selfware 0.2.2

Your personal AI workshop — software you own, software that lasts
Documentation
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
# Selfware

[![Sponsor](https://img.shields.io/badge/Sponsor-%E2%9D%A4-pink?logo=github)](https://github.com/sponsors/architehc)
[![CI](https://github.com/architehc/selfware/actions/workflows/ci.yml/badge.svg)](https://github.com/architehc/selfware/actions/workflows/ci.yml)
[![Crates.io](https://img.shields.io/crates/v/selfware)](https://crates.io/crates/selfware)
[![Docs.rs](https://docs.rs/selfware/badge.svg)](https://docs.rs/selfware)
[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
[![codecov](https://codecov.io/gh/architehc/selfware/branch/main/graph/badge.svg)](https://codecov.io/gh/architehc/selfware)

```
       /\___/\
      ( o   o )    ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
      (  =^=  )    selfware — Your Personal AI Workshop
       )     (     Software you own. Software that knows you.
      (       )    Software that lasts.
     ( |     | )   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
      \|     |/
```

An **agentic coding harness** for local LLMs that runs entirely on your hardware. 70+ tools, multi-agent swarm, evolution engine, hooks, MCP integration, LSP intelligence, ZED extension, TUI dashboard, and a fox mascot — all local-first, no cloud required.

> **TL;DR** — Point it at any OpenAI-compatible endpoint (vLLM, Ollama, llama.cpp, LM Studio), give it a task, and watch it autonomously read, plan, edit, test, and commit code. Then let the evolution engine improve itself.

---

## What It Looks Like

### Interactive Chat (`selfware chat`)

```
╭─── selfware workshop ────────────────────────────────────╮
│                                                           │
│   /\___/\                                                 │
│  ( o   o )  Welcome to your workshop!                     │
│  (  =^=  )  What shall we tend to today?                  │
│   )     (                                                 │
│                                                           │
│  you> Add unit tests for the auth module                  │
│                                                           │
│  🌿 Planning...                                           │
│  🔍 Reading src/auth/mod.rs                               │
│  ✍️  Writing tests/auth_test.rs                            │
│  🧪 Running cargo test... 12 passed                       │
│  📦 Committing: "Add 12 unit tests for auth module"       │
│                                                           │
│  🌸 BLOOM — Task complete!                                │
╰───────────────────────────────────────────────────────────╯
```

### TUI Dashboard (`selfware --tui`)

```
┌─ Selfware Dashboard ──────────────────────────────────────────────┐
│ ┌─ Agent Status ─────────┐  ┌─ Token Usage ──────────────────┐   │
│ │ State: WORKING         │  │ ████████████░░░░ 75% (37k/50k) │   │
│ │ Tool:  file_edit       │  │ Budget remaining: 13,000 tokens │   │
│ │ Step:  7 / 100         │  └─────────────────────────────────┘   │
│ │ Time:  2m 34s          │                                        │
│ └────────────────────────┘  ┌─ Digital Garden ───────────────┐   │
│ ┌─ Message Stream ───────┐  │ src/                           │   │
│ │ Reading auth/mod.rs... │  │  🌳 mod.rs      [THRIVING]    │   │
│ │ Found 3 functions      │  │  🌿 handler.rs  [GROWING]     │   │
│ │ Writing test file...   │  │  🌱 utils.rs    [SEEDLING]    │   │
│ │ Running tests...       │  │                                │   │
│ └────────────────────────┘  └────────────────────────────────┘   │
└───────────────────────────────────────────────────────────────────┘
```

### Evolution Engine (`selfware evolve`)

```
╭─── Evolution Daemon ──────────────────────────────────────╮
│                                                           │
│  Generation 1 / 3                                         │
│  ├─ Hypothesis 1: Cache token lookups     → 🌸 BLOOM     │
│  ├─ Hypothesis 2: Optimize FIM joining    → 🌸 BLOOM     │
│  ├─ Hypothesis 3: Refactor parse logic    → ❄️  FROST     │
│  └─ Hypothesis 4: Inline hot path         → 🌸 BLOOM     │
│                                                           │
│  SAB Fitness: 50 → 60 (+10.0)                            │
│  Committed: "Gen 1 BLOOM: Cache token lookups"            │
│                                                           │
│  3/4 edits applied · 3/3 compiled · 3/3 tests passed     │
╰───────────────────────────────────────────────────────────╯
```

### Multi-Agent Swarm (`selfware multi-chat`)

```
╭─── Swarm: 4 agents active ───────────────────────────────╮
│                                                           │
│  🏗️  Architect  → Designing module structure              │
│  💻 Coder      → Implementing auth handler                │
│  🧪 Tester     → Writing integration tests                │
│  🔍 Reviewer   → Reviewing PR #42                         │
│                                                           │
│  Progress: ████████████████░░░░ 80%                       │
╰───────────────────────────────────────────────────────────╯
```

> **Screenshots & GIFs**: See the [`docs/`]docs/ directory for full-resolution screenshots and animated GIFs of each mode in action.

---

## Quick Start

### 1. Install Selfware

**Option A: Download prebuilt binary (recommended)**

```bash
# Linux one-liner
ARCH=$(uname -m | sed 's/arm64/aarch64/')
curl -fsSL "https://github.com/architehc/selfware/releases/latest/download/selfware-linux-${ARCH}.tar.gz" | tar -xz
sudo mv selfware /usr/local/bin/

# macOS one-liner
ARCH=$(uname -m | sed 's/arm64/aarch64/')
curl -fsSL -o /tmp/selfware.zip "https://github.com/architehc/selfware/releases/latest/download/selfware-macos-${ARCH}.zip"
unzip -o /tmp/selfware.zip -d /tmp/selfware && sudo mv /tmp/selfware/selfware /usr/local/bin/
```

| Platform | Architecture | Download |
|----------|--------------|----------|
| **Linux** | x86_64 (Intel/AMD) | [selfware-linux-x86_64.tar.gz]https://github.com/architehc/selfware/releases/latest |
| **Linux** | aarch64 (ARM64) | [selfware-linux-aarch64.tar.gz]https://github.com/architehc/selfware/releases/latest |
| **macOS** | Apple Silicon (M1–M4) | [selfware-macos-aarch64.zip]https://github.com/architehc/selfware/releases/latest |
| **macOS** | Intel | [selfware-macos-x86_64.zip]https://github.com/architehc/selfware/releases/latest |
| **Windows** | x86_64 | [selfware-windows-x86_64.zip]https://github.com/architehc/selfware/releases/latest |

**Option B: Install via Cargo**

```bash
cargo install selfware
```

**Option C: Build from source**

```bash
git clone https://github.com/architehc/selfware.git
cd selfware
cargo build --release --all-features
./target/release/selfware --help
```

**Option D: Docker**

```bash
docker build -t selfware .
docker run --rm -it -v $(pwd):/workspace selfware chat
```

### 2. Set Up a Local LLM

Selfware needs an **OpenAI-compatible API endpoint**. Pick any backend:

| Backend | Best For | One-liner |
|---------|----------|-----------|
| **[vLLM]https://docs.vllm.ai/** | Fast inference, GPU servers | `vllm serve Qwen/Qwen3-Coder-Next-FP8` |
| **[Ollama]https://ollama.ai/** | Easy setup, any hardware | `ollama run qwen3.5:4b` |
| **[llama.cpp]https://github.com/ggerganov/llama.cpp** | GGUF models, minimal deps | `./llama-server -m model.gguf -c 65536` |
| **[LM Studio]https://lmstudio.ai/** | GUI, Windows/Mac | Download → load model → start server |
| **[MLX]https://github.com/ml-explore/mlx-examples** | Apple Silicon native | `mlx_lm.server --model mlx-community/Qwen3.5-Coder-35B-A3B-4bit` |
| **[SGLang]https://github.com/architehc/sglang** | High throughput, native tool calling | `python -m sglang.launch_server --model Qwen/Qwen3.5-4B --tool-call-parser qwen --reasoning-parser qwen3` |

> For finding and downloading the best local models, see **[Unsloth Model Zoo]https://unsloth.ai/docs/models/qwen3.5** — they provide optimized quantized versions ready to run.

> **Mac + LM Studio?** See the dedicated **[LM Studio Mac Setup Guide]docs/LM_STUDIO_MAC_GUIDE.md** for step-by-step setup with RAM-based model recommendations.

### 3. Configure

Create `selfware.toml` in your project directory:

```toml
# Your local workshop
endpoint = "http://localhost:8000/v1"    # Your LLM backend
model = "Qwen/Qwen3-Coder-Next-FP8"     # Model name
max_tokens = 65536
temperature = 0.7

[safety]
allowed_paths = ["./**", "/home/*/projects/**"]
denied_paths = ["**/.env", "**/secrets/**"]
protected_branches = ["main"]

[agent]
max_iterations = 100
step_timeout_secs = 600         # 10 min per step

[continuous_work]
enabled = true
checkpoint_interval_tools = 10  # Checkpoint every 10 tool calls
auto_recovery = true

[retry]
max_retries = 5
base_delay_ms = 1000
max_delay_ms = 60000
```

Or use the setup wizard:

```bash
selfware init
```

### 4. Start Coding

```bash
# Interactive chat
selfware chat

# Run a specific task
selfware run "Add unit tests for the auth module"

# Multi-agent mode (4 concurrent agents)
selfware multi-chat

# Analyze your codebase
selfware analyze ./src

# View your code as a living garden
selfware garden

# Full TUI dashboard
selfware --tui
```

---

## Recommended Models & Hardware

### Qwen3.5 — Hardware Requirements

[Qwen3.5](https://unsloth.ai/docs/models/qwen3.5) is highly recommended for selfware. It's a strong coder with excellent instruction following and thinking capabilities. Here are the total **VRAM + RAM** requirements at different quantization levels:

| Qwen3.5 Model | 3-bit | 4-bit | 6-bit | 8-bit | BF16 |
|----------------|-------|-------|-------|-------|------|
| **0.8B + 2B** | 3 GB | 3.5 GB | 5 GB | 7.5 GB | 9 GB |
| **4B** | 4.5 GB | 5.5 GB | 7 GB | 10 GB | 14 GB |
| **9B** | 5.5 GB | 6.5 GB | 9 GB | 13 GB | 19 GB |
| **27B** | 14 GB | 17 GB | 24 GB | 30 GB | 54 GB |
| **35B-A3B** (MoE) | 17 GB | 22 GB | 30 GB | 38 GB | 70 GB |
| **122B-A10B** (MoE) | 60 GB | 70 GB | 106 GB | 132 GB | 245 GB |
| **397B-A17B** (MoE) | 180 GB | 214 GB | 340 GB | 512 GB | 810 GB |

> Source: [Unsloth — Qwen3.5 Inference Requirements]https://unsloth.ai/docs/models/qwen3.5

The **MoE models** (35B-A3B, 122B-A10B, 397B-A17B) only activate a fraction of parameters per token, making them significantly faster at inference despite their large parameter count.

### GPU Servers (vLLM / llama.cpp / SGLang)

| Model | Quant | VRAM | Recommended GPU | Context | SAB Score |
|-------|-------|------|-----------------|---------|-----------|
| **Qwen3-Coder-Next-FP8** | FP8 | 80 GB | H100 / A100 80 GB | 1M | **90/100** (27 rounds) |
| **Qwen3.5-Coder 35B-A3B** | Q4_K_M | 22 GB | **RTX 5090** (32 GB) | 32–128K | Best value |
| **Qwen3.5 27B** | Q4 | 17 GB | RTX 4090 / 3090 (24 GB) | 32–64K | Strong |
| **LFM2 24B-A2B** | 4-bit | 13 GB | RTX 4090 / 3090 (24 GB) | 32–64K | Good |
| **Qwen3.5 9B** | Q4 | 6.5 GB | RTX 4060 Ti (16 GB) | 16–32K | Decent |
| **LFM2.5 1.2B** | Q8 | 1.25 GB | Any GPU | 8–16K | Prototyping |

### Apple Silicon (MLX / Ollama / llama.cpp)

Mac uses unified memory — your total RAM determines what you can run:

| RAM | Recommended Model | Quant | Context | Use Case |
|-----|-------------------|-------|---------|----------|
| **96–128 GB** | Qwen3.5 35B-A3B | Q8 | 64–128K | Full SAB, production coding |
| **64 GB** | Qwen3.5 35B-A3B | Q4_K_M | 32–64K | Most scenarios, good context |
| **32 GB** | Qwen3.5 27B or LFM2 24B-A2B | 4-bit | 16–32K | Everyday coding |
| **24 GB** | Qwen3.5 9B | Q4 | 16–32K | Moderate tasks |
| **16 GB** | Qwen3.5 4B or LFM2.5 1.2B | Q8 | 8–16K | Lightweight, fast feedback |

> **Context window matters.** SAB scenarios work best with >=32K context. Adjust `max_tokens` in `selfware.toml` to match your model's context.

### Quick Setup Examples

```bash
# H100 with vLLM (reference setup, 90/100 SAB)
vllm serve Qwen/Qwen3-Coder-Next-FP8 --max-model-len 131072

# RTX 5090 with Qwen3.5 35B MoE (llama.cpp)
./llama-server -m qwen3.5-coder-35b-a3b-q4_k_m.gguf \
  -c 65536 -ngl 99 --port 8000

# RTX 4090 / 3090 with SGLang (recommended — native tool calling)
python -m sglang.launch_server --model-path Qwen/Qwen3.5-4B \
  --context-length 131072 --kv-cache-dtype fp8_e4m3 \
  --reasoning-parser qwen3 --tool-call-parser qwen --port 8000

# RTX 4090 with Qwen3.5 27B (vLLM)
vllm serve Qwen/Qwen3.5-27B-AWQ --max-model-len 32768

# Mac M2/M3/M4 with MLX
mlx_lm.server --model mlx-community/Qwen3.5-Coder-35B-A3B-4bit \
  --port 8000

# Any machine with Ollama
ollama run qwen3.5:4b

# Ultra-light (CPU or weak GPU)
ollama run qwen3.5:0.8b
```

### SGLang

SGLang provides native tool calling support with `--tool-call-parser qwen` and `--reasoning-parser qwen3`, which is the recommended way to run Qwen models with selfware. This gives you proper OpenAI-compatible function calling instead of XML-based parsing.

**Single RTX 4090 / 3090 (24 GB) — Qwen3.5-4B:**

```bash
python -m sglang.launch_server \
  --model-path Qwen/Qwen3.5-4B \
  --trust-remote-code \
  --tensor-parallel-size 1 \
  --context-length 131072 \
  --attention-backend flashinfer \
  --mem-fraction-static 0.90 \
  --max-running-requests 32 \
  --chunked-prefill-size 8192 \
  --max-prefill-tokens 65536 \
  --kv-cache-dtype fp8_e4m3 \
  --disable-custom-all-reduce \
  --cuda-graph-max-bs 8 \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen \
  --port 8000 \
  --host 0.0.0.0
```

**Single RTX 4090 / 3090 — Qwen3.5-9B (Q8):**

```bash
python -m sglang.launch_server \
  --model-path Qwen/Qwen3.5-9B \
  --trust-remote-code \
  --tensor-parallel-size 1 \
  --context-length 65536 \
  --attention-backend flashinfer \
  --mem-fraction-static 0.90 \
  --max-running-requests 16 \
  --chunked-prefill-size 8192 \
  --max-prefill-tokens 32768 \
  --kv-cache-dtype fp8_e4m3 \
  --disable-custom-all-reduce \
  --cuda-graph-max-bs 8 \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen \
  --port 8000 \
  --host 0.0.0.0
```

**Dual RTX 4090 — Qwen3-VL-30B-A3B (vision + tool calling):**

```bash
python -m sglang.launch_server \
  --model-path Qwen/Qwen3-VL-30B-A3B-Thinking-FP8 \
  --trust-remote-code \
  --tensor-parallel-size 2 \
  --enable-multimodal \
  --context-length 131072 \
  --attention-backend flashinfer \
  --mem-fraction-static 0.85 \
  --max-running-requests 32 \
  --chunked-prefill-size 8192 \
  --max-prefill-tokens 65536 \
  --kv-cache-dtype fp8_e5m2 \
  --disable-custom-all-reduce \
  --cuda-graph-max-bs 32 \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen \
  --port 8000 \
  --host 0.0.0.0
```

> **Tip:** When using SGLang with tool calling, set `native_function_calling = true` in your `selfware.toml`. This uses OpenAI-compatible function calling instead of XML parsing, which is more reliable with small models.

### vLLM

```bash
vllm serve Qwen/Qwen3.5-4B --port 8000 --tensor-parallel-size 1 \
  --max-model-len 262144 --reasoning-parser qwen3 \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder
```

### llama.cpp

**Kimi K2.5 Thinking on RTX 6000 Pro (96 GB VRAM + 1 TB RAM):**

```bash
llama.cpp/build/bin/llama-server \
    --model models/unsloth/Kimi-K2-Thinking-GGUF/Q4_K_S/Kimi-K2-Thinking-Q4_K_S-00001-of-00013.gguf \
    --alias "unsloth/Kimi-K2-Thinking" \
    --threads -1 \
    --n-gpu-layers 1999 \
    --temp 1 \
    --min_p 0.01 \
    --ctx-size 198304 \
    --seed 3407 \
    -fa on \
    --cache-type-k q4_0 \
    --cache-type-v q4_0 \
    --port 8000 \
    -ot ".ffn_.*_exps.=CPU" \
    --chat-template chatml \
    --special
```

**Qwen3.5-122B-A10B on RTX 6000 Pro:**

```bash
LLAMA_SET_ROWS=1 llama.cpp/build/bin/llama-server \
    --model models/models/unsloth/Qwen3.5-122B-A10B-GGUF/UD-Q4_K_XL/Qwen3.5-122B-A10B-UD-Q4_K_XL-00001-of-00003.gguf \
    --mmproj models/unsloth/Qwen3.5-122B-A10B-GGUF/mmproj-F16.gguf \
    --alias "Qwen3.5-122B-A10B" \
    --threads 64 \
    --n-gpu-layers 999 \
    --ctx-size 2097152 \
    --seed 3407 \
    -fa on \
    --cache-type-k q4_0 \
    --cache-type-v q4_0 \
    --port 8000 \
    --special \
    --chat-template chatml \
    -np 16
```

### LM Studio

LM Studio provides a GUI for running local models on Mac and Windows.

> **Important:** Set the Prompt Template to **Manual → ChatML** (not Jinja) to ensure tool calling works correctly. See the full **[LM Studio Mac Setup Guide]docs/LM_STUDIO_MAC_GUIDE.md** for detailed instructions.

Enable **KV cache quantization** (set to Q4) to fit larger context windows in limited RAM.

---

## Features

### 70+ Built-in Tools

Selfware gives the LLM a full toolkit for autonomous coding:

| Category | Tools | Examples |
|----------|-------|---------|
| **File Tending** | Read, write, edit, search, tree | `file_read`, `file_write`, `file_edit`, `directory_tree` |
| **Git Cultivation** | Status, diff, commit, branch, log | `git_status`, `git_diff`, `git_commit`, `git_checkpoint` |
| **Cargo Workshop** | Test, check, clippy, fmt, build | `cargo_test`, `cargo_check`, `cargo_clippy`, `cargo_fmt` |
| **Code Foraging** | Grep, glob, symbol search | `grep_search`, `glob_find`, `symbol_search` |
| **Shell** | Execute commands with safety checks | `shell_exec` |
| **PTY Shell** | Persistent interactive sessions | `pty_shell` |
| **Analysis** | AST parsing, complexity, BM25 | `code_analysis`, `bm25_search` |
| **Knowledge** | Web fetch, documentation lookup | `web_fetch`, `knowledge_query` |
| **FIM Editing** | Fill-in-the-Middle AI code replacement | `file_fim_edit` |
| **Computer Control** | Mouse, keyboard, screen, window management | `computer_mouse`, `computer_keyboard`, `computer_screen`, `computer_window` |
| **LSP** | Semantic code intelligence | `lsp_goto_definition`, `lsp_find_references`, `lsp_document_symbols`, `lsp_hover` |
| **Browser Automation** | 28-action Playwright controller | `page_control` |
| **MCP Server** | Expose selfware tools to other AI systems | `selfware mcp-server` |

### Multi-Agent Swarm

Up to **16 concurrent agents** with role specialization:

```bash
selfware multi-chat -n 8
```

Roles: **Architect**, **Coder**, **Tester**, **Reviewer**, **DevOps**, **Security** — each with its own context and tool access. The swarm coordinator distributes tasks and merges results.

### Task Persistence & Recovery

Tasks survive crashes via automatic checkpointing:

```bash
# Start a long task
selfware run "Refactor the entire authentication system"

# Power outage? System crash? No problem.
selfware journal          # Browse saved checkpoints
selfware resume <task-id> # Pick up exactly where you left off
```

### Cognitive Architecture

The agent thinks in PDVR cycles with working memory:

```
    ╭─────────╮         ╭─────────╮
    │  PLAN   │────────▶│   DO    │
    ╰─────────╯         ╰─────────╯
         ▲                    │
         │                    ▼
    ╭─────────╮         ╭─────────╮
    │ REFLECT │◀────────│ VERIFY  │
    ╰─────────╯         ╰─────────╯
```

**Working Memory** tracks current plan, active hypothesis, open questions, and discovered facts. **Episodic Memory** learns from past sessions — what worked, your preferences, project patterns.

### Multi-Layer Safety

```
Request → Path Guardian → Command Sentinel → Protected Groves → Execute
```

- **Path validation**: Allowed/denied path globs, no escape from workspace
- **Command filtering**: Dangerous commands blocked by default
- **Protected branches**: Prevent force-push to main
- **SSRF protection**: URL validation on web requests
- **Evolution safety**: Cannot modify its own fitness function, SAB suite, or safety module

### Warm Terminal Aesthetic

Four color themes for your workshop:

| Theme | Style | Flag |
|-------|-------|------|
| **Amber** (default) | Warm amber, soil brown, garden green | `--theme amber` |
| **Ocean** | Cool blues and teals | `--theme ocean` |
| **Minimal** | Clean grayscale | `--theme minimal` |
| **High Contrast** | Accessibility-focused | `--theme high-contrast` |

Status messages use garden metaphors:
- **BLOOM** — Success, fresh growth
- **GROW** — Progress, on the right track
- **WILT** — Warning, needs attention
- **FROST** — Error, needs warmth

### Hooks System

Event-driven automation with three hook points: **PreToolUse**, **PostToolUse**, and **Stop**. Built-in presets for auto-commit, auto-format, and lint-on-edit. Configure hooks in `selfware.toml` or toggle them at runtime with `/hooks`.

```toml
[hooks]
enabled = true
presets = ["auto-commit", "auto-format", "lint-on-edit"]
```

### MCP Integration

Selfware supports the **Model Context Protocol** as both client and server. Connect to external MCP servers (GitHub, Playwright, databases) to extend the agent's capabilities, or expose selfware's own tools to other AI systems via `selfware mcp-server`.

```toml
[mcp]
servers = [
    { name = "github", command = "npx", args = ["-y", "@modelcontextprotocol/server-github"] },
    { name = "playwright", command = "npx", args = ["-y", "@playwright/mcp-server"] },
]
```

### LSP Integration

Semantic code intelligence via language servers (rust-analyzer, pyright, tsserver, gopls). Go-to-definition, find-references, document-symbols, and hover information are all available as agent tools, giving the LLM deep understanding of code structure.

### Doctor Mode

`selfware doctor` checks 30+ system dependencies (git, cargo, rustc, node, python, docker, etc.) and reports what is available. `selfware llm-doctor` analyzes your LLM backend, model configuration, template setup, and gives optimization recommendations.

### Interview Mode

Structured pre-task questions (language, framework, scope, testing preference) with smart defaults and auto-detection. Launch with `selfware chat --interview` to guide the agent before it begins work.

### Claude Code-like UI

ESC to interrupt generation, fixed input line for typing anytime, work queue with delayed execution (`@5m run tests`), and full input history. The interactive experience is designed to feel responsive even with slow local models.

### Visual Verification

VLM-powered screenshot analysis for UI testing. The agent can capture screenshots and use a vision-language model to verify that UI changes look correct.

### Active Selections

Guided wizard with recommendations for project template, architecture, database, testing framework, and deployment strategy. The agent walks you through choices with opinionated defaults.

### Swarm Visualization

Terminal panels showing agent status, consensus log, and activity timeline for multi-agent swarm sessions. See what each agent is doing in real time.

### Inline Diff Viewer

Colored unified and side-by-side diffs with word-level highlighting before applying edits. Review every change the agent proposes before it touches your code.

### ZED Extension

IDE integration via the ZED editor extension (WASM-based). Use selfware directly from ZED with full tool access.

---

## Evolution Engine — Recursive Self-Improvement

The evolution engine is selfware's most unique feature: it uses an LLM to generate code improvements to itself, then verifies them through compilation and testing. Only improvements that pass `cargo check` + `cargo test` survive.

### How It Works

```
┌──────────────┐     ┌──────────────┐     ┌──────────────┐
│  Generate    │────▶│  Apply       │────▶│  Verify      │
│  Hypotheses  │     │  Edits       │     │  (compile +  │
│  (LLM call)  │     │  (search/    │     │   test)      │
└──────────────┘     │   replace)   │     └──────┬───────┘
       ▲             └──────────────┘            │
       │                                         ▼
┌──────┴───────┐                          ┌──────────────┐
│  History +   │◀─────────────────────────│  Select or   │
│  Telemetry   │     fitness improved?    │  Rollback    │
└──────────────┘                          └──────────────┘
```

1. **Generate**: LLM reads your mutation target files and proposes N hypotheses as search-and-replace edits
2. **Apply**: Each hypothesis's edits are applied with fuzzy whitespace matching
3. **Verify**: `cargo check``cargo test` — if either fails, the hypothesis is rolled back
4. **Select**: If all tests pass and fitness improves, the change is committed as a **BLOOM**

### Running Evolution

```bash
# Build with the self-improvement feature
cargo build --release --features self-improvement

# Run 3 generations with 4 hypotheses each
./target/release/selfware evolve --generations 3 --population 4

# Dry run (show config, don't execute)
./target/release/selfware evolve --dry-run
```

### Configuring Mutation Targets

In `selfware.toml`, specify which files the evolution engine is allowed to modify:

```toml
[evolution]
# Prompt construction logic
prompt_logic = [
    "src/agent/planning.rs",
    "src/agent/loop_control.rs",
    "src/orchestration/planning.rs",
]

# Tool implementations
tool_code = [
    "src/tools/file.rs",
    "src/tools/search.rs",
    "src/tool_parser.rs",
]

# Cognitive architecture
cognitive = [
    "src/cognitive/memory_hierarchy.rs",
    "src/cognitive/episodic.rs",
    "src/memory.rs",
]

# Config keys the agent may tune
config_keys = ["temperature", "max_tokens", "token_budget"]
```

### Safety Invariants

The evolution engine **CANNOT** modify:

| Protected Path | Reason |
|----------------|--------|
| `src/evolution/` | Cannot modify its own evolution logic |
| `src/safety/` | Cannot weaken safety checks |
| `system_tests/` | Cannot modify its own benchmark suite |
| `benches/sab_*` | Cannot game fitness measurements |

These are enforced at the code level via `PROTECTED_PATHS` in `src/evolution/mod.rs`.

### Evolution Output

The engine writes a JSONL event log to `.evolution-log.jsonl` for every generation:

```jsonl
{"event":"generation_start","generation":1,"timestamp":"2026-03-04T12:25:00Z"}
{"event":"hypothesis_result","generation":1,"hypothesis":"Cache token lookups","applied":true,"compiled":true,"tests_passed":true,"rating":"BLOOM"}
{"event":"generation_end","generation":1,"blooms":3,"frosts":1,"fitness_delta":10.0}
```

Successful improvements are auto-committed to the repo with descriptive messages:

```
Gen 1 BLOOM: Cache token lookups in FIM string joining
Gen 2 BLOOM: Optimize search-replace dispatch
Gen 3 BLOOM: Inline hot path in token counter
```

---

## SAB — Selfware Agentic Benchmark

A **12-scenario agentic coding benchmark** that measures how well a local LLM can autonomously fix bugs, write tests, refactor code, and optimize performance through selfware's agent loop.

### Scenarios

| Difficulty | Scenario | What It Tests |
|------------|----------|---------------|
| Easy | `easy_calculator` | Simple arithmetic bug fixes (3–4 bugs) |
| Easy | `easy_string_ops` | String manipulation bugs |
| Medium | `medium_json_merge` | JSON deep merge logic |
| Medium | `medium_bitset` | Bitwise operations and edge cases |
| Medium | `testgen_ringbuf` | Write 15+ tests for an untested ring buffer |
| Medium | `refactor_monolith` | Split a 210-line monolith into 4 modules |
| Hard | `hard_scheduler` | Multi-file scheduler with duration parsing |
| Hard | `hard_event_bus` | Event system with async subscribers |
| Hard | `security_audit` | Replace 5 vulnerable functions with secure alternatives |
| Hard | `perf_optimization` | Fix 5 O(n^2)/exponential algorithms |
| Hard | `codegen_task_runner` | Implement 12 `todo!()` method stubs |
| Expert | `expert_async_race` | Fix 4 concurrency bugs in a Tokio task pool |

### Scoring

Each scenario scores 0–100:
- **70 pts** — all tests pass after agent edits
- **20 pts** — agent also fixes intentionally broken tests
- **10 pts** — clean exit (no crash, no timeout)

Round ratings: **BLOOM** (>=85) · **GROW** (>=60) · **WILT** (>=30) · **FROST** (<30)

### Benchmark Results — Qwen3-Coder-Next-FP8

Tested on RTX 6000 Pro via custom [SGLang](https://github.com/architehc/sglang), 6 parallel scenarios, 27 rounds (323 scenario runs):

| Metric | Value |
|--------|-------|
| Steady-state average (R2–R27) | **90/100** |
| Peak phase (R9–R27) | **91/100** |
| Best round | **96/100** (achieved 8 times) |
| Perfect rounds (12/12 pass) | **16 out of 27** |
| BLOOM rounds (>=85) | **22 out of 27** |
| S-tier scenarios (100% reliable) | 5 of 12 |

<details>
<summary>Full round-by-round results</summary>

| Round | Score | Rating | Passed |
|-------|-------|--------|--------|
| R1 | 60/100 | GROW | 7/11 |
| R2 | 96/100 | BLOOM | 12/12 |
| R3 | 70/100 | GROW | 9/12 |
| R4 | 87/100 | BLOOM | 11/12 |
| R5 | 79/100 | GROW | 10/12 |
| R6 | 81/100 | GROW | 10/12 |
| R7 | 87/100 | BLOOM | 11/12 |
| R8 | 89/100 | BLOOM | 11/12 |
| R9 | 95/100 | BLOOM | 12/12 |
| R10 | 95/100 | BLOOM | 12/12 |
| R11 | 96/100 | BLOOM | 12/12 |
| R12 | 87/100 | BLOOM | 11/12 |
| R13 | 96/100 | BLOOM | 12/12 |
| R14 | 88/100 | BLOOM | 11/12 |
| R15 | 95/100 | BLOOM | 12/12 |
| R16 | 95/100 | BLOOM | 12/12 |
| R17 | 95/100 | BLOOM | 12/12 |
| R18 | 96/100 | BLOOM | 12/12 |
| R19 | 96/100 | BLOOM | 12/12 |
| R20 | 96/100 | BLOOM | 12/12 |
| R21 | 89/100 | BLOOM | 11/12 |
| R22 | 87/100 | BLOOM | 11/12 |
| R23 | 96/100 | BLOOM | 12/12 |
| R24 | 87/100 | BLOOM | 11/12 |
| R25 | 90/100 | BLOOM | 11/12 |
| R26 | 95/100 | BLOOM | 12/12 |
| R27 | 73/100 | GROW | 9/12 |

</details>

### Scenario Reliability

| Tier | Scenarios | Pass Rate |
|------|-----------|-----------|
| **S** (100%) | `easy_calculator`, `easy_string_ops`, `medium_json_merge`, `perf_optimization`, `codegen_task_runner` | 100% |
| **A** (>80%) | `hard_scheduler`, `hard_event_bus`, `expert_async_race`, `medium_bitset` | 89–96% |
| **B** (50–80%) | `security_audit`, `testgen_ringbuf`, `refactor_monolith` | 70–74% |

### Running Your Own Benchmark

```bash
export ENDPOINT="http://localhost:8000/v1"
export MODEL="Qwen/Qwen3-Coder-Next-FP8"
export MAX_PARALLEL=6

bash system_tests/projecte2e/run_full_sab.sh

# Results in system_tests/projecte2e/reports/<timestamp>/
```

---

## CLI Reference

| Command | Alias | Description |
|---------|-------|-------------|
| `selfware chat` | `c` | Interactive chat session |
| `selfware multi-chat` | `m` | Multi-agent swarm chat |
| `selfware run <task>` | `r` | Execute a specific task |
| `selfware analyze <path>` | `a` | Survey codebase structure |
| `selfware garden` | | View code as a digital garden |
| `selfware journal` | `j` | Browse checkpoint entries |
| `selfware resume <id>` | | Resume from checkpoint |
| `selfware status` | | Show workshop stats |
| `selfware workflow <file>` | `w` | Run a YAML workflow |
| `selfware init` | | Setup wizard |
| `selfware evolve` | | Run evolution engine* |
| `selfware improve` | | Self-improvement pass* |
| `selfware doctor` | | System dependency check |
| `selfware mcp-server` | | Run as MCP server |
| `selfware lsp` | | Run as LSP server (stub) |
| `selfware demo` | | Run animated demo** |
| `selfware dashboard` | | Launch TUI dashboard** |

\* Requires `--features self-improvement`
\*\* Requires `--features tui`

### Global Flags

| Flag | Description |
|------|-------------|
| `-p <PROMPT>` | Headless mode: run prompt and exit |
| `-C <DIR>` | Set working directory |
| `-m <MODE>` | Execution mode: `normal`, `auto-edit`, `yolo`, `daemon` |
| `-y` | Shortcut for `--mode=yolo` |
| `--tui` | Launch TUI dashboard |
| `--theme <THEME>` | Color theme: `amber`, `ocean`, `minimal`, `high-contrast` |
| `--compact` | Dense output, less chrome |
| `-v, --verbose` | Detailed tool output |
| `--show-tokens` | Display token usage after each response |
| `--ascii` | ASCII-only output (no emoji) |
| `--plan` | Plan mode (read-only, no edits) |
| `--resume-session <name>` | Resume a named session |
| `--interview` | Pre-task interview mode |
| `--no-color` | Disable colored output |

### Environment Variables

| Variable | Description | Default |
|----------|-------------|---------|
| `SELFWARE_ENDPOINT` | LLM API endpoint | `http://localhost:8000/v1` |
| `SELFWARE_MODEL` | Model name | `Qwen/Qwen3-Coder-Next-FP8` |
| `SELFWARE_API_KEY` | API key (if required) | None |
| `SELFWARE_MAX_TOKENS` | Max tokens per response | `65536` |
| `SELFWARE_TEMPERATURE` | Sampling temperature | `0.7` |
| `SELFWARE_TIMEOUT` | Request timeout (seconds) | `600` |
| `SELFWARE_DEBUG` | Enable debug logging | Disabled |
| `SELFWARE_ASCII` | Force ASCII-only mode | Disabled |
| `NO_COLOR` | Disable colors (standard) | Disabled |

### Interactive Commands

During a chat session, use slash commands to control the agent:

| Command | Description |
|---------|-------------|
| `/plan` | Toggle plan mode (read-only, no edits) |
| `/think` | Toggle extended thinking |
| `/hooks` | Toggle hook presets on/off |
| `/queue` | View and manage the work queue |
| `/interview` | Run the pre-task interview |

---

## Slow Model Support

Designed for local LLMs on consumer hardware. The agent will wait patiently:

```
Model Speed          Timeout Setting
─────────────────────────────────────
> 10 tok/s           300s (5 min)
1-10 tok/s           3600s (1 hour)
< 1 tok/s            14400s (4 hours)
0.08 tok/s           Works! Be patient.
```

---

## Project Structure

```
src/
├── agent/          Core agent logic, checkpointing, execution
├── tools/          70+ tool implementations (file, git, cargo, search, shell, FIM, computer, LSP, browser)
├── api/            LLM client with timeout, retry, streaming
├── ui/             Terminal aesthetic (themes, animations, banners, fox mascot)
│   ├── tui/        Full ratatui dashboard (garden view, swarm widgets, particles)
│   ├── task_display.rs   Task progress display
│   ├── diff_viewer.rs    Inline colored diff viewer (unified + side-by-side)
│   ├── input_handler.rs  Claude Code-like input handling (ESC, history, queue)
│   ├── selections.rs     Active selection wizard
│   └── swarm_viz.rs      Swarm visualization panels
├── analysis/       Code analysis, BM25 search, vector store
├── cognitive/      PDVR cycle, working/episodic memory, RAG, token budget
├── config/         Configuration management (TOML + env + CLI)
├── hooks/          Event-driven hook system (PreToolUse, PostToolUse, Stop)
├── mcp/            MCP client + server (JSON-RPC 2.0, stdio transport)
├── lsp/            LSP client (rust-analyzer, pyright, tsserver, gopls)
├── computer/       Desktop automation (mouse, keyboard, screen capture, window management)
├── devops/         Container support, process manager
├── evolution/      Recursive self-improvement engine (feature-gated)
│   ├── daemon.rs   Main evolution loop + LLM hypothesis generation
│   ├── fitness.rs  SAB-based fitness scoring
│   ├── sandbox.rs  Isolated evaluation environments
│   └── tournament.rs  Parallel hypothesis evaluation
├── observability/  OpenTelemetry tracing, Prometheus metrics
├── orchestration/  Multi-agent swarm, planning, workflows
├── safety/         Path validation, command filtering, sandboxing, JSONL audit logging
├── self_healing/   Error classification, recovery, exponential backoff
├── session/        Checkpoint persistence
├── testing/        Verification, contract testing, workflow DSL, multi-language QA
├── doctor.rs       System dependency diagnostics (30+ checks)
├── llm_doctor.rs   LLM configuration diagnostics and optimization
├── interview.rs    Pre-task interview with smart defaults
├── memory.rs       Memory management
├── tool_parser.rs  Robust multi-format XML parser
└── token_count.rs  Token estimation

zed-extension/      ZED editor extension (WASM-based)
docs/               User documentation (8 guides)
```

---

## Development

### Run Tests

```bash
# All tests (6,000+ tests)
cargo test --all-features

# Quick unit tests only
cargo test --lib --all-features

# Evolution engine tests (95 tests)
cargo test --features self-improvement evolution::
cargo test --features self-improvement --test evolution_integration_test

# With resilience features
cargo test --features resilience

# Integration tests with real LLM
cargo test --features integration
```

### Test Coverage

| Metric | Value |
|--------|-------|
| **Total Tests** | 6,000+ |
| **Line Coverage** | ~75% |
| **Test Targets** | lib + external + integration + doc + property |

### Code Quality

```bash
cargo clippy --all-features -- -D warnings
cargo fmt -- --check
cargo llvm-cov --lib --all-features --summary-only
```

---

## Documentation

Full guides are available in the [`docs/`](docs/) directory:

| Guide | Description |
|-------|-------------|
| [Getting Started]docs/GETTING_STARTED.md | Installation, first run, basic usage |
| [Configuration]docs/CONFIGURATION.md | All config options, TOML reference |
| [Tools Reference]docs/TOOLS_REFERENCE.md | Complete tool catalog with examples |
| [Interactive Commands]docs/INTERACTIVE_COMMANDS.md | Slash commands and shortcuts |
| [Hooks]docs/HOOKS.md | Event-driven automation setup |
| [MCP]docs/MCP.md | MCP client and server configuration |
| [Doctor]docs/DOCTOR.md | System and LLM diagnostics |
| [ZED Extension]docs/ZED_EXTENSION.md | IDE integration via ZED |

---

## Troubleshooting

**"Connection refused"** — Is your LLM backend running?
```bash
curl http://localhost:8000/v1/models
```

**"Request timeout"** — Increase timeout for slow models:
```toml
[agent]
step_timeout_secs = 14400  # 4 hours
```

**"Safety check failed"** — Check `allowed_paths` in your config. The agent only accesses paths you permit.

**Evolution produces no BLOOMs** — Common causes:
- Model response truncated → increase `max_tokens` in config
- Thinking mode consuming tokens → the engine disables it automatically with `/no_think`
- Patch context mismatch → the engine uses fuzzy whitespace matching to handle this

---

## License

MIT License

## Sponsors

Sponsored by **[Trebuchet Network](https://trebuchet.network)**

## Acknowledgments

- Built for [Qwen3-Coder]https://qwenlm.github.io/, [Kimi K2.5]https://kimi.moonshot.cn/, [LFM2]https://www.liquid.ai/, and other local LLMs
- Model downloads and quantizations via [Unsloth]https://unsloth.ai/docs/models/qwen3.5
- Inspired by the [AiSocratic]https://aisocratic.org/ movement
- UI philosophy: software should feel like a warm workshop, not a cold datacenter

---

```
    "Tend your garden. The code will grow."
                                    — selfware proverb
```