@caveman-ai/cli 1.0.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -0,0 +1,3030 @@
1
+ // Code generated from the validated practice registry. DO NOT EDIT.
2
+ // Public artifact intentionally omits sources, Cave Agent scope, and private paths.
3
+ export const PRACTICE_REGISTRY = [
4
+ {
5
+ "id": "prompt-prefix-stability",
6
+ "family": "cache",
7
+ "title": "Cache stable prompt prefixes",
8
+ "predicate": {
9
+ "harness": [
10
+ "any"
11
+ ],
12
+ "payload_shape": "repeated stable system, tool, or reference prefix",
13
+ "wire_protocol": [
14
+ "any"
15
+ ]
16
+ },
17
+ "fix": {
18
+ "prose": "Keep reusable content before volatile input and use provider-native cache controls supported by the active protocol.",
19
+ "apply": "enable_gateway_optimizer"
20
+ },
21
+ "grader": {
22
+ "type": "token_threshold",
23
+ "fixture_template": "Repeated held request with identical model-visible prefix and a token ceiling for uncached input."
24
+ },
25
+ "experiment": {
26
+ "method": "trial",
27
+ "metric": "cache_hit_ratio"
28
+ },
29
+ "evidence": {
30
+ "status": "unmeasured"
31
+ },
32
+ "never_claim": "Never attribute an automatic or caller-created cache hit to Caveman without a provider-causal active path.",
33
+ "skill_render": {
34
+ "eligible": true,
35
+ "skill_name": "caveman-practice-prompt-prefix-stability"
36
+ }
37
+ },
38
+ {
39
+ "id": "context-compression",
40
+ "family": "input_bloat",
41
+ "title": "Compress or checkpoint growing context",
42
+ "predicate": {
43
+ "harness": [
44
+ "any"
45
+ ],
46
+ "payload_shape": "large growing history with small downstream output",
47
+ "wire_protocol": [
48
+ "any"
49
+ ]
50
+ },
51
+ "fix": {
52
+ "prose": "Checkpoint stable history or use the approved recoverable compression path, preserving originals on every failure.",
53
+ "apply": "enable_sdk_config"
54
+ },
55
+ "grader": {
56
+ "type": "llm_judge",
57
+ "fixture_template": "Held long-context tasks comparing task completion before and after context reduction."
58
+ },
59
+ "experiment": {
60
+ "method": "replay",
61
+ "metric": "tokens_in"
62
+ },
63
+ "evidence": {
64
+ "status": "unmeasured"
65
+ },
66
+ "never_claim": "Never call local tokenizer deltas measured spend or verified savings.",
67
+ "skill_render": {
68
+ "eligible": false,
69
+ "skill_name": "caveman-practice-context-compression"
70
+ }
71
+ },
72
+ {
73
+ "id": "tool-result-truncation",
74
+ "family": "input_bloat",
75
+ "title": "Bound oversized tool results",
76
+ "predicate": {
77
+ "harness": [
78
+ "any"
79
+ ],
80
+ "payload_shape": "tool result larger than the task needs",
81
+ "wire_protocol": [
82
+ "any"
83
+ ]
84
+ },
85
+ "fix": {
86
+ "prose": "Project, paginate, summarize, or artifact-page large tool results before they re-enter model context.",
87
+ "apply": "enable_sdk_config"
88
+ },
89
+ "grader": {
90
+ "type": "llm_judge",
91
+ "fixture_template": "Held tool tasks where required facts appear near the start, middle, and end of a large result."
92
+ },
93
+ "experiment": {
94
+ "method": "replay",
95
+ "metric": "tokens_in"
96
+ },
97
+ "evidence": {
98
+ "status": "unmeasured"
99
+ },
100
+ "never_claim": "Never truncate without proving required facts and task outcomes remain available.",
101
+ "skill_render": {
102
+ "eligible": true,
103
+ "skill_name": "caveman-practice-tool-result-truncation"
104
+ }
105
+ },
106
+ {
107
+ "id": "tool-schema-deferral",
108
+ "family": "input_bloat",
109
+ "title": "Load tool schemas on demand",
110
+ "predicate": {
111
+ "harness": [
112
+ "any"
113
+ ],
114
+ "payload_shape": "large tool catalog with sparse per-task use",
115
+ "wire_protocol": [
116
+ "any"
117
+ ]
118
+ },
119
+ "fix": {
120
+ "prose": "Expose a small relevant tool set first and retrieve deferred schemas only when the task requires them.",
121
+ "apply": "enable_sdk_config"
122
+ },
123
+ "grader": {
124
+ "type": "tool_called",
125
+ "fixture_template": "Held tasks requiring common, deferred, and no-tool paths with expected tool names."
126
+ },
127
+ "experiment": {
128
+ "method": "replay",
129
+ "metric": "tokens_in"
130
+ },
131
+ "evidence": {
132
+ "status": "unmeasured"
133
+ },
134
+ "never_claim": "Never remove a tool schema unless deferred discovery still selects it when needed.",
135
+ "skill_render": {
136
+ "eligible": true,
137
+ "skill_name": "caveman-practice-tool-schema-deferral"
138
+ }
139
+ },
140
+ {
141
+ "id": "toon-reencoding",
142
+ "family": "input_bloat",
143
+ "title": "Re-encode uniform tabular JSON as TOON",
144
+ "predicate": {
145
+ "harness": [
146
+ "any"
147
+ ],
148
+ "payload_shape": "uniform tabular JSON measured smaller after TOON encoding",
149
+ "wire_protocol": [
150
+ "any"
151
+ ]
152
+ },
153
+ "fix": {
154
+ "prose": "Use TOON only when forced or flag-gated and measured smaller, never for tool-call arguments.",
155
+ "apply": "enable_sdk_config"
156
+ },
157
+ "grader": {
158
+ "type": "json_schema",
159
+ "fixture_template": "Held uniform tables decoded back to the expected schema before task-quality grading."
160
+ },
161
+ "experiment": {
162
+ "method": "replay",
163
+ "metric": "tokens_in"
164
+ },
165
+ "evidence": {
166
+ "status": "unmeasured"
167
+ },
168
+ "never_claim": "forced-only, never tool-call args, not byte-safe, inferred until causal proof.",
169
+ "skill_render": {
170
+ "eligible": false,
171
+ "skill_name": "caveman-practice-toon-reencoding"
172
+ }
173
+ },
174
+ {
175
+ "id": "anthropic-cache-breakpoints",
176
+ "family": "cache",
177
+ "title": "Place an Anthropic explicit cache breakpoint",
178
+ "predicate": {
179
+ "harness": [
180
+ "any"
181
+ ],
182
+ "payload_shape": "stable tool definitions or system instructions repeated across requests",
183
+ "wire_protocol": [
184
+ "anthropic"
185
+ ]
186
+ },
187
+ "fix": {
188
+ "prose": "Add an explicit 5-minute cache marker to the last non-deferred tool, or to the last system block when no eligible tool is loaded; the marker changes only the upstream request.",
189
+ "apply": "enable_gateway_optimizer"
190
+ },
191
+ "grader": {
192
+ "type": "token_threshold",
193
+ "fixture_template": "Repeated held request pair with an identical tools-and-system prefix, provider-reported cache usage, and an unchanged model-visible projection."
194
+ },
195
+ "experiment": {
196
+ "method": "trial",
197
+ "metric": "cache_hit_ratio"
198
+ },
199
+ "evidence": {
200
+ "status": "unmeasured"
201
+ },
202
+ "never_claim": "Never describe this explicit static-prefix marker as Anthropic automatic caching or claim a saving from placement alone; its write can be net-negative without reuse.",
203
+ "skill_render": {
204
+ "eligible": true,
205
+ "skill_name": "caveman-practice-anthropic-cache-breakpoints"
206
+ }
207
+ },
208
+ {
209
+ "id": "anthropic-static-prefix-breakpoint",
210
+ "family": "cache",
211
+ "title": "Place Anthropic breakpoint after static content",
212
+ "predicate": {
213
+ "harness": [
214
+ "any"
215
+ ],
216
+ "payload_shape": "static prefix followed by request-specific suffix",
217
+ "wire_protocol": [
218
+ "anthropic"
219
+ ]
220
+ },
221
+ "fix": {
222
+ "prose": "Place cache control on the last static content block and keep timestamps, request identifiers, and live input after it.",
223
+ "apply": "stabilize_prefix"
224
+ },
225
+ "grader": {
226
+ "type": "token_threshold",
227
+ "fixture_template": "Repeated held requests varying only suffix fields with bounded cache writes."
228
+ },
229
+ "experiment": {
230
+ "method": "trial",
231
+ "metric": "cache_hit_ratio"
232
+ },
233
+ "evidence": {
234
+ "status": "unmeasured"
235
+ },
236
+ "never_claim": "Never describe a breakpoint as effective until provider usage shows reuse.",
237
+ "skill_render": {
238
+ "eligible": true,
239
+ "skill_name": "caveman-practice-anthropic-static-prefix-breakpoint"
240
+ }
241
+ },
242
+ {
243
+ "id": "anthropic-tool-definitions-cache",
244
+ "family": "cache",
245
+ "title": "Cache stable Anthropic tool definitions",
246
+ "predicate": {
247
+ "harness": [
248
+ "any"
249
+ ],
250
+ "payload_shape": "stable tool definitions repeated across agent turns",
251
+ "wire_protocol": [
252
+ "anthropic"
253
+ ]
254
+ },
255
+ "fix": {
256
+ "prose": "Mark the last stable tool definition as cacheable and defer rarely used tools without breaking prefix order.",
257
+ "apply": "stabilize_prefix"
258
+ },
259
+ "grader": {
260
+ "type": "tool_called",
261
+ "fixture_template": "Held tasks selecting tools before and after cache placement, including one deferred tool."
262
+ },
263
+ "experiment": {
264
+ "method": "trial",
265
+ "metric": "cache_hit_ratio"
266
+ },
267
+ "evidence": {
268
+ "status": "unmeasured"
269
+ },
270
+ "never_claim": "Never change tool availability merely to improve cache metrics.",
271
+ "skill_render": {
272
+ "eligible": true,
273
+ "skill_name": "caveman-practice-anthropic-tool-definitions-cache"
274
+ }
275
+ },
276
+ {
277
+ "id": "anthropic-cache-ttl-right-sizing",
278
+ "family": "cache",
279
+ "title": "Match Anthropic cache TTL to request cadence",
280
+ "predicate": {
281
+ "harness": [
282
+ "any"
283
+ ],
284
+ "payload_shape": "stable prefix reused on a known cadence",
285
+ "wire_protocol": [
286
+ "anthropic"
287
+ ]
288
+ },
289
+ "fix": {
290
+ "prose": "Use the shorter lifetime for frequent reuse and test the longer lifetime only when gaps regularly exceed the short window.",
291
+ "apply": "stabilize_prefix"
292
+ },
293
+ "grader": {
294
+ "type": "cost_threshold",
295
+ "fixture_template": "Held cadence replay comparing supported TTL configurations under one cost ceiling."
296
+ },
297
+ "experiment": {
298
+ "method": "replay",
299
+ "metric": "$/day band"
300
+ },
301
+ "evidence": {
302
+ "status": "unmeasured"
303
+ },
304
+ "never_claim": "Never select a longer TTL from intuition; cache-write cost and actual cadence must support it.",
305
+ "skill_render": {
306
+ "eligible": true,
307
+ "skill_name": "caveman-practice-anthropic-cache-ttl-right-sizing"
308
+ }
309
+ },
310
+ {
311
+ "id": "anthropic-mixed-ttl-ordering",
312
+ "family": "cache",
313
+ "title": "Order mixed Anthropic cache lifetimes",
314
+ "predicate": {
315
+ "harness": [
316
+ "any"
317
+ ],
318
+ "payload_shape": "request using both long and short cache lifetimes",
319
+ "wire_protocol": [
320
+ "anthropic"
321
+ ]
322
+ },
323
+ "fix": {
324
+ "prose": "Place longer-lived cache entries before shorter-lived entries and keep each boundary aligned with its content cadence.",
325
+ "apply": "stabilize_prefix"
326
+ },
327
+ "grader": {
328
+ "type": "http_status",
329
+ "fixture_template": "Held mixed-TTL request expected to succeed and report separate cache-write buckets."
330
+ },
331
+ "experiment": {
332
+ "method": "trial",
333
+ "metric": "cache_hit_ratio"
334
+ },
335
+ "evidence": {
336
+ "status": "unmeasured"
337
+ },
338
+ "never_claim": "Never reorder model-visible content solely to satisfy TTL ordering without an outcome gate.",
339
+ "skill_render": {
340
+ "eligible": true,
341
+ "skill_name": "caveman-practice-anthropic-mixed-ttl-ordering"
342
+ }
343
+ },
344
+ {
345
+ "id": "anthropic-cache-observability",
346
+ "family": "cache",
347
+ "title": "Measure Anthropic cache reads and writes",
348
+ "predicate": {
349
+ "harness": [
350
+ "any"
351
+ ],
352
+ "payload_shape": "cache-enabled traffic without read and write accounting",
353
+ "wire_protocol": [
354
+ "anthropic"
355
+ ]
356
+ },
357
+ "fix": {
358
+ "prose": "Record provider-reported cache creation, cache read, and lifetime-specific write fields before tuning placement.",
359
+ "apply": "manual"
360
+ },
361
+ "grader": {
362
+ "type": "json_path_assertion",
363
+ "fixture_template": "Held cached response requiring cache usage fields to exist with non-negative values."
364
+ },
365
+ "experiment": {
366
+ "method": "trial",
367
+ "metric": "cache_hit_ratio"
368
+ },
369
+ "evidence": {
370
+ "status": "unmeasured"
371
+ },
372
+ "never_claim": "Never estimate cache effectiveness from request count when provider usage fields are available.",
373
+ "skill_render": {
374
+ "eligible": true,
375
+ "skill_name": "caveman-practice-anthropic-cache-observability"
376
+ }
377
+ },
378
+ {
379
+ "id": "anthropic-message-batch-offload",
380
+ "family": "routing",
381
+ "title": "Batch non-interactive Anthropic work",
382
+ "predicate": {
383
+ "harness": [
384
+ "any"
385
+ ],
386
+ "payload_shape": "bulk or scheduled Messages work with no interactive latency need",
387
+ "wire_protocol": [
388
+ "anthropic"
389
+ ]
390
+ },
391
+ "fix": {
392
+ "prose": "Send eligible independent requests through Message Batches and preserve per-item identifiers, errors, and completion checks.",
393
+ "apply": "enable_sdk_config"
394
+ },
395
+ "grader": {
396
+ "type": "exact_match",
397
+ "fixture_template": "Held synchronous and batch outputs normalized to the same expected task result."
398
+ },
399
+ "experiment": {
400
+ "method": "local_ab",
401
+ "metric": "$/day band"
402
+ },
403
+ "evidence": {
404
+ "status": "unmeasured"
405
+ },
406
+ "never_claim": "Never batch latency-sensitive work or treat accepted batch jobs as completed results.",
407
+ "skill_render": {
408
+ "eligible": true,
409
+ "skill_name": "caveman-practice-anthropic-message-batch-offload"
410
+ }
411
+ },
412
+ {
413
+ "id": "anthropic-token-count-preflight",
414
+ "family": "input_bloat",
415
+ "title": "Preflight large Anthropic requests with provider token counts",
416
+ "predicate": {
417
+ "harness": [
418
+ "any"
419
+ ],
420
+ "payload_shape": "large request near a policy, cost, or context threshold",
421
+ "wire_protocol": [
422
+ "anthropic"
423
+ ]
424
+ },
425
+ "fix": {
426
+ "prose": "Call the provider token-count endpoint for sampled or boundary requests before choosing a reduction path.",
427
+ "apply": "manual"
428
+ },
429
+ "grader": {
430
+ "type": "token_threshold",
431
+ "fixture_template": "Held request whose provider count must remain below the configured input ceiling."
432
+ },
433
+ "experiment": {
434
+ "method": "shadow",
435
+ "metric": "tokens_in"
436
+ },
437
+ "evidence": {
438
+ "status": "unmeasured"
439
+ },
440
+ "never_claim": "Never substitute a local tokenizer estimate for provider-measured baseline tokens.",
441
+ "skill_render": {
442
+ "eligible": true,
443
+ "skill_name": "caveman-practice-anthropic-token-count-preflight"
444
+ }
445
+ },
446
+ {
447
+ "id": "anthropic-strict-tool-use",
448
+ "family": "reliability",
449
+ "title": "Constrain Anthropic tool arguments",
450
+ "predicate": {
451
+ "harness": [
452
+ "any"
453
+ ],
454
+ "payload_shape": "tool calls retried after invalid arguments",
455
+ "wire_protocol": [
456
+ "anthropic"
457
+ ]
458
+ },
459
+ "fix": {
460
+ "prose": "Use strict tool schemas for stable contracts and reject invalid arguments without replaying unrelated context.",
461
+ "apply": "enable_sdk_config"
462
+ },
463
+ "grader": {
464
+ "type": "tool_argument_assertion",
465
+ "fixture_template": "Held tool calls covering valid, boundary, and invalid argument shapes."
466
+ },
467
+ "experiment": {
468
+ "method": "replay",
469
+ "metric": "$/day band"
470
+ },
471
+ "evidence": {
472
+ "status": "unmeasured"
473
+ },
474
+ "never_claim": "Never assume schema-valid arguments are semantically correct.",
475
+ "skill_render": {
476
+ "eligible": true,
477
+ "skill_name": "caveman-practice-anthropic-strict-tool-use"
478
+ }
479
+ },
480
+ {
481
+ "id": "anthropic-output-budget",
482
+ "family": "routing",
483
+ "title": "Bound Anthropic output for narrow tasks",
484
+ "predicate": {
485
+ "harness": [
486
+ "any"
487
+ ],
488
+ "payload_shape": "narrow task with output allowance far above observed need",
489
+ "wire_protocol": [
490
+ "anthropic"
491
+ ]
492
+ },
493
+ "fix": {
494
+ "prose": "Set a workload-specific output ceiling and keep completion checks that fail when required content is cut.",
495
+ "apply": "enable_sdk_config"
496
+ },
497
+ "grader": {
498
+ "type": "llm_judge",
499
+ "fixture_template": "Held short and longest-valid responses judged for completeness under the new ceiling."
500
+ },
501
+ "experiment": {
502
+ "method": "canary",
503
+ "metric": "tokens_out"
504
+ },
505
+ "evidence": {
506
+ "status": "unmeasured"
507
+ },
508
+ "never_claim": "Never lower an output ceiling from averages alone; tail cases must still complete.",
509
+ "skill_render": {
510
+ "eligible": true,
511
+ "skill_name": "caveman-practice-anthropic-output-budget"
512
+ }
513
+ },
514
+ {
515
+ "id": "openai-prefix-ordering",
516
+ "family": "cache",
517
+ "title": "Put reusable OpenAI content first",
518
+ "predicate": {
519
+ "harness": [
520
+ "any"
521
+ ],
522
+ "payload_shape": "long request with stable and variable sections",
523
+ "wire_protocol": [
524
+ "openai"
525
+ ]
526
+ },
527
+ "fix": {
528
+ "prose": "Place stable instructions, examples, tools, and reference blocks before request-specific content so exact prefix matching can reuse them.",
529
+ "apply": "stabilize_prefix"
530
+ },
531
+ "grader": {
532
+ "type": "token_threshold",
533
+ "fixture_template": "Repeated held requests with a stable prefix and changing suffix under an uncached-token ceiling."
534
+ },
535
+ "experiment": {
536
+ "method": "trial",
537
+ "metric": "cache_hit_ratio"
538
+ },
539
+ "evidence": {
540
+ "status": "unmeasured"
541
+ },
542
+ "never_claim": "Never attribute an automatic OpenAI cache hit to Caveman on a per-request basis.",
543
+ "skill_render": {
544
+ "eligible": true,
545
+ "skill_name": "caveman-practice-openai-prefix-ordering"
546
+ }
547
+ },
548
+ {
549
+ "id": "openai-prompt-cache-key",
550
+ "family": "cache",
551
+ "title": "Partition OpenAI cache affinity keys",
552
+ "predicate": {
553
+ "harness": [
554
+ "any"
555
+ ],
556
+ "payload_shape": "repeated prefixes split across unrelated cache routing keys",
557
+ "wire_protocol": [
558
+ "openai"
559
+ ]
560
+ },
561
+ "fix": {
562
+ "prose": "Use a stable low-cardinality cache key per reusable prompt template, without embedding request identifiers or secrets.",
563
+ "apply": "enable_gateway_optimizer"
564
+ },
565
+ "grader": {
566
+ "type": "token_threshold",
567
+ "fixture_template": "Repeated held template requests sharing one safe key and bounded uncached input."
568
+ },
569
+ "experiment": {
570
+ "method": "trial",
571
+ "metric": "cache_hit_ratio"
572
+ },
573
+ "evidence": {
574
+ "status": "unmeasured"
575
+ },
576
+ "never_claim": "Never describe routing affinity as a guaranteed or provider-causal cache hit.",
577
+ "skill_render": {
578
+ "eligible": true,
579
+ "skill_name": "caveman-practice-openai-prompt-cache-key"
580
+ }
581
+ },
582
+ {
583
+ "id": "openai-explicit-cache-breakpoint",
584
+ "family": "cache",
585
+ "title": "Mark explicit OpenAI cache boundaries",
586
+ "predicate": {
587
+ "harness": [
588
+ "any"
589
+ ],
590
+ "payload_shape": "GPT-5.6 request needing caller-controlled cache placement",
591
+ "wire_protocol": [
592
+ "openai"
593
+ ]
594
+ },
595
+ "fix": {
596
+ "prose": "Use explicit mode and mark the end of reusable content when automatic placement would write volatile or low-reuse prefixes.",
597
+ "apply": "stabilize_prefix"
598
+ },
599
+ "grader": {
600
+ "type": "token_threshold",
601
+ "fixture_template": "Repeated held GPT-5.6 requests with explicit breakpoint and bounded cache-write tokens."
602
+ },
603
+ "experiment": {
604
+ "method": "trial",
605
+ "metric": "cache_hit_ratio"
606
+ },
607
+ "evidence": {
608
+ "status": "unmeasured"
609
+ },
610
+ "never_claim": "Never enable explicit mode on unsupported models or claim causality before end-to-end provider proof.",
611
+ "skill_render": {
612
+ "eligible": true,
613
+ "skill_name": "caveman-practice-openai-explicit-cache-breakpoint"
614
+ }
615
+ },
616
+ {
617
+ "id": "openai-cache-write-read-observability",
618
+ "family": "cache",
619
+ "title": "Measure OpenAI cache writes and reads",
620
+ "predicate": {
621
+ "harness": [
622
+ "any"
623
+ ],
624
+ "payload_shape": "cache-eligible OpenAI traffic without cached and write token tracking",
625
+ "wire_protocol": [
626
+ "openai"
627
+ ]
628
+ },
629
+ "fix": {
630
+ "prose": "Record provider-reported cached tokens and cache-write tokens before changing cache mode, boundaries, or retention.",
631
+ "apply": "manual"
632
+ },
633
+ "grader": {
634
+ "type": "json_path_assertion",
635
+ "fixture_template": "Held response requiring cache usage fields to exist and contain non-negative values."
636
+ },
637
+ "experiment": {
638
+ "method": "trial",
639
+ "metric": "cache_hit_ratio"
640
+ },
641
+ "evidence": {
642
+ "status": "unmeasured"
643
+ },
644
+ "never_claim": "Never treat cached token counters as Caveman-caused savings on automatic caching paths.",
645
+ "skill_render": {
646
+ "eligible": true,
647
+ "skill_name": "caveman-practice-openai-cache-write-read-observability"
648
+ }
649
+ },
650
+ {
651
+ "id": "openai-batch-offload",
652
+ "family": "routing",
653
+ "title": "Batch non-interactive OpenAI work",
654
+ "predicate": {
655
+ "harness": [
656
+ "any"
657
+ ],
658
+ "payload_shape": "bulk or scheduled requests with no interactive latency need",
659
+ "wire_protocol": [
660
+ "openai"
661
+ ]
662
+ },
663
+ "fix": {
664
+ "prose": "Move independent background requests to Batch, preserve custom identifiers, and reconcile every success and error.",
665
+ "apply": "enable_sdk_config"
666
+ },
667
+ "grader": {
668
+ "type": "exact_match",
669
+ "fixture_template": "Held synchronous and batch outputs normalized to the same expected task result."
670
+ },
671
+ "experiment": {
672
+ "method": "local_ab",
673
+ "metric": "$/day band"
674
+ },
675
+ "evidence": {
676
+ "status": "unmeasured"
677
+ },
678
+ "never_claim": "Never count queued work as completed or batch latency as acceptable without an explicit workload SLA.",
679
+ "skill_render": {
680
+ "eligible": true,
681
+ "skill_name": "caveman-practice-openai-batch-offload"
682
+ }
683
+ },
684
+ {
685
+ "id": "openai-flex-processing",
686
+ "family": "routing",
687
+ "title": "Use Flex for delay-tolerant OpenAI work",
688
+ "predicate": {
689
+ "harness": [
690
+ "any"
691
+ ],
692
+ "payload_shape": "non-production or asynchronous work tolerant of slower and variable service",
693
+ "wire_protocol": [
694
+ "openai"
695
+ ]
696
+ },
697
+ "fix": {
698
+ "prose": "Select Flex only for workloads whose latency and availability contract permits it, with fallback and queue timing observed.",
699
+ "apply": "enable_sdk_config"
700
+ },
701
+ "grader": {
702
+ "type": "latency_threshold",
703
+ "fixture_template": "Held asynchronous workload with a declared completion-time ceiling."
704
+ },
705
+ "experiment": {
706
+ "method": "local_ab",
707
+ "metric": "$/day band"
708
+ },
709
+ "evidence": {
710
+ "status": "unmeasured"
711
+ },
712
+ "never_claim": "Never move interactive or availability-sensitive traffic to Flex from price alone.",
713
+ "skill_render": {
714
+ "eligible": true,
715
+ "skill_name": "caveman-practice-openai-flex-processing"
716
+ }
717
+ },
718
+ {
719
+ "id": "reasoning-effort",
720
+ "family": "routing",
721
+ "title": "Tune OpenAI reasoning effort by workload",
722
+ "predicate": {
723
+ "harness": [
724
+ "any"
725
+ ],
726
+ "payload_shape": "fixed high reasoning effort across easy and hard requests",
727
+ "wire_protocol": [
728
+ "openai"
729
+ ]
730
+ },
731
+ "fix": {
732
+ "prose": "Evaluate the current effort and one lower setting on representative cases, then scope the lower setting to passing work.",
733
+ "apply": "review_reasoning"
734
+ },
735
+ "grader": {
736
+ "type": "llm_judge",
737
+ "fixture_template": "Held easy, boundary, and hard cases judged by a different model family."
738
+ },
739
+ "experiment": {
740
+ "method": "canary",
741
+ "metric": "tokens_out"
742
+ },
743
+ "evidence": {
744
+ "status": "unmeasured"
745
+ },
746
+ "never_claim": "Never choose reasoning effort from task labels without outcome evaluation.",
747
+ "skill_render": {
748
+ "eligible": true,
749
+ "skill_name": "caveman-practice-reasoning-effort"
750
+ }
751
+ },
752
+ {
753
+ "id": "openai-persisted-reasoning",
754
+ "family": "input_bloat",
755
+ "title": "Reuse relevant OpenAI reasoning state",
756
+ "predicate": {
757
+ "harness": [
758
+ "any"
759
+ ],
760
+ "payload_shape": "multi-turn task that re-derives stable prior reasoning",
761
+ "wire_protocol": [
762
+ "openai"
763
+ ]
764
+ },
765
+ "fix": {
766
+ "prose": "Preserve supported reasoning items across related turns and stop carrying them when prior reasoning is no longer relevant.",
767
+ "apply": "enable_sdk_config"
768
+ },
769
+ "grader": {
770
+ "type": "llm_judge",
771
+ "fixture_template": "Held multi-turn task comparing continued and fresh reasoning paths for answer quality."
772
+ },
773
+ "experiment": {
774
+ "method": "replay",
775
+ "metric": "tokens_in"
776
+ },
777
+ "evidence": {
778
+ "status": "unmeasured"
779
+ },
780
+ "never_claim": "Never reuse reasoning across unrelated tasks or assume reuse lowers total cost without measurement.",
781
+ "skill_render": {
782
+ "eligible": true,
783
+ "skill_name": "caveman-practice-openai-persisted-reasoning"
784
+ }
785
+ },
786
+ {
787
+ "id": "openai-structured-output",
788
+ "family": "reliability",
789
+ "title": "Replace parse retries with strict OpenAI output",
790
+ "predicate": {
791
+ "harness": [
792
+ "any"
793
+ ],
794
+ "payload_shape": "free-form response reparsed or retried into a known schema",
795
+ "wire_protocol": [
796
+ "openai"
797
+ ]
798
+ },
799
+ "fix": {
800
+ "prose": "Use strict structured output for stable schemas and retain semantic validation for fields the schema cannot prove.",
801
+ "apply": "enable_sdk_config"
802
+ },
803
+ "grader": {
804
+ "type": "json_schema",
805
+ "fixture_template": "Held outputs covering required, optional, boundary, and refused result shapes."
806
+ },
807
+ "experiment": {
808
+ "method": "replay",
809
+ "metric": "$/day band"
810
+ },
811
+ "evidence": {
812
+ "status": "unmeasured"
813
+ },
814
+ "never_claim": "Never equate schema validity with factual or business correctness.",
815
+ "skill_render": {
816
+ "eligible": true,
817
+ "skill_name": "caveman-practice-openai-structured-output"
818
+ }
819
+ },
820
+ {
821
+ "id": "output-brevity",
822
+ "family": "routing",
823
+ "title": "Bound OpenAI output for narrow tasks",
824
+ "predicate": {
825
+ "harness": [
826
+ "any"
827
+ ],
828
+ "payload_shape": "narrow task with output allowance far above observed need",
829
+ "wire_protocol": [
830
+ "openai"
831
+ ]
832
+ },
833
+ "fix": {
834
+ "prose": "Set a workload-specific output ceiling and gate longest-valid cases for completeness.",
835
+ "apply": "enable_sdk_config"
836
+ },
837
+ "grader": {
838
+ "type": "llm_judge",
839
+ "fixture_template": "Held short and longest-valid responses judged for completeness under the new ceiling."
840
+ },
841
+ "experiment": {
842
+ "method": "canary",
843
+ "metric": "tokens_out"
844
+ },
845
+ "evidence": {
846
+ "status": "unmeasured"
847
+ },
848
+ "never_claim": "Never lower an output ceiling from median behavior alone.",
849
+ "skill_render": {
850
+ "eligible": true,
851
+ "skill_name": "caveman-practice-output-brevity"
852
+ }
853
+ },
854
+ {
855
+ "id": "gemini-prefix-ordering",
856
+ "family": "cache",
857
+ "title": "Put reusable Gemini content first",
858
+ "predicate": {
859
+ "harness": [
860
+ "any"
861
+ ],
862
+ "payload_shape": "repeated long context with a stable prefix",
863
+ "wire_protocol": [
864
+ "gemini"
865
+ ]
866
+ },
867
+ "fix": {
868
+ "prose": "Place stable context before changing user content so implicit caching can match a longer prefix.",
869
+ "apply": "stabilize_prefix"
870
+ },
871
+ "grader": {
872
+ "type": "token_threshold",
873
+ "fixture_template": "Repeated held GenerateContent requests with bounded non-cached input."
874
+ },
875
+ "experiment": {
876
+ "method": "trial",
877
+ "metric": "cache_hit_ratio"
878
+ },
879
+ "evidence": {
880
+ "status": "unmeasured"
881
+ },
882
+ "never_claim": "Never attribute implicit Gemini cache hits to Caveman on a per-request basis.",
883
+ "skill_render": {
884
+ "eligible": true,
885
+ "skill_name": "caveman-practice-gemini-prefix-ordering"
886
+ }
887
+ },
888
+ {
889
+ "id": "gemini-cache-observability",
890
+ "family": "cache",
891
+ "title": "Measure Gemini cached token use",
892
+ "predicate": {
893
+ "harness": [
894
+ "any"
895
+ ],
896
+ "payload_shape": "cache-enabled Gemini traffic without cached token accounting",
897
+ "wire_protocol": [
898
+ "gemini"
899
+ ]
900
+ },
901
+ "fix": {
902
+ "prose": "Record provider usage metadata for cached content, prompt, candidates, and thoughts before tuning caching.",
903
+ "apply": "manual"
904
+ },
905
+ "grader": {
906
+ "type": "json_path_assertion",
907
+ "fixture_template": "Held response requiring cached-content and total token usage fields to exist."
908
+ },
909
+ "experiment": {
910
+ "method": "trial",
911
+ "metric": "cache_hit_ratio"
912
+ },
913
+ "evidence": {
914
+ "status": "unmeasured"
915
+ },
916
+ "never_claim": "Never infer caused savings from implicit cached-token counters.",
917
+ "skill_render": {
918
+ "eligible": true,
919
+ "skill_name": "caveman-practice-gemini-cache-observability"
920
+ }
921
+ },
922
+ {
923
+ "id": "gemini-batch-offload",
924
+ "family": "routing",
925
+ "title": "Batch non-interactive Gemini work",
926
+ "predicate": {
927
+ "harness": [
928
+ "any"
929
+ ],
930
+ "payload_shape": "bulk or scheduled GenerateContent work with no interactive latency need",
931
+ "wire_protocol": [
932
+ "gemini"
933
+ ]
934
+ },
935
+ "fix": {
936
+ "prose": "Move independent background requests to Batch and reconcile each result or error before downstream use.",
937
+ "apply": "enable_sdk_config"
938
+ },
939
+ "grader": {
940
+ "type": "exact_match",
941
+ "fixture_template": "Held synchronous and batch outputs normalized to the same expected task result."
942
+ },
943
+ "experiment": {
944
+ "method": "local_ab",
945
+ "metric": "$/day band"
946
+ },
947
+ "evidence": {
948
+ "status": "unmeasured"
949
+ },
950
+ "never_claim": "Never treat job submission as task completion or ignore per-item failures.",
951
+ "skill_render": {
952
+ "eligible": true,
953
+ "skill_name": "caveman-practice-gemini-batch-offload"
954
+ }
955
+ },
956
+ {
957
+ "id": "gemini-token-count-preflight",
958
+ "family": "input_bloat",
959
+ "title": "Preflight large Gemini requests with provider counts",
960
+ "predicate": {
961
+ "harness": [
962
+ "any"
963
+ ],
964
+ "payload_shape": "large multimodal or text request near a threshold",
965
+ "wire_protocol": [
966
+ "gemini"
967
+ ]
968
+ },
969
+ "fix": {
970
+ "prose": "Use countTokens for sampled or boundary requests before deciding whether to trim, split, or reject them.",
971
+ "apply": "manual"
972
+ },
973
+ "grader": {
974
+ "type": "token_threshold",
975
+ "fixture_template": "Held request whose provider token count must remain below the configured input ceiling."
976
+ },
977
+ "experiment": {
978
+ "method": "shadow",
979
+ "metric": "tokens_in"
980
+ },
981
+ "evidence": {
982
+ "status": "unmeasured"
983
+ },
984
+ "never_claim": "Never price multimodal input with a local text tokenizer.",
985
+ "skill_render": {
986
+ "eligible": true,
987
+ "skill_name": "caveman-practice-gemini-token-count-preflight"
988
+ }
989
+ },
990
+ {
991
+ "id": "gemini-thinking-budget-right-sizing",
992
+ "family": "routing",
993
+ "title": "Tune Gemini thinking budget",
994
+ "predicate": {
995
+ "harness": [
996
+ "any"
997
+ ],
998
+ "payload_shape": "fixed thinking budget across easy and hard requests",
999
+ "wire_protocol": [
1000
+ "gemini"
1001
+ ]
1002
+ },
1003
+ "fix": {
1004
+ "prose": "Evaluate lower thinking settings on representative tasks and scope any change to workloads that preserve quality.",
1005
+ "apply": "review_reasoning"
1006
+ },
1007
+ "grader": {
1008
+ "type": "llm_judge",
1009
+ "fixture_template": "Held easy, boundary, and hard cases judged for correctness under each thinking setting."
1010
+ },
1011
+ "experiment": {
1012
+ "method": "canary",
1013
+ "metric": "tokens_out"
1014
+ },
1015
+ "evidence": {
1016
+ "status": "unmeasured"
1017
+ },
1018
+ "never_claim": "Never lower thinking from visible answer length alone.",
1019
+ "skill_render": {
1020
+ "eligible": true,
1021
+ "skill_name": "caveman-practice-gemini-thinking-budget-right-sizing"
1022
+ }
1023
+ },
1024
+ {
1025
+ "id": "gemini-flash-right-sizing",
1026
+ "family": "routing",
1027
+ "title": "Evaluate Gemini Flash for narrow work",
1028
+ "predicate": {
1029
+ "harness": [
1030
+ "any"
1031
+ ],
1032
+ "payload_shape": "Gemini Pro-class model serving narrow high-volume work",
1033
+ "wire_protocol": [
1034
+ "gemini"
1035
+ ]
1036
+ },
1037
+ "fix": {
1038
+ "prose": "Evaluate a compatible Flash candidate on held cases and canary only the passing workload.",
1039
+ "apply": "right_size_model"
1040
+ },
1041
+ "grader": {
1042
+ "type": "llm_judge",
1043
+ "fixture_template": "Held representative cases with candidate and baseline outputs judged outside their model family."
1044
+ },
1045
+ "experiment": {
1046
+ "method": "canary",
1047
+ "metric": "$/day band"
1048
+ },
1049
+ "evidence": {
1050
+ "status": "unmeasured"
1051
+ },
1052
+ "never_claim": "Never infer parity from model-family naming or unit price.",
1053
+ "skill_render": {
1054
+ "eligible": true,
1055
+ "skill_name": "caveman-practice-gemini-flash-right-sizing"
1056
+ }
1057
+ },
1058
+ {
1059
+ "id": "bedrock-cachepoint-placement",
1060
+ "family": "cache",
1061
+ "title": "Place Bedrock cache checkpoints",
1062
+ "predicate": {
1063
+ "harness": [
1064
+ "any"
1065
+ ],
1066
+ "payload_shape": "supported Bedrock model with repeated stable context",
1067
+ "wire_protocol": [
1068
+ "any"
1069
+ ]
1070
+ },
1071
+ "fix": {
1072
+ "prose": "Add supported cache checkpoints after stable tools, system content, or messages while respecting model-specific minimums and order.",
1073
+ "apply": "enable_gateway_optimizer"
1074
+ },
1075
+ "grader": {
1076
+ "type": "token_threshold",
1077
+ "fixture_template": "Repeated held Converse or InvokeModel requests with bounded uncached input."
1078
+ },
1079
+ "experiment": {
1080
+ "method": "trial",
1081
+ "metric": "cache_hit_ratio"
1082
+ },
1083
+ "evidence": {
1084
+ "status": "unmeasured"
1085
+ },
1086
+ "never_claim": "Never mark Bedrock savings verified until Caveman proves the provider-causal path end to end.",
1087
+ "skill_render": {
1088
+ "eligible": true,
1089
+ "skill_name": "caveman-practice-bedrock-cachepoint-placement"
1090
+ }
1091
+ },
1092
+ {
1093
+ "id": "bedrock-cache-observability",
1094
+ "family": "cache",
1095
+ "title": "Measure Bedrock cache reads and writes",
1096
+ "predicate": {
1097
+ "harness": [
1098
+ "any"
1099
+ ],
1100
+ "payload_shape": "Bedrock cache-enabled traffic without normalized cache counters",
1101
+ "wire_protocol": [
1102
+ "any"
1103
+ ]
1104
+ },
1105
+ "fix": {
1106
+ "prose": "Record cache read, cache write, lifetime detail, and non-cached input from provider responses before tuning checkpoints.",
1107
+ "apply": "manual"
1108
+ },
1109
+ "grader": {
1110
+ "type": "json_path_assertion",
1111
+ "fixture_template": "Held cached response requiring read, write, and non-cached usage fields."
1112
+ },
1113
+ "experiment": {
1114
+ "method": "trial",
1115
+ "metric": "cache_hit_ratio"
1116
+ },
1117
+ "evidence": {
1118
+ "status": "unmeasured"
1119
+ },
1120
+ "never_claim": "Never sum Bedrock input and cache counters without following the provider response contract.",
1121
+ "skill_render": {
1122
+ "eligible": true,
1123
+ "skill_name": "caveman-practice-bedrock-cache-observability"
1124
+ }
1125
+ },
1126
+ {
1127
+ "id": "claude-code-claude-md-budget",
1128
+ "family": "input_bloat",
1129
+ "title": "Keep CLAUDE.md load-bearing and small",
1130
+ "predicate": {
1131
+ "harness": [
1132
+ "claude-code"
1133
+ ],
1134
+ "payload_shape": "large CLAUDE.md loaded into every turn",
1135
+ "wire_protocol": [
1136
+ "anthropic"
1137
+ ]
1138
+ },
1139
+ "fix": {
1140
+ "prose": "Keep durable project rules in CLAUDE.md and move procedures or long references into on-demand skills.",
1141
+ "apply": "enable_sdk_config"
1142
+ },
1143
+ "grader": {
1144
+ "type": "exact_match",
1145
+ "fixture_template": "Held repository tasks requiring each retained rule after the context edit."
1146
+ },
1147
+ "experiment": {
1148
+ "method": "local_ab",
1149
+ "metric": "tokens_in"
1150
+ },
1151
+ "evidence": {
1152
+ "status": "unmeasured"
1153
+ },
1154
+ "never_claim": "Never remove a rule without proving its behavior remains enforced or available on demand.",
1155
+ "skill_render": {
1156
+ "eligible": true,
1157
+ "skill_name": "caveman-practice-claude-code-claude-md-budget"
1158
+ }
1159
+ },
1160
+ {
1161
+ "id": "claude-code-scoped-rules",
1162
+ "family": "input_bloat",
1163
+ "title": "Scope Claude Code rules to relevant paths",
1164
+ "predicate": {
1165
+ "harness": [
1166
+ "claude-code"
1167
+ ],
1168
+ "payload_shape": "repository-wide instructions relevant to one subtree",
1169
+ "wire_protocol": [
1170
+ "anthropic"
1171
+ ]
1172
+ },
1173
+ "fix": {
1174
+ "prose": "Move subtree-specific guidance to the narrowest supported project location so unrelated work does not carry it.",
1175
+ "apply": "enable_sdk_config"
1176
+ },
1177
+ "grader": {
1178
+ "type": "exact_match",
1179
+ "fixture_template": "Held tasks inside and outside the target subtree verifying rule presence only where needed."
1180
+ },
1181
+ "experiment": {
1182
+ "method": "local_ab",
1183
+ "metric": "tokens_in"
1184
+ },
1185
+ "evidence": {
1186
+ "status": "unmeasured"
1187
+ },
1188
+ "never_claim": "Never narrow scope when the rule protects repository-wide correctness or safety.",
1189
+ "skill_render": {
1190
+ "eligible": true,
1191
+ "skill_name": "caveman-practice-claude-code-scoped-rules"
1192
+ }
1193
+ },
1194
+ {
1195
+ "id": "claude-code-skill-deferral",
1196
+ "family": "input_bloat",
1197
+ "title": "Load Claude Code procedures as skills",
1198
+ "predicate": {
1199
+ "harness": [
1200
+ "claude-code"
1201
+ ],
1202
+ "payload_shape": "long workflow instructions always loaded but rarely used",
1203
+ "wire_protocol": [
1204
+ "anthropic"
1205
+ ]
1206
+ },
1207
+ "fix": {
1208
+ "prose": "Move reusable procedures into focused skills whose bodies load only when invoked or matched.",
1209
+ "apply": "enable_sdk_config"
1210
+ },
1211
+ "grader": {
1212
+ "type": "exact_match",
1213
+ "fixture_template": "Held task that invokes the skill and an unrelated task that must not load its body."
1214
+ },
1215
+ "experiment": {
1216
+ "method": "local_ab",
1217
+ "metric": "tokens_in"
1218
+ },
1219
+ "evidence": {
1220
+ "status": "unmeasured"
1221
+ },
1222
+ "never_claim": "Never treat skill discovery metadata as free; keep names and descriptions concise.",
1223
+ "skill_render": {
1224
+ "eligible": true,
1225
+ "skill_name": "caveman-practice-claude-code-skill-deferral"
1226
+ }
1227
+ },
1228
+ {
1229
+ "id": "claude-code-explorer-subagent",
1230
+ "family": "input_bloat",
1231
+ "title": "Use isolated Claude Code exploration",
1232
+ "predicate": {
1233
+ "harness": [
1234
+ "claude-code"
1235
+ ],
1236
+ "payload_shape": "broad repository reads polluting solver history",
1237
+ "wire_protocol": [
1238
+ "anthropic"
1239
+ ]
1240
+ },
1241
+ "fix": {
1242
+ "prose": "Use a read-only isolated explorer for broad localization and return only compact evidence to the main solver.",
1243
+ "apply": "offload_exploration"
1244
+ },
1245
+ "grader": {
1246
+ "type": "localization_f1",
1247
+ "fixture_template": "Held repository localization tasks with expected files and line neighborhoods."
1248
+ },
1249
+ "experiment": {
1250
+ "method": "local_ab",
1251
+ "metric": "tokens_in"
1252
+ },
1253
+ "evidence": {
1254
+ "status": "unmeasured"
1255
+ },
1256
+ "never_claim": "Never delegate a one-file lookup or assume isolated work lowers combined tokens.",
1257
+ "skill_render": {
1258
+ "eligible": true,
1259
+ "skill_name": "caveman-practice-claude-code-explorer-subagent"
1260
+ }
1261
+ },
1262
+ {
1263
+ "id": "claude-code-command-output-hook",
1264
+ "family": "input_bloat",
1265
+ "title": "Rewrite noisy Claude Code commands through shrink",
1266
+ "predicate": {
1267
+ "harness": [
1268
+ "claude-code"
1269
+ ],
1270
+ "payload_shape": "shell command likely to return large repetitive output",
1271
+ "wire_protocol": [
1272
+ "anthropic"
1273
+ ]
1274
+ },
1275
+ "fix": {
1276
+ "prose": "Use the verified PreToolUse rewrite to route eligible shell commands through recoverable output shrinking.",
1277
+ "apply": "enable_sdk_config"
1278
+ },
1279
+ "grader": {
1280
+ "type": "exact_match",
1281
+ "fixture_template": "Held command outputs whose compact view and byte-exact recovery both match expected content."
1282
+ },
1283
+ "experiment": {
1284
+ "method": "local_ab",
1285
+ "metric": "tokens_in"
1286
+ },
1287
+ "evidence": {
1288
+ "status": "unmeasured"
1289
+ },
1290
+ "never_claim": "Never rewrite commands outside the verified hook contract or lose access to original output.",
1291
+ "skill_render": {
1292
+ "eligible": true,
1293
+ "skill_name": "caveman-practice-claude-code-command-output-hook"
1294
+ }
1295
+ },
1296
+ {
1297
+ "id": "claude-code-subagent-model-right-sizing",
1298
+ "family": "routing",
1299
+ "title": "Right-size Claude Code subagent models",
1300
+ "predicate": {
1301
+ "harness": [
1302
+ "claude-code"
1303
+ ],
1304
+ "payload_shape": "read-only or narrow subagent using the main solver model",
1305
+ "wire_protocol": [
1306
+ "anthropic"
1307
+ ]
1308
+ },
1309
+ "fix": {
1310
+ "prose": "Assign a cheaper compatible model to narrow subagents and gate the combined workflow on final task quality.",
1311
+ "apply": "right_size_model"
1312
+ },
1313
+ "grader": {
1314
+ "type": "localization_f1",
1315
+ "fixture_template": "Held subagent tasks with expected evidence and final solver outcomes."
1316
+ },
1317
+ "experiment": {
1318
+ "method": "local_ab",
1319
+ "metric": "$/day band"
1320
+ },
1321
+ "evidence": {
1322
+ "status": "unmeasured"
1323
+ },
1324
+ "never_claim": "Never count main-agent savings without including subagent tokens and calls.",
1325
+ "skill_render": {
1326
+ "eligible": true,
1327
+ "skill_name": "caveman-practice-claude-code-subagent-model-right-sizing"
1328
+ }
1329
+ },
1330
+ {
1331
+ "id": "codex-agents-md-budget",
1332
+ "family": "input_bloat",
1333
+ "title": "Keep AGENTS.md concise",
1334
+ "predicate": {
1335
+ "harness": [
1336
+ "codex"
1337
+ ],
1338
+ "payload_shape": "large AGENTS.md guidance loaded into Codex turns",
1339
+ "wire_protocol": [
1340
+ "openai"
1341
+ ]
1342
+ },
1343
+ "fix": {
1344
+ "prose": "Keep durable rules in AGENTS.md and move repeatable procedures or references into on-demand skills.",
1345
+ "apply": "enable_sdk_config"
1346
+ },
1347
+ "grader": {
1348
+ "type": "exact_match",
1349
+ "fixture_template": "Held repository tasks requiring each retained rule after the context edit."
1350
+ },
1351
+ "experiment": {
1352
+ "method": "local_ab",
1353
+ "metric": "tokens_in"
1354
+ },
1355
+ "evidence": {
1356
+ "status": "unmeasured"
1357
+ },
1358
+ "never_claim": "Never trim instructions whose absence changes required scope, safety, or verification behavior.",
1359
+ "skill_render": {
1360
+ "eligible": true,
1361
+ "skill_name": "caveman-practice-codex-agents-md-budget"
1362
+ }
1363
+ },
1364
+ {
1365
+ "id": "codex-scoped-agents",
1366
+ "family": "input_bloat",
1367
+ "title": "Scope Codex instructions by directory",
1368
+ "predicate": {
1369
+ "harness": [
1370
+ "codex"
1371
+ ],
1372
+ "payload_shape": "root instructions relevant only to a nested path",
1373
+ "wire_protocol": [
1374
+ "openai"
1375
+ ]
1376
+ },
1377
+ "fix": {
1378
+ "prose": "Move path-specific rules into the nearest applicable AGENTS.md while preserving parent precedence.",
1379
+ "apply": "enable_sdk_config"
1380
+ },
1381
+ "grader": {
1382
+ "type": "exact_match",
1383
+ "fixture_template": "Held tasks in sibling paths verifying only applicable instructions are active."
1384
+ },
1385
+ "experiment": {
1386
+ "method": "local_ab",
1387
+ "metric": "tokens_in"
1388
+ },
1389
+ "evidence": {
1390
+ "status": "unmeasured"
1391
+ },
1392
+ "never_claim": "Never move a repository-wide rule into a subtree.",
1393
+ "skill_render": {
1394
+ "eligible": true,
1395
+ "skill_name": "caveman-practice-codex-scoped-agents"
1396
+ }
1397
+ },
1398
+ {
1399
+ "id": "codex-skill-deferral",
1400
+ "family": "input_bloat",
1401
+ "title": "Load Codex procedures as skills",
1402
+ "predicate": {
1403
+ "harness": [
1404
+ "codex"
1405
+ ],
1406
+ "payload_shape": "long repeatable workflow embedded in always-loaded instructions",
1407
+ "wire_protocol": [
1408
+ "openai"
1409
+ ]
1410
+ },
1411
+ "fix": {
1412
+ "prose": "Move the workflow into a focused SKILL.md and keep only concise triggering guidance in persistent instructions.",
1413
+ "apply": "enable_sdk_config"
1414
+ },
1415
+ "grader": {
1416
+ "type": "exact_match",
1417
+ "fixture_template": "Held matching and unrelated tasks verifying skill activation and non-activation."
1418
+ },
1419
+ "experiment": {
1420
+ "method": "local_ab",
1421
+ "metric": "tokens_in"
1422
+ },
1423
+ "evidence": {
1424
+ "status": "unmeasured"
1425
+ },
1426
+ "never_claim": "Never create a skill whose discovery description costs more than the procedure it replaces.",
1427
+ "skill_render": {
1428
+ "eligible": true,
1429
+ "skill_name": "caveman-practice-codex-skill-deferral"
1430
+ }
1431
+ },
1432
+ {
1433
+ "id": "codex-command-output-nudge",
1434
+ "family": "input_bloat",
1435
+ "title": "Prefer shrink for noisy Codex commands",
1436
+ "predicate": {
1437
+ "harness": [
1438
+ "codex"
1439
+ ],
1440
+ "payload_shape": "shell commands producing large repetitive output",
1441
+ "wire_protocol": [
1442
+ "openai"
1443
+ ]
1444
+ },
1445
+ "fix": {
1446
+ "prose": "Install the reversible AGENTS.md instruction note because Codex lacks a verified byte-safe hard rewrite surface for this path.",
1447
+ "apply": "enable_sdk_config"
1448
+ },
1449
+ "grader": {
1450
+ "type": "token_threshold",
1451
+ "fixture_template": "Held noisy command task with an input-token ceiling and byte-exact recovery check."
1452
+ },
1453
+ "experiment": {
1454
+ "method": "local_ab",
1455
+ "metric": "tokens_in"
1456
+ },
1457
+ "evidence": {
1458
+ "status": "unmeasured"
1459
+ },
1460
+ "never_claim": "Never describe this instruction note as a deterministic command rewrite.",
1461
+ "skill_render": {
1462
+ "eligible": true,
1463
+ "skill_name": "caveman-practice-codex-command-output-nudge"
1464
+ }
1465
+ },
1466
+ {
1467
+ "id": "codex-compaction-threshold",
1468
+ "family": "input_bloat",
1469
+ "title": "Tune Codex auto-compaction threshold",
1470
+ "predicate": {
1471
+ "harness": [
1472
+ "codex"
1473
+ ],
1474
+ "payload_shape": "long sessions compacting too late or repeatedly",
1475
+ "wire_protocol": [
1476
+ "openai"
1477
+ ]
1478
+ },
1479
+ "fix": {
1480
+ "prose": "Set a threshold from the model window and observed session shape, then verify required state survives compaction.",
1481
+ "apply": "enable_sdk_config"
1482
+ },
1483
+ "grader": {
1484
+ "type": "llm_judge",
1485
+ "fixture_template": "Held long session requiring earlier decisions and identifiers after compaction."
1486
+ },
1487
+ "experiment": {
1488
+ "method": "local_ab",
1489
+ "metric": "tokens_in"
1490
+ },
1491
+ "evidence": {
1492
+ "status": "unmeasured"
1493
+ },
1494
+ "never_claim": "Never tune compaction from token reduction alone; state retention must pass.",
1495
+ "skill_render": {
1496
+ "eligible": true,
1497
+ "skill_name": "caveman-practice-codex-compaction-threshold"
1498
+ }
1499
+ },
1500
+ {
1501
+ "id": "gemini-cli-gemini-md-budget",
1502
+ "family": "input_bloat",
1503
+ "title": "Keep GEMINI.md concise",
1504
+ "predicate": {
1505
+ "harness": [
1506
+ "gemini"
1507
+ ],
1508
+ "payload_shape": "large GEMINI.md hierarchy loaded into every turn",
1509
+ "wire_protocol": [
1510
+ "gemini"
1511
+ ]
1512
+ },
1513
+ "fix": {
1514
+ "prose": "Keep durable project rules in context files and move long procedures into on-demand skills.",
1515
+ "apply": "enable_sdk_config"
1516
+ },
1517
+ "grader": {
1518
+ "type": "exact_match",
1519
+ "fixture_template": "Held repository tasks requiring each retained rule after the context edit."
1520
+ },
1521
+ "experiment": {
1522
+ "method": "local_ab",
1523
+ "metric": "tokens_in"
1524
+ },
1525
+ "evidence": {
1526
+ "status": "unmeasured"
1527
+ },
1528
+ "never_claim": "Never remove project guidance without proving required behavior remains.",
1529
+ "skill_render": {
1530
+ "eligible": true,
1531
+ "skill_name": "caveman-practice-gemini-cli-gemini-md-budget"
1532
+ }
1533
+ },
1534
+ {
1535
+ "id": "gemini-cli-context-imports",
1536
+ "family": "input_bloat",
1537
+ "title": "Modularize Gemini CLI context",
1538
+ "predicate": {
1539
+ "harness": [
1540
+ "gemini"
1541
+ ],
1542
+ "payload_shape": "monolithic context file mixing unrelated project guidance",
1543
+ "wire_protocol": [
1544
+ "gemini"
1545
+ ]
1546
+ },
1547
+ "fix": {
1548
+ "prose": "Split context into stable modules and import only modules that remain globally load-bearing.",
1549
+ "apply": "enable_sdk_config"
1550
+ },
1551
+ "grader": {
1552
+ "type": "exact_match",
1553
+ "fixture_template": "Held tasks requiring each imported module and unrelated tasks checking omitted modules."
1554
+ },
1555
+ "experiment": {
1556
+ "method": "local_ab",
1557
+ "metric": "tokens_in"
1558
+ },
1559
+ "evidence": {
1560
+ "status": "unmeasured"
1561
+ },
1562
+ "never_claim": "Imports improve maintainability, not tokens, unless the loaded set becomes smaller.",
1563
+ "skill_render": {
1564
+ "eligible": true,
1565
+ "skill_name": "caveman-practice-gemini-cli-context-imports"
1566
+ }
1567
+ },
1568
+ {
1569
+ "id": "gemini-cli-skill-deferral",
1570
+ "family": "input_bloat",
1571
+ "title": "Load Gemini CLI procedures as skills",
1572
+ "predicate": {
1573
+ "harness": [
1574
+ "gemini"
1575
+ ],
1576
+ "payload_shape": "rare workflow stored in always-loaded context",
1577
+ "wire_protocol": [
1578
+ "gemini"
1579
+ ]
1580
+ },
1581
+ "fix": {
1582
+ "prose": "Move reusable procedures into Agent Skills and keep persistent context focused on rules needed every turn.",
1583
+ "apply": "enable_sdk_config"
1584
+ },
1585
+ "grader": {
1586
+ "type": "exact_match",
1587
+ "fixture_template": "Held matching and unrelated tasks verifying skill activation and non-activation."
1588
+ },
1589
+ "experiment": {
1590
+ "method": "local_ab",
1591
+ "metric": "tokens_in"
1592
+ },
1593
+ "evidence": {
1594
+ "status": "unmeasured"
1595
+ },
1596
+ "never_claim": "Never count deferred skill bodies as removed if their discovery metadata still dominates context.",
1597
+ "skill_render": {
1598
+ "eligible": true,
1599
+ "skill_name": "caveman-practice-gemini-cli-skill-deferral"
1600
+ }
1601
+ },
1602
+ {
1603
+ "id": "gemini-cli-explorer-subagent",
1604
+ "family": "input_bloat",
1605
+ "title": "Use Gemini CLI isolated exploration",
1606
+ "predicate": {
1607
+ "harness": [
1608
+ "gemini"
1609
+ ],
1610
+ "payload_shape": "broad investigation generating large read-only context",
1611
+ "wire_protocol": [
1612
+ "gemini"
1613
+ ]
1614
+ },
1615
+ "fix": {
1616
+ "prose": "Delegate broad codebase investigation to an isolated read-only subagent and return compact evidence.",
1617
+ "apply": "offload_exploration"
1618
+ },
1619
+ "grader": {
1620
+ "type": "localization_f1",
1621
+ "fixture_template": "Held repository localization tasks with expected files and line neighborhoods."
1622
+ },
1623
+ "experiment": {
1624
+ "method": "local_ab",
1625
+ "metric": "tokens_in"
1626
+ },
1627
+ "evidence": {
1628
+ "status": "unmeasured"
1629
+ },
1630
+ "never_claim": "Never delegate small lookups or omit combined subagent usage from measurement.",
1631
+ "skill_render": {
1632
+ "eligible": true,
1633
+ "skill_name": "caveman-practice-gemini-cli-explorer-subagent"
1634
+ }
1635
+ },
1636
+ {
1637
+ "id": "gemini-cli-command-output-hook",
1638
+ "family": "input_bloat",
1639
+ "title": "Rewrite noisy Gemini CLI commands through shrink",
1640
+ "predicate": {
1641
+ "harness": [
1642
+ "gemini"
1643
+ ],
1644
+ "payload_shape": "run_shell_command likely to return large repetitive output",
1645
+ "wire_protocol": [
1646
+ "gemini"
1647
+ ]
1648
+ },
1649
+ "fix": {
1650
+ "prose": "Use the verified BeforeTool rewrite to route eligible shell commands through recoverable output shrinking.",
1651
+ "apply": "enable_sdk_config"
1652
+ },
1653
+ "grader": {
1654
+ "type": "exact_match",
1655
+ "fixture_template": "Held command outputs whose compact view and byte-exact recovery match expected content."
1656
+ },
1657
+ "experiment": {
1658
+ "method": "local_ab",
1659
+ "metric": "tokens_in"
1660
+ },
1661
+ "evidence": {
1662
+ "status": "unmeasured"
1663
+ },
1664
+ "never_claim": "Never emit non-JSON hook stdout or rewrite unsupported tool inputs.",
1665
+ "skill_render": {
1666
+ "eligible": true,
1667
+ "skill_name": "caveman-practice-gemini-cli-command-output-hook"
1668
+ }
1669
+ },
1670
+ {
1671
+ "id": "gemini-cli-api-key-token-cache",
1672
+ "family": "cache",
1673
+ "title": "Preserve Gemini CLI token caching eligibility",
1674
+ "predicate": {
1675
+ "harness": [
1676
+ "gemini"
1677
+ ],
1678
+ "payload_shape": "Gemini CLI session using an authentication mode without cached-content support",
1679
+ "wire_protocol": [
1680
+ "gemini"
1681
+ ]
1682
+ },
1683
+ "fix": {
1684
+ "prose": "Use a supported API-key or Vertex authentication path when token caching matters and policy permits it.",
1685
+ "apply": "manual"
1686
+ },
1687
+ "grader": {
1688
+ "type": "json_path_assertion",
1689
+ "fixture_template": "Held CLI session requiring cached-token usage to appear in stats under the selected authentication mode."
1690
+ },
1691
+ "experiment": {
1692
+ "method": "local_ab",
1693
+ "metric": "cache_hit_ratio"
1694
+ },
1695
+ "evidence": {
1696
+ "status": "unmeasured"
1697
+ },
1698
+ "never_claim": "Never change authentication mode solely for cache savings without security and account-policy approval.",
1699
+ "skill_render": {
1700
+ "eligible": true,
1701
+ "skill_name": "caveman-practice-gemini-cli-api-key-token-cache"
1702
+ }
1703
+ },
1704
+ {
1705
+ "id": "opencode-instruction-budget",
1706
+ "family": "input_bloat",
1707
+ "title": "Keep OpenCode instructions concise",
1708
+ "predicate": {
1709
+ "harness": [
1710
+ "opencode"
1711
+ ],
1712
+ "payload_shape": "large always-loaded project instructions",
1713
+ "wire_protocol": [
1714
+ "openai"
1715
+ ]
1716
+ },
1717
+ "fix": {
1718
+ "prose": "Keep persistent rules short and move task procedures into focused commands, skills, or agents.",
1719
+ "apply": "enable_sdk_config"
1720
+ },
1721
+ "grader": {
1722
+ "type": "exact_match",
1723
+ "fixture_template": "Held project tasks requiring retained rules after the instruction edit."
1724
+ },
1725
+ "experiment": {
1726
+ "method": "local_ab",
1727
+ "metric": "tokens_in"
1728
+ },
1729
+ "evidence": {
1730
+ "status": "unmeasured"
1731
+ },
1732
+ "never_claim": "Never remove instructions that define required repository behavior.",
1733
+ "skill_render": {
1734
+ "eligible": true,
1735
+ "skill_name": "caveman-practice-opencode-instruction-budget"
1736
+ }
1737
+ },
1738
+ {
1739
+ "id": "opencode-tool-permission-pruning",
1740
+ "family": "input_bloat",
1741
+ "title": "Restrict OpenCode agent tools",
1742
+ "predicate": {
1743
+ "harness": [
1744
+ "opencode"
1745
+ ],
1746
+ "payload_shape": "specialized agent receiving unrelated tool schemas",
1747
+ "wire_protocol": [
1748
+ "openai"
1749
+ ]
1750
+ },
1751
+ "fix": {
1752
+ "prose": "Use agent-specific permissions to expose only tools required for the role.",
1753
+ "apply": "enable_sdk_config"
1754
+ },
1755
+ "grader": {
1756
+ "type": "tool_not_called",
1757
+ "fixture_template": "Held specialized tasks verifying denied tools remain unavailable and required tools still work."
1758
+ },
1759
+ "experiment": {
1760
+ "method": "local_ab",
1761
+ "metric": "tokens_in"
1762
+ },
1763
+ "evidence": {
1764
+ "status": "unmeasured"
1765
+ },
1766
+ "never_claim": "Never prune a tool from the main agent without a reliable route to regain it when needed.",
1767
+ "skill_render": {
1768
+ "eligible": true,
1769
+ "skill_name": "caveman-practice-opencode-tool-permission-pruning"
1770
+ }
1771
+ },
1772
+ {
1773
+ "id": "opencode-cheap-subagent-model",
1774
+ "family": "routing",
1775
+ "title": "Right-size OpenCode subagent models",
1776
+ "predicate": {
1777
+ "harness": [
1778
+ "opencode"
1779
+ ],
1780
+ "payload_shape": "narrow plan or read-only agent using the main model",
1781
+ "wire_protocol": [
1782
+ "openai"
1783
+ ]
1784
+ },
1785
+ "fix": {
1786
+ "prose": "Assign a cheaper compatible model to the narrow agent and gate the combined workflow on final task quality.",
1787
+ "apply": "right_size_model"
1788
+ },
1789
+ "grader": {
1790
+ "type": "llm_judge",
1791
+ "fixture_template": "Held delegated tasks comparing final outcomes with baseline and candidate subagent models."
1792
+ },
1793
+ "experiment": {
1794
+ "method": "local_ab",
1795
+ "metric": "$/day band"
1796
+ },
1797
+ "evidence": {
1798
+ "status": "unmeasured"
1799
+ },
1800
+ "never_claim": "Never omit parent-agent and delegated-call usage from the comparison.",
1801
+ "skill_render": {
1802
+ "eligible": true,
1803
+ "skill_name": "caveman-practice-opencode-cheap-subagent-model"
1804
+ }
1805
+ },
1806
+ {
1807
+ "id": "opencode-command-output-plugin",
1808
+ "family": "input_bloat",
1809
+ "title": "Rewrite noisy OpenCode commands through shrink",
1810
+ "predicate": {
1811
+ "harness": [
1812
+ "opencode"
1813
+ ],
1814
+ "payload_shape": "bash tool command likely to return large repetitive output",
1815
+ "wire_protocol": [
1816
+ "openai"
1817
+ ]
1818
+ },
1819
+ "fix": {
1820
+ "prose": "Use the verified plugin hook to route eligible commands through recoverable output shrinking.",
1821
+ "apply": "enable_sdk_config"
1822
+ },
1823
+ "grader": {
1824
+ "type": "exact_match",
1825
+ "fixture_template": "Held command outputs whose compact view and byte-exact recovery match expected content."
1826
+ },
1827
+ "experiment": {
1828
+ "method": "local_ab",
1829
+ "metric": "tokens_in"
1830
+ },
1831
+ "evidence": {
1832
+ "status": "unmeasured"
1833
+ },
1834
+ "never_claim": "Never claim hard rewrite support outside the tested plugin event.",
1835
+ "skill_render": {
1836
+ "eligible": true,
1837
+ "skill_name": "caveman-practice-opencode-command-output-plugin"
1838
+ }
1839
+ },
1840
+ {
1841
+ "id": "opencode-mcp-tool-pruning",
1842
+ "family": "input_bloat",
1843
+ "title": "Hide unused OpenCode MCP tools",
1844
+ "predicate": {
1845
+ "harness": [
1846
+ "opencode"
1847
+ ],
1848
+ "payload_shape": "large MCP tool surface irrelevant to one agent",
1849
+ "wire_protocol": [
1850
+ "openai"
1851
+ ]
1852
+ },
1853
+ "fix": {
1854
+ "prose": "Deny unrelated MCP tool groups on specialized agents and keep an explicit route for tasks that need them.",
1855
+ "apply": "enable_sdk_config"
1856
+ },
1857
+ "grader": {
1858
+ "type": "tool_not_called",
1859
+ "fixture_template": "Held specialized tasks verifying hidden MCP tools stay unavailable and required tools remain callable."
1860
+ },
1861
+ "experiment": {
1862
+ "method": "local_ab",
1863
+ "metric": "tokens_in"
1864
+ },
1865
+ "evidence": {
1866
+ "status": "unmeasured"
1867
+ },
1868
+ "never_claim": "Never hide tools from workflows that lack another discovery or delegation path.",
1869
+ "skill_render": {
1870
+ "eligible": true,
1871
+ "skill_name": "caveman-practice-opencode-mcp-tool-pruning"
1872
+ }
1873
+ },
1874
+ {
1875
+ "id": "hermes-context-file-budget",
1876
+ "family": "input_bloat",
1877
+ "title": "Bound Hermes project context files",
1878
+ "predicate": {
1879
+ "harness": [
1880
+ "hermes"
1881
+ ],
1882
+ "payload_shape": "large project context files loaded into every session",
1883
+ "wire_protocol": [
1884
+ "openai"
1885
+ ]
1886
+ },
1887
+ "fix": {
1888
+ "prose": "Keep always-loaded project facts concise and move procedures into on-demand skills.",
1889
+ "apply": "enable_sdk_config"
1890
+ },
1891
+ "grader": {
1892
+ "type": "exact_match",
1893
+ "fixture_template": "Held project tasks requiring retained context after the edit."
1894
+ },
1895
+ "experiment": {
1896
+ "method": "local_ab",
1897
+ "metric": "tokens_in"
1898
+ },
1899
+ "evidence": {
1900
+ "status": "unmeasured"
1901
+ },
1902
+ "never_claim": "Never lower context limits to hide truncation of load-bearing instructions.",
1903
+ "skill_render": {
1904
+ "eligible": true,
1905
+ "skill_name": "caveman-practice-hermes-context-file-budget"
1906
+ }
1907
+ },
1908
+ {
1909
+ "id": "hermes-context-compression-tuning",
1910
+ "family": "input_bloat",
1911
+ "title": "Tune Hermes context compression",
1912
+ "predicate": {
1913
+ "harness": [
1914
+ "hermes"
1915
+ ],
1916
+ "payload_shape": "sessions compressing too early, too late, or too often",
1917
+ "wire_protocol": [
1918
+ "openai"
1919
+ ]
1920
+ },
1921
+ "fix": {
1922
+ "prose": "Tune compression threshold, retained tail, and protected messages from observed sessions, then gate state retention.",
1923
+ "apply": "enable_sdk_config"
1924
+ },
1925
+ "grader": {
1926
+ "type": "llm_judge",
1927
+ "fixture_template": "Held long sessions requiring prior decisions, identifiers, and current task state after compression."
1928
+ },
1929
+ "experiment": {
1930
+ "method": "local_ab",
1931
+ "metric": "tokens_in"
1932
+ },
1933
+ "evidence": {
1934
+ "status": "unmeasured"
1935
+ },
1936
+ "never_claim": "Never optimize only for fewer tokens when compression drops task state.",
1937
+ "skill_render": {
1938
+ "eligible": true,
1939
+ "skill_name": "caveman-practice-hermes-context-compression-tuning"
1940
+ }
1941
+ },
1942
+ {
1943
+ "id": "hermes-compression-model-right-sizing",
1944
+ "family": "routing",
1945
+ "title": "Right-size Hermes compression model",
1946
+ "predicate": {
1947
+ "harness": [
1948
+ "hermes"
1949
+ ],
1950
+ "payload_shape": "context summarization using the primary solver model",
1951
+ "wire_protocol": [
1952
+ "openai"
1953
+ ]
1954
+ },
1955
+ "fix": {
1956
+ "prose": "Evaluate a cheaper auxiliary model for compression and keep it only when state-retention checks pass.",
1957
+ "apply": "right_size_model"
1958
+ },
1959
+ "grader": {
1960
+ "type": "llm_judge",
1961
+ "fixture_template": "Held long sessions comparing retained facts and task completion across compression models."
1962
+ },
1963
+ "experiment": {
1964
+ "method": "local_ab",
1965
+ "metric": "$/day band"
1966
+ },
1967
+ "evidence": {
1968
+ "status": "unmeasured"
1969
+ },
1970
+ "never_claim": "Never count cheaper summary calls when they cause more solver turns or lost state.",
1971
+ "skill_render": {
1972
+ "eligible": true,
1973
+ "skill_name": "caveman-practice-hermes-compression-model-right-sizing"
1974
+ }
1975
+ },
1976
+ {
1977
+ "id": "hermes-file-read-dedup",
1978
+ "family": "input_bloat",
1979
+ "title": "Preserve Hermes file-read deduplication",
1980
+ "predicate": {
1981
+ "harness": [
1982
+ "hermes"
1983
+ ],
1984
+ "payload_shape": "same unchanged file region read repeatedly",
1985
+ "wire_protocol": [
1986
+ "openai"
1987
+ ]
1988
+ },
1989
+ "fix": {
1990
+ "prose": "Keep built-in read deduplication enabled and avoid workflows that re-inline unchanged regions after lightweight stubs are available.",
1991
+ "apply": "enable_sdk_config"
1992
+ },
1993
+ "grader": {
1994
+ "type": "token_threshold",
1995
+ "fixture_template": "Held repeated read of one unchanged region with an input-token ceiling on later turns."
1996
+ },
1997
+ "experiment": {
1998
+ "method": "local_ab",
1999
+ "metric": "tokens_in"
2000
+ },
2001
+ "evidence": {
2002
+ "status": "unmeasured"
2003
+ },
2004
+ "never_claim": "Never deduplicate a region after its file content changes.",
2005
+ "skill_render": {
2006
+ "eligible": true,
2007
+ "skill_name": "caveman-practice-hermes-file-read-dedup"
2008
+ }
2009
+ },
2010
+ {
2011
+ "id": "hermes-skill-deferral",
2012
+ "family": "input_bloat",
2013
+ "title": "Keep Hermes procedures in on-demand skills",
2014
+ "predicate": {
2015
+ "harness": [
2016
+ "hermes"
2017
+ ],
2018
+ "payload_shape": "long procedural guidance loaded outside its task",
2019
+ "wire_protocol": [
2020
+ "openai"
2021
+ ]
2022
+ },
2023
+ "fix": {
2024
+ "prose": "Use progressive-disclosure skills for procedures and keep persistent memory for short durable facts.",
2025
+ "apply": "enable_sdk_config"
2026
+ },
2027
+ "grader": {
2028
+ "type": "exact_match",
2029
+ "fixture_template": "Held matching and unrelated tasks verifying skill activation and non-activation."
2030
+ },
2031
+ "experiment": {
2032
+ "method": "local_ab",
2033
+ "metric": "tokens_in"
2034
+ },
2035
+ "evidence": {
2036
+ "status": "unmeasured"
2037
+ },
2038
+ "never_claim": "Never let skill discovery metadata grow into another always-loaded manual.",
2039
+ "skill_render": {
2040
+ "eligible": true,
2041
+ "skill_name": "caveman-practice-hermes-skill-deferral"
2042
+ }
2043
+ },
2044
+ {
2045
+ "id": "hermes-toolset-pruning",
2046
+ "family": "input_bloat",
2047
+ "title": "Select only needed Hermes toolsets",
2048
+ "predicate": {
2049
+ "harness": [
2050
+ "hermes"
2051
+ ],
2052
+ "payload_shape": "session or delegate exposed to unrelated toolsets",
2053
+ "wire_protocol": [
2054
+ "openai"
2055
+ ]
2056
+ },
2057
+ "fix": {
2058
+ "prose": "Enable only toolsets required by the platform and task, with a separate route for broader work.",
2059
+ "apply": "enable_sdk_config"
2060
+ },
2061
+ "grader": {
2062
+ "type": "tool_not_called",
2063
+ "fixture_template": "Held narrow tasks proving hidden tools stay unused and required tools remain available."
2064
+ },
2065
+ "experiment": {
2066
+ "method": "local_ab",
2067
+ "metric": "tokens_in"
2068
+ },
2069
+ "evidence": {
2070
+ "status": "unmeasured"
2071
+ },
2072
+ "never_claim": "Never prune a toolset without checking all task paths that depend on it.",
2073
+ "skill_render": {
2074
+ "eligible": true,
2075
+ "skill_name": "caveman-practice-hermes-toolset-pruning"
2076
+ }
2077
+ },
2078
+ {
2079
+ "id": "hermes-command-output-plugin",
2080
+ "family": "input_bloat",
2081
+ "title": "Rewrite noisy Hermes commands through shrink",
2082
+ "predicate": {
2083
+ "harness": [
2084
+ "hermes"
2085
+ ],
2086
+ "payload_shape": "terminal tool output likely to be large and repetitive",
2087
+ "wire_protocol": [
2088
+ "openai"
2089
+ ]
2090
+ },
2091
+ "fix": {
2092
+ "prose": "Use the verified Hermes plugin path to shrink eligible output while retaining byte-exact recovery.",
2093
+ "apply": "enable_sdk_config"
2094
+ },
2095
+ "grader": {
2096
+ "type": "exact_match",
2097
+ "fixture_template": "Held command outputs whose compact view and recovery match expected content."
2098
+ },
2099
+ "experiment": {
2100
+ "method": "local_ab",
2101
+ "metric": "tokens_in"
2102
+ },
2103
+ "evidence": {
2104
+ "status": "unmeasured"
2105
+ },
2106
+ "never_claim": "Never discard original output or claim rewrite support outside the tested plugin path.",
2107
+ "skill_render": {
2108
+ "eligible": true,
2109
+ "skill_name": "caveman-practice-hermes-command-output-plugin"
2110
+ }
2111
+ },
2112
+ {
2113
+ "id": "openclaw-bootstrap-budget",
2114
+ "family": "input_bloat",
2115
+ "title": "Bound OpenClaw bootstrap context",
2116
+ "predicate": {
2117
+ "harness": [
2118
+ "openclaw"
2119
+ ],
2120
+ "payload_shape": "large workspace bootstrap files injected every run",
2121
+ "wire_protocol": [
2122
+ "openai"
2123
+ ]
2124
+ },
2125
+ "fix": {
2126
+ "prose": "Trim bootstrap files to load-bearing facts and move procedures into on-demand skills or memory.",
2127
+ "apply": "enable_sdk_config"
2128
+ },
2129
+ "grader": {
2130
+ "type": "exact_match",
2131
+ "fixture_template": "Held workspace tasks requiring retained bootstrap rules after the edit."
2132
+ },
2133
+ "experiment": {
2134
+ "method": "local_ab",
2135
+ "metric": "tokens_in"
2136
+ },
2137
+ "evidence": {
2138
+ "status": "unmeasured"
2139
+ },
2140
+ "never_claim": "Never lower bootstrap caps to silently truncate required instructions.",
2141
+ "skill_render": {
2142
+ "eligible": true,
2143
+ "skill_name": "caveman-practice-openclaw-bootstrap-budget"
2144
+ }
2145
+ },
2146
+ {
2147
+ "id": "openclaw-tool-schema-pruning",
2148
+ "family": "input_bloat",
2149
+ "title": "Reduce OpenClaw tool schema surface",
2150
+ "predicate": {
2151
+ "harness": [
2152
+ "openclaw"
2153
+ ],
2154
+ "payload_shape": "tool schema JSON dominates the context breakdown",
2155
+ "wire_protocol": [
2156
+ "openai"
2157
+ ]
2158
+ },
2159
+ "fix": {
2160
+ "prose": "Expose only tools needed by each agent and use separate agents or skills for broader capabilities.",
2161
+ "apply": "enable_sdk_config"
2162
+ },
2163
+ "grader": {
2164
+ "type": "tool_not_called",
2165
+ "fixture_template": "Held narrow tasks proving hidden tools stay unavailable and required tools remain callable."
2166
+ },
2167
+ "experiment": {
2168
+ "method": "local_ab",
2169
+ "metric": "tokens_in"
2170
+ },
2171
+ "evidence": {
2172
+ "status": "unmeasured"
2173
+ },
2174
+ "never_claim": "Never remove a tool without a tested route for tasks that need it.",
2175
+ "skill_render": {
2176
+ "eligible": true,
2177
+ "skill_name": "caveman-practice-openclaw-tool-schema-pruning"
2178
+ }
2179
+ },
2180
+ {
2181
+ "id": "openclaw-session-pruning",
2182
+ "family": "input_bloat",
2183
+ "title": "Prune stale OpenClaw tool results",
2184
+ "predicate": {
2185
+ "harness": [
2186
+ "openclaw"
2187
+ ],
2188
+ "payload_shape": "long session retaining old large tool results",
2189
+ "wire_protocol": [
2190
+ "openai"
2191
+ ]
2192
+ },
2193
+ "fix": {
2194
+ "prose": "Enable session pruning for old tool results while preserving the persisted transcript and current task evidence.",
2195
+ "apply": "enable_sdk_config"
2196
+ },
2197
+ "grader": {
2198
+ "type": "llm_judge",
2199
+ "fixture_template": "Held long session requiring recent and previously extracted facts after pruning."
2200
+ },
2201
+ "experiment": {
2202
+ "method": "local_ab",
2203
+ "metric": "tokens_in"
2204
+ },
2205
+ "evidence": {
2206
+ "status": "unmeasured"
2207
+ },
2208
+ "never_claim": "Never prune a tool result still needed for current task correctness.",
2209
+ "skill_render": {
2210
+ "eligible": true,
2211
+ "skill_name": "caveman-practice-openclaw-session-pruning"
2212
+ }
2213
+ },
2214
+ {
2215
+ "id": "openclaw-compaction-tail-budget",
2216
+ "family": "input_bloat",
2217
+ "title": "Tune OpenClaw compaction tail",
2218
+ "predicate": {
2219
+ "harness": [
2220
+ "openclaw"
2221
+ ],
2222
+ "payload_shape": "long sessions retaining too much or too little recent history",
2223
+ "wire_protocol": [
2224
+ "openai"
2225
+ ]
2226
+ },
2227
+ "fix": {
2228
+ "prose": "Tune recent-token and reserve budgets from observed sessions, preserving tool-call pairs and current task state.",
2229
+ "apply": "enable_sdk_config"
2230
+ },
2231
+ "grader": {
2232
+ "type": "llm_judge",
2233
+ "fixture_template": "Held long session requiring recent tool linkage, identifiers, and earlier decisions after compaction."
2234
+ },
2235
+ "experiment": {
2236
+ "method": "local_ab",
2237
+ "metric": "tokens_in"
2238
+ },
2239
+ "evidence": {
2240
+ "status": "unmeasured"
2241
+ },
2242
+ "never_claim": "Never reduce the retained tail without proving current task continuity.",
2243
+ "skill_render": {
2244
+ "eligible": true,
2245
+ "skill_name": "caveman-practice-openclaw-compaction-tail-budget"
2246
+ }
2247
+ },
2248
+ {
2249
+ "id": "openclaw-compaction-model-right-sizing",
2250
+ "family": "routing",
2251
+ "title": "Right-size OpenClaw compaction model",
2252
+ "predicate": {
2253
+ "harness": [
2254
+ "openclaw"
2255
+ ],
2256
+ "payload_shape": "compaction or memory flush using the primary conversation model",
2257
+ "wire_protocol": [
2258
+ "openai"
2259
+ ]
2260
+ },
2261
+ "fix": {
2262
+ "prose": "Evaluate a cheaper exact model override for housekeeping and retain it only when state checks pass.",
2263
+ "apply": "right_size_model"
2264
+ },
2265
+ "grader": {
2266
+ "type": "llm_judge",
2267
+ "fixture_template": "Held long sessions comparing retained facts and task completion across compaction models."
2268
+ },
2269
+ "experiment": {
2270
+ "method": "local_ab",
2271
+ "metric": "$/day band"
2272
+ },
2273
+ "evidence": {
2274
+ "status": "unmeasured"
2275
+ },
2276
+ "never_claim": "Never accept cheaper housekeeping when it causes lost state or paid fallback calls.",
2277
+ "skill_render": {
2278
+ "eligible": true,
2279
+ "skill_name": "caveman-practice-openclaw-compaction-model-right-sizing"
2280
+ }
2281
+ },
2282
+ {
2283
+ "id": "openclaw-lightweight-subagent-context",
2284
+ "family": "input_bloat",
2285
+ "title": "Use lightweight OpenClaw subagent context",
2286
+ "predicate": {
2287
+ "harness": [
2288
+ "openclaw"
2289
+ ],
2290
+ "payload_shape": "isolated subagent receiving unnecessary parent context state",
2291
+ "wire_protocol": [
2292
+ "openai"
2293
+ ]
2294
+ },
2295
+ "fix": {
2296
+ "prose": "Use isolated lightweight context for self-contained delegated work and return only the result needed by the parent.",
2297
+ "apply": "offload_exploration"
2298
+ },
2299
+ "grader": {
2300
+ "type": "exact_match",
2301
+ "fixture_template": "Held delegated tasks comparing required result fields under full and lightweight context."
2302
+ },
2303
+ "experiment": {
2304
+ "method": "local_ab",
2305
+ "metric": "tokens_in"
2306
+ },
2307
+ "evidence": {
2308
+ "status": "unmeasured"
2309
+ },
2310
+ "never_claim": "Never use lightweight context when the child depends on parent decisions or hidden state.",
2311
+ "skill_render": {
2312
+ "eligible": true,
2313
+ "skill_name": "caveman-practice-openclaw-lightweight-subagent-context"
2314
+ }
2315
+ },
2316
+ {
2317
+ "id": "openclaw-command-output-plugin",
2318
+ "family": "input_bloat",
2319
+ "title": "Rewrite noisy OpenClaw commands through shrink",
2320
+ "predicate": {
2321
+ "harness": [
2322
+ "openclaw"
2323
+ ],
2324
+ "payload_shape": "tool execution output likely to be large and repetitive",
2325
+ "wire_protocol": [
2326
+ "openai"
2327
+ ]
2328
+ },
2329
+ "fix": {
2330
+ "prose": "Use the verified plugin interception path to shrink eligible output while retaining byte-exact recovery.",
2331
+ "apply": "enable_sdk_config"
2332
+ },
2333
+ "grader": {
2334
+ "type": "exact_match",
2335
+ "fixture_template": "Held command outputs whose compact view and recovery match expected content."
2336
+ },
2337
+ "experiment": {
2338
+ "method": "local_ab",
2339
+ "metric": "tokens_in"
2340
+ },
2341
+ "evidence": {
2342
+ "status": "unmeasured"
2343
+ },
2344
+ "never_claim": "Never claim command rewrite support outside the tested OpenClaw plugin path.",
2345
+ "skill_render": {
2346
+ "eligible": true,
2347
+ "skill_name": "caveman-practice-openclaw-command-output-plugin"
2348
+ }
2349
+ },
2350
+ {
2351
+ "id": "stable-map-serialization",
2352
+ "family": "cache",
2353
+ "title": "Serialize stable maps deterministically",
2354
+ "predicate": {
2355
+ "harness": [
2356
+ "any"
2357
+ ],
2358
+ "payload_shape": "semantically identical prefix with changing object key order",
2359
+ "wire_protocol": [
2360
+ "any"
2361
+ ]
2362
+ },
2363
+ "fix": {
2364
+ "prose": "Sort keys and use deterministic serialization for cacheable request sections without changing their semantics.",
2365
+ "apply": "stabilize_prefix"
2366
+ },
2367
+ "grader": {
2368
+ "type": "exact_match",
2369
+ "fixture_template": "Held objects with varied insertion order that must serialize identically and produce equivalent outputs."
2370
+ },
2371
+ "experiment": {
2372
+ "method": "local_ab",
2373
+ "metric": "cache_hit_ratio"
2374
+ },
2375
+ "evidence": {
2376
+ "status": "unmeasured"
2377
+ },
2378
+ "never_claim": "Never normalize fields whose order is semantically meaningful.",
2379
+ "skill_render": {
2380
+ "eligible": true,
2381
+ "skill_name": "caveman-practice-stable-map-serialization"
2382
+ }
2383
+ },
2384
+ {
2385
+ "id": "volatile-prefix-isolation",
2386
+ "family": "cache",
2387
+ "title": "Move volatile fields after stable prefixes",
2388
+ "predicate": {
2389
+ "harness": [
2390
+ "any"
2391
+ ],
2392
+ "payload_shape": "timestamp, request identifier, nonce, or live state inside reusable prefix",
2393
+ "wire_protocol": [
2394
+ "any"
2395
+ ]
2396
+ },
2397
+ "fix": {
2398
+ "prose": "Move volatile fields after the cache boundary or omit them when the model does not need them.",
2399
+ "apply": "stabilize_prefix"
2400
+ },
2401
+ "grader": {
2402
+ "type": "exact_match",
2403
+ "fixture_template": "Held requests varying volatile fields while retaining equivalent task outputs and stable prefix bytes."
2404
+ },
2405
+ "experiment": {
2406
+ "method": "local_ab",
2407
+ "metric": "cache_hit_ratio"
2408
+ },
2409
+ "evidence": {
2410
+ "status": "unmeasured"
2411
+ },
2412
+ "never_claim": "Never remove live state that affects model correctness merely to improve cache reuse.",
2413
+ "skill_render": {
2414
+ "eligible": true,
2415
+ "skill_name": "caveman-practice-volatile-prefix-isolation"
2416
+ }
2417
+ },
2418
+ {
2419
+ "id": "prompt-literal-dedup",
2420
+ "family": "input_bloat",
2421
+ "title": "Remove duplicated prompt instructions",
2422
+ "predicate": {
2423
+ "harness": [
2424
+ "any"
2425
+ ],
2426
+ "payload_shape": "same instruction or example repeated inside one request",
2427
+ "wire_protocol": [
2428
+ "any"
2429
+ ]
2430
+ },
2431
+ "fix": {
2432
+ "prose": "State each load-bearing instruction once and remove duplicated literals from composed prompt sections.",
2433
+ "apply": "enable_sdk_config"
2434
+ },
2435
+ "grader": {
2436
+ "type": "exact_match",
2437
+ "fixture_template": "Held cases requiring every unique instruction after duplicate removal."
2438
+ },
2439
+ "experiment": {
2440
+ "method": "local_ab",
2441
+ "metric": "tokens_in"
2442
+ },
2443
+ "evidence": {
2444
+ "status": "unmeasured"
2445
+ },
2446
+ "never_claim": "Never merge repetitions that intentionally reinforce conflicting or high-risk constraints without an eval.",
2447
+ "skill_render": {
2448
+ "eligible": true,
2449
+ "skill_name": "caveman-practice-prompt-literal-dedup"
2450
+ }
2451
+ },
2452
+ {
2453
+ "id": "few-shot-example-pruning",
2454
+ "family": "input_bloat",
2455
+ "title": "Keep only useful few-shot examples",
2456
+ "predicate": {
2457
+ "harness": [
2458
+ "any"
2459
+ ],
2460
+ "payload_shape": "large example set with redundant or non-discriminating cases",
2461
+ "wire_protocol": [
2462
+ "any"
2463
+ ]
2464
+ },
2465
+ "fix": {
2466
+ "prose": "Remove one example group at a time and retain only examples that protect a measured behavior.",
2467
+ "apply": "enable_sdk_config"
2468
+ },
2469
+ "grader": {
2470
+ "type": "exact_match",
2471
+ "fixture_template": "Held cases mapped to each retained example behavior plus unseen boundary cases."
2472
+ },
2473
+ "experiment": {
2474
+ "method": "local_ab",
2475
+ "metric": "tokens_in"
2476
+ },
2477
+ "evidence": {
2478
+ "status": "unmeasured"
2479
+ },
2480
+ "never_claim": "Never prune examples without measuring the behavior each example protects.",
2481
+ "skill_render": {
2482
+ "eligible": true,
2483
+ "skill_name": "caveman-practice-few-shot-example-pruning"
2484
+ }
2485
+ },
2486
+ {
2487
+ "id": "lost-in-middle-context-ordering",
2488
+ "family": "input_bloat",
2489
+ "title": "Keep critical evidence out of the context middle",
2490
+ "predicate": {
2491
+ "harness": [
2492
+ "any"
2493
+ ],
2494
+ "payload_shape": "long multi-document context with critical evidence buried centrally",
2495
+ "wire_protocol": [
2496
+ "any"
2497
+ ]
2498
+ },
2499
+ "fix": {
2500
+ "prose": "Retrieve and place the most relevant evidence near the query or another tested high-attention position.",
2501
+ "apply": "enable_sdk_config"
2502
+ },
2503
+ "grader": {
2504
+ "type": "exact_match",
2505
+ "fixture_template": "Held retrieval tasks with required facts placed at beginning, middle, and end."
2506
+ },
2507
+ "experiment": {
2508
+ "method": "local_ab",
2509
+ "metric": "tokens_in"
2510
+ },
2511
+ "evidence": {
2512
+ "status": "unmeasured"
2513
+ },
2514
+ "never_claim": "Never treat one universal position as optimal across models and task shapes.",
2515
+ "skill_render": {
2516
+ "eligible": true,
2517
+ "skill_name": "caveman-practice-lost-in-middle-context-ordering"
2518
+ }
2519
+ },
2520
+ {
2521
+ "id": "retrieval-top-k-tuning",
2522
+ "family": "input_bloat",
2523
+ "title": "Tune retrieval depth on held queries",
2524
+ "predicate": {
2525
+ "harness": [
2526
+ "any"
2527
+ ],
2528
+ "payload_shape": "retriever returns a fixed large candidate set for every query",
2529
+ "wire_protocol": [
2530
+ "any"
2531
+ ]
2532
+ },
2533
+ "fix": {
2534
+ "prose": "Select retrieval depth against held recall and task quality instead of filling the available window.",
2535
+ "apply": "enable_sdk_config"
2536
+ },
2537
+ "grader": {
2538
+ "type": "localization_f1",
2539
+ "fixture_template": "Held queries with expected supporting documents or code locations across candidate depths."
2540
+ },
2541
+ "experiment": {
2542
+ "method": "local_ab",
2543
+ "metric": "tokens_in"
2544
+ },
2545
+ "evidence": {
2546
+ "status": "unmeasured"
2547
+ },
2548
+ "never_claim": "Never lower retrieval depth from token count alone when support recall falls.",
2549
+ "skill_render": {
2550
+ "eligible": true,
2551
+ "skill_name": "caveman-practice-retrieval-top-k-tuning"
2552
+ }
2553
+ },
2554
+ {
2555
+ "id": "retrieval-chunk-sizing",
2556
+ "family": "input_bloat",
2557
+ "title": "Tune retrieval chunk size",
2558
+ "predicate": {
2559
+ "harness": [
2560
+ "any"
2561
+ ],
2562
+ "payload_shape": "retrieved chunks contain large irrelevant neighborhoods or split required facts",
2563
+ "wire_protocol": [
2564
+ "any"
2565
+ ]
2566
+ },
2567
+ "fix": {
2568
+ "prose": "Evaluate chunk size and overlap against held support recall and total injected tokens.",
2569
+ "apply": "enable_sdk_config"
2570
+ },
2571
+ "grader": {
2572
+ "type": "localization_f1",
2573
+ "fixture_template": "Held queries whose supporting spans cross normal and boundary chunk positions."
2574
+ },
2575
+ "experiment": {
2576
+ "method": "local_ab",
2577
+ "metric": "tokens_in"
2578
+ },
2579
+ "evidence": {
2580
+ "status": "unmeasured"
2581
+ },
2582
+ "never_claim": "Never select chunk size from corpus averages without measuring retrieval and answer quality.",
2583
+ "skill_render": {
2584
+ "eligible": true,
2585
+ "skill_name": "caveman-practice-retrieval-chunk-sizing"
2586
+ }
2587
+ },
2588
+ {
2589
+ "id": "exact-response-cache",
2590
+ "family": "cache",
2591
+ "title": "Cache exact deterministic responses",
2592
+ "predicate": {
2593
+ "harness": [
2594
+ "any"
2595
+ ],
2596
+ "payload_shape": "identical deterministic requests repeated with stable dependencies",
2597
+ "wire_protocol": [
2598
+ "any"
2599
+ ]
2600
+ },
2601
+ "fix": {
2602
+ "prose": "Cache responses by exact normalized request and dependency version, with explicit expiry and invalidation.",
2603
+ "apply": "enable_gateway_optimizer"
2604
+ },
2605
+ "grader": {
2606
+ "type": "exact_match",
2607
+ "fixture_template": "Held repeated requests plus changed dependency and expiry cases."
2608
+ },
2609
+ "experiment": {
2610
+ "method": "shadow",
2611
+ "metric": "$/day band"
2612
+ },
2613
+ "evidence": {
2614
+ "status": "unmeasured"
2615
+ },
2616
+ "never_claim": "Never serve cached output when user, tool, policy, or source state can change the answer.",
2617
+ "skill_render": {
2618
+ "eligible": true,
2619
+ "skill_name": "caveman-practice-exact-response-cache"
2620
+ }
2621
+ },
2622
+ {
2623
+ "id": "semantic-response-cache",
2624
+ "family": "cache",
2625
+ "title": "Gate semantic response caching",
2626
+ "predicate": {
2627
+ "harness": [
2628
+ "any"
2629
+ ],
2630
+ "payload_shape": "near-duplicate requests in a context-stable workload",
2631
+ "wire_protocol": [
2632
+ "any"
2633
+ ]
2634
+ },
2635
+ "fix": {
2636
+ "prose": "Use a context-aware similarity gate, explicit tenancy and version keys, and a false-hit eval before any canary.",
2637
+ "apply": "enable_gateway_optimizer"
2638
+ },
2639
+ "grader": {
2640
+ "type": "llm_judge",
2641
+ "fixture_template": "Held similar and adversarially close requests with fresh and cached outputs judged for interchangeability."
2642
+ },
2643
+ "experiment": {
2644
+ "method": "canary",
2645
+ "metric": "$/day band"
2646
+ },
2647
+ "evidence": {
2648
+ "status": "unmeasured"
2649
+ },
2650
+ "never_claim": "Never treat embedding similarity as proof that two answers are interchangeable.",
2651
+ "skill_render": {
2652
+ "eligible": true,
2653
+ "skill_name": "caveman-practice-semantic-response-cache"
2654
+ }
2655
+ },
2656
+ {
2657
+ "id": "retry-request-deduplication",
2658
+ "family": "reliability",
2659
+ "title": "Deduplicate in-flight request retries",
2660
+ "predicate": {
2661
+ "harness": [
2662
+ "any"
2663
+ ],
2664
+ "payload_shape": "same logical request replayed concurrently after timeout",
2665
+ "wire_protocol": [
2666
+ "any"
2667
+ ]
2668
+ },
2669
+ "fix": {
2670
+ "prose": "Coalesce identical in-flight work under a scoped request key and return the completed result to waiters.",
2671
+ "apply": "review_retries"
2672
+ },
2673
+ "grader": {
2674
+ "type": "exact_match",
2675
+ "fixture_template": "Held concurrent duplicate calls requiring one provider execution and identical waiter results."
2676
+ },
2677
+ "experiment": {
2678
+ "method": "replay",
2679
+ "metric": "$/day band"
2680
+ },
2681
+ "evidence": {
2682
+ "status": "unmeasured"
2683
+ },
2684
+ "never_claim": "Never deduplicate requests whose authorization, tool state, or side effects differ.",
2685
+ "skill_render": {
2686
+ "eligible": true,
2687
+ "skill_name": "caveman-practice-retry-request-deduplication"
2688
+ }
2689
+ },
2690
+ {
2691
+ "id": "retry-after-backoff",
2692
+ "family": "reliability",
2693
+ "title": "Honor retry timing and add jitter",
2694
+ "predicate": {
2695
+ "harness": [
2696
+ "any"
2697
+ ],
2698
+ "payload_shape": "rate-limit or transient retry loop without server timing or jitter",
2699
+ "wire_protocol": [
2700
+ "any"
2701
+ ]
2702
+ },
2703
+ "fix": {
2704
+ "prose": "Treat server retry timing as a floor, add bounded exponential backoff and jitter, and retry only recoverable classes.",
2705
+ "apply": "review_retries"
2706
+ },
2707
+ "grader": {
2708
+ "type": "latency_threshold",
2709
+ "fixture_template": "Held transient failures with declared minimum delay, bounded completion time, and no retry on permanent errors."
2710
+ },
2711
+ "experiment": {
2712
+ "method": "replay",
2713
+ "metric": "$/day band"
2714
+ },
2715
+ "evidence": {
2716
+ "status": "unmeasured"
2717
+ },
2718
+ "never_claim": "Never retry permanent client errors or ignore provider timing guidance.",
2719
+ "skill_render": {
2720
+ "eligible": true,
2721
+ "skill_name": "caveman-practice-retry-after-backoff"
2722
+ }
2723
+ },
2724
+ {
2725
+ "id": "retry-attempt-cap",
2726
+ "family": "reliability",
2727
+ "title": "Cap repeated provider attempts",
2728
+ "predicate": {
2729
+ "harness": [
2730
+ "any"
2731
+ ],
2732
+ "payload_shape": "retry policy with no maximum attempt or elapsed-time bound",
2733
+ "wire_protocol": [
2734
+ "any"
2735
+ ]
2736
+ },
2737
+ "fix": {
2738
+ "prose": "Set workload-specific attempt and elapsed-time caps, then surface terminal failure instead of looping.",
2739
+ "apply": "review_retries"
2740
+ },
2741
+ "grader": {
2742
+ "type": "http_status",
2743
+ "fixture_template": "Held persistent failure requiring bounded attempts and the expected terminal error."
2744
+ },
2745
+ "experiment": {
2746
+ "method": "replay",
2747
+ "metric": "$/day band"
2748
+ },
2749
+ "evidence": {
2750
+ "status": "unmeasured"
2751
+ },
2752
+ "never_claim": "Never hide terminal failure behind synthetic success after the cap.",
2753
+ "skill_render": {
2754
+ "eligible": true,
2755
+ "skill_name": "caveman-practice-retry-attempt-cap"
2756
+ }
2757
+ },
2758
+ {
2759
+ "id": "provider-circuit-breaker",
2760
+ "family": "reliability",
2761
+ "title": "Open a circuit on sustained provider failure",
2762
+ "predicate": {
2763
+ "harness": [
2764
+ "any"
2765
+ ],
2766
+ "payload_shape": "repeated transient failures continuing across requests",
2767
+ "wire_protocol": [
2768
+ "any"
2769
+ ]
2770
+ },
2771
+ "fix": {
2772
+ "prose": "Stop calls after a measured failure threshold, probe recovery conservatively, and use only approved fallback behavior.",
2773
+ "apply": "review_retries"
2774
+ },
2775
+ "grader": {
2776
+ "type": "http_status",
2777
+ "fixture_template": "Held failure burst, open interval, probe, and recovery sequence with expected statuses."
2778
+ },
2779
+ "experiment": {
2780
+ "method": "replay",
2781
+ "metric": "$/day band"
2782
+ },
2783
+ "evidence": {
2784
+ "status": "unmeasured"
2785
+ },
2786
+ "never_claim": "Never return a fabricated model result while the circuit is open.",
2787
+ "skill_render": {
2788
+ "eligible": true,
2789
+ "skill_name": "caveman-practice-provider-circuit-breaker"
2790
+ }
2791
+ },
2792
+ {
2793
+ "id": "tool-result-pagination",
2794
+ "family": "input_bloat",
2795
+ "title": "Page large tool results",
2796
+ "predicate": {
2797
+ "harness": [
2798
+ "any"
2799
+ ],
2800
+ "payload_shape": "large list or document returned in one tool result",
2801
+ "wire_protocol": [
2802
+ "any"
2803
+ ]
2804
+ },
2805
+ "fix": {
2806
+ "prose": "Return a bounded first page plus a stable continuation handle and fetch later pages only when needed.",
2807
+ "apply": "enable_sdk_config"
2808
+ },
2809
+ "grader": {
2810
+ "type": "tool_sequence",
2811
+ "fixture_template": "Held tasks requiring first-page facts, later-page facts, and no unnecessary continuation."
2812
+ },
2813
+ "experiment": {
2814
+ "method": "replay",
2815
+ "metric": "tokens_in"
2816
+ },
2817
+ "evidence": {
2818
+ "status": "unmeasured"
2819
+ },
2820
+ "never_claim": "Never truncate silently; expose continuation and preserve result order.",
2821
+ "skill_render": {
2822
+ "eligible": true,
2823
+ "skill_name": "caveman-practice-tool-result-pagination"
2824
+ }
2825
+ },
2826
+ {
2827
+ "id": "tool-result-field-projection",
2828
+ "family": "input_bloat",
2829
+ "title": "Project only needed tool fields",
2830
+ "predicate": {
2831
+ "harness": [
2832
+ "any"
2833
+ ],
2834
+ "payload_shape": "wide structured tool result with sparsely used fields",
2835
+ "wire_protocol": [
2836
+ "any"
2837
+ ]
2838
+ },
2839
+ "fix": {
2840
+ "prose": "Select task-relevant fields before model injection and retain a handle for omitted data.",
2841
+ "apply": "enable_sdk_config"
2842
+ },
2843
+ "grader": {
2844
+ "type": "json_path_assertion",
2845
+ "fixture_template": "Held tasks requiring each retained field and one omitted field recovered through the handle."
2846
+ },
2847
+ "experiment": {
2848
+ "method": "replay",
2849
+ "metric": "tokens_in"
2850
+ },
2851
+ "evidence": {
2852
+ "status": "unmeasured"
2853
+ },
2854
+ "never_claim": "Never drop fields without a tested recovery path for tasks that need them.",
2855
+ "skill_render": {
2856
+ "eligible": true,
2857
+ "skill_name": "caveman-practice-tool-result-field-projection"
2858
+ }
2859
+ },
2860
+ {
2861
+ "id": "document-summary-memoization",
2862
+ "family": "input_bloat",
2863
+ "title": "Reuse document summaries",
2864
+ "predicate": {
2865
+ "harness": [
2866
+ "any"
2867
+ ],
2868
+ "payload_shape": "same versioned document summarized repeatedly",
2869
+ "wire_protocol": [
2870
+ "any"
2871
+ ]
2872
+ },
2873
+ "fix": {
2874
+ "prose": "Memoize summaries by document content hash, prompt version, and model configuration, then invalidate on any dependency change.",
2875
+ "apply": "enable_sdk_config"
2876
+ },
2877
+ "grader": {
2878
+ "type": "exact_match",
2879
+ "fixture_template": "Held unchanged, changed-content, changed-prompt, and changed-model cases."
2880
+ },
2881
+ "experiment": {
2882
+ "method": "replay",
2883
+ "metric": "tokens_in"
2884
+ },
2885
+ "evidence": {
2886
+ "status": "unmeasured"
2887
+ },
2888
+ "never_claim": "Never reuse a summary after its document or summarization contract changes.",
2889
+ "skill_render": {
2890
+ "eligible": true,
2891
+ "skill_name": "caveman-practice-document-summary-memoization"
2892
+ }
2893
+ },
2894
+ {
2895
+ "id": "structured-output-retry-elimination",
2896
+ "family": "reliability",
2897
+ "title": "Remove schema-repair re-asks",
2898
+ "predicate": {
2899
+ "harness": [
2900
+ "any"
2901
+ ],
2902
+ "payload_shape": "invalid structured output repaired by another full model request",
2903
+ "wire_protocol": [
2904
+ "any"
2905
+ ]
2906
+ },
2907
+ "fix": {
2908
+ "prose": "Use provider-native strict output where available, validate once, and handle refusal or invalid results without blind full-context re-asks.",
2909
+ "apply": "enable_sdk_config"
2910
+ },
2911
+ "grader": {
2912
+ "type": "json_schema",
2913
+ "fixture_template": "Held valid, boundary, refusal, and invalid outputs against the required schema."
2914
+ },
2915
+ "experiment": {
2916
+ "method": "replay",
2917
+ "metric": "$/day band"
2918
+ },
2919
+ "evidence": {
2920
+ "status": "unmeasured"
2921
+ },
2922
+ "never_claim": "Never treat schema conformance as semantic correctness.",
2923
+ "skill_render": {
2924
+ "eligible": true,
2925
+ "skill_name": "caveman-practice-structured-output-retry-elimination"
2926
+ }
2927
+ },
2928
+ {
2929
+ "id": "output-length-contract",
2930
+ "family": "routing",
2931
+ "title": "Set explicit output length contracts",
2932
+ "predicate": {
2933
+ "harness": [
2934
+ "any"
2935
+ ],
2936
+ "payload_shape": "verbose free-form output where a bounded structure satisfies the task",
2937
+ "wire_protocol": [
2938
+ "any"
2939
+ ]
2940
+ },
2941
+ "fix": {
2942
+ "prose": "Specify the required structure and concise length, then gate completeness and boundary cases.",
2943
+ "apply": "enable_sdk_config"
2944
+ },
2945
+ "grader": {
2946
+ "type": "llm_judge",
2947
+ "fixture_template": "Held tasks covering minimum, typical, and longest-valid complete answers."
2948
+ },
2949
+ "experiment": {
2950
+ "method": "canary",
2951
+ "metric": "tokens_out"
2952
+ },
2953
+ "evidence": {
2954
+ "status": "unmeasured"
2955
+ },
2956
+ "never_claim": "Never call shorter output better when required content is missing.",
2957
+ "skill_render": {
2958
+ "eligible": true,
2959
+ "skill_name": "caveman-practice-output-length-contract"
2960
+ }
2961
+ },
2962
+ {
2963
+ "id": "model-cascade",
2964
+ "family": "routing",
2965
+ "title": "Escalate only hard requests",
2966
+ "predicate": {
2967
+ "harness": [
2968
+ "any"
2969
+ ],
2970
+ "payload_shape": "mixed-difficulty workload always routed to the strongest model",
2971
+ "wire_protocol": [
2972
+ "any"
2973
+ ]
2974
+ },
2975
+ "fix": {
2976
+ "prose": "Run a cheaper candidate first and escalate only under a calibrated uncertainty or validation rule.",
2977
+ "apply": "right_size_model"
2978
+ },
2979
+ "grader": {
2980
+ "type": "llm_judge",
2981
+ "fixture_template": "Held easy, ambiguous, and hard cases testing answer quality and escalation decisions."
2982
+ },
2983
+ "experiment": {
2984
+ "method": "canary",
2985
+ "metric": "$/day band"
2986
+ },
2987
+ "evidence": {
2988
+ "status": "unmeasured"
2989
+ },
2990
+ "never_claim": "Never count cascade savings without including escalated requests and failed first attempts.",
2991
+ "skill_render": {
2992
+ "eligible": true,
2993
+ "skill_name": "caveman-practice-model-cascade"
2994
+ }
2995
+ },
2996
+ {
2997
+ "id": "subagent-fanout-guard",
2998
+ "family": "cache",
2999
+ "title": "Bound sub-agent fan-out with operator-approved limits",
3000
+ "predicate": {
3001
+ "harness": [
3002
+ "any"
3003
+ ],
3004
+ "payload_shape": "concurrent sibling sub-agent spans sharing one parent and a large request prefix without cache reuse",
3005
+ "wire_protocol": [
3006
+ "any"
3007
+ ]
3008
+ },
3009
+ "fix": {
3010
+ "prose": "Review the framework callsite and propose explicit root and child limits selected by an operator only after paired outcome replay.",
3011
+ "apply": "add_subagent_guard"
3012
+ },
3013
+ "grader": {
3014
+ "type": "tool_sequence",
3015
+ "fixture_template": "Paired held tasks asserting the guarded arm preserves required child results and failure handling under the operator-selected cap."
3016
+ },
3017
+ "experiment": {
3018
+ "method": "replay",
3019
+ "metric": "task_outcome_pass_rate"
3020
+ },
3021
+ "evidence": {
3022
+ "status": "unmeasured"
3023
+ },
3024
+ "never_claim": "Never derive a cap from sibling count or recorded spend, and never present the proposal, merge, or active guard as savings; this report-only recipe mints nothing.",
3025
+ "skill_render": {
3026
+ "eligible": false,
3027
+ "skill_name": "caveman-recipe-subagent-fanout-guard"
3028
+ }
3029
+ }
3030
+ ];