zc2 0.0.46

P2P compute broker with yen-denominated billing, WAL, and broker mesh support
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
# zc2 Usage Guide

This guide covers the complete usage of `zc` (built from `zc2`), the Zakuro AI command-line tool. `zc` is a P2P compute broker that routes compute requests to workers across a distributed network, with yen-denominated billing settled through the Zakuro hub HTTP API (ADR ZC-0005), P2P broker-to-broker communication, write-ahead logging, automatic worker discovery, and real-time monitoring.

---

## Table of Contents

- [Prerequisites](#prerequisites)
- [Installation](#installation)
- [Quick Start](#quick-start)
- [Broker](#broker)
  - [Starting the Broker](#starting-the-broker)
  - [Foreground Mode](#foreground-mode)
  - [Daemon Mode](#daemon-mode)
- [Worker Discovery](#worker-discovery)
  - [Mesh Mode](#mesh-mode)
  - [Local Mode](#local-mode)
  - [Manual Registration](#manual-registration)
- [API Reference](#api-reference)
  - [Health & Status](#health--status)
  - [Worker Management](#worker-management)
  - [Compute Execution](#compute-execution)
  - [Balance Management](#balance-management)
  - [Ledger Administration](#ledger-administration)
  - [P2P Broker Endpoints](#p2p-broker-endpoints)
- [Routing Strategies](#routing-strategies)
- [Benchmarking](#benchmarking)
  - [Basic Benchmark](#basic-benchmark)
  - [Strategy Comparison](#strategy-comparison)
  - [Benchmark Options](#benchmark-options)
- [Remote Monitoring](#remote-monitoring)
- [System Diagnostics](#system-diagnostics)
- [Balance System](#balance-system)
  - [Hub Ledger (Credit Authority)](#hub-ledger-credit-authority)
  - [Reserve-Commit Model](#reserve-commit-model)
  - [API Keys](#api-keys)
  - [P2P Broker-to-Broker Communication](#p2p-broker-to-broker-communication)
  - [Write-Ahead Log (WAL)](#write-ahead-log-wal)
- [Cluster Commands](#cluster-commands)
- [Zakuro Drive Hooks](#zakuro-drive-hooks)
- [Environment Variables](#environment-variables)
- [Configuration Reference](#configuration-reference)

---

## Prerequisites

- **Rust** toolchain (recent stable; the dependency tree's MSRV is 1.88) for building from source
- **Docker** and **Docker Compose** (for containerized deployments)
- **WireGuard** (optional, for P2P network discovery)
- `ZAKURO_API_KEY` environment variable must be set

## Installation

### From Source

```bash
git clone https://github.com/zakuro-ai/zc.git
cd zc
task build:bin
```

This compiles the release binary and installs it to `/usr/local/bin/zc`.

### With Docker

```bash
task build
```

This builds the binary, then builds and starts the Docker services via `docker compose`.

### Verify Installation

```bash
zc --version
```

Prints the version, the commit, and the SHA-256 of the binary itself — the
same hash each broker discloses, so a local install can be matched against
what is running on the mesh.

---

## Quick Start

1. Set your authentication token:

```bash
export ZAKURO_API_KEY="your-token-here"
```

2. Start the broker:

```bash
zc broker
```

3. The broker will automatically discover workers on your network and begin accepting compute requests on port `9000`.

---

## Broker

The broker is the core component of `zc`. It accepts compute requests from clients, selects an optimal worker, forwards the request, and handles billing.

### Starting the Broker

```bash
# Default: foreground mode on 0.0.0.0:9000
zc broker

# Custom port
zc broker 8080

# Custom host and port
zc broker 0.0.0.0 9000
```

### Foreground Mode

```bash
zc broker
```

Foreground mode displays a startup banner listing all API endpoints and then prints a live transaction log to stdout. Each request is shown with its status, user, cost, and latency. Press `Ctrl+C` to stop.

Example output:

```
  ╔═══════════════════════════════════════════╗
  ║          Zakuro Compute Broker            ║
  ╚═══════════════════════════════════════════╝

  Listening:  http://0.0.0.0:9000

  API Endpoints:
  ─────────────────────────────────────────────
    GET  /health              Health check
    GET  /workers             List workers
    POST /workers             Register worker
    POST /workers/heartbeat   Worker heartbeat
    DEL  /workers/:id         Unregister worker
    POST /execute             Execute request
    POST /price               Estimate price
    GET  /balance/:user       Get balance
    POST /balance/:user/add   Add funds
  ─────────────────────────────────────────────

  Press Ctrl+C to stop.
```

### Daemon Mode

```bash
zc -d broker
zc --daemon broker
zc -d broker 8080     # Daemon on custom port
```

Daemon mode runs the broker in the background with minimal output. Useful for production deployments or running the broker as a service.

---

## Worker Discovery

The broker automatically discovers compute workers based on the available network interfaces.

### Mesh Mode

**Trigger:** A `zakuro0` or `wg*` interface is detected holding an address in `10.13.13.0/24`, or `ZAKURO_WIREGUARD_IP` is set. See `get_mesh_ip` in `src/broker/discovery.rs`.

**Behavior:**

- Scans the mesh subnet (`10.13.13.2` through `10.13.13.254`) on port `3960` (configurable via `ZAKURO_WORKER_PORT`)
- Also checks extra ports: `3961`, `3962`
- Probes each address for `/health` and `/info` endpoints
- Only registers services that return a valid `worker_type` in their `/info` response
- Refreshes heartbeats for existing workers every 15 seconds (configurable via `ZAKURO_SCAN_INTERVAL`)
- **Balances are enforced** in this mode — all executions are billed

### Local Mode

**Trigger:** No mesh interface is detected.

**Behavior:**

- Scans `127.0.0.1` on ports `3960`, `3961`, `3962` (or `ZAKURO_SCAN_RANGE` if set)
- Same probing logic as mesh mode
- **Execution is free** — no money is charged
- Useful for development and testing

### Manual Registration

Workers can always register themselves via the REST API regardless of discovery mode. When `ZAKURO_WORKER_KEY` is configured, include the key in the `X-Worker-Key` header:

```bash
curl -X POST http://localhost:9000/workers \
  -H "Content-Type: application/json" \
  -H "X-Worker-Key: $ZAKURO_WORKER_KEY" \
  -d '{
    "name": "my-worker",
    "uri": "http://10.13.13.5:3960",
    "worker_type": "zakuro",
    "resources": {
      "cpus_total": 8.0,
      "cpus_available": 8.0,
      "memory_total": 17179869184,
      "memory_available": 17179869184,
      "gpus_total": 1,
      "gpus_available": 1
    },
    "pricing": {
      "cpu_price": 0.001,
      "memory_price": 0.0001,
      "gpu_price": 0.01,
      "min_charge": 0.15
    },
    "tags": ["gpu", "cuda"]
  }'
```

Workers must send periodic heartbeats to remain healthy:

```bash
curl -X POST http://localhost:9000/workers/heartbeat \
  -H "Content-Type: application/json" \
  -H "X-Worker-Key: $ZAKURO_WORKER_KEY" \
  -d '{"worker_id": "uuid-of-worker"}'
```

Workers that miss heartbeats for longer than `worker_timeout` (default: 30 seconds) are marked unhealthy and eventually removed.

---

## API Reference

The broker exposes an HTTP API on its configured host and port (default: `0.0.0.0:9000`).

### Health & Status

#### `GET /health`

Returns `200 OK` if the broker is running, plus the identity of the binary
answering — the semver, the commit it was built from, and the SHA-256 of the
executable file:

```json
{
  "status": "healthy",
  "service": "zakuro-broker",
  "node_id": "zc://node-dd715b7f6d94d44e",
  "mesh_tunnel": "up",
  "build": {
    "version": "0.0.33",
    "git": "0447049",
    "binary_sha256": "0130785371eae09b5816ced91616ce48a4419f912809e757307459eb7dd0b6be"
  }
}
```

The hash is what distinguishes two brokers reporting the same version from the
same commit but running different artifacts. A broker built before build
disclosure omits `build` entirely. The same block rides `GET /peer/advert` and
`GET /peer/identity`, and `GET /brokers` reports it per peer as `build`
alongside this broker's own `self_build`.

Run `zc brokers` to reconcile it across the mesh, or `zc --version` to print
the local binary's hash for comparison.

#### `GET /stats`

Returns broker statistics in JSON format.

**Authentication:** Requires `Authorization: Bearer <key>`. Non-admin callers only see their own transactions; admin callers see all.

**Response fields:** transactions, workers, metrics (request count, RPS, latency percentiles, yen spent).

### Worker Management

#### `GET /workers`

List all registered workers and their current status.

**Response:**

```json
{
  "total": 2,
  "healthy": 2,
  "workers": [
    {
      "id": "uuid",
      "name": "worker-1",
      "uri": "http://10.13.13.5:3960",
      "status": "Healthy",
      "resources": { "cpus_total": 8, "cpus_available": 6, ... },
      "pricing": { "cpu_price": 0.001, ... },
      "avg_latency_ms": 12.5
    }
  ]
}
```

#### `POST /workers`

Register a new worker. See [Manual Registration](#manual-registration) for the request body schema.

**Authentication:** Requires `X-Worker-Key` header matching `ZAKURO_WORKER_KEY` env var (when configured). Without a configured key, registration is open (local dev).

#### `POST /workers/heartbeat`

Refresh a worker's heartbeat to keep it marked as healthy.

**Authentication:** Same as `POST /workers` — requires `X-Worker-Key` when configured.

#### `DELETE /workers/:id`

Unregister a worker by its UUID.

**Authentication:** Same as `POST /workers` — requires `X-Worker-Key` when configured.

### Compute Execution

#### `POST /execute`

The primary endpoint. Sends a compute request that the broker routes to the best available worker.

**Request headers:**

| Header | Required | Description |
|--------|----------|-------------|
| `Authorization` | Yes (remote mode) | `Bearer <api_key>` — resolves user ID for billing |
| `X-Zakuro-User` | No | Fallback user ID in local mode only (defaults to `anonymous`) |
| `X-Zakuro-Requirements` | No | JSON resource requirements (see below) |
| `Content-Type` | No | `application/octet-stream` for binary payloads |

**Requirements JSON:**

```json
{
  "cpus": 1.0,
  "memory_bytes": 1073741824,
  "gpus": 0,
  "estimated_duration_secs": 1.0,
  "strategy": "best_price",
  "worker_type": "zakuro",
  "tags": ["gpu"],
  "budget": 150
}
```

All fields are optional and have sensible defaults (1 CPU, 1 GiB memory, 1 second duration, `best_price` strategy), with one exception: `budget`.

**Remote dispatch needs a budget.** When the broker requires vouchers (`ZAKURO_REQUIRE_VOUCHER=1`, which a CLI broker with billing credentials sets by default), a job it offers to a peer broker carries a hub-signed voucher minted for `budget`. A request that would go remote without a positive `budget` fails fast with `400 BUDGET_REQUIRED`; there is no silent default. A job that runs on a local worker does not need one. The default will come from the node's `policy.toml` `[budget]` once ZC-0017 is implemented (`docs/adr/0017-one-sharing-policy.md`).

**Response headers:**

| Header | Description |
|--------|-------------|
| `X-Zakuro-Request-Id` | Unique request ID |
| `X-Zakuro-Cost` | Actual cost in yen |
| `X-Zakuro-Balance` | User's balance after the transaction |
| `X-Zakuro-Duration-Ms` | Total request duration in milliseconds |

#### `POST /price`

Estimate the cost of a request without executing it.

**Request body:** Same `requirements` JSON as `/execute`.

**Response:**

```json
{
  "min_cost": 0.15,
  "max_cost": 0.005,
  "workers_available": 3
}
```

### Balance Management

#### `GET /balance/:user`

Get the balance for a user.

**Authentication:** Requires `Authorization: Bearer <key>`. Non-admin callers can only view their own balance (403 otherwise).

```bash
curl http://localhost:9000/balance/my-user \
  -H "Authorization: Bearer $API_KEY"
```

#### `POST /balance/:user/add`

Add funds to a user's account. Requires an API key.

```bash
curl -X POST http://localhost:9000/balance/my-user/add \
  -H "Content-Type: application/json" \
  -H "X-Api-Key: your-api-key" \
  -d '{"amount": 100.0, "description": "Initial deposit"}'
```

### Ledger Administration

#### `GET /ledger/status`

Report which ledger the broker runs: `{"api_mode": true}` when balances go through the hub API (`ZAKURO_API_URL` + `ZAKURO_API_KEY`, no `ZAKURO_MASTER_KEY`), `false` in standalone mode (`ZAKURO_MASTER_KEY` set: local in-memory ledger). There is no database connection to report: the broker has no PostgreSQL client (ADR ZC-0005).

**Authentication:** Requires `Authorization: Bearer <master_key>` (admin only).

```bash
curl http://localhost:9000/ledger/status \
  -H "Authorization: Bearer $ZAKURO_MASTER_KEY"
```

### P2P Broker Endpoints

These endpoints are **only available when P2P is on** (the default for a CLI-started broker with a mesh route, or `ZAKURO_P2P=true`; see [P2P Broker-to-Broker Communication](#p2p-broker-to-broker-communication)) and are used for broker-to-broker communication.

#### `GET /peer/health`

Check if the peer broker is healthy and reachable.

**Authentication:** Requires `X-Peer-Key` header matching `ZAKURO_PEER_KEY`.

```bash
curl http://peer-broker:9000/peer/health \
  -H "X-Peer-Key: $ZAKURO_PEER_KEY"
```

#### `POST /peer/reserve`

Reserve funds for a user owned by this broker (called by non-authoritative brokers).

**Authentication:** Requires `X-Peer-Key` header.

**Request body:**
```json
{
  "user_id": "9000000001",
  "amount": 0.005,
  "reservation_id": "res-uuid-123"
}
```

**Response:**
```json
{
  "success": true,
  "balance_remaining": 99.995
}
```

#### `POST /peer/commit`

Commit a transaction after worker execution (finalizes the reservation).

**Authentication:** Requires `X-Peer-Key` header.

**Request body:**
```json
{
  "reservation_id": "res-uuid-123",
  "actual_cost": 0.003,
  "worker_id": "worker-uuid",
  "duration_ms": 250.0
}
```

#### `POST /peer/cancel`

Cancel a reservation and refund the full amount.

**Authentication:** Requires `X-Peer-Key` header.

**Request body:**
```json
{
  "reservation_id": "res-uuid-123"
}
```

#### `GET /peer/balance`

Get the balance for a user owned by this broker.

**Authentication:** Requires `X-Peer-Key` header.

**Query parameters:** `user_id=9000000001`

```bash
curl "http://peer-broker:9000/peer/balance?user_id=9000000001" \
  -H "X-Peer-Key: $ZAKURO_PEER_KEY"
```

**Response:**
```json
{
  "user_id": "9000000001",
  "balance": 99.995
}
```

#### `POST /peer/earn`

Pay earnings to the broker owner's account (used when a remote broker pays for execution on a local worker).

**Authentication:** Requires `X-Peer-Key` header.

**Request body:**
```json
{
  "request_id": "req-uuid-123",
  "amount": 0.003
}
```

#### `GET /peer/workers`

List workers that are local to this broker (at `127.0.0.1` or `localhost`). Used by peer brokers to discover available workers across the mesh.

**Authentication:** Requires `X-Peer-Key` header.

---

## Routing Strategies

When sending a compute request via `/execute`, you can specify a routing strategy in the `X-Zakuro-Requirements` header. The strategy determines how the broker selects which worker handles the request.

| Strategy | Aliases | Description |
|----------|---------|-------------|
| `best_price` | `price`, `cheap`, `cheapest` | Selects the worker with the lowest estimated cost. **(Default)** |
| `best_latency` | `latency`, `fast`, `fastest` | Selects the worker with the lowest average response time. |
| `best_availability` | `availability`, `available` | Selects the worker with the most available resources and fewest active requests. |
| `round_robin` | `robin`, `rr` | Distributes requests evenly across all healthy workers in order. |
| `random` | `rand` | Selects a random healthy worker. |
| `weighted_capacity` | `weighted`, `capacity` | Randomly selects a worker weighted by available CPU and memory. |

### Adaptive Routing (opt-in)

`ZAKURO_ADAPTIVE_ROUTING` is **off by default**. With `1`/`true`/`yes`/`on` the adaptive policy (`src/broker/policy.rs`: a Thompson sample of each worker's service time, best of `d` draws) makes the routing decision instead of the strategies above. Off, it runs in shadow: it is shown every decision the normal strategy made and keeps its ledger, but routes nothing. See ADR ZC-0011 (`docs/adr/0011-adaptive-scheduler-promotion.md`). On `zc agent`, the variable is forwarded to the broker the agent starts.

### Worker Scoring Algorithm

For the default `best_price` strategy, the overall worker score is computed as:

```
score = price_score * 0.5 + load_score * 0.3 + latency_score * 0.2
```

Where:

- **Price score (50%):** `cpu_price * 0.6 + memory_price * 0.3 + gpu_price * 0.1`
- **Load score (30%):** `(1 - cpus_available/total + 1 - memory_available/total) / 2`
- **Latency score (20%):** `avg_latency_ms / 100`

Lower scores indicate better workers.

---

## Benchmarking

`zc bench` measures the broker's throughput and latency under load.

### Basic Benchmark

```bash
# Default: 10 concurrent workers, 1000 requests, best_price strategy
zc bench

# Benchmark with a specific strategy
zc bench -s best_latency

# High-load benchmark
zc bench -c 100 -n 10000

# Benchmark a remote broker
zc bench zc://node-fd29ea8fe9ffa2e0
```

### Finding where a slow request went

`X-Zakuro-Duration-Ms` is one number for everything up to settlement, so a
broker that spends seconds somewhere can only report that it did. Send
`X-Zakuro-Trace: 1` and the reply carries `X-Zakuro-Phases`: each phase's own
cost in milliseconds, in the order the request crossed them.

```bash
curl -s -o /dev/null -X POST "$BROKER/execute" \
  -H 'X-Zakuro-Trace: 1' -H 'Authorization: Bearer $ZAKURO_API_KEY' \
  -H 'Content-Type: application/octet-stream' --data-binary @payload.pkl \
  -w '%{time_total}s  [%header{X-Zakuro-Phases}]\n'
# 0.004s  [queue=0.06 auth=0.01 balance=0.06 route=0.07 voucher=0.00
#          reserve=0.00 body=0.00 forward=3.44 finish_queue=0.01 settle=0.12]
```

| phase | what it covers |
|---|---|
| `queue` | waiting for a blocking-pool slot before prepare ran — load, not work |
| `auth` | resolving the caller from the Bearer token |
| `balance` | the authority lookup and balance fetch |
| `route` | choosing a worker |
| `voucher` | minting a dashboard voucher, when one is required |
| `reserve` | the reservation and its WAL record |
| `body` | reading the request body |
| `forward` | the worker round trip — the only phase that is the actual job |
| `finish_queue` | waiting for a blocking-pool slot again, after the worker replied |
| `settle` | commit, WAL, ledger, voucher redeem — all of it after `X-Zakuro-Duration-Ms` is taken |

Everything but `forward` is overhead. On a healthy broker they are all under a
millisecond; anything else is the thing to fix.

Tracing is opt-in per request and costs nothing when it is not asked for.

### Benchmarking one node

Naming a node measures **that node**: the run addresses its broker and asks
for the work to stay there, so the "where it ran" table should show one row.
The report says which it was, and names the nodes if the work spread anyway.

```bash
# One node, and only it
zc bench zc://node-fd29ea8fe9ffa2e0
```

Two flags decide where the work may NOT run, and they mean different things:

| Flag | Means | Use it to |
|---|---|---|
| `--remote` | not on **this machine** | keep a run off your laptop. A broker already out on the mesh satisfies it as-is, so it is a no-op there. |
| `--remote-only` | the **broker** must skip its own workers and offer the job to a peer | measure peer dispatch itself. It sends work AWAY from a node you named, so it releases the pin. |

`--remote` used to be sent as the broker-side `remote_only`, which made
`zc bench zc://node-… --remote` scatter its work over every peer of the node
it had just named — slower, mostly timeouts, and never a measurement of that
node.

A pin is only put on the wire for a broker running 0.0.33 or newer; older
brokers mishandle it, and bench relies on their local-first routing instead.
`zc brokers` shows each broker's build.

The benchmark:

1. Verifies broker connectivity via `/health`
2. Checks available workers via `/workers`
3. Sends requests concurrently to `/execute`
4. Displays a progress spinner during execution
5. Prints a detailed report with throughput and latency percentiles

**Example output:**

```
  ═══════════════════════════════════════════════
  ◆  Benchmark Results
  ═══════════════════════════════════════════════

  Summary
  ────────────────────────────────
    Total Requests:    1000
    Successful:        987 (98.7%)
    Failed:            13 (1.3%)
    Duration:          4.21s
    Throughput:        237.53 req/s

  Latency (ms)
  ────────────────────────────────
    Min:               0.84
    Avg:               3.21
    Max:               42.10
    p50:               2.15
    p75:               3.80
    p90:               6.20
    p95:               9.45
    p99:               28.30

  Latency Histogram
  ────────────────────────────────
      ≤1ms [██████                        ]    120 ( 12.2%)
      ≤5ms [██████████████████████████████]    650 ( 65.9%)
     ≤10ms [████████████                  ]    150 ( 15.2%)
     ≤25ms [██                            ]     40 (  4.1%)
     ≤50ms [█                             ]     27 (  2.7%)
```

### Strategy Comparison

Compare all six routing strategies side by side:

```bash
zc bench --compare
zc bench --compare -c 50 -n 5000
```

This runs the benchmark separately for each strategy and outputs a comparison table highlighting the best throughput and lowest latency.

### Mesh Warmup / Calibration

`zc bench mesh` probes a worker set and emits a JSON report that mirrors the
Python warmup primitive used by `AdaptiveCompute`.

```bash
zc bench mesh --workers a:3960,b:3960,c:3960 | jq .
```

The report includes per-worker `latency_p95`, `bandwidth_bps`, and a
`recommended_backpressure` threshold for the mesh as a whole.

### Benchmark Options

| Option | Short | Default | Description |
|--------|-------|---------|-------------|
| `--concurrency` | `-c` | `10` | Number of concurrent workers |
| `--requests` | `-n` | `1000` | Total requests to send |
| `--url` | `-u` | `http://127.0.0.1:9000` | Broker URL |
| `--user` | | `bench-user` | User ID for requests |
| `--strategy` | `-s` | `best_price` | Routing strategy to test |
| `--compare` | | | Compare all strategies |
| `--list-strategies` | | | List available strategies |
| `--help` | `-h` | | Show help |

---

## Remote Monitoring

Read a running broker from another machine:

```bash
# Its workers, their health and prices
zc workers localhost:9000
zc workers http://broker.mesh:9000

# Transactions, request rate and latency percentiles, as JSON
curl -s -H "Authorization: Bearer $ZAKURO_API_KEY" http://broker.mesh:9000/stats
```

See [`GET /stats`](#get-stats) for the response.

---

## System Diagnostics

View system information, network status, and detected compute clusters:

```bash
zc info
```

**Output sections:**

| Section | Content |
|---------|---------|
| **Network** | Hostname, operating mode (P2P / Local), local IP, mesh IP |
| **Compute Clusters** | Availability of Ray, Dask, and Spark clusters with version and node count |
| **Configuration** | The hub API URL in use and whether `ZAKURO_API_KEY` is set (the key itself is never printed) |
| **Services** | Reachability of the production hub (`hub.zakuro-ai.com`), the local broker (9000) and the local worker (3960) |

---

## Balance System

The broker uses a yen-denominated billing system to charge users for compute requests.

Every amount (balances, prices, charges, earnings, budgets) is yen, a decimal
with up to 6 fractional digits. Default worker price: ¥540/hour (¥0.15/s),
minimum charge ¥0.15 per request.

Peers settle only in the same currency: `/peer/advert` and `/peer/identity`
carry `"currency": "jpy"`, every `/peer/*` call sends `X-Zakuro-Currency: jpy`,
and a node refuses (409 `CURRENCY_MISMATCH`) to reserve, commit or earn with a
peer that does not.

### Hub Ledger (Credit Authority)

The hub is the balance authority (ADR ZC-0005, `docs/adr/0005-hub-balance-authority-and-wal.md`). The broker has **no database client**: it never holds database credentials, so it can run on a user's machine on a public mesh. Every balance read and settlement goes through the hub HTTP API, authenticated with `X-Broker-Api-Key` (`src/broker/ledger.rs`, `src/broker/node_sync.rs`):

| Hub endpoint | Used for |
|--------------|----------|
| `GET /api/broker/balance/{uid}` | Balance read (cached, see below) |
| `POST /api/broker/transaction` | Single settlement |
| `POST /api/broker/batch-sync` | Buffered settlements, flushed in chunks |
| `POST /api/broker/sync-workers` | Worker registry sync |
| `/api/broker/node/register`, `/api/broker/node/roster` | Node registration and mesh roster |
| `/api/broker/mesh/owner-ips` | Owner IPs on the mesh |
| `POST /api/broker/voucher`, `GET /api/broker/voucher/pubkey` | Hub-signed job vouchers (ADR ZC-0006) |
| `/api/broker/config/mesh`, `/api/broker/config/wireguard` | Mesh and WireGuard profile configuration |

The ledger runs in one of three modes:

| Mode | Selected by | Balances |
|------|-------------|----------|
| **API mode** | `ZAKURO_API_URL` + `ZAKURO_API_KEY`, no `ZAKURO_MASTER_KEY` | The hub. Reads are cached, settlements are buffered and flushed asynchronously |
| **Standalone** | `ZAKURO_MASTER_KEY` set (wins over API mode) | Local in-memory ledger, billing still enforced; lost on restart |
| **Local / free** | Neither | No billing: local execution is free |

**Balance reads** are cached for `ZAKURO_BALANCE_TTL_MS` (default 1000 ms; `0` asks the hub every time), served stale while a single in-flight refresh runs. For 120 s after a user last spent on this broker (`SPEND_GRACE`), a refresh may lower the cached balance but not raise it, because the hub may not have seen the spend yet; a hub top-up can therefore take up to two minutes to show.

**Settlement** is buffered in a `TransactionBuffer` (`src/broker/flush.rs`) and flushed to `batch-sync` by a dedicated `zc-flush` thread on every health-check tick (`health_check_interval`, default 5 s) and once more on shutdown. Each POST is chunked to at most `MAX_FLUSH_BATCH = 500` transactions **and** `MAX_FLUSH_BYTES = 512 KiB` of body (the hub's proxy refuses bodies over ~1 MiB with 413, #330); a row larger than the byte cap travels alone. A chunk that fails is re-queued; after `MAX_FLUSH_ATTEMPTS = 5` failures its transactions are quarantined to `wal-rejected.jsonl` next to the WAL (rotated to `.1` past its size cap). The hub deduplicates on `request_id`, so a retried chunk is not charged twice.

### Reserve-Commit Model

Each compute request follows a reserve-commit flow:

1. **Reserve:** The estimated cost is atomically deducted from the user's balance and stored as a reservation (5-minute TTL).
2. **Execute:** The request is forwarded to the selected worker.
3. **Commit:** On success, the reservation is finalized with the actual cost, and any difference is refunded. On failure, the full amount is refunded (cancel).

This ensures users are never overcharged and money is not lost on failures.

### API Keys

- **Master Key:** Set via `ZAKURO_MASTER_KEY`. Has full permissions (admin operations like adding funds, viewing ledger status).
- **User API Keys:** Follow the `zk_{user_id}_{hex}` format. In API mode the broker does not trust the key's shape: it asks the hub who the key belongs to (`GET /api/auth/me/api-key`, `src/broker/key_auth.rs`) and caches the verdict by the key's SHA-256 (positive 300 s, negative 30 s). A hub that cannot answer yields 503, never a fallback to the shape. `ZAKURO_TRUST_KEY_SHAPE=1` restores shape-only resolution as an emergency opt-out.
- **Storage:** The broker stores no API keys; keys are held and checked by the hub.

### Hub Unavailable

There is no database fallback. When the hub cannot be reached, a user whose balance was never loaded gets `BalanceUnavailable` (not a zero balance), key checks fail closed with 503, and settlement flushes fail. A transport error counts as a failed attempt, so a transaction that fails `MAX_FLUSH_ATTEMPTS = 5` flushes in a row (about 25 s at the default 5 s tick) is quarantined to `wal-rejected.jsonl` rather than retried forever. Work already admitted keeps executing.

### Local Mode (Free Execution)

When running in local discovery mode (no mesh), all executions are free and no money is charged.

### Per-Worker Local Detection (P2P Mesh)

In a multi-node P2P mesh, each broker detects its own mesh IP at startup and uses it to distinguish local vs remote workers:

- **Local workers** (at `127.0.0.1` or matching the broker's own mesh IP) execute for **free** — no reservation or billing
- **Remote workers** (on other mesh IPs) are **charged** via the normal reserve-commit flow

This enables a mesh topology where each node acts as both broker and worker. When `best_price` routing is used, the local worker is always preferred (cost=0). Other strategies like `round_robin` distribute across local and remote workers.

### P2P Broker-to-Broker Communication

P2P is on by default for a CLI-started broker with a mesh route, and `ZAKURO_P2P` overrides it (see the table below). This mode adds broker-to-broker communication for money operations, so the authoritative broker for a user answers reserve/commit from memory instead of asking the hub on the hot path.

**Architecture:**

Each user is assigned an **authoritative broker** using a consistent hash (FNV-1a):
```
authority_index = hash(zakuro_user_id) % num_brokers
```

**Authority modes:**

| Mode | Trigger | Hot Path | Behavior |
|------|---------|----------|----------|
| **Authority::Local** | User owned by this broker | **No hub call** | DashMap in-memory money ops (reserve, commit, balance) |
| **Authority::Peer(url)** | User owned by remote broker | **~1ms HTTP** | Forward money ops to authoritative broker via `/peer/*` endpoints |
| **Authority::Standalone** | Peer broker unreachable | **Hub API** | Falls back to the hub ledger path |

**Transaction flow (P2P authoritative):**
1. Auth (DashMap cache) → 2. Balance (DashMap) → 3. Reserve (DashMap atomic deduct)
4. WAL: RESERVED → 5. Forward to worker → 6. WAL: EXECUTED
7. Commit (DashMap refund diff) → 8. Queue in flush buffer → 9. WAL: COMMITTED → 10. Return

**Transaction flow (P2P non-authoritative):**
1. Auth (DashMap cache) → 2. HTTP `/peer/reserve` to authoritative (~1ms)
3. WAL: RESERVED → 4. Forward to worker → 5. WAL: EXECUTED
6. HTTP `/peer/commit` to authoritative (~1ms) → 7. WAL: COMMITTED → 8. Return

**Peer endpoints:**

All peer endpoints require `X-Peer-Key` header authentication:

| Endpoint | Method | Purpose |
|----------|--------|---------|
| `/peer/health` | GET | Peer broker health check |
| `/peer/reserve` | POST | Reserve funds for remote user (body: `{user_id, amount, reservation_id}`) |
| `/peer/commit` | POST | Commit transaction (body: `{reservation_id, actual_cost, worker_id, duration_ms}`) |
| `/peer/cancel` | POST | Cancel reservation (body: `{reservation_id}`) |
| `/peer/earn` | POST | Pay earnings to broker owner (body: `{request_id, amount}`) |
| `/peer/balance` | GET | Get balance (query: `?user_id=...`) |
| `/peer/workers` | GET | List this broker's local workers |

**Flush thread:**

In P2P mode, authoritative brokers maintain a `TransactionBuffer` that batches writes every **5 seconds** (every health-check tick) to the hub `batch-sync` API, chunked as described in [Hub Ledger](#hub-ledger-credit-authority):
- Pending transactions (all state transitions)
- Balance snapshots (for dashboard/audit)

This keeps the hub round trip off the hot path; balances are eventually consistent with the hub.

**Environment variables:**

| Variable | Required | Default | Description |
|----------|----------|---------|-------------|
| `ZAKURO_P2P` | No | on with a mesh route, else off | Unset: `zc broker` / `zc share` / `zc up` enable P2P when the machine reaches the mesh (local mesh interface or `zc connect` proxy sidecar), see `p2p_default_from` in `src/broker/mod.rs`. `true`/`1` forces it on, any other value forces it off. |
| `ZAKURO_PEER_KEY` | Yes (in P2P) | — | Shared secret for `X-Peer-Key` auth |
| `ZAKURO_OWNER_ID` | Yes (in P2P) | — | This broker's unique ID (used in authority hash) |
| `ZAKURO_PEERS` | Yes (in P2P) | — | Comma-separated peer broker URLs: `http://10.13.13.5:9000,http://10.13.13.6:9000` |

**Resilience:**

- If authoritative peer is down → `Authority::Standalone` → hub ledger API
- WAL replay on crash: `reserved→cancel`, `executed→commit` (same as before)
- Peer discovery via `ZAKURO_PEERS` + health probes on startup

---

## Write-Ahead Log (WAL)

The broker maintains a write-ahead log for crash recovery. Candidate paths (first writable wins): `ZAKURO_WAL_PATH` env → `$HOME/.zakuro/wal.jsonl` → `/tmp/zakuro-wal.jsonl`. Each transaction progresses through these states (`WalStatus` in `src/broker/wal.rs`):

| Status | Description |
|--------|-------------|
| `Reserved` | Funds reserved, request sent to worker |
| `Executed` | Worker returned successfully, actual cost calculated |
| `Earned` | A peer earning applied in memory and queued for the hub, not yet confirmed flushed |
| `Committed` | Charge committed and settled |
| `Failed` | Request failed; reservation refunded |

**Implementation:**

- **Append-only JSONL file** — one JSON object per line, never modified in-place
- **DashMap in-memory index** — O(1) lookup by `request_id` for status checks and updates
- **Group commit off the request path** — `append` only pushes the serialized line onto a lock-free queue (`SegQueue`); it never takes the file lock or calls `fsync`. A dedicated flush worker (`run_flush_worker`) drains the queue and `fsync`s when the oldest entry reaches **100 ms** (`WRITE_BUFFER_MAX_AGE_MS`) or sooner once **1024 entries** are buffered (`WRITE_BUFFER_MAX_ENTRIES`), and drains once more on shutdown

Durability is time-bounded (an acknowledged entry reaches disk within about 100 ms) while no request thread ever blocks on `fsync`.

### Crash Recovery

On startup, the broker replays any uncommitted WAL entries via `recovery.rs`:

- **Reserved** entries: funds are refunded (worker may have crashed before responding)
- **Executed** entries: charges are committed at the actual cost and published to the hub (worker succeeded but broker crashed before committing)
- **Earned** entries: re-delivered to the hub through the normal flush

**P2P mode:** WAL replay works seamlessly with P2P — authoritative brokers apply state changes to their DashMap, non-authoritative brokers skip or retry peer calls.

The WAL is **automatically compacted** every ~5 minutes: committed and failed entries are removed from disk, keeping only uncommitted entries. The in-memory index is also cleared of completed entries.

### WAL Format

Append-only JSONL file. Each line is a JSON object:

```json
{"request_id":"uuid","user_id":"user","reservation_id":"res-id","estimated_cost":0.15,"actual_cost":0.12,"worker_id":"worker-uuid","duration_ms":250.0,"timestamp":"2024-01-01T00:00:00Z","status":"Committed","unit":"jpy"}
```

`unit` is always `"jpy"`: every amount is yen, a decimal with up to 6 fractional
digits (ADR 0027 in zak-marketplace). A line without `unit` was written before
the switch; on start the broker multiplies its amounts by 150 once, tags it
`jpy` and rewrites the WAL, so it is never converted twice.

---

## P2P Mesh Deployment

Deploy a multi-node mesh where each node runs a broker + worker pair reachable
over WireGuard.

### Architecture

```
  Node 1 (zk0-node01)                       Node 2 (zk0-node02)
  ┌─────────────────────┐                   ┌─────────────────────┐
  │  Broker (:9000)     │◄══ WireGuard ════►│  Broker (:9000)     │
  │    ├ WAL + DashMap  │    10.13.13.0/24  │    ├ WAL + DashMap  │
  │    ├ hub API client │                   │    ├ hub API client │
  │    └ Discovery      │                   │    └ Discovery      │
  │                     │                   │                     │
  │  Worker (:3960)     │                   │  Worker (:3960)     │
  │    type: standard   │                   │    type: premium    │
  │    (free to self)   │                   │    (free to self)   │
  └─────────────────────┘                   └─────────────────────┘
```

The tunnel is brought up on the **host**, not in a sidecar container: each host
holds a `zakuro0` (or `wg*`) interface with an address in `10.13.13.0/24`, and
the broker finds it through `get_mesh_ip` (`src/broker/discovery.rs`).
`zc vpn` manages that interface, and the dashboard hands out a per-user peer
profile via `GET /api/broker/config/wireguard`.

### Prerequisites

- A WireGuard peer profile per host (see `zc vpn --help`, or fetch one from the
  dashboard with your broker API key)
- Docker and Docker Compose

### How discovery works in the mesh

1. Each broker resolves its own mesh IP — `ZAKURO_WIREGUARD_IP` if set,
   otherwise the `zakuro0`/`wg*` interface address
2. Discovery scans localhost ports first (finds the local worker)
3. Then probes explicit `ZAKURO_PEERS` (finds remote workers across the tunnel)
4. Existing workers get their heartbeat refreshed by TCP connect only, with no
   full HTTP probe
5. Workers are deduplicated by name, so the same worker never appears twice

### Environment Variables (Mesh)

| Variable | Description |
|----------|-------------|
| `ZAKURO_WIREGUARD_IP` | This node's mesh IP (for local worker detection). Named for the mesh we used to run; it applies to WireGuard too. |
| `ZAKURO_PEERS` | Comma-separated `ip:port` of remote workers or brokers to probe |

> **The two-node Docker recipe was removed.** It stood up a VPN sidecar per
> node under `zak-zakuro/docker/` and shared its network namespace with the
> broker and worker, which only worked on the mesh we no longer run. There is
> no drop-in Compose replacement yet: bring WireGuard up on each host, then run
> the brokers and workers with `ZAKURO_PEERS` pointing at the other host's mesh
> address.

---

## Cluster Commands

| Command | Description |
|---------|-------------|
| `zc up` | Start 1 worker + broker locally (foreground) |
| `zc up --workers N` | Start N workers + broker (ports 3960–3960+N) |
| `zc up --workers N -d` | Start N workers + broker in the background (daemon) |
| `zc down` | Stop all workers + broker started by `zc up` |
| `zc workers` | List workers registered with the local broker |
| `zc me` / `zc balance` | Show authenticated user info and money |
| `zc update` | Update the `zc` CLI from `get.zakuro-ai.com` |
| `zc pull` | Pull updated Docker images |
| `zc images` | List Zakuro images built on the machine |
| `zc ps` | List currently running Zakuro containers |
| `zc kill` | Stop and remove running Zakuro containers |
| `zc restart` | Restart containers with updated images |

---

## Zakuro Drive Hooks

`zc hooks` manages file-event hooks on a running [zakuro-drive](https://github.com/zakuro-ai/zakuro-drive) daemon (`zakurod`) over its QUIC API (default `127.0.0.1:4719`; override with `--daemon HOST:PORT` or the `ZAKURO_DRIVE_DAEMON` env var). A hook fires on file events matching a glob and either POSTs an HMAC-signed JSON payload to a webhook or runs a local command.

| Command | Description |
|---------|-------------|
| `zc hooks add` | Create a hook (see flags below) |
| `zc hooks get <id\|name>` | Show full config for a hook (secrets redacted by the daemon) |
| `zc hooks list [--all]` | List hooks; `--all` includes disabled ones |
| `zc hooks update <id\|name>` | `--enable`/`--disable`, or tune `--retry-*` options |
| `zc hooks remove <id\|name>` | Delete a hook |
| `zc hooks logs [<id\|name>]` | Delivery history; `--dead` (dead letters), `--failed`, `--from 2h`, `--to`, `--limit N`, `--verbose` |

### `zc hooks add` flags

- `--name NAME` (required), `--glob GLOB` (required)
- `--on EVENT` (repeatable, required): `file.created` | `file.modified` | `file.deleted`
- Exactly one handler: `--webhook URL` or `--exec CMD`
- Webhook secret, exactly one of: `--secret-stdin` (recommended — reads one line from stdin), `--secret-from-env VAR`, `--secret VALUE` (discouraged: visible in `ps`/shell history)
- Optional: `--timeout SECS` (default: webhook 30, exec 60), `--tls-insecure` (webhook only; prints a warning), `--retry-attempts N`, `--retry-initial-delay MS`, `--retry-max-delay MS`, `--disabled`

```bash
echo "$WEBHOOK_SECRET" | zc hooks add --name slack-notify \
  --on file.created --glob "reports/**/*.pdf" \
  --webhook https://hooks.example.com/zakuro --secret-stdin

zc hooks logs slack-notify --from 7d --verbose
```

`--from`/`--to` accept relative durations (`90s`, `30m`, `2h`, `7d`), ISO dates (`2026-06-01`), or RFC3339 timestamps.

---

## Environment Variables

| Variable | Required | Default | Description |
|----------|----------|---------|-------------|
| `ZAKURO_API_KEY` | Yes | — | Authentication token. Also used for dashboard API calls in API mode. |
| `ZAKURO_API_URL` | No | — | Dashboard API URL (e.g., `https://hub.zakuro-ai.com`). When set with `ZAKURO_API_KEY`, enables API mode: the hub is the balance authority. |
| `ZAKURO_MASTER_KEY` | No | — | Master API key for admin operations (adding funds, ledger status). When set, the broker runs the standalone in-memory ledger even if `ZAKURO_API_URL` is set. |
| `ZAKURO_BALANCE_TTL_MS` | No | `1000` | How long a hub balance read is cached, in ms. `0` asks the hub on every request. |
| `ZAKURO_WAL_PATH` | No | `$HOME/.zakuro/wal.jsonl` | WAL location; falls back to `/tmp/zakuro-wal.jsonl` when the default is not writable. |
| `ZAKURO_WORKER_KEY` | No | — | Shared secret for worker management endpoints (`POST/DELETE /workers`, heartbeat). When unset, these endpoints are open. |
| `ZAKURO_WIREGUARD_IP` | No | — | Override mesh IP detection. Named for the mesh we used to run; it applies to WireGuard too. |
| `ZAKURO_PEERS` | No | — | Comma-separated `ip:port` peers for worker discovery, or `http://host:port` for broker peers in P2P mode. |
| `ZAKURO_P2P` | No | on with a mesh route, else off | P2P broker-to-broker communication. Unset, a CLI-started broker enables it when the machine reaches the mesh (local mesh interface or `zc connect` proxy sidecar); `true`/`1` forces it on, any other value forces it off. |
| `ZAKURO_ADAPTIVE_ROUTING` | No | off | `1`/`true`/`yes`/`on` lets the adaptive (Thompson-sampling) policy route (`adaptive_routing_enabled`, `src/broker/server.rs`; ADR ZC-0011). Off, it runs in shadow and only observes the normal strategy's decisions. Set on `zc agent`, it is forwarded to the broker the agent starts. |
| `ZAKURO_PEER_KEY` | No (Yes in P2P) | — | Shared secret for P2P broker authentication (`X-Peer-Key` header). |
| `ZAKURO_OWNER_ID` | No (Yes in P2P) | — | This broker's unique ID for authority assignment (FNV-1a hash). Auto-derived from `ZAKURO_API_KEY` if in `zk_{user}_{hex}` format. |
| `ZAKURO_NODE_NAME` | No | Auto-detected | Human-readable node name (mesh hostname → system hostname). |
| `ZAKURO_WORKER_PORT` | No | `3960` | Primary port workers listen on (discovery). |
| `ZAKURO_SCAN_RANGE` | No | — | Port range for discovery scan, e.g. `3960-3999`. |
| `ZAKURO_SCAN_INTERVAL` | No | `15` | Seconds between discovery scans. |
| `ZAKURO_AUTO_UPDATE` | No | off | Opt-in broker self-update from crates.io (crate `zc2`). `1`/`true`/`yes`/`on` enables it; anything else is off. Never set it on k8s fleets (their image is managed by CD). When on, the broker reads the crates.io sparse index; for a strictly newer non-yanked release it runs `cargo install zc2 --version X --locked` into a scratch dir next to the executable while it keeps serving, checks the built binary runs and reports exactly X, waits until idle (no in-flight requests, local workers or pending settlements, on 3 polls 5 s apart), swaps it in atomically and exits with code 75 so systemd (`Restart=always`) starts it. Needs `cargo` (see `ZAKURO_CARGO`) and a writable executable directory; failures are logged and retried next interval. Status (`source`, `building`, `last_error`, ...) is in `GET /health` under `auto_update`. `zc update` still uses the release mirror. |
| `ZAKURO_AUTO_UPDATE_INTERVAL` | No | `600` | Time between index checks, in seconds or with an `s`/`m`/`h` suffix (`10m`, `1h`). Minimum 300 s. +up to 10% jitter; the first check is after a random 60 s..interval delay. |
| `ZAKURO_AUTO_UPDATE_INDEX` | No | `https://index.crates.io/3/z/zc2` | Sparse-index URL for `zc2` (https; http only for `127.0.0.1`/`localhost`). For tests. |
| `ZAKURO_CARGO` | No | `cargo` on PATH, else `$HOME/.cargo/bin/cargo` | The cargo used by auto-update builds. |
| `ZAKURO_ENV` | No | `prod` | Environment profile: `lab`, `staging` (`stg`, `stage`) or `prod`. Sets the hub API, the get/update mirror, the registry and the trust roots together (`zc doctor` prints them). `lab` = `http://hub.lab.zakuro.internal`, `http://get.lab.zakuro.internal`, `registry.lab.zakuro.internal:5000`, and skips the legacy `/env` fetch into `/opt/zakuro/etc/zk0.env`. Also read from `env=` in `~/.zakuro/credentials` or `ZAKURO_ENV=` in `~/.zakuro/env`. `ZAKURO_API_URL`, `ZAKURO_UPDATE_URL` and `ZAKURO_REGISTRY` still override their part. |
| `ZAKURO_REGISTRY` | No | the profile's | OCI registry override for the active `ZAKURO_ENV` profile. |
| `ZAKURO_UPDATE_URL` | No | the profile's (`https://get.zakuro-ai.com/zc`) | Mirror base for `zc update` (the manual path; broker auto-update uses crates.io). Either a bare host base (`https://stg.get.zakuro-ai.com`) or one ending in `/zc`; trailing `/` is trimmed and `/zc` appended when missing. Must be `https://` (plain `http://` only for `127.0.0.1`/`localhost`); an invalid value is an error, never silently ignored. |

---

## Configuration Reference

### BrokerConfig

| Field | Default | Description |
|-------|---------|-------------|
| `host` | `0.0.0.0` | Host to bind the broker server |
| `port` | `9000` | Port to bind the broker server |
| `health_check_interval` | `5` | Worker health check interval in seconds |
| `worker_timeout` | `30` | Seconds before marking a worker unhealthy |
| `min_balance` | `0.15` | Minimum balance (yen) required for any request |
| `daemon` | `false` | Run in daemon (background) mode |
| `verbose` | `true` | Show live transaction logs (foreground mode) |
| `enable_discovery` | `true` | Enable automatic worker discovery |

### DiscoveryConfig

| Field | Default | Description |
|-------|---------|-------------|
| `subnet` | `10.13.13` | Mesh subnet to scan |
| `worker_port` | `3960` | Primary port workers listen on (from `ZAKURO_WORKER_PORT` env) |
| `extra_ports` | `[3961, 3962]` | Additional ports to scan |
| `scan_port_range` | `None` | Full port range scan, e.g. `(3960, 3999)` (from `ZAKURO_SCAN_RANGE` env) |
| `interval_secs` | `15` | Seconds between discovery scans (from `ZAKURO_SCAN_INTERVAL` env) |
| `enable_scan` | `true` | Enable active network scanning |
| `enable_dns` | `true` | Enable DNS-based discovery |
| `peers` | `[]` | Explicit peer addresses from `ZAKURO_PEERS` env var |

### Ledger configuration

There is no ledger config struct or database URL. `Ledger::new(api_url, api_key)` (`src/broker/ledger.rs`) takes the broker's `api_url` / `api_key`, which come from the environment:

| Input | Source | Effect |
|-------|--------|--------|
| `api_url` | `ZAKURO_API_URL` (or the `ZAKURO_ENV` profile's hub) | Hub base URL for balance and settlement calls |
| `api_key` | `ZAKURO_API_KEY` | Sent as `X-Broker-Api-Key`; with `api_url`, selects API mode |
| master key | `ZAKURO_MASTER_KEY` | Standalone in-memory ledger; overrides API mode |
| balance cache | `ZAKURO_BALANCE_TTL_MS` | Balance read cache window (default 1000 ms) |

### Worker Pricing Model

Workers advertise their pricing when registering:

| Field | Description |
|-------|-------------|
| `cpu_price` | Cost per CPU-core per second |
| `memory_price` | Cost per GiB of memory per second |
| `gpu_price` | Cost per GPU per second |
| `min_charge` | Minimum charge per request |

The estimated cost for a request is:

```
cost = max(
  cpu_price * cpus * duration_secs +
  memory_price * memory_gib * duration_secs +
  gpu_price * gpus * duration_secs,
  min_charge
)
```