PID eq %d", pid))
		out, err := cmd.Output()
		if err != nil {
			return false
		}
		return strings.Contains(string(out), strconv.Itoa(pid))
	}
	process, err := os.FindProcess(pid)
	if err != nil {
		return false
	}
	return process.Signal(syscall.Signal(0)) == nil
}

// --- UI & Navigation Helpers ---
func clearScreen() {
	cmd := exec.Command("clear")
	if runtime.GOOS == "windows" {
		cmd = exec.Command("cmd", "/c", "cls")
	}
	cmd.Stdout = os.Stdout
	cmd.Run()
}

func pause() {
	fmt.Print(colorYellow + "\nPress Enter to continue..." + colorReset)
	bufio.NewReader(os.Stdin).ReadBytes('\n')
}

func promptInput(prompt string) string {
	fmt.Print(prompt)
	reader := bufio.NewReader(os.Stdin)
	text, _ := reader.ReadString('\n')
	return strings.TrimSpace(text)
}

func generateAPIKey() string {
	bytes := make([]byte, 16)
	rand.Read(bytes)
	return hex.EncodeToString(bytes)
}

// --- Intelligent Dependency Advisor ---
func adviseDependencies(missingTool string) {
	fmt.Println(colorRed + "\n[Error] Missing Compilation Dependency: " + missingTool + colorReset)
	fmt.Println("To compile an optimized engine from source, your system needs '" + missingTool + "'.")

	switch runtime.GOOS {
	case "windows":
		fmt.Println(colorCyan + "\nWindows Installation Instructions:" + colorReset)
		fmt.Println("Open PowerShell as Administrator and run:")
		if missingTool == "git" {
			fmt.Println("  winget install --id Git.Git -e --source winget")
		} else if missingTool == "cmake" {
			fmt.Println("  winget install --id Kitware.CMake -e --source winget")
			fmt.Println("\nNote: You also need a C++ Compiler for Windows.")
			fmt.Println("Run: winget install --id Microsoft.VisualStudio.2022.BuildTools")
			fmt.Println("Then open Visual Studio Installer and check 'Desktop development with C++'.")
		}
	case "darwin":
		fmt.Println(colorCyan + "\nmacOS Installation Instructions:" + colorReset)
		fmt.Println("Open Terminal and run:")
		fmt.Println("  xcode-select --install")
		fmt.Println("  brew install " + missingTool)
	case "linux":
		fmt.Println(colorCyan + "\nLinux Installation Instructions:" + colorReset)
		fmt.Println("Run the command for your specific distribution:")
		fmt.Println("  Debian/Ubuntu:  sudo apt update && sudo apt install " + missingTool + " build-essential")
		fmt.Println("  Fedora/RHEL:    sudo dnf install " + missingTool + " gcc-c++")
		fmt.Println("  Arch Linux:     sudo pacman -S " + missingTool + " base-devel")
	}
	fmt.Println(colorYellow + "\nAfter installing, please close this terminal, open a new one, and try again." + colorReset)
}

// --- Resumable Download Engine ---
type progressWriter struct {
	io.Writer
	TotalDownloaded int64
	TotalSize       int64
	LastPercent     int64
}

func (pw *progressWriter) Write(p []byte) (int, error) {
	n, err := pw.Writer.Write(p)
	pw.TotalDownloaded += int64(n)
	if pw.TotalSize > 0 {
		percent := (pw.TotalDownloaded * 100) / pw.TotalSize
		if percent > pw.LastPercent {
			fmt.Printf("\r\033[KDownloading... %d%% (%d MB / %d MB) - Press Ctrl+C to Pause",
				percent, pw.TotalDownloaded/(1024*1024), pw.TotalSize/(1024*1024))
			pw.LastPercent = percent
		}
	}
	return n, err
}

func downloadWithResume(ctx context.Context, dlURL, finalDest string) error {
	partDest := finalDest + ".part"
	var startBytes int64 = 0

	if info, err := os.Stat(partDest); err == nil {
		startBytes = info.Size()
	} else if _, err := os.Stat(finalDest); err == nil {
		fmt.Println(colorGreen + "File already fully downloaded!" + colorReset)
		return nil
	}

	req, err := http.NewRequestWithContext(ctx, "GET", dlURL, nil)
	if err != nil {
		return err
	}

	if startBytes > 0 {
		req.Header.Set("Range", fmt.Sprintf("bytes=%d-", startBytes))
		fmt.Printf("Resuming incomplete download from %d MB...\n", startBytes/(1024*1024))
	}

	dlClient := &http.Client{}
	resp, err := dlClient.Do(req)
	if err != nil {
		if ctx.Err() == context.Canceled {
			return fmt.Errorf("paused by user")
		}
		return fmt.Errorf("connection error: %v", err)
	}
	defer resp.Body.Close()

	if resp.StatusCode == http.StatusRequestedRangeNotSatisfiable {
		os.Rename(partDest, finalDest)
		return nil
	}
	if resp.StatusCode != http.StatusOK && resp.StatusCode != http.StatusPartialContent {
		return fmt.Errorf("HTTP %d: %s", resp.StatusCode, resp.Status)
	}

	flags := os.O_CREATE | os.O_WRONLY
	if resp.StatusCode == http.StatusPartialContent {
		flags |= os.O_APPEND
	} else {
		flags |= os.O_TRUNC
		startBytes = 0
	}

	outFile, err := os.OpenFile(partDest, flags, 0644)
	if err != nil {
		return fmt.Errorf("file access error: %v", err)
	}

	pw := &progressWriter{
		Writer:          outFile,
		TotalDownloaded: startBytes,
		TotalSize:       startBytes + resp.ContentLength,
	}

	_, err = io.Copy(pw, resp.Body)
	outFile.Close()
	fmt.Println()

	if err != nil {
		if ctx.Err() == context.Canceled {
			return fmt.Errorf("paused by user")
		}
		return fmt.Errorf("interrupted (network drop or sleep): %v", err)
	}

	return os.Rename(partDest, finalDest)
}

// --- Smart Hardware Detection ---
func detectHardware() (backend string, name string) {
	if runtime.GOOS == "darwin" {
		return "metal", "Apple Silicon/Metal"
	}
	if err := exec.Command("nvidia-smi").Run(); err == nil {
		return "cuda", "NVIDIA GPU (CUDA)"
	}
	if err := exec.Command("rocm-smi").Run(); err == nil {
		return "rocm", "AMD GPU (ROCm)"
	}
	return "vulkan", "Standard/APU (Vulkan)"
}

func getServerExePath() string {
	exeName := "llama-server"
	if runtime.GOOS == "windows" {
		exeName = "llama-server.exe"
	}

	localPath := filepath.Join(binDir, exeName)
	if _, err := os.Stat(localPath); err == nil {
		return localPath
	}

	if path, err := exec.LookPath(exeName); err == nil {
		return path
	}
	return ""
}

// --- Dynamic Memory Capacity Scanner ---
func getMemoryCapacity() (float64, string) {
	out, err := exec.Command("nvidia-smi", "--query-gpu=memory.total", "--format=csv,noheader,nounits").Output()
	if err == nil {
		lines := strings.Split(strings.TrimSpace(string(out)), "\n")
		if len(lines) > 0 {
			vramMB, _ := strconv.ParseFloat(strings.TrimSpace(lines[0]), 64)
			if vramMB > 0 {
				return vramMB / 1024.0, "Dedicated NVIDIA VRAM"
			}
		}
	}

	out, err = exec.Command("rocm-smi", "--showmeminfo", "vram").Output()
	if err == nil {
		lines := strings.Split(string(out), "\n")
		for _, line := range lines {
			if strings.Contains(line, "VRAM Total Memory") {
				parts := strings.Split(line, ":")
				if len(parts) == 2 {
					bytes, _ := strconv.ParseFloat(strings.TrimSpace(parts[1]), 64)
					if bytes > 0 {
						return bytes / (1024 * 1024 * 1024), "Dedicated AMD VRAM"
					}
				}
			}
		}
	}

	var ramBytes int64 = 0
	if runtime.GOOS == "windows" {
		out, err := exec.Command("powershell", "-NoProfile", "-Command", "(Get-CimInstance Win32_ComputerSystem).TotalPhysicalMemory").Output()
		if err == nil {
			ramBytes, _ = strconv.ParseInt(strings.TrimSpace(string(out)), 10, 64)
		}
	} else if runtime.GOOS == "linux" {
		out, err := os.ReadFile("/proc/meminfo")
		if err == nil {
			for _, line := range strings.Split(string(out), "\n") {
				if strings.HasPrefix(line, "MemTotal:") {
					parts := strings.Fields(line)
					if len(parts) >= 2 {
						kb, _ := strconv.ParseInt(parts[1], 10, 64)
						ramBytes = kb * 1024
						break
					}
				}
			}
		}
	} else if runtime.GOOS == "darwin" {
		out, err := exec.Command("sysctl", "-n", "hw.memsize").Output()
		if err == nil {
			ramBytes, _ = strconv.ParseInt(strings.TrimSpace(string(out)), 10, 64)
		}
	}

	if ramBytes > 0 {
		return float64(ramBytes) / (1024 * 1024 * 1024), "Shared System RAM"
	}
	return 8.0, "Unknown RAM (Fallback Default)"
}

// --- Network Helpers ---
func getTailscaleIP() string {
	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Second)
	defer cancel()
	out, err := exec.CommandContext(ctx, "tailscale", "ip", "-4").Output()
	if err != nil {
		return ""
	}
	return strings.TrimSpace(string(out))
}

func getAvailablePort(startPort int) int {
	for port := startPort; port < startPort+100; port++ {
		ln, err := net.Listen("tcp", fmt.Sprintf(":%d", port))
		if err == nil {
			ln.Close()
			return port
		}
	}
	return startPort
}

// --- Extraction Helper ---
func isTargetFile(filename string) bool {
	base := strings.ToLower(filepath.Base(filename))
	if base == "llama-server" || base == "llama-server.exe" {
		return true
	}
	if strings.HasSuffix(base, ".dll") || strings.Contains(base, ".so") || strings.HasSuffix(base, ".dylib") {
		return true
	}
	return false
}

func clearOldBinaries() {
	files, _ := filepath.Glob(filepath.Join(binDir, "*"))
	for _, f := range files {
		if isTargetFile(f) {
			os.Remove(f)
		}
	}
}

func extractArchive(src, dest string) error {
	if strings.HasSuffix(src, ".zip") {
		r, err := zip.OpenReader(src)
		if err != nil {
			return err
		}
		defer r.Close()
		for _, f := range r.File {
			if isTargetFile(f.Name) {
				out, err := os.Create(filepath.Join(dest, filepath.Base(f.Name)))
				if err != nil {
					return fmt.Errorf("failed to write %s (Is the server running?): %v", f.Name, err)
				}
				rc, err := f.Open()
				if err != nil {
					out.Close()
					return fmt.Errorf("failed to open file inside zip: %v", err)
				}
				_, err = io.Copy(out, rc)
				rc.Close()
				out.Close()
				if err != nil {
					return fmt.Errorf("failed to copy file contents: %v", err)
				}
			}
		}
	} else if strings.HasSuffix(src, ".tar.gz") {
		f, err := os.Open(src)
		if err != nil {
			return err
		}
		defer f.Close()
		gzr, err := gzip.NewReader(f)
		if err != nil {
			return err
		}
		defer gzr.Close()
		tr := tar.NewReader(gzr)
		for {
			hdr, err := tr.Next()
			if err == io.EOF {
				break
			}
			if isTargetFile(hdr.Name) {
				out, err := os.Create(filepath.Join(dest, filepath.Base(hdr.Name)))
				if err != nil {
					return fmt.Errorf("failed to write %s: %v", hdr.Name, err)
				}
				_, err = io.Copy(out, tr)
				out.Close()
				if err != nil {
					return fmt.Errorf("failed to extract file contents: %v", err)
				}
			}
		}
	}
	return nil
}

// --- Core Logic ---
func smartInstallAndCheck() {
	targetBackend, hwName := detectHardware()
	fmt.Println(colorCyan + "\n--- System Hardware Check ---" + colorReset)
	fmt.Printf("OS/Arch: %s/%s\nActive Profile: %s\n", runtime.GOOS, runtime.GOARCH, hwName)

	req, _ := http.NewRequest("GET", "https://api.github.com/repos/ggml-org/llama.cpp/releases/latest", nil)
	resp, err := httpClient.Do(req)
	if err != nil {
		fmt.Println(colorRed + "Failed to connect to GitHub." + colorReset)
		pause()
		return
	}
	defer resp.Body.Close()

	var release struct {
		TagName string                                      `json:"tag_name"`
		Assets  []struct{ Name, BrowserDownloadURL string } `json:"assets"`
	}
	json.NewDecoder(resp.Body).Decode(&release)
	latestGitHubTag = release.TagName

	localVerBytes, _ := os.ReadFile(filepath.Join(binDir, "version.txt"))
	localBackBytes, _ := os.ReadFile(filepath.Join(binDir, "backend.txt"))
	localVer := strings.TrimSpace(string(localVerBytes))
	localBack := strings.TrimSpace(string(localBackBytes))

	exePath := getServerExePath()
	isSystemInstall := exePath != "" && !strings.HasPrefix(exePath, binDir)
	isCompiledLocally := strings.HasPrefix(localVer, "source-")

	if isSystemInstall {
		fmt.Printf(colorRed+"\n[Warning] Suboptimal System-wide Installation Detected:\n     %s\n"+colorReset, exePath)
		fmt.Println(colorYellow + "System package managers usually install generic, CPU-only versions." + colorReset)
		fmt.Printf("For your %s, you need the optimized %s backend for maximum performance.\n\n", hwName, targetBackend)

		lowerPath := strings.ToLower(exePath)
		if strings.Contains(lowerPath, "windowsapps") || strings.Contains(lowerPath, "winget") {
			fmt.Println("To remove it, open PowerShell and run:  " + colorRed + "winget uninstall llama.cpp" + colorReset)
		} else if strings.Contains(lowerPath, "homebrew") || strings.Contains(lowerPath, "linuxbrew") {
			fmt.Println("To remove it, open your terminal and run:  " + colorRed + "brew uninstall llama.cpp" + colorReset)
		} else if strings.Contains(lowerPath, "nix") {
			fmt.Println("To remove it, please uninstall it from your active Nix profile.")
		} else if strings.Contains(lowerPath, "apt") || strings.Contains(lowerPath, "dpkg") || runtime.GOOS == "linux" {
			fmt.Println("To remove it, run:  " + colorRed + "sudo apt remove llama.cpp" + colorReset + " (or distro equivalent).")
		}

		fmt.Println(colorGreen + "\nWe highly recommend letting this manager download or compile the hardware-optimized version instead." + colorReset)
	}

	needsUpdate := exePath == "" || (!isSystemInstall && localBack != targetBackend) || (!isSystemInstall && !isCompiledLocally && localVer != release.TagName) || isSystemInstall

	if !needsUpdate {
		if isCompiledLocally {
			fmt.Println(colorGreen + "\n[OK] Engine is compiled natively for maximum performance." + colorReset)
		} else if !isSystemInstall {
			fmt.Println(colorGreen + "\n[OK] Pre-built Engine is up to date." + colorReset)
		}

		choice := promptInput("\nForce update / re-install local binary? (y/n/0): ")
		if choice == "0" || strings.ToLower(choice) != "y" {
			return
		}
	} else {
		var msg string
		if isSystemInstall {
			msg = "\nInstall the optimized local engine now? (y/n/0 to go back): "
		} else {
			msg = "\nUpdates or Hardware Changes Detected! Update now? (y/n/0 to go back): "
		}
		choice := promptInput(msg)
		if choice == "0" || strings.ToLower(choice) != "y" {
			return
		}
	}

	var downloadURL, fileName, fallbackURL, fallbackName string
	osStr, archStr := runtime.GOOS, runtime.GOARCH

	for _, asset := range release.Assets {
		name := strings.ToLower(asset.Name)
		if strings.Contains(name, "cudart") || strings.HasSuffix(name, ".sig") || strings.HasSuffix(name, ".sha256") {
			continue
		}
		if !strings.HasSuffix(name, ".zip") && !strings.HasSuffix(name, ".tar.gz") {
			continue
		}

		isWin := osStr == "windows" && (strings.Contains(name, "win") || strings.Contains(name, "windows"))
		isMac := osStr == "darwin" && (strings.Contains(name, "macos") || strings.Contains(name, "darwin") || strings.Contains(name, "mac"))
		isLin := osStr == "linux" && (strings.Contains(name, "linux") || strings.Contains(name, "ubuntu"))

		isX64 := archStr == "amd64" && (strings.Contains(name, "x64") || strings.Contains(name, "x86_64") || strings.Contains(name, "amd64") || strings.Contains(name, "avx"))
		isArm := archStr == "arm64" && (strings.Contains(name, "arm64") || strings.Contains(name, "aarch64"))

		if (isWin || isLin || isMac) && (isX64 || isArm) {
			isSpecialBackend := strings.Contains(name, "cuda") || strings.Contains(name, "vulkan") || strings.Contains(name, "rocm") || strings.Contains(name, "sycl")
			if !isSpecialBackend {
				fallbackURL, fallbackName = asset.BrowserDownloadURL, asset.Name
			}

			if isMac {
				downloadURL, fileName = asset.BrowserDownloadURL, asset.Name
				break
			} else if strings.Contains(name, targetBackend) {
				if targetBackend == "cuda" && strings.Contains(name, "cu11") {
					continue
				}
				downloadURL, fileName = asset.BrowserDownloadURL, asset.Name
				break
			}
		}
	}

	if downloadURL == "" {
		if targetBackend == "cuda" && runtime.GOOS == "linux" {
			fmt.Println(colorYellow + "\nNote: Pre-built Linux CUDA binaries are not officially provided." + colorReset)
		} else {
			fmt.Println(colorYellow + "\nOptimized binary not found for your exact GPU backend." + colorReset)
		}

		fmt.Println("Compiling from source yields 10-20% faster generation as it optimizes for your specific CPU.")
		compChoice := promptInput(colorCyan + "Would you like to compile it from source now? (y/n): " + colorReset)

		if strings.ToLower(compChoice) == "y" {
			buildFromSource(targetBackend, release.TagName)
			return
		}

		if fallbackURL != "" {
			fmt.Println(colorYellow + "Falling back to universally compatible CPU binary." + colorReset)
			downloadURL, fileName = fallbackURL, fallbackName
			targetBackend = "cpu-fallback"
		} else {
			fmt.Println(colorRed + "\n[Error] Pre-built binary not found for your OS/Arch." + colorReset)
			pause()
			return
		}
	}

	archivePath := filepath.Join(binDir, fileName)
	fmt.Printf("\nDownloading %s...\n", fileName)

	ctx, cancel := context.WithCancel(context.Background())
	defer cancel()
	sigChan := make(chan os.Signal, 1)
	signal.Notify(sigChan, os.Interrupt, syscall.SIGTERM)
	go func() { <-sigChan; cancel() }()

	if err := downloadWithResume(ctx, downloadURL, archivePath); err != nil {
		fmt.Println(colorYellow + "\nDownload stopped: " + err.Error() + colorReset)
		signal.Stop(sigChan)
		pause()
		return
	}
	signal.Stop(sigChan)

	fmt.Println("Extracting binaries...")
	clearOldBinaries()
	if err := extractArchive(archivePath, binDir); err != nil {
		fmt.Println(colorRed + "\nExtraction Error: " + err.Error() + colorReset)
		pause()
		return
	}
	os.Remove(archivePath)
	if runtime.GOOS != "windows" {
		os.Chmod(filepath.Join(binDir, "llama-server"), 0755)
	}

	os.WriteFile(filepath.Join(binDir, "version.txt"), []byte(release.TagName), 0644)
	os.WriteFile(filepath.Join(binDir, "backend.txt"), []byte(targetBackend), 0644)
	fmt.Println(colorGreen + "Installation successful!" + colorReset)
	pause()
}

func buildFromSource(targetBackend string, tagName string) {
	fmt.Println(colorCyan + "\n--- Compiling llama.cpp from Source ---" + colorReset)

	if _, err := exec.LookPath("git"); err != nil {
		adviseDependencies("git")
		pause()
		return
	}
	if _, err := exec.LookPath("cmake"); err != nil {
		adviseDependencies("cmake")
		pause()
		return
	}

	if targetBackend == "vulkan" && runtime.GOOS == "windows" && os.Getenv("VULKAN_SDK") == "" {
		fmt.Println(colorRed + "\n[Error] Vulkan SDK not found!" + colorReset)
		fmt.Println("To compile for Vulkan on Windows, you must install the Vulkan SDK.")
		fmt.Println("Download it here: https://vulkan.lunarg.com/sdk/home")
		fmt.Println("Install it, completely restart this terminal, and try again.")
		pause()
		return
	}

	srcDir := filepath.Join(baseDir, "src")
	if _, err := os.Stat(srcDir); os.IsNotExist(err) {
		fmt.Println("Cloning repository...")
		cmd := exec.Command("git", "clone", "https://github.com/ggml-org/llama.cpp.git", srcDir)
		cmd.Stdout, cmd.Stderr = os.Stdout, os.Stderr
		if err := cmd.Run(); err != nil {
			fmt.Println(colorRed + "Clone failed." + colorReset)
			pause()
			return
		}
	} else {
		fmt.Println("Fetching latest updates...")
		exec.Command("git", "-C", srcDir, "fetch", "--all", "--tags").Run()
	}

	fmt.Printf(colorYellow+"Locking onto stable release: %s\n"+colorReset, tagName)
	exec.Command("git", "-C", srcDir, "checkout", "-f", tagName).Run()

	buildDir := filepath.Join(srcDir, "build")
	os.RemoveAll(buildDir)

	cmakeArgs := []string{"-B", "build"}
	if runtime.GOOS == "windows" {
		cmakeArgs = append(cmakeArgs, "-G", "Visual Studio 17 2022", "-A", "x64")
	}

	if targetBackend == "cuda" {
		cmakeArgs = append(cmakeArgs, "-DGGML_CUDA=ON")
	} else if targetBackend == "vulkan" {
		cmakeArgs = append(cmakeArgs, "-DGGML_VULKAN=ON")
	} else if targetBackend == "rocm" {
		cmakeArgs = append(cmakeArgs, "-DGGML_HIP=ON")
	} else if targetBackend == "metal" {
		cmakeArgs = append(cmakeArgs, "-DGGML_METAL=ON")
	}

	fmt.Println(colorYellow + "\nConfiguring CMake Environment..." + colorReset)
	configCmd := exec.Command("cmake", cmakeArgs...)
	configCmd.Dir = srcDir
	configCmd.Stdout, configCmd.Stderr = os.Stdout, os.Stderr
	if err := configCmd.Run(); err != nil {
		fmt.Println(colorRed + "CMake configuration failed." + colorReset)
		pause()
		return
	}

	cores := strconv.Itoa(runtime.NumCPU())
	fmt.Printf(colorYellow+"\nCompiling with %s threads (this will take several minutes)...\n"+colorReset, cores)
	buildCmd := exec.Command("cmake", "--build", "build", "--config", "Release", "-j", cores)
	buildCmd.Dir = srcDir
	buildCmd.Stdout, buildCmd.Stderr = os.Stdout, os.Stderr
	if err := buildCmd.Run(); err != nil {
		fmt.Println(colorRed + "Compilation failed." + colorReset)
		pause()
		return
	}

	exeName := "llama-server"
	if runtime.GOOS == "windows" {
		exeName = "llama-server.exe"
	}

	possiblePaths := []string{
		filepath.Join(srcDir, "build", "bin", exeName),
		filepath.Join(srcDir, "build", "bin", "Release", exeName),
		filepath.Join(srcDir, "build", exeName),
	}

	compiledPath := ""
	for _, p := range possiblePaths {
		if _, err := os.Stat(p); err == nil {
			compiledPath = p
			break
		}
	}

	if compiledPath == "" {
		fmt.Println(colorRed + "Could not locate compiled binary." + colorReset)
		pause()
		return
	}

	clearOldBinaries()
	destPath := filepath.Join(binDir, exeName)
	if err := copyFile(compiledPath, destPath); err != nil {
		fmt.Println(colorRed + "Install failed: " + err.Error() + colorReset)
		pause()
		return
	}

	filepath.Walk(filepath.Join(srcDir, "build"), func(path string, info os.FileInfo, err error) error {
		if err == nil && !info.IsDir() && isTargetFile(path) && filepath.Base(path) != exeName {
			copyFile(path, filepath.Join(binDir, filepath.Base(path)))
		}
		return nil
	})

	if runtime.GOOS != "windows" {
		os.Chmod(destPath, 0755)
	}

	hashCmd := exec.Command("git", "rev-parse", "--short", "HEAD")
	hashCmd.Dir = srcDir
	hash, _ := hashCmd.Output()
	versionStr := "source-" + strings.TrimSpace(string(hash))

	os.WriteFile(filepath.Join(binDir, "version.txt"), []byte(versionStr), 0644)
	os.WriteFile(filepath.Join(binDir, "backend.txt"), []byte(targetBackend), 0644)

	fmt.Println(colorGreen + "\nCompilation and installation successful!" + colorReset)
	pause()
}

func copyFile(src, dst string) error {
	in, err := os.Open(src)
	if err != nil {
		return err
	}
	defer in.Close()
	out, err := os.Create(dst)
	if err != nil {
		return err
	}
	defer out.Close()
	_, err = io.Copy(out, in)
	return err
}

func searchAndDownloadModel() {
	for {
		clearScreen()
		fmt.Println(colorCyan + "=== Search & Download Models ===" + colorReset)
		rawQuery := promptInput("Enter model search term (e.g., 'qwen 2.5 7b q4_k_m') or '0' to return to Main Menu: ")

		if rawQuery == "0" || rawQuery == "" {
			return
		}

		query := url.QueryEscape(rawQuery)
		apiURL := fmt.Sprintf("https://huggingface.co/api/models?search=%s&filter=gguf&sort=downloads&direction=-1&limit=5", query)

		resp, err := httpClient.Get(apiURL)
		if err != nil {
			fmt.Println(colorRed + "\nNetwork error: " + err.Error() + colorReset)
			pause()
			continue
		}

		var models []struct {
			ID        string `json:"id"`
			Downloads int    `json:"downloads"`
		}
		json.NewDecoder(resp.Body).Decode(&models)
		resp.Body.Close()

		if len(models) == 0 {
			fmt.Println(colorYellow + "\nNo GGUF models found." + colorReset)
			pause()
			continue
		}

		fmt.Println(colorCyan + "\n--- Search Results ---" + colorReset)
		fmt.Println("0. Cancel & Start New Search")
		for i, m := range models {
			fmt.Printf("%d. %s (Downloads: %d)\n", i+1, m.ID, m.Downloads)
		}

		choice, _ := strconv.Atoi(promptInput("\nSelect model number: "))
		if choice == 0 || choice < 1 || choice > len(models) {
			continue
		}

		selectedModel := models[choice-1].ID
		resp2, err := httpClient.Get(fmt.Sprintf("https://huggingface.co/api/models/%s/tree/main", selectedModel))
		if err != nil {
			continue
		}

		var files []struct {
			Path string `json:"path"`
			Size int64  `json:"size"`
		}
		json.NewDecoder(resp2.Body).Decode(&files)
		resp2.Body.Close()

		var ggufFiles []struct {
			Path string `json:"path"`
			Size int64  `json:"size"`
		}
		for _, f := range files {
			if strings.HasSuffix(f.Path, ".gguf") {
				ggufFiles = append(ggufFiles, f)
			}
		}

		if len(ggufFiles) == 0 {
			fmt.Println(colorYellow + "\nNo GGUF files found in this repository." + colorReset)
			pause()
			continue
		}

		fmt.Println(colorCyan + "\n--- Available Files ---" + colorReset)
		fmt.Println("0. Cancel & Start New Search")
		for i, f := range ggufFiles {
			fmt.Printf("%d. %s (%.2f MB)\n", i+1, f.Path, float64(f.Size)/(1024*1024))
		}

		fChoice, _ := strconv.Atoi(promptInput("\nSelect file to download: "))
		if fChoice == 0 || fChoice < 1 || fChoice > len(ggufFiles) {
			continue
		}

		selectedFile := ggufFiles[fChoice-1]
		fileSizeGB := float64(selectedFile.Size) / (1024 * 1024 * 1024)

		memCapGB, memType := getMemoryCapacity()
		maxSafeGB := memCapGB - 2.5
		if maxSafeGB < 0 {
			maxSafeGB = 0
		}

		if fileSizeGB > maxSafeGB {
			fmt.Println(colorRed + "\n[WARNING] Hardware Capacity Exceeded!" + colorReset)
			fmt.Printf("You are trying to download a %.2f GB model.\n", fileSizeGB)
			fmt.Printf("Your system is reporting %.2f GB of %s.\n", memCapGB, memType)
			fmt.Printf("To leave room for your OS and the context window, your safe maximum is ~%.2f GB.\n", maxSafeGB)
			fmt.Println(colorYellow + "If you proceed, this model will likely crash the server or run extremely slowly via pagefile swapping." + colorReset)

			proceed := promptInput("\nProceed anyway? (y/n): ")
			if strings.ToLower(proceed) != "y" {
				continue
			}
		}

		dlURL := fmt.Sprintf("https://huggingface.co/%s/resolve/main/%s", selectedModel, selectedFile.Path)
		destPath := filepath.Join(modelsDir, strings.ReplaceAll(selectedModel, "/", "_")+"_"+selectedFile.Path)

		ctx, cancel := context.WithCancel(context.Background())
		sigChan := make(chan os.Signal, 1)
		signal.Notify(sigChan, os.Interrupt, syscall.SIGTERM)
		go func() { <-sigChan; cancel() }()

		err = downloadWithResume(ctx, dlURL, destPath)
		signal.Stop(sigChan)
		cancel()

		if err != nil {
			fmt.Println(colorYellow + "\nDownload stopped: " + err.Error() + colorReset)
			fmt.Println("The incomplete file has been saved. Select it again later to resume.")
		} else {
			fmt.Println(colorGreen + "Download complete!" + colorReset)
		}
		pause()
	}
}

func manageModels() {
	entries, _ := os.ReadDir(modelsDir)
	var models []os.DirEntry
	for _, e := range entries {
		if strings.HasSuffix(e.Name(), ".gguf") || strings.HasSuffix(e.Name(), ".part") {
			models = append(models, e)
		}
	}

	if len(models) == 0 {
		fmt.Println(colorYellow + "No models found." + colorReset)
		pause()
		return
	}

	fmt.Println(colorCyan + "\n--- Local Models ---" + colorReset)
	fmt.Println("0. Go Back")
	for i, m := range models {
		info, _ := m.Info()
		status := ""
		if strings.HasSuffix(m.Name(), ".part") {
			status = colorYellow + " [Incomplete]" + colorReset
		}
		fmt.Printf("%d. %s (%.2f MB)%s\n", i+1, m.Name(), float64(info.Size())/(1024*1024), status)
	}

	if choiceStr := promptInput("\nEnter number to DELETE (or 0 to cancel): "); choiceStr != "0" {
		if choice, err := strconv.Atoi(choiceStr); err == nil && choice >= 1 && choice <= len(models) {
			target := models[choice-1].Name()
			if strings.ToLower(promptInput(fmt.Sprintf("Delete %s? (y/n): ", target))) == "y" {
				os.Remove(filepath.Join(modelsDir, target))
				fmt.Println(colorGreen + "Deleted." + colorReset)
			}
		}
	}
}

// --- Daemon Server Management ---
func stopWebUI() {
	pidBytes, err := os.ReadFile(pidFile)
	if err != nil {
		fmt.Println(colorYellow + "No background server is currently running." + colorReset)
		pause()
		return
	}

	pid, err := strconv.Atoi(strings.TrimSpace(string(pidBytes)))
	if err != nil {
		os.Remove(pidFile)
		return
	}

	if !isProcessAlive(pid) {
		fmt.Println(colorYellow + "Server is not actively running (Cleaning up stale PID file)." + colorReset)
		os.Remove(pidFile)
		pause()
		return
	}

	fmt.Printf("Stopping background server (PID: %d)...\n", pid)

	if runtime.GOOS == "windows" {
		exec.Command("taskkill", "/PID", strconv.Itoa(pid), "/F", "/T").Run()
	} else {
		proc, err := os.FindProcess(pid)
		if err == nil {
			proc.Signal(os.Interrupt)
		}
	}

	os.Remove(pidFile)
	fmt.Println(colorGreen + "Server successfully stopped and VRAM has been freed." + colorReset)
	pause()
}

func startWebUI() {
	if pidBytes, err := os.ReadFile(pidFile); err == nil {
		pid, _ := strconv.Atoi(strings.TrimSpace(string(pidBytes)))
		if isProcessAlive(pid) {
			fmt.Println(colorYellow + "A background server is already running! Please stop it using Option 5 before deploying a new one." + colorReset)
			pause()
			return
		}
		os.Remove(pidFile) // Auto-heal ghost process
	}

	serverExe := getServerExePath()
	if serverExe == "" {
		fmt.Println(colorRed + "llama.cpp not installed." + colorReset)
		pause()
		return
	}

	entries, _ := os.ReadDir(modelsDir)
	var models []string
	for _, e := range entries {
		if strings.HasSuffix(e.Name(), ".gguf") {
			models = append(models, e.Name())
		}
	}

	if len(models) == 0 {
		fmt.Println(colorYellow + "No models found!" + colorReset)
		pause()
		return
	}

	fmt.Println(colorCyan + "\n--- Select Model ---" + colorReset)
	fmt.Println("0. Go Back")
	for i, m := range models {
		fmt.Printf("%d. %s\n", i+1, m)
	}

	choice, _ := strconv.Atoi(promptInput("\nChoice: "))
	if choice == 0 || choice < 1 || choice > len(models) {
		return
	}
	selectedModel := filepath.Join(modelsDir, models[choice-1])

	// --- 0. HARDWARE & MODEL ASSESSMENT ---
	modelInfo, err := os.Stat(selectedModel)
	fileSizeGB := 0.0
	if err == nil {
		fileSizeGB = float64(modelInfo.Size()) / (1024 * 1024 * 1024)
	}

	lowerName := strings.ToLower(filepath.Base(selectedModel))
	quantName := "Unknown/FP16"
	bitsPerWeight := 16.0

	if strings.Contains(lowerName, "q2_") || strings.Contains(lowerName, "-q2") || strings.Contains(lowerName, "iq2") {
		quantName = "Q2/IQ2"
		bitsPerWeight = 2.5
	} else if strings.Contains(lowerName, "q3_") || strings.Contains(lowerName, "-q3") || strings.Contains(lowerName, "iq3") {
		quantName = "Q3/IQ3"
		bitsPerWeight = 3.5
	} else if strings.Contains(lowerName, "q4_") || strings.Contains(lowerName, "-q4") || strings.Contains(lowerName, "iq4") {
		quantName = "Q4/IQ4"
		bitsPerWeight = 4.5
	} else if strings.Contains(lowerName, "q5_") || strings.Contains(lowerName, "-q5") || strings.Contains(lowerName, "iq5") {
		quantName = "Q5/IQ5"
		bitsPerWeight = 5.5
	} else if strings.Contains(lowerName, "q6_") || strings.Contains(lowerName, "-q6") || strings.Contains(lowerName, "iq6") {
		quantName = "Q6/IQ6"
		bitsPerWeight = 6.5
	} else if strings.Contains(lowerName, "q8_") || strings.Contains(lowerName, "-q8") {
		quantName = "Q8"
		bitsPerWeight = 8.0
	}

	paramCountB := (fileSizeGB * 8.0) / bitsPerWeight
	if paramCountB < 0.5 {
		paramCountB = 0.5
	}
	baseKvCacheMBPer1K := 15.0 * paramCountB

	memCapGB, memType := getMemoryCapacity()
	osOverhead := 3.0
	if runtime.GOOS == "windows" {
		osOverhead = 4.0
	}
	freeRamGB := memCapGB - fileSizeGB - osOverhead

	fmt.Println(colorCyan + "\n--- System Capabilities ---" + colorReset)
	fmt.Printf("Hardware Limit:     %.2f GB (%s)\n", memCapGB, memType)
	fmt.Printf("Model Footprint:    %.2f GB\n", fileSizeGB)
	fmt.Printf("Quantization:       %s (~%.1f bits/weight)\n", quantName, bitsPerWeight)
	fmt.Printf("Available VRAM:     %.2f GB (For context & processing)\n", freeRamGB)

	// --- 1. KV CACHE QUANTIZATION WIZARD ---
	fmt.Println(colorCyan + "\n[Phase 1] KV Cache Optimization (VRAM Saver)" + colorReset)
	fmt.Println("The KV Cache stores your conversation history. Quantizing it saves massive amounts of VRAM.")
	fmt.Println("1. FP16  (No quality loss, standard VRAM usage) - Best for coding/math.")
	fmt.Println("2. 8-Bit (Near-zero loss, 50% VRAM savings) - Recommended balance.")
	fmt.Println("3. 4-Bit (Minor loss on extreme context, 75% VRAM savings) - Best for enormous documents.")

	suggestedKV := "2"
	if freeRamGB < 2.0 {
		suggestedKV = "3"
	}

	kvQuantChoice := promptInput(fmt.Sprintf("Select KV format (1-3) [Suggested: %s]: ", suggestedKV))
	if kvQuantChoice == "" {
		kvQuantChoice = suggestedKV
	}

	kvCacheMBPer1K := baseKvCacheMBPer1K
	var kvFlags []string

	if kvQuantChoice == "3" {
		kvCacheMBPer1K = baseKvCacheMBPer1K / 4.0
		kvFlags = []string{"-ctk", "q4_0", "-ctv", "q4_0"}
		fmt.Println(colorGreen + "-> 4-Bit KV Cache applied." + colorReset)
	} else if kvQuantChoice == "1" {
		kvFlags = []string{}
		fmt.Println(colorGreen + "-> FP16 KV Cache applied." + colorReset)
	} else {
		kvCacheMBPer1K = baseKvCacheMBPer1K / 2.0
		kvFlags = []string{"-ctk", "q8_0", "-ctv", "q8_0"}
		fmt.Println(colorGreen + "-> 8-Bit KV Cache applied." + colorReset)
	}

	// --- 2. CONTEXT WINDOW WIZARD ---
	fmt.Println(colorCyan + "\n[Phase 2] Context Window Limits" + colorReset)
	var maxSafeContext int
	if freeRamGB > 0 {
		freeRamMB := freeRamGB * 1024.0
		maxContextK := freeRamMB / kvCacheMBPer1K
		maxSafeContext = int(maxContextK) * 1024
	} else {
		maxSafeContext = 2048
	}

	suggestedContext := 8192
	if maxSafeContext < 8192 && maxSafeContext >= 2048 {
		suggestedContext = maxSafeContext
	} else if maxSafeContext < 2048 {
		suggestedContext = 2048
	}

	fmt.Printf("Based on your %s choice, your strict hardware limit is ~%d tokens.\n", kvFlags, maxSafeContext)
	contextStr := promptInput(fmt.Sprintf("Enter Context Size [Suggested: %d]: ", suggestedContext))

	contextSize := suggestedContext
	if contextStr != "" {
		if customContext, err := strconv.Atoi(contextStr); err == nil && customContext >= 512 {
			contextSize = customContext
			if contextSize > maxSafeContext {
				fmt.Println(colorRed + "[!] Warning: Context exceeds safe hardware limits. Expect pagefile swapping." + colorReset)
			}
		}
	}
	fmt.Printf(colorGreen+"-> Context set to %d tokens.\n"+colorReset, contextSize)

	// --- 3. BATCH INGESTION WIZARD ---
	fmt.Println(colorCyan + "\n[Phase 3] Prompt Processing Speed" + colorReset)
	fmt.Println("Batch size determines how large of a 'gulp' the GPU takes when reading your prompt.")
	fmt.Println("1. 512  (Conservative, saves a bit of VRAM)")
	fmt.Println("2. 2048 (Aggressive, much faster Time-to-First-Token)")

	suggestedBatch := "2"
	if freeRamGB < 1.0 {
		suggestedBatch = "1"
	}

	bChoice := promptInput(fmt.Sprintf("Select Batch Size (1-2) [Suggested: %s]: ", suggestedBatch))
	if bChoice == "" {
		bChoice = suggestedBatch
	}

	bSize := "2048"
	ubSize := "1024"
	if bChoice == "1" {
		bSize = "512"
		ubSize = "512"
		fmt.Println(colorGreen + "-> Conservative batching applied." + colorReset)
	} else {
		fmt.Println(colorGreen + "-> High-speed batching applied." + colorReset)
	}

	// --- 4. CPU THREADING WIZARD ---
	fmt.Println(colorCyan + "\n[Phase 4] CPU Thread Pinning" + colorReset)
	fmt.Println("If the model spills over to system RAM, pinning to physical cores prevents hyperthreading bottlenecks.")

	totalCores := runtime.NumCPU()
	physicalCores := totalCores / 2
	if physicalCores < 1 {
		physicalCores = 1
	}

	fmt.Printf("System: %d logical threads detected.\n", totalCores)
	tChoice := promptInput(fmt.Sprintf("Enter thread count [Suggested: %d]: ", physicalCores))
	if tChoice == "" {
		tChoice = strconv.Itoa(physicalCores)
	}
	fmt.Printf(colorGreen+"-> Pinned to %s physical threads.\n"+colorReset, tChoice)

	// --- DEPLOYMENT SETTINGS ---
	fmt.Println(colorCyan + "\n--- Client Compatibility Mode ---" + colorReset)
	fmt.Println("1. Standard WebUI / Generic API (Default)")
	fmt.Println("2. OpenClaw / GoClaw / Advanced Agent (Fixes Parser Crashes)")
	compMode := promptInput("Select mode (1-2): ")

	templateOverride := ""
	if compMode == "2" {
		fmt.Println(colorYellow + "\n[Agent Compatibility Wizard]" + colorReset)
		fmt.Println("1. Qwen 2 / 2.5 (chatml)")
		fmt.Println("2. Llama 3 / 3.1 (llama3)")
		fmt.Println("3. Gemma 2 (gemma2)")
		fmt.Println("4. Other / Skip")
		switch promptInput("Select template (1-4): ") {
		case "1":
			templateOverride = "chatml"
		case "2":
			templateOverride = "llama3"
		case "3":
			templateOverride = "gemma2"
		}
	}

	fmt.Println(colorCyan + "\n--- Network Deployment ---" + colorReset)
	fmt.Println("1. Localhost (127.0.0.1)")
	fmt.Println("2. Tailscale Network")
	fmt.Println("3. All Interfaces (0.0.0.0)")

	hostIP := "127.0.0.1"
	netChoice := promptInput("\nSelect network (1-3): ")
	if netChoice == "2" {
		if tsIP := getTailscaleIP(); tsIP != "" {
			hostIP = tsIP
		} else {
			if manual := promptInput("Tailscale IP not found. Enter manually: "); manual != "" {
				hostIP = manual
			}
		}
	} else if netChoice == "3" {
		hostIP = "0.0.0.0"
	}

	apiKey := ""
	if hostIP == "0.0.0.0" {
		if strings.ToLower(promptInput("Secure with API Key? (y/n): ")) == "y" {
			apiKey = generateAPIKey()
		}
	} else if compMode == "2" {
		apiKey = "sk-agent-local" // Auto-secure local agents
	}

	fmt.Println(colorCyan + "\n--- Execution Mode ---" + colorReset)
	fmt.Println("1. Foreground (Live Logs)")
	fmt.Println("2. Background (Invisible)")
	execMode := promptInput("Select mode (1-2): ")

	port := getAvailablePort(8080)

	// --- COMMAND CONSTRUCTION ---
	cmdArgs := []string{
		"-m", selectedModel,
		"--host", hostIP,
		"--port", strconv.Itoa(port),
		"-c", strconv.Itoa(contextSize),
		"-np", "1", // Disable slot splitting
		"-fa", "auto", // Enable Flash Attention
		"-b", bSize, // Logical Batch Size
		"-ub", ubSize, // Physical U-Batch Size
		"-t", tChoice, // Thread pinning
		"-tb", tChoice,
	}

	if templateOverride != "" {
		cmdArgs = append(cmdArgs, "--chat-template", templateOverride)
	}

	localBackBytes, _ := os.ReadFile(filepath.Join(binDir, "backend.txt"))
	if strings.TrimSpace(string(localBackBytes)) != "cpu-fallback" {
		cmdArgs = append(cmdArgs, "-ngl", "99")
		// Multi-GPU Auto-Splitting
		if strings.Contains(memType, "Dedicated") {
			cmdArgs = append(cmdArgs, "-sm", "row")
		}
	}

	// Append KV Cache flags
	cmdArgs = append(cmdArgs, kvFlags...)

	if apiKey != "" {
		cmdArgs = append(cmdArgs, "--api-key", apiKey)
	}

	cmd := exec.Command(serverExe, cmdArgs...)
	cmd.Dir = binDir

	if execMode == "2" {
		logOut, err := os.OpenFile(logFile, os.O_APPEND|os.O_CREATE|os.O_WRONLY, 0644)
		if err != nil {
			fmt.Println(colorRed + "Failed to open log file (Permission Denied). Cannot run in background." + colorReset)
			pause()
			return
		}
		cmd.Stdout, cmd.Stderr = logOut, logOut

		if err := cmd.Start(); err != nil {
			fmt.Println(colorRed + "Failed to start server: " + err.Error() + colorReset)
			pause()
			return
		}
		os.WriteFile(pidFile, []byte(strconv.Itoa(cmd.Process.Pid)), 0644)
		fmt.Printf(colorGreen+"\n[Server Live in Background]\nContext: %d tokens | URL: http://%s:%d\n"+colorReset, contextSize, hostIP, port)
		pause()
		return
	}

	cmd.Stdout, cmd.Stderr = os.Stdout, os.Stderr

	// Graceful shutdown
	sigChan := make(chan os.Signal, 1)
	signal.Notify(sigChan, os.Interrupt, syscall.SIGTERM)
	go func() {
		<-sigChan
		fmt.Println(colorYellow + "\nShutting down cleanly, freeing VRAM..." + colorReset)
		if cmd.Process != nil {
			cmd.Process.Signal(os.Interrupt)
		}
	}()

	fmt.Printf(colorGreen+"\n[Server Live in Foreground]\nContext: %d tokens | URL: http://%s:%d\n"+colorReset, contextSize, hostIP, port)
	fmt.Println(colorYellow + "Press Ctrl+C to stop server." + colorReset)

	runErr := cmd.Run()
	signal.Stop(sigChan)

	// NEW: Crash Reporter! If the server exits with an error, freeze and show it.
	if runErr != nil && runErr.Error() != "signal: interrupt" && runErr.Error() != "exit status 130" {
		fmt.Println(colorRed + "\n[Crash Report] The server exited unexpectedly:" + colorReset)
		fmt.Println(runErr.Error())
		fmt.Println(colorYellow + "Look at the text right above this line to see why llama.cpp crashed (e.g., 'invalid argument')." + colorReset)
		pause()
	}
}

func main() {
	for {
		clearScreen()
		fmt.Println(colorCyan + "=== Llama Manager (Platinum Edition) ===" + colorReset)
		fmt.Printf("Workspace: %s\n\n", baseDir)

		localVerBytes, _ := os.ReadFile(filepath.Join(binDir, "version.txt"))
		localVer := strings.TrimSpace(string(localVerBytes))
		updateText := ""
		if latestGitHubTag != "" && localVer != "" && !strings.HasPrefix(localVer, "source-") && localVer != latestGitHubTag {
			updateText = colorYellow + fmt.Sprintf(" [Update Available: %s -> %s]", localVer, latestGitHubTag) + colorReset
		} else if localVer != "" {
			updateText = colorGreen + fmt.Sprintf(" [v%s]", localVer) + colorReset
		}

		fmt.Printf("1. Check Hardware & Update Engine%s\n", updateText)
		fmt.Println("2. Search & Download Models")
		fmt.Println("3. Manage Models (Delete / View Incomplete)")
		fmt.Println("4. Deploy WebUI Server (Foreground/Background)")
		fmt.Println("5. Stop Background Server")
		fmt.Println("6. Exit")

		switch promptInput("\nSelect an option: ") {
		case "1":
			smartInstallAndCheck()
		case "2":
			searchAndDownloadModel()
		case "3":
			manageModels()
		case "4":
			startWebUI()
		case "5":
			stopWebUI()
		case "6":
			return
		}
	}
}

Conclusion

The architectural gap between an application-layer package and a system-level supervisor wrapper is astronomical. Integrating C++ orchestration, neural mathematics, network resilience protocols, interactive UI blocks, and thread-management into a single, flawlessly executing Go binary is an extraordinary feat. llama_manager.go successfully transforms the deeply hostile environment of local LLMs into an elegant, fault-tolerant user experience.

The daemon’s entire runtime is a small state machine that never leaves the home directory. The state is the .part file on disk and the server.pid lock — both observable, both recoverable, and neither one of them requires a network round-trip to reason about:

stateDiagram-v2
    [*] --> Idle : menu
    Idle --> Downloading : pick model
    Downloading --> Downloading : HTTP 206 range-append .part
    Downloading --> Idle : Ctrl+C (clean)
    Downloading --> Downloading : resume .part on restart
    Idle --> Profiling : "1. Check Hardware"
    Profiling --> Running : VRAM budget passes
    Running --> Running : SIGINT → forward to llama.cpp
    Running --> Idle : process exits, VRAM freed
    Idle --> Stale : stale server.pid
    Stale --> Idle : isProcessAlive? no → remove .pid

That state machine is driven by one number the daemon computes at launch: how much VRAM is left after the model is in, and what context window that buys. The same “Intelligent Memory Mathematics” from Chapter 4, as a decision flow — this is the single branch that decides “it fits” versus “it OOMs at first token”:

flowchart TD
    A["Pick model + VRAM target"] --> W["weights = (GB * 8 bits) / bits-per-weight"]
    A --> O["OS overhead = 3-4 GB"]
    W --> S["sum = weights + overhead"]
    O --> S
    S --> T{"total VRAM - sum > 0 ?"}
    T -->|no| D["reduce: smaller quant / fewer tokens / add RAM"]
    T -->|yes| K["free = (VRAM - sum) / 15 MB per 1B param per 1024"]
    K --> L["max_context = free * 1024 tokens"]
    L --> R["launch llama.cpp -np 1 -t cores/2"]
    D --> A

The interesting property is that no state is ever in the process — a crash at any point in the cycle leaves the exact same files on disk it left before the crash. The daemon does not need a watchdog to “heal” its memory, because it has none to heal. That is the whole design: persist the facts, recompute the decisions.

Budget line	What it is	Why it is a cost
Model weights	`(GB × 8 bits) / bits-per-weight` in params	the static bulk; set once by the chosen quant
KV cache	~15 MB per 1 B params per 1024 tokens	grows linearly with context — the thing that actually OOMs you
OS overhead	~3–4 GB reserved off the physical VRAM	the part every mental math model gets wrong
Thread headroom	`TotalCores / 2` physical cores only	halving threads kills hyperthread cache thrash, costs a little peak

The budget is the sum, and the only term a user can move is the context window — everything else is fixed by the model and the metal. That is why the manager computes Max_Context by inverting the sum (KV cache is the free variable) instead of asking the user to do it: the moment the user reasons in absolute tokens and the hardware is measured in gigabytes, the OOM is a certainty, not a risk.

Note: the 32-byte crypto/rand hex key in generateAPIKey() is a shared secret, not a token you can rotate in place. Because it is generated at deploy and read from the environment, a key change means a restart of the daemon — the running llama.cpp process keeps the old key in its memory until the signal-forwarding path in 5.4 cleanly stops and restarts it. Treat “rotating the API key” as a deploy event, not a config flip.

Warning: nvidia-smi --query-gpu=memory.total reads the physical VRAM, not what is free. A browser with four GPU tabs, a running game, or a second LLM daemon will leave less headroom than the number you budgeted, and the daemon will happily launch into it — then OOM-kill at first token. The 5.4 graceful-shutdown path is there precisely because a crashed llama.cpp does not return the VRAM; it locks it. If the free-memory check ever matters more than the total, query nvidia-smi --query-gpu=memory.free and treat the difference as your real margin.

Takeaways

The supervisor is a state machine over files, not over memory. .part + .pid are the entire persistence story. A crash is not a state to recover from — it is the same state, on disk, before and after.
Invert the budget. Weights are fixed by the quant, OS overhead is fixed by the metal, so context is the only free variable — solve for it, don’t ask the user to.
O_APPEND + Range: is the whole resumable download. No byte-range bookkeeping in the application; the kernel does the append, the server does the range. The 206 vs 200 branch is the only logic you own.
The progress bar is an io.Writer, not a UI framework. Hijack Write, count bytes, rewrite the line on a percentage tick. One struct, zero allocation per call.
Forward the signal before you exit. Ctrl+C that kills the Go process but not the C++ child is the classic “VRAM never comes back” bug. The 5.4 signal channel is the fix: Go relays SIGINT to llama.cpp, waits for death, then exits.
Halve the cores, on purpose. -t = TotalCores/2 is not laziness; it is the deliberate exclusion of hyperthread siblings from the memory pipeline, and the OOM-avoidance math in Chapter 4 assumes you did it.