swarm-lib/DESIGN.html at main · dpdanpittman/swarm-lib · GitHub

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
<!doctype html>
<html lang="en">
  <head>
    <meta charset="UTF-8" />
    <title>swarm-lib — Design Spec v0.1</title>
    <style>
      :root {
        --bg: #0f1115;
        --panel: #15181f;
        --panel-soft: #1a1e26;
        --text: #d8dde7;
        --muted: #8a93a3;
        --accent: #6fd3ff;
        --accent-warm: #f0b46a;
        --accent-green: #95d36b;
        --border: #232833;
        --code-bg: #11141a;
        --code-border: #2a2f3a;
      }
      * {
        box-sizing: border-box;
      }
      html,
      body {
        background: var(--bg);
        color: var(--text);
        margin: 0;
        padding: 0;
      }
      body {
        font:
          15px/1.55 -apple-system,
          BlinkMacSystemFont,
          "Segoe UI",
          system-ui,
          sans-serif;
        max-width: 980px;
        margin: 0 auto;
        padding: 48px 32px 96px;
      }
      h1 {
        font-size: 28px;
        line-height: 1.2;
        border-bottom: 1px solid var(--border);
        padding-bottom: 12px;
        margin: 0 0 12px;
      }
      .subtitle {
        color: var(--muted);
        margin: 0 0 24px;
        font-size: 14px;
      }
      h2 {
        font-size: 22px;
        margin-top: 44px;
        color: var(--accent);
        border-bottom: 1px solid var(--border);
        padding-bottom: 6px;
      }
      h3 {
        font-size: 17px;
        margin-top: 24px;
        color: var(--accent-warm);
      }
      h4 {
        font-size: 15px;
        margin-top: 18px;
        margin-bottom: 6px;
      }
      p {
        margin: 8px 0 12px;
      }
      hr {
        border: 0;
        border-top: 1px solid var(--border);
        margin: 32px 0;
      }
      ul,
      ol {
        padding-left: 22px;
      }
      li {
        margin: 4px 0;
      }
      code {
        font:
          13px/1.5 "JetBrains Mono",
          "Fira Code",
          Consolas,
          monospace;
        background: var(--code-bg);
        border: 1px solid var(--code-border);
        border-radius: 3px;
        padding: 1px 5px;
      }
      pre {
        background: var(--code-bg);
        border: 1px solid var(--code-border);
        border-radius: 6px;
        padding: 14px 16px;
        overflow-x: auto;
        font:
          13px/1.5 "JetBrains Mono",
          "Fira Code",
          Consolas,
          monospace;
      }
      pre code {
        background: none;
        border: 0;
        padding: 0;
      }
      table {
        border-collapse: collapse;
        margin: 12px 0 20px;
        width: 100%;
        font-size: 14px;
      }
      th,
      td {
        border: 1px solid var(--border);
        padding: 8px 12px;
        text-align: left;
        vertical-align: top;
      }
      th {
        background: var(--panel-soft);
        color: var(--accent);
        font-weight: 600;
      }
      tr:nth-child(even) td {
        background: var(--panel);
      }
      .tldr {
        background: var(--panel);
        border-left: 3px solid var(--accent);
        padding: 14px 18px;
        margin: 16px 0 24px;
        border-radius: 0 6px 6px 0;
      }
      blockquote {
        border-left: 3px solid var(--accent-warm);
        background: var(--panel);
        margin: 10px 0;
        padding: 8px 14px;
        color: var(--muted);
        font-size: 14px;
      }
      .footer-note {
        color: var(--muted);
        font-size: 13px;
        border-top: 1px solid var(--border);
        padding-top: 14px;
        margin-top: 48px;
      }
    </style>
  </head>
  <body>
    <h1>swarm-lib — Design Spec v0.1</h1>
    <p class="subtitle">
      Status: <b>Draft — pending review</b> · Inspired by the Inkcloud
      Architecture Post-Mortem ("Unix Swarm Blueprint")
    </p>

    <div class="tldr">
      A small Python + Bash library that gives agentic workflows three
      primitives:
      <ol style="margin: 8px 0 0 0">
        <li>
          <b>Atomic-rename task queueing</b> — multiple workers can race for
          tasks; exactly one wins, no external broker
        </li>
        <li>
          <b><code>status.json</code> checkpointing</b> — durable handoff so any
          agent can resume from a file, never from chat history
        </li>
        <li>
          <b>A generic <code>worker_loop.sh</code></b> — polls a queue, claims a
          task, invokes a handler, moves the result
        </li>
      </ol>
      v0.1 substrate only. Triage (HMD), orphan cleanup, multi-host federation,
      and a Kanban UI deferred to v0.2+. Goal: prove the substrate by porting
      Tribunal's 7-skill chain onto it.
    </div>

    <h2>Why this exists</h2>

    <table>
      <tr>
        <th>Problem</th>
        <th>Manifestation</th>
        <th>swarm-lib answer</th>
      </tr>
      <tr>
        <td><b>Context starvation</b></td>
        <td>
          Long conversation accumulates context, hits compaction, in-flight work
          dies
        </td>
        <td>
          Tasks become claimable by fresh contexts; no single conversation owns
          the whole chain
        </td>
      </tr>
      <tr>
        <td><b>Sync tool-call blocking</b></td>
        <td>
          Planner holds an expensive context window open while subprocess churns
          for 30 min
        </td>
        <td>
          The "Yield Rule": planner decomposes → writes payload to queue →
          exits. Consumers claim asynchronously.
        </td>
      </tr>
      <tr>
        <td><b>Chat-history-as-state</b></td>
        <td>If the conversation dies, the work dies</td>
        <td>
          <code>status.json</code> is the canonical state. Chat is volatile; the
          file is the contract.
        </td>
      </tr>
    </table>

    <p>
      UNIX shops have had this discipline for 30 years (Maildir, cron + lock
      files, <code>/var/spool/</code>). swarm-lib applies it to LLM-driven agent
      work.
    </p>

    <h2>Architectural principles</h2>

    <h3>1. The file system IS the orchestrator</h3>
    <p>
      No daemon, no broker, no external database. Directory layout + POSIX
      atomic rename + JSON files do all coordination. Any agent with read/write
      access to the queue directory is a participant.
    </p>

    <h3>2. The Yield Rule</h3>
    <p>
      A high-context agent (planner) NEVER blocks waiting on subprocess output.
      It decomposes → enqueues atomically → <b>exits immediately</b>, freeing
      its context. Fresh consumers claim tasks in their own short-lived
      contexts. The planner can be re-invoked later from
      <code>status.json</code>.
    </p>

    <h3>3. Atomic-rename queueing (Maildir physics)</h3>
    <p>
      Producers stage tasks under <code>pending/</code>. Consumers claim by
      <code
        >os.replace(pending/&lt;id&gt;.json,
        claimed/&lt;worker_id&gt;/&lt;id&gt;.json)</code
      >. POSIX guarantees the rename is atomic on the same filesystem — two
      consumers racing for the same task, exactly one wins.
    </p>

    <h3>4. Agents are interchangeable</h3>
    <p>
      A worker is anything that can read a JSON task file, do work, and write a
      JSON result file:
    </p>
    <ul>
      <li>Claude Code (Mabus) with a specific skill</li>
      <li>
        Ollama-backed local agents (via <code>ollama-assistant</code> MCP)
      </li>
      <li>n8n flows</li>
      <li>Plain shell scripts</li>
      <li>Future agents that don't exist yet</li>
    </ul>
    <p>
      Task ownership lives in the filesystem path (the
      <code>claimed/&lt;worker_id&gt;/</code> directory), not in a database row
      tied to a specific agent identity.
    </p>

    <h3>5. Cold-start recoverability</h3>
    <p>Any new agent picks up an in-flight workflow by reading two files:</p>
    <ul>
      <li>
        <code>portrait.md</code> — identity continuity (from session-essence)
      </li>
      <li><code>status.json</code> — task-state continuity (from swarm-lib)</li>
    </ul>
    <p>
      Compactions, crashes, rate limits, multi-day pauses — all become
      indistinguishable from a clean restart.
    </p>

    <h2>Directory layout</h2>

    <pre><code>&lt;consumer-repo&gt;/.swarm/&lt;run-id&gt;/
├── BACKLOG.md           # original human intent (optional, free-form)
├── status.json          # current checkpoint state (durable handoff)
├── pending/             # tasks waiting to be claimed
│   └── &lt;task_id&gt;.json
├── claimed/             # tasks currently being worked on, namespaced by worker
│   └── &lt;worker_id&gt;/
│       └── &lt;task_id&gt;.json
├── done/                # successfully completed tasks
│   └── &lt;task_id&gt;.json
├── failed/              # quarantined tasks (worker death without clean handoff)
│   └── &lt;task_id&gt;.json
└── artifacts/           # task outputs — markdown, JSON, logs, anything
    └── &lt;task_id&gt;.&lt;ext&gt;</code></pre>

    <p>
      <b>Cross-project federation</b>: when a task spans multiple consumer repos
      (Mabus → n8n → Mabus), the run directory lives under
      <code>~/.swarm/&lt;run-id&gt;/</code> instead. v0.1 doesn't enforce this
      distinction — <code>worker_loop.sh</code> takes a
      <code>--run-dir</code> arg.
    </p>

    <h2><code>status.json</code> schema</h2>

    <pre><code>{
  "schema_version": "0.1",
  "run_id": "tribunal-r42x",
  "checkpoint": {
    "summary": "Completed plan; ready for implement stage",
    "next_step": "Invoke implement skill with plan output as input",
    "next_task_id": "t.r42x.implement",
    "resume_command": null,
    "risk": "Plan references xion-stytch-proxy hostnames; ensure consumer has env access",
    "completed_tasks": ["t.r42x.intent", "t.r42x.plan"],
    "current_worker": null,
    "timestamp": "2026-05-20T18:45:11Z"
  },
  "metadata": {
    "created_by": "mabus",
    "consumer": "tribunal",
    "tier_hint": "opus"
  }
}</code></pre>

    <table>
      <tr>
        <th>Field</th>
        <th>Type</th>
        <th>Required</th>
        <th>Notes</th>
      </tr>
      <tr>
        <td><code>schema_version</code></td>
        <td>string</td>
        <td>yes</td>
        <td>semver; lib refuses to read unknown majors</td>
      </tr>
      <tr>
        <td><code>run_id</code></td>
        <td>string</td>
        <td>yes</td>
        <td>unique per workflow run</td>
      </tr>
      <tr>
        <td><code>checkpoint.summary</code></td>
        <td>string</td>
        <td>yes</td>
        <td>human-readable "what was just finished"</td>
      </tr>
      <tr>
        <td><code>checkpoint.next_step</code></td>
        <td>string</td>
        <td>yes</td>
        <td>concrete "what to do next" — agent-readable</td>
      </tr>
      <tr>
        <td><code>checkpoint.next_task_id</code></td>
        <td>string | null</td>
        <td>yes</td>
        <td>the task_id to claim next; null = workflow complete</td>
      </tr>
      <tr>
        <td><code>checkpoint.resume_command</code></td>
        <td>string | null</td>
        <td>no</td>
        <td>
          optional; <b>must be validated against allowlist before exec</b>
        </td>
      </tr>
      <tr>
        <td><code>checkpoint.risk</code></td>
        <td>string</td>
        <td>yes</td>
        <td>known blockers or fragile state</td>
      </tr>
      <tr>
        <td><code>checkpoint.completed_tasks</code></td>
        <td>string[]</td>
        <td>yes</td>
        <td>task_ids that have run</td>
      </tr>
      <tr>
        <td><code>checkpoint.current_worker</code></td>
        <td>string | null</td>
        <td>yes</td>
        <td>worker_id currently holding a claim</td>
      </tr>
      <tr>
        <td><code>checkpoint.timestamp</code></td>
        <td>ISO-8601</td>
        <td>yes</td>
        <td>last update</td>
      </tr>
      <tr>
        <td><code>metadata.*</code></td>
        <td>object</td>
        <td>no</td>
        <td>freeform; consumer-specific</td>
      </tr>
    </table>

    <p>
      <b>Evolution policy</b>: minor bumps additive only. Major bumps breaking;
      lib supports reading the previous major for one cycle to allow migration.
    </p>

    <blockquote>
      Pushback note: I removed the <code>resume_command</code> "blindly execute"
      framing from the slideshow. If consumers want a shell-resumable workflow,
      the command must be validated against an allowlist before exec. Default is
      <code>null</code>; LLM-driven workflows ignore it.
    </blockquote>

    <h2>Task schema</h2>

    <pre><code>{
  "task_id": "t.r42x.implement",
  "task_type": "implement",
  "run_id": "tribunal-r42x",
  "created_at": "2026-05-20T18:42:11Z",
  "created_by": "mabus",
  "depends_on": ["t.r42x.plan"],
  "payload": {
    "skill": "tribunal-implement",
    "input_artifacts": ["artifacts/t.r42x.plan.md"],
    "output_artifact": "artifacts/t.r42x.implement.md"
  },
  "tier_hint": "sonnet",
  "deadline": null
}</code></pre>

    <ul>
      <li>
        <code>payload</code> is consumer-specific; lib doesn't validate its
        shape
      </li>
      <li>
        <code>tier_hint</code> is an optional cost-routing signal (used in v0.2)
      </li>
      <li>
        <code>depends_on</code> is honored by <code>worker_loop.sh</code> —
        won't claim until deps are in <code>completed_tasks</code>
      </li>
      <li>
        <code>deadline</code> is informational in v0.1; v0.2 may route to a
        scheduler worker
      </li>
    </ul>

    <h2>Claim protocol</h2>

    <h3>Producer (enqueue)</h3>
    <pre><code>from swarm_lib import claims

claims.enqueue(
    run_dir="~/src/tribunal/run/r42x",
    task_id="t.r42x.implement",
    task_type="implement",
    payload={...},
    depends_on=["t.r42x.plan"],
)</code></pre>

    <p>
      Internally: write JSON to temp file → fsync →
      <code>os.replace(tmp, pending/&lt;id&gt;.json)</code> → atomic publish.
    </p>

    <h3>Consumer (claim)</h3>
    <pre><code>task = claims.try_claim(
    run_dir="~/src/tribunal/run/r42x",
    worker_id="mabus-tribunal-1",
    task_type_filter=["implement", "verify"],
)
if task is None:
    return  # no claimable task

do_work(task)
claims.complete(task, success=True, artifact_path="...")</code></pre>

    <p>
      Internally <code>try_claim</code> scans <code>pending/</code>, attempts
      <code>os.replace</code> per candidate, returns the first successful claim.
      <code>OSError</code> means another worker won — try next candidate.
    </p>

    <h3>Orphan recovery (v0.2)</h3>
    <p>
      If a worker dies mid-claim, the task is stuck in
      <code>claimed/&lt;worker_id&gt;/</code>. v0.2 adds cron-driven cleanup.
      <b>v0.1 ignores this</b>; manual cleanup is fine for now.
    </p>

    <h2><code>worker_loop.sh</code> contract</h2>

    <pre><code>worker_loop.sh \
  --run-dir ~/src/tribunal/run/r42x \
  --worker-id mabus-tribunal-1 \
  --handler tribunal_handle.sh \
  --task-type-filter implement,verify \
  --poll-interval 5 \
  --max-iterations 100</code></pre>

    <p>
      Loop body: try_claim → if claimed, invoke handler with task JSON on stdin
      → on handler exit, mark done/failed → repeat.
    </p>

    <p>
      A handler is any executable that reads task JSON from stdin, writes its
      output to <code>$SWARM_ARTIFACT_PATH</code>, optionally enqueues follow-up
      tasks, updates <code>status.json</code>, and exits 0/non-zero.
    </p>

    <h2>Python API surface (v0.1)</h2>

    <pre><code># swarm_lib/claims.py
def enqueue(run_dir, task_id, task_type, payload,
            depends_on=None, tier_hint=None, created_by=None): ...
def try_claim(run_dir, worker_id, task_type_filter=None): ...
def complete(task, success, artifact_path=None): ...

# swarm_lib/status.py
def read(run_dir): ...
def write(run_dir, summary, next_step, next_task_id,
          risk="", current_worker=None, metadata=None): ...
def append_completed(run_dir, task_id): ...</code></pre>

    <h2>Tribunal port plan</h2>

    <h3>Today's shape</h3>
    <p>
      <code>/tribunal &lt;prompt&gt;</code> runs intent → plan → implement →
      review → verify → classify → incentive sequentially in one Claude Code
      context. Long probes blow the window. Compaction loses the chain.
    </p>

    <h3>Ported shape</h3>
    <pre><code>~/src/tribunal/run/&lt;run-id&gt;/
├── BACKLOG.md           # user's /tribunal prompt verbatim
├── status.json
├── pending/ claimed/ done/ failed/
└── artifacts/
    ├── 01-intent.md
    ├── 02-plan.md
    ├── 03-implement.md
    ├── 04-review.md
    ├── 05-verify.md
    ├── 06-classify.md
    └── 07-incentive.md</code></pre>

    <h3>Flow</h3>
    <ol>
      <li>
        <code>/tribunal &lt;prompt&gt;</code> writes <code>BACKLOG.md</code> +
        enqueues <code>intent</code> task + writes initial
        <code>status.json</code> → exits
      </li>
      <li>
        Worker loop polls <code>pending/</code> → claims intent → invokes
        <code>claude -p --skill tribunal-intent</code>
      </li>
      <li>
        Intent handler writes <code>artifacts/01-intent.md</code>, enqueues
        <code>plan</code>, updates status, exits
      </li>
      <li>
        Worker loop continues through plan → implement → review → verify →
        classify → incentive, each in a fresh context
      </li>
      <li>
        Final task completes;
        <code>status.json::next_task_id = null</code> signals workflow complete
      </li>
    </ol>

    <h3>Model tier mapping</h3>
    <table>
      <tr>
        <th>Skill</th>
        <th>Tier</th>
        <th>Why</th>
      </tr>
      <tr>
        <td>intent</td>
        <td>haiku</td>
        <td>Cheap, just structures the user's input</td>
      </tr>
      <tr>
        <td>plan</td>
        <td>opus</td>
        <td>Hardest reasoning step</td>
      </tr>
      <tr>
        <td>implement</td>
        <td>sonnet</td>
        <td>Mechanical from plan</td>
      </tr>
      <tr>
        <td>review</td>
        <td>opus</td>
        <td>Adversarial probe quality matters here</td>
      </tr>
      <tr>
        <td>verify</td>
        <td>sonnet</td>
        <td>Pattern matching</td>
      </tr>
      <tr>
        <td>classify</td>
        <td>haiku</td>
        <td>Just categorization</td>
      </tr>
      <tr>
        <td>incentive</td>
        <td>sonnet</td>
        <td>Generative but bounded</td>
      </tr>
    </table>

    <h3>Compaction survival</h3>
    <p>
      At any point, if the worker dies, the next agent reads
      <code>status.json</code>, sees <code>next_task_id</code>, claims it from
      <code>pending/</code>, resumes. No conversation history needed.
    </p>

    <h2>Out of scope for v0.1</h2>
    <ul>
      <li><b>HMD triage layer</b> — v0.2</li>
      <li><b>Orphan cleanup daemon</b> — v0.2</li>
      <li><b>Multi-host coordination</b> — v0.3</li>
      <li><b>Kanban UI</b> — v0.3 if a second contributor lands</li>
      <li><b>n8n federation</b> — v0.3</li>
      <li>
        <b>Full agents-are-interchangeable abstraction</b> — applied gradually
      </li>
    </ul>

    <h2>Build plan</h2>

    <table>
      <tr>
        <th>Day</th>
        <th>Deliverable</th>
      </tr>
      <tr>
        <td>1</td>
        <td>This design doc (v0.1 spec)</td>
      </tr>
      <tr>
        <td>2</td>
        <td><code>swarm_lib/claims.py</code> + pytest</td>
      </tr>
      <tr>
        <td>3</td>
        <td><code>swarm_lib/status.py</code> + pytest</td>
      </tr>
      <tr>
        <td>4</td>
        <td><code>worker_loop.sh</code> + integration smoke test</td>
      </tr>
      <tr>
        <td>5</td>
        <td>Tribunal port: <code>tribunal-driver</code> + handler shells</td>
      </tr>
      <tr>
        <td>6</td>
        <td>Tribunal port: end-to-end run + iterate</td>
      </tr>
      <tr>
        <td>7</td>
        <td>README + minimal docs + v0.1.0 tag</td>
      </tr>
    </table>

    <h2>Open questions for review</h2>

    <ol>
      <li>
        <b>Worker ID convention</b>: PID-based (auto, transient) vs named
        (<code>mabus-tribunal-1</code>, durable). I lean named.
      </li>
      <li>
        <b>Run directory location</b>: per-consumer-repo
        <code>&lt;repo&gt;/.swarm/</code> vs shared
        <code>~/.swarm/&lt;run-id&gt;/</code>. Lean per-consumer-repo for v0.1.
      </li>
      <li>
        <b>Task ID format</b>: short prefix + run-id + type
        (<code>t.r42x.implement</code>) vs UUID vs ISO timestamp. Lean
        short-prefix-with-run-id.
      </li>
      <li>
        <b>Atomic write pattern</b>: temp-then-rename (safer) vs direct write
        (faster). Lean temp-then-rename.
      </li>
      <li>
        <b>Schema versioning</b>: strict refuse-unknown vs warn-and-attempt.
        Lean strict.
      </li>
      <li><b>License</b>: confirmed MIT.</li>
      <li>
        <b>Single source of truth for completion</b>:
        <code>status.json::completed_tasks</code> (cached) +
        <code>done/</code> directory (artifact-of-record). They should match;
        drift is a bug.
      </li>
    </ol>

    <h2>Risks</h2>
    <ul>
      <li>
        <b>Partial writes</b>: mitigated by temp+rename pattern, lib-enforced
      </li>
      <li>
        <b>Filesystem mismatch</b>: lib validates source + dest on same mount at
        startup
      </li>
      <li>
        <b>Race between scan and claim</b>: expected; <code>OSError</code> =
        lost race, loop continues
      </li>
      <li>
        <b>Compaction during status write</b>: low risk via temp-rename
        atomicity
      </li>
      <li>
        <b>Schema drift</b>: <code>metadata.*</code> preserved on
        read-modify-write
      </li>
    </ul>

    <h2>Inspiration / prior art</h2>
    <ul>
      <li>
        <b>Inkcloud Architecture Post-Mortem</b> — direct inspiration for Yield
        Rule + filesystem-as-orchestrator
      </li>
      <li>
        <b>Maildir</b> (1995) — atomic rename for email delivery; same physics
      </li>
      <li>
        <b>cron + lock files</b> + <code>/var/spool/</code> — UNIX queue
        patterns since the 90s
      </li>
      <li>
        <b>GitHub Actions reusable workflows + <code>workflow_call</code></b> —
        same yield rule applied to CI/CD (per the DO-302 work)
      </li>
    </ul>

    <div class="footer-note">
      Source: <code>~/src/swarm-lib/DESIGN.md</code><br />
      Once this gets review/redirect, <code>claims.py</code> work starts per the
      build plan. No code lands until the design is greenlit.
    </div>
  </body>
</html>