@yahyahammad/venastine 0.2.0 → 0.3.9
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.
- package/.env.example +9 -4
- package/CONFIG_ARCHITECTURE.md +2030 -0
- package/LICENSE +201 -201
- package/NOTICE +1 -1
- package/README.md +1034 -895
- package/agents/builtin/build.md +73 -0
- package/agents/builtin/compactor.md +75 -75
- package/agents/builtin/explore.md +77 -77
- package/agents/builtin/general.md +80 -0
- package/agents/builtin/pipeline-reviewer.md +100 -100
- package/agents/builtin/plan.md +106 -88
- package/agents/builtin/review.md +74 -74
- package/agents/builtin/test.md +65 -0
- package/agents/builtin/writer.md +67 -0
- package/agents/manager.py +68 -2
- package/agents/subagent_tool.py +472 -238
- package/agents/tui_commands.py +238 -151
- package/bin/venastine.mjs +150 -4
- package/bin/venastine.py +71 -0
- package/cli_commands.py +838 -0
- package/config.py +95 -1198
- package/config.yaml +1103 -0
- package/config_edit.py +1301 -0
- package/config_schema.py +971 -0
- package/config_update.py +749 -0
- package/core/agent_activity.py +256 -118
- package/core/ansi.py +452 -0
- package/core/approval.py +137 -115
- package/core/call_view.py +364 -0
- package/core/cancel.py +121 -0
- package/core/chatgpt_auth.py +747 -0
- package/core/claims_view.py +107 -0
- package/core/client.py +2715 -1559
- package/core/commands.py +460 -0
- package/core/compaction.py +1735 -971
- package/core/config_loader.py +1982 -1563
- package/core/connect.py +686 -0
- package/core/events.py +123 -61
- package/core/file_mentions.py +280 -0
- package/core/interaction.py +411 -328
- package/core/line_pattern.py +128 -0
- package/core/loop.py +2566 -1490
- package/core/memory.py +482 -370
- package/core/pipeline_models.py +190 -176
- package/core/provider_errors.py +234 -0
- package/core/reasoning/authorization.py +130 -130
- package/core/reasoning/base.py +219 -219
- package/core/reasoning/confidence_scoring.py +484 -468
- package/core/reasoning/events.py +149 -149
- package/core/reasoning/json_retry.py +19 -1
- package/core/reasoning/orchestrator.py +1561 -1382
- package/core/reasoning/output_writer.py +345 -286
- package/core/reasoning/payload_validation.py +524 -398
- package/core/reasoning/pipeline_storage.py +88 -22
- package/core/reasoning/review.py +659 -615
- package/core/reasoning/scholar.py +458 -434
- package/core/reasoning/source_corpus.py +433 -327
- package/core/reasoning/source_scoring.py +855 -855
- package/core/replay.py +315 -183
- package/core/session.py +36 -12
- package/core/session_view.py +259 -0
- package/core/session_wake.py +259 -0
- package/core/shell_sessions.py +1553 -0
- package/core/workspace_trust.py +1 -1
- package/credentials.py +148 -20
- package/database.py +333 -252
- package/{PRIVACY.md → docs/PRIVACY.md} +78 -66
- package/docs/SECURITY.md +108 -0
- package/docs/THIRD_PARTY_NOTICES.md +148 -0
- package/env_secrets.py +50 -9
- package/json_store.py +343 -123
- package/logging_setup.py +295 -185
- package/main.py +3433 -1961
- package/mcp_client/client.py +733 -499
- package/mcp_client/config.py +126 -21
- package/mcp_client/registration.py +343 -182
- package/memories/manager.py +147 -132
- package/memories/tui_commands.py +319 -244
- package/package.json +303 -73
- package/project_init/config_files.py +282 -278
- package/project_init/generator.py +728 -564
- package/project_init/manifest.py +18 -6
- package/project_init/tui_commands.py +147 -144
- package/prompts/critic_pass.md +24 -24
- package/prompts/revalidate.md +25 -25
- package/prompts/revise.md +2 -0
- package/prompts/source_grounding.md +66 -66
- package/prompts/system_prompts.py +316 -262
- package/providers.json.example +4 -2
- package/pyproject.toml +149 -100
- package/requirements.txt +295 -72
- package/safety/policy_enforcement.py +1082 -746
- package/security/askpass.py +419 -0
- package/security/capability.py +207 -177
- package/security/confine.py +484 -0
- package/security/hosted.py +182 -0
- package/security/permissions.py +237 -190
- package/security/posture.py +124 -12
- package/security/protected_paths.py +501 -309
- package/security/sandbox.py +4108 -985
- package/security/secrets.py +855 -0
- package/skills/builtin/math/formal-specification.md +67 -67
- package/skills/builtin/math/numerical-methods.md +66 -66
- package/skills/builtin/research/experiment-design.md +62 -62
- package/skills/builtin/research/reproducibility.md +55 -55
- package/skills/builtin/research/source-evaluation.md +70 -70
- package/skills/builtin/research/statistical-inference.md +63 -63
- package/skills/builtin/research/technical-writing.md +59 -59
- package/skills/builtin/software/code-review.md +61 -61
- package/skills/builtin/software/debugging.md +61 -61
- package/skills/builtin/software/test-design.md +67 -67
- package/skills/tui_commands.py +110 -111
- package/storage.py +1338 -1005
- package/tools/base.py +161 -1
- package/tools/builtin/_math_common.py +667 -377
- package/tools/builtin/_net_common.py +225 -19
- package/tools/builtin/arxiv.py +497 -39
- package/tools/builtin/ask_user.py +231 -213
- package/tools/builtin/discrete_math.py +35 -4
- package/tools/builtin/fetch_url.py +371 -153
- package/tools/builtin/file_ops.py +720 -412
- package/tools/builtin/geometry.py +48 -12
- package/tools/builtin/get_time.py +2 -2
- package/tools/builtin/linear_algebra.py +66 -23
- package/tools/builtin/load_skill.py +76 -66
- package/tools/builtin/logic.py +171 -12
- package/tools/builtin/pin.py +227 -193
- package/tools/builtin/probability_stats.py +12 -16
- package/tools/builtin/project_docs.py +464 -405
- package/tools/builtin/remember.py +175 -158
- package/tools/builtin/shell.py +980 -461
- package/tools/builtin/shell_sessions.py +613 -0
- package/tools/builtin/symbolic_math.py +80 -10
- package/tools/builtin/todo.py +149 -147
- package/tools/builtin/web_search.py +191 -13
- package/tools/context.py +159 -136
- package/tools/isolation.py +107 -34
- package/tools/registry.py +1242 -925
- package/tui/app.py +8318 -4100
- package/tui/app.tcss +747 -556
- package/tui/commands.py +26 -170
- package/tui/diffs.py +34 -13
- package/tui/history.py +128 -128
- package/tui/hosted_driver.py +102 -0
- package/tui/markdown.py +1329 -1279
- package/tui/meters.py +364 -268
- package/tui/ravens.py +146 -146
- package/tui/screens.py +1255 -943
- package/tui/themes.py +599 -559
- package/tui/widgets.py +3813 -2332
- package/SECURITY.md +0 -88
- package/THIRD_PARTY_NOTICES.md +0 -100
package/core/client.py
CHANGED
|
@@ -1,1559 +1,2715 @@
|
|
|
1
|
-
|
|
2
|
-
|
|
3
|
-
|
|
4
|
-
import
|
|
5
|
-
import
|
|
6
|
-
import os
|
|
7
|
-
|
|
8
|
-
import
|
|
9
|
-
from
|
|
10
|
-
|
|
11
|
-
import config
|
|
12
|
-
|
|
13
|
-
|
|
14
|
-
|
|
15
|
-
|
|
16
|
-
|
|
17
|
-
|
|
18
|
-
|
|
19
|
-
#
|
|
20
|
-
#
|
|
21
|
-
#
|
|
22
|
-
#
|
|
23
|
-
#
|
|
24
|
-
#
|
|
25
|
-
# the
|
|
26
|
-
#
|
|
27
|
-
#
|
|
28
|
-
#
|
|
29
|
-
#
|
|
30
|
-
#
|
|
31
|
-
|
|
32
|
-
|
|
33
|
-
|
|
34
|
-
#
|
|
35
|
-
#
|
|
36
|
-
|
|
37
|
-
|
|
38
|
-
|
|
39
|
-
|
|
40
|
-
|
|
41
|
-
|
|
42
|
-
|
|
43
|
-
|
|
44
|
-
|
|
45
|
-
|
|
46
|
-
|
|
47
|
-
|
|
48
|
-
|
|
49
|
-
|
|
50
|
-
|
|
51
|
-
#
|
|
52
|
-
#
|
|
53
|
-
|
|
54
|
-
|
|
55
|
-
|
|
56
|
-
|
|
57
|
-
|
|
58
|
-
|
|
59
|
-
|
|
60
|
-
|
|
61
|
-
|
|
62
|
-
|
|
63
|
-
|
|
64
|
-
|
|
65
|
-
|
|
66
|
-
|
|
67
|
-
|
|
68
|
-
|
|
69
|
-
|
|
70
|
-
|
|
71
|
-
|
|
72
|
-
|
|
73
|
-
|
|
74
|
-
|
|
75
|
-
|
|
76
|
-
|
|
77
|
-
|
|
78
|
-
#
|
|
79
|
-
#
|
|
80
|
-
#
|
|
81
|
-
|
|
82
|
-
|
|
83
|
-
|
|
84
|
-
|
|
85
|
-
|
|
86
|
-
|
|
87
|
-
|
|
88
|
-
|
|
89
|
-
|
|
90
|
-
|
|
91
|
-
|
|
92
|
-
|
|
93
|
-
|
|
94
|
-
|
|
95
|
-
|
|
96
|
-
|
|
97
|
-
|
|
98
|
-
|
|
99
|
-
|
|
100
|
-
|
|
101
|
-
|
|
102
|
-
|
|
103
|
-
|
|
104
|
-
|
|
105
|
-
|
|
106
|
-
|
|
107
|
-
|
|
108
|
-
|
|
109
|
-
|
|
110
|
-
|
|
111
|
-
|
|
112
|
-
|
|
113
|
-
|
|
114
|
-
|
|
115
|
-
|
|
116
|
-
|
|
117
|
-
|
|
118
|
-
|
|
119
|
-
|
|
120
|
-
|
|
121
|
-
|
|
122
|
-
|
|
123
|
-
|
|
124
|
-
|
|
125
|
-
|
|
126
|
-
|
|
127
|
-
|
|
128
|
-
|
|
129
|
-
|
|
130
|
-
|
|
131
|
-
|
|
132
|
-
|
|
133
|
-
|
|
134
|
-
|
|
135
|
-
|
|
136
|
-
|
|
137
|
-
|
|
138
|
-
|
|
139
|
-
|
|
140
|
-
|
|
141
|
-
|
|
142
|
-
|
|
143
|
-
|
|
144
|
-
|
|
145
|
-
|
|
146
|
-
|
|
147
|
-
|
|
148
|
-
|
|
149
|
-
|
|
150
|
-
|
|
151
|
-
|
|
152
|
-
|
|
153
|
-
|
|
154
|
-
|
|
155
|
-
|
|
156
|
-
|
|
157
|
-
|
|
158
|
-
|
|
159
|
-
|
|
160
|
-
|
|
161
|
-
|
|
162
|
-
|
|
163
|
-
|
|
164
|
-
|
|
165
|
-
|
|
166
|
-
|
|
167
|
-
|
|
168
|
-
|
|
169
|
-
#
|
|
170
|
-
# a
|
|
171
|
-
#
|
|
172
|
-
#
|
|
173
|
-
#
|
|
174
|
-
#
|
|
175
|
-
#
|
|
176
|
-
#
|
|
177
|
-
#
|
|
178
|
-
#
|
|
179
|
-
|
|
180
|
-
|
|
181
|
-
|
|
182
|
-
|
|
183
|
-
|
|
184
|
-
|
|
185
|
-
|
|
186
|
-
|
|
187
|
-
|
|
188
|
-
|
|
189
|
-
|
|
190
|
-
|
|
191
|
-
|
|
192
|
-
|
|
193
|
-
|
|
194
|
-
|
|
195
|
-
|
|
196
|
-
|
|
197
|
-
|
|
198
|
-
|
|
199
|
-
|
|
200
|
-
|
|
201
|
-
|
|
202
|
-
|
|
203
|
-
#
|
|
204
|
-
#
|
|
205
|
-
#
|
|
206
|
-
#
|
|
207
|
-
|
|
208
|
-
|
|
209
|
-
|
|
210
|
-
|
|
211
|
-
|
|
212
|
-
|
|
213
|
-
|
|
214
|
-
|
|
215
|
-
|
|
216
|
-
|
|
217
|
-
|
|
218
|
-
|
|
219
|
-
|
|
220
|
-
|
|
221
|
-
|
|
222
|
-
|
|
223
|
-
|
|
224
|
-
|
|
225
|
-
#
|
|
226
|
-
#
|
|
227
|
-
#
|
|
228
|
-
|
|
229
|
-
|
|
230
|
-
|
|
231
|
-
|
|
232
|
-
|
|
233
|
-
|
|
234
|
-
|
|
235
|
-
|
|
236
|
-
|
|
237
|
-
|
|
238
|
-
|
|
239
|
-
|
|
240
|
-
|
|
241
|
-
|
|
242
|
-
|
|
243
|
-
|
|
244
|
-
|
|
245
|
-
|
|
246
|
-
|
|
247
|
-
|
|
248
|
-
|
|
249
|
-
|
|
250
|
-
|
|
251
|
-
|
|
252
|
-
|
|
253
|
-
|
|
254
|
-
|
|
255
|
-
|
|
256
|
-
|
|
257
|
-
|
|
258
|
-
|
|
259
|
-
|
|
260
|
-
|
|
261
|
-
|
|
262
|
-
|
|
263
|
-
|
|
264
|
-
|
|
265
|
-
|
|
266
|
-
|
|
267
|
-
|
|
268
|
-
|
|
269
|
-
|
|
270
|
-
|
|
271
|
-
|
|
272
|
-
|
|
273
|
-
|
|
274
|
-
|
|
275
|
-
|
|
276
|
-
|
|
277
|
-
|
|
278
|
-
|
|
279
|
-
|
|
280
|
-
|
|
281
|
-
|
|
282
|
-
|
|
283
|
-
|
|
284
|
-
|
|
285
|
-
|
|
286
|
-
|
|
287
|
-
|
|
288
|
-
|
|
289
|
-
|
|
290
|
-
|
|
291
|
-
|
|
292
|
-
|
|
293
|
-
|
|
294
|
-
|
|
295
|
-
|
|
296
|
-
|
|
297
|
-
|
|
298
|
-
"
|
|
299
|
-
"
|
|
300
|
-
|
|
301
|
-
|
|
302
|
-
|
|
303
|
-
|
|
304
|
-
|
|
305
|
-
|
|
306
|
-
|
|
307
|
-
|
|
308
|
-
#
|
|
309
|
-
#
|
|
310
|
-
#
|
|
311
|
-
|
|
312
|
-
|
|
313
|
-
|
|
314
|
-
|
|
315
|
-
|
|
316
|
-
|
|
317
|
-
|
|
318
|
-
|
|
319
|
-
|
|
320
|
-
|
|
321
|
-
|
|
322
|
-
|
|
323
|
-
|
|
324
|
-
|
|
325
|
-
|
|
326
|
-
|
|
327
|
-
|
|
328
|
-
|
|
329
|
-
|
|
330
|
-
|
|
331
|
-
|
|
332
|
-
|
|
333
|
-
|
|
334
|
-
|
|
335
|
-
|
|
336
|
-
|
|
337
|
-
|
|
338
|
-
|
|
339
|
-
|
|
340
|
-
|
|
341
|
-
|
|
342
|
-
|
|
343
|
-
|
|
344
|
-
|
|
345
|
-
|
|
346
|
-
|
|
347
|
-
|
|
348
|
-
|
|
349
|
-
|
|
350
|
-
def
|
|
351
|
-
|
|
352
|
-
|
|
353
|
-
|
|
354
|
-
|
|
355
|
-
|
|
356
|
-
|
|
357
|
-
|
|
358
|
-
|
|
359
|
-
|
|
360
|
-
|
|
361
|
-
|
|
362
|
-
|
|
363
|
-
|
|
364
|
-
|
|
365
|
-
|
|
366
|
-
|
|
367
|
-
|
|
368
|
-
|
|
369
|
-
|
|
370
|
-
|
|
371
|
-
|
|
372
|
-
|
|
373
|
-
|
|
374
|
-
|
|
375
|
-
|
|
376
|
-
|
|
377
|
-
|
|
378
|
-
|
|
379
|
-
|
|
380
|
-
|
|
381
|
-
|
|
382
|
-
|
|
383
|
-
|
|
384
|
-
|
|
385
|
-
|
|
386
|
-
|
|
387
|
-
|
|
388
|
-
|
|
389
|
-
|
|
390
|
-
|
|
391
|
-
|
|
392
|
-
|
|
393
|
-
|
|
394
|
-
|
|
395
|
-
|
|
396
|
-
|
|
397
|
-
|
|
398
|
-
|
|
399
|
-
|
|
400
|
-
|
|
401
|
-
|
|
402
|
-
|
|
403
|
-
|
|
404
|
-
|
|
405
|
-
|
|
406
|
-
|
|
407
|
-
|
|
408
|
-
|
|
409
|
-
|
|
410
|
-
|
|
411
|
-
|
|
412
|
-
|
|
413
|
-
|
|
414
|
-
|
|
415
|
-
|
|
416
|
-
|
|
417
|
-
|
|
418
|
-
|
|
419
|
-
|
|
420
|
-
|
|
421
|
-
|
|
422
|
-
|
|
423
|
-
|
|
424
|
-
|
|
425
|
-
|
|
426
|
-
|
|
427
|
-
|
|
428
|
-
|
|
429
|
-
|
|
430
|
-
|
|
431
|
-
|
|
432
|
-
|
|
433
|
-
|
|
434
|
-
|
|
435
|
-
|
|
436
|
-
|
|
437
|
-
|
|
438
|
-
)
|
|
439
|
-
|
|
440
|
-
|
|
441
|
-
|
|
442
|
-
|
|
443
|
-
|
|
444
|
-
|
|
445
|
-
|
|
446
|
-
|
|
447
|
-
|
|
448
|
-
|
|
449
|
-
|
|
450
|
-
|
|
451
|
-
|
|
452
|
-
|
|
453
|
-
|
|
454
|
-
|
|
455
|
-
|
|
456
|
-
|
|
457
|
-
|
|
458
|
-
|
|
459
|
-
|
|
460
|
-
|
|
461
|
-
|
|
462
|
-
|
|
463
|
-
|
|
464
|
-
|
|
465
|
-
|
|
466
|
-
|
|
467
|
-
|
|
468
|
-
|
|
469
|
-
|
|
470
|
-
|
|
471
|
-
|
|
472
|
-
|
|
473
|
-
|
|
474
|
-
|
|
475
|
-
|
|
476
|
-
|
|
477
|
-
|
|
478
|
-
|
|
479
|
-
|
|
480
|
-
|
|
481
|
-
|
|
482
|
-
|
|
483
|
-
|
|
484
|
-
|
|
485
|
-
|
|
486
|
-
|
|
487
|
-
|
|
488
|
-
|
|
489
|
-
|
|
490
|
-
|
|
491
|
-
|
|
492
|
-
|
|
493
|
-
|
|
494
|
-
|
|
495
|
-
|
|
496
|
-
|
|
497
|
-
|
|
498
|
-
|
|
499
|
-
|
|
500
|
-
|
|
501
|
-
|
|
502
|
-
|
|
503
|
-
|
|
504
|
-
|
|
505
|
-
|
|
506
|
-
|
|
507
|
-
|
|
508
|
-
|
|
509
|
-
|
|
510
|
-
|
|
511
|
-
|
|
512
|
-
|
|
513
|
-
|
|
514
|
-
|
|
515
|
-
|
|
516
|
-
|
|
517
|
-
|
|
518
|
-
|
|
519
|
-
|
|
520
|
-
|
|
521
|
-
|
|
522
|
-
|
|
523
|
-
|
|
524
|
-
|
|
525
|
-
|
|
526
|
-
|
|
527
|
-
|
|
528
|
-
|
|
529
|
-
|
|
530
|
-
|
|
531
|
-
|
|
532
|
-
|
|
533
|
-
|
|
534
|
-
|
|
535
|
-
|
|
536
|
-
|
|
537
|
-
|
|
538
|
-
|
|
539
|
-
|
|
540
|
-
|
|
541
|
-
|
|
542
|
-
|
|
543
|
-
|
|
544
|
-
|
|
545
|
-
|
|
546
|
-
|
|
547
|
-
#
|
|
548
|
-
#
|
|
549
|
-
|
|
550
|
-
|
|
551
|
-
|
|
552
|
-
|
|
553
|
-
|
|
554
|
-
|
|
555
|
-
|
|
556
|
-
|
|
557
|
-
|
|
558
|
-
|
|
559
|
-
|
|
560
|
-
|
|
561
|
-
|
|
562
|
-
|
|
563
|
-
|
|
564
|
-
|
|
565
|
-
|
|
566
|
-
|
|
567
|
-
|
|
568
|
-
|
|
569
|
-
|
|
570
|
-
|
|
571
|
-
|
|
572
|
-
|
|
573
|
-
|
|
574
|
-
|
|
575
|
-
|
|
576
|
-
|
|
577
|
-
|
|
578
|
-
|
|
579
|
-
|
|
580
|
-
|
|
581
|
-
|
|
582
|
-
|
|
583
|
-
|
|
584
|
-
|
|
585
|
-
|
|
586
|
-
|
|
587
|
-
|
|
588
|
-
|
|
589
|
-
|
|
590
|
-
|
|
591
|
-
|
|
592
|
-
|
|
593
|
-
|
|
594
|
-
|
|
595
|
-
|
|
596
|
-
|
|
597
|
-
|
|
598
|
-
|
|
599
|
-
|
|
600
|
-
""
|
|
601
|
-
|
|
602
|
-
|
|
603
|
-
|
|
604
|
-
|
|
605
|
-
|
|
606
|
-
|
|
607
|
-
|
|
608
|
-
|
|
609
|
-
|
|
610
|
-
|
|
611
|
-
|
|
612
|
-
|
|
613
|
-
|
|
614
|
-
|
|
615
|
-
|
|
616
|
-
|
|
617
|
-
|
|
618
|
-
|
|
619
|
-
|
|
620
|
-
|
|
621
|
-
|
|
622
|
-
|
|
623
|
-
|
|
624
|
-
|
|
625
|
-
|
|
626
|
-
|
|
627
|
-
|
|
628
|
-
|
|
629
|
-
|
|
630
|
-
|
|
631
|
-
|
|
632
|
-
|
|
633
|
-
|
|
634
|
-
|
|
635
|
-
|
|
636
|
-
|
|
637
|
-
|
|
638
|
-
|
|
639
|
-
|
|
640
|
-
|
|
641
|
-
|
|
642
|
-
|
|
643
|
-
|
|
644
|
-
|
|
645
|
-
|
|
646
|
-
|
|
647
|
-
|
|
648
|
-
|
|
649
|
-
|
|
650
|
-
|
|
651
|
-
|
|
652
|
-
|
|
653
|
-
|
|
654
|
-
|
|
655
|
-
|
|
656
|
-
|
|
657
|
-
|
|
658
|
-
|
|
659
|
-
|
|
660
|
-
|
|
661
|
-
|
|
662
|
-
|
|
663
|
-
|
|
664
|
-
|
|
665
|
-
|
|
666
|
-
|
|
667
|
-
|
|
668
|
-
"""
|
|
669
|
-
return
|
|
670
|
-
|
|
671
|
-
|
|
672
|
-
|
|
673
|
-
|
|
674
|
-
|
|
675
|
-
|
|
676
|
-
|
|
677
|
-
|
|
678
|
-
|
|
679
|
-
the
|
|
680
|
-
|
|
681
|
-
|
|
682
|
-
|
|
683
|
-
|
|
684
|
-
|
|
685
|
-
|
|
686
|
-
|
|
687
|
-
|
|
688
|
-
|
|
689
|
-
|
|
690
|
-
|
|
691
|
-
|
|
692
|
-
return []
|
|
693
|
-
if provider_name
|
|
694
|
-
|
|
695
|
-
|
|
696
|
-
|
|
697
|
-
|
|
698
|
-
|
|
699
|
-
|
|
700
|
-
|
|
701
|
-
|
|
702
|
-
|
|
703
|
-
|
|
704
|
-
|
|
705
|
-
|
|
706
|
-
|
|
707
|
-
|
|
708
|
-
|
|
709
|
-
|
|
710
|
-
|
|
711
|
-
|
|
712
|
-
|
|
713
|
-
|
|
714
|
-
|
|
715
|
-
|
|
716
|
-
|
|
717
|
-
|
|
718
|
-
|
|
719
|
-
|
|
720
|
-
|
|
721
|
-
|
|
722
|
-
|
|
723
|
-
|
|
724
|
-
|
|
725
|
-
|
|
726
|
-
|
|
727
|
-
|
|
728
|
-
|
|
729
|
-
|
|
730
|
-
|
|
731
|
-
|
|
732
|
-
|
|
733
|
-
|
|
734
|
-
|
|
735
|
-
|
|
736
|
-
|
|
737
|
-
|
|
738
|
-
|
|
739
|
-
|
|
740
|
-
|
|
741
|
-
|
|
742
|
-
|
|
743
|
-
|
|
744
|
-
|
|
745
|
-
|
|
746
|
-
|
|
747
|
-
|
|
748
|
-
|
|
749
|
-
|
|
750
|
-
|
|
751
|
-
|
|
752
|
-
|
|
753
|
-
|
|
754
|
-
|
|
755
|
-
|
|
756
|
-
|
|
757
|
-
|
|
758
|
-
|
|
759
|
-
|
|
760
|
-
|
|
761
|
-
|
|
762
|
-
|
|
763
|
-
|
|
764
|
-
|
|
765
|
-
|
|
766
|
-
|
|
767
|
-
|
|
768
|
-
|
|
769
|
-
|
|
770
|
-
|
|
771
|
-
|
|
772
|
-
|
|
773
|
-
|
|
774
|
-
|
|
775
|
-
|
|
776
|
-
|
|
777
|
-
|
|
778
|
-
|
|
779
|
-
|
|
780
|
-
|
|
781
|
-
|
|
782
|
-
|
|
783
|
-
|
|
784
|
-
|
|
785
|
-
|
|
786
|
-
|
|
787
|
-
|
|
788
|
-
|
|
789
|
-
|
|
790
|
-
|
|
791
|
-
|
|
792
|
-
|
|
793
|
-
|
|
794
|
-
|
|
795
|
-
|
|
796
|
-
|
|
797
|
-
|
|
798
|
-
|
|
799
|
-
|
|
800
|
-
|
|
801
|
-
|
|
802
|
-
|
|
803
|
-
|
|
804
|
-
|
|
805
|
-
|
|
806
|
-
|
|
807
|
-
|
|
808
|
-
|
|
809
|
-
|
|
810
|
-
|
|
811
|
-
|
|
812
|
-
|
|
813
|
-
|
|
814
|
-
|
|
815
|
-
|
|
816
|
-
|
|
817
|
-
|
|
818
|
-
|
|
819
|
-
|
|
820
|
-
|
|
821
|
-
|
|
822
|
-
|
|
823
|
-
|
|
824
|
-
|
|
825
|
-
|
|
826
|
-
|
|
827
|
-
|
|
828
|
-
|
|
829
|
-
|
|
830
|
-
|
|
831
|
-
|
|
832
|
-
|
|
833
|
-
|
|
834
|
-
|
|
835
|
-
|
|
836
|
-
|
|
837
|
-
|
|
838
|
-
|
|
839
|
-
|
|
840
|
-
|
|
841
|
-
|
|
842
|
-
|
|
843
|
-
|
|
844
|
-
|
|
845
|
-
|
|
846
|
-
|
|
847
|
-
|
|
848
|
-
|
|
849
|
-
|
|
850
|
-
|
|
851
|
-
|
|
852
|
-
|
|
853
|
-
|
|
854
|
-
|
|
855
|
-
|
|
856
|
-
|
|
857
|
-
|
|
858
|
-
|
|
859
|
-
|
|
860
|
-
|
|
861
|
-
|
|
862
|
-
|
|
863
|
-
|
|
864
|
-
|
|
865
|
-
|
|
866
|
-
|
|
867
|
-
|
|
868
|
-
|
|
869
|
-
|
|
870
|
-
"
|
|
871
|
-
|
|
872
|
-
|
|
873
|
-
|
|
874
|
-
|
|
875
|
-
|
|
876
|
-
|
|
877
|
-
|
|
878
|
-
|
|
879
|
-
|
|
880
|
-
|
|
881
|
-
|
|
882
|
-
|
|
883
|
-
|
|
884
|
-
|
|
885
|
-
|
|
886
|
-
|
|
887
|
-
|
|
888
|
-
|
|
889
|
-
|
|
890
|
-
|
|
891
|
-
|
|
892
|
-
|
|
893
|
-
|
|
894
|
-
|
|
895
|
-
|
|
896
|
-
|
|
897
|
-
|
|
898
|
-
|
|
899
|
-
|
|
900
|
-
|
|
901
|
-
|
|
902
|
-
|
|
903
|
-
|
|
904
|
-
|
|
905
|
-
|
|
906
|
-
|
|
907
|
-
|
|
908
|
-
|
|
909
|
-
|
|
910
|
-
|
|
911
|
-
|
|
912
|
-
|
|
913
|
-
|
|
914
|
-
|
|
915
|
-
|
|
916
|
-
|
|
917
|
-
|
|
918
|
-
|
|
919
|
-
|
|
920
|
-
|
|
921
|
-
|
|
922
|
-
|
|
923
|
-
|
|
924
|
-
|
|
925
|
-
|
|
926
|
-
|
|
927
|
-
|
|
928
|
-
|
|
929
|
-
|
|
930
|
-
|
|
931
|
-
|
|
932
|
-
|
|
933
|
-
|
|
934
|
-
|
|
935
|
-
|
|
936
|
-
|
|
937
|
-
|
|
938
|
-
|
|
939
|
-
|
|
940
|
-
|
|
941
|
-
|
|
942
|
-
|
|
943
|
-
|
|
944
|
-
|
|
945
|
-
|
|
946
|
-
|
|
947
|
-
|
|
948
|
-
|
|
949
|
-
|
|
950
|
-
|
|
951
|
-
|
|
952
|
-
|
|
953
|
-
|
|
954
|
-
|
|
955
|
-
|
|
956
|
-
|
|
957
|
-
|
|
958
|
-
|
|
959
|
-
|
|
960
|
-
|
|
961
|
-
|
|
962
|
-
|
|
963
|
-
|
|
964
|
-
|
|
965
|
-
|
|
966
|
-
|
|
967
|
-
|
|
968
|
-
|
|
969
|
-
|
|
970
|
-
|
|
971
|
-
|
|
972
|
-
|
|
973
|
-
|
|
974
|
-
|
|
975
|
-
|
|
976
|
-
|
|
977
|
-
|
|
978
|
-
|
|
979
|
-
|
|
980
|
-
|
|
981
|
-
|
|
982
|
-
|
|
983
|
-
|
|
984
|
-
|
|
985
|
-
|
|
986
|
-
|
|
987
|
-
|
|
988
|
-
|
|
989
|
-
|
|
990
|
-
|
|
991
|
-
|
|
992
|
-
|
|
993
|
-
|
|
994
|
-
|
|
995
|
-
|
|
996
|
-
|
|
997
|
-
|
|
998
|
-
|
|
999
|
-
|
|
1000
|
-
|
|
1001
|
-
|
|
1002
|
-
|
|
1003
|
-
|
|
1004
|
-
|
|
1005
|
-
|
|
1006
|
-
|
|
1007
|
-
|
|
1008
|
-
|
|
1009
|
-
|
|
1010
|
-
|
|
1011
|
-
|
|
1012
|
-
|
|
1013
|
-
|
|
1014
|
-
|
|
1015
|
-
|
|
1016
|
-
|
|
1017
|
-
|
|
1018
|
-
|
|
1019
|
-
|
|
1020
|
-
|
|
1021
|
-
|
|
1022
|
-
|
|
1023
|
-
|
|
1024
|
-
|
|
1025
|
-
|
|
1026
|
-
|
|
1027
|
-
|
|
1028
|
-
|
|
1029
|
-
|
|
1030
|
-
|
|
1031
|
-
|
|
1032
|
-
|
|
1033
|
-
|
|
1034
|
-
|
|
1035
|
-
|
|
1036
|
-
|
|
1037
|
-
|
|
1038
|
-
|
|
1039
|
-
|
|
1040
|
-
|
|
1041
|
-
|
|
1042
|
-
|
|
1043
|
-
|
|
1044
|
-
|
|
1045
|
-
|
|
1046
|
-
|
|
1047
|
-
|
|
1048
|
-
|
|
1049
|
-
|
|
1050
|
-
|
|
1051
|
-
|
|
1052
|
-
|
|
1053
|
-
|
|
1054
|
-
|
|
1055
|
-
#
|
|
1056
|
-
#
|
|
1057
|
-
#
|
|
1058
|
-
|
|
1059
|
-
|
|
1060
|
-
|
|
1061
|
-
|
|
1062
|
-
|
|
1063
|
-
|
|
1064
|
-
|
|
1065
|
-
|
|
1066
|
-
|
|
1067
|
-
|
|
1068
|
-
|
|
1069
|
-
|
|
1070
|
-
|
|
1071
|
-
|
|
1072
|
-
|
|
1073
|
-
|
|
1074
|
-
|
|
1075
|
-
|
|
1076
|
-
|
|
1077
|
-
|
|
1078
|
-
|
|
1079
|
-
|
|
1080
|
-
|
|
1081
|
-
|
|
1082
|
-
|
|
1083
|
-
|
|
1084
|
-
|
|
1085
|
-
|
|
1086
|
-
|
|
1087
|
-
|
|
1088
|
-
|
|
1089
|
-
|
|
1090
|
-
|
|
1091
|
-
|
|
1092
|
-
|
|
1093
|
-
|
|
1094
|
-
|
|
1095
|
-
|
|
1096
|
-
|
|
1097
|
-
|
|
1098
|
-
|
|
1099
|
-
|
|
1100
|
-
|
|
1101
|
-
|
|
1102
|
-
|
|
1103
|
-
|
|
1104
|
-
|
|
1105
|
-
#
|
|
1106
|
-
|
|
1107
|
-
|
|
1108
|
-
|
|
1109
|
-
|
|
1110
|
-
|
|
1111
|
-
|
|
1112
|
-
|
|
1113
|
-
|
|
1114
|
-
|
|
1115
|
-
|
|
1116
|
-
|
|
1117
|
-
|
|
1118
|
-
|
|
1119
|
-
|
|
1120
|
-
|
|
1121
|
-
|
|
1122
|
-
|
|
1123
|
-
|
|
1124
|
-
|
|
1125
|
-
|
|
1126
|
-
|
|
1127
|
-
|
|
1128
|
-
#
|
|
1129
|
-
|
|
1130
|
-
|
|
1131
|
-
|
|
1132
|
-
|
|
1133
|
-
|
|
1134
|
-
|
|
1135
|
-
|
|
1136
|
-
|
|
1137
|
-
|
|
1138
|
-
|
|
1139
|
-
|
|
1140
|
-
|
|
1141
|
-
|
|
1142
|
-
|
|
1143
|
-
|
|
1144
|
-
|
|
1145
|
-
|
|
1146
|
-
|
|
1147
|
-
|
|
1148
|
-
|
|
1149
|
-
|
|
1150
|
-
|
|
1151
|
-
|
|
1152
|
-
|
|
1153
|
-
|
|
1154
|
-
|
|
1155
|
-
|
|
1156
|
-
|
|
1157
|
-
|
|
1158
|
-
|
|
1159
|
-
|
|
1160
|
-
|
|
1161
|
-
|
|
1162
|
-
|
|
1163
|
-
|
|
1164
|
-
|
|
1165
|
-
|
|
1166
|
-
|
|
1167
|
-
|
|
1168
|
-
|
|
1169
|
-
|
|
1170
|
-
|
|
1171
|
-
|
|
1172
|
-
|
|
1173
|
-
|
|
1174
|
-
|
|
1175
|
-
|
|
1176
|
-
|
|
1177
|
-
|
|
1178
|
-
|
|
1179
|
-
|
|
1180
|
-
|
|
1181
|
-
|
|
1182
|
-
|
|
1183
|
-
|
|
1184
|
-
|
|
1185
|
-
|
|
1186
|
-
|
|
1187
|
-
|
|
1188
|
-
|
|
1189
|
-
|
|
1190
|
-
|
|
1191
|
-
|
|
1192
|
-
|
|
1193
|
-
|
|
1194
|
-
|
|
1195
|
-
|
|
1196
|
-
|
|
1197
|
-
|
|
1198
|
-
|
|
1199
|
-
|
|
1200
|
-
|
|
1201
|
-
|
|
1202
|
-
|
|
1203
|
-
|
|
1204
|
-
|
|
1205
|
-
|
|
1206
|
-
|
|
1207
|
-
|
|
1208
|
-
|
|
1209
|
-
|
|
1210
|
-
|
|
1211
|
-
|
|
1212
|
-
|
|
1213
|
-
|
|
1214
|
-
|
|
1215
|
-
|
|
1216
|
-
|
|
1217
|
-
|
|
1218
|
-
|
|
1219
|
-
|
|
1220
|
-
|
|
1221
|
-
|
|
1222
|
-
|
|
1223
|
-
|
|
1224
|
-
|
|
1225
|
-
|
|
1226
|
-
|
|
1227
|
-
|
|
1228
|
-
|
|
1229
|
-
|
|
1230
|
-
|
|
1231
|
-
|
|
1232
|
-
|
|
1233
|
-
|
|
1234
|
-
|
|
1235
|
-
|
|
1236
|
-
|
|
1237
|
-
|
|
1238
|
-
|
|
1239
|
-
|
|
1240
|
-
|
|
1241
|
-
|
|
1242
|
-
|
|
1243
|
-
|
|
1244
|
-
|
|
1245
|
-
|
|
1246
|
-
|
|
1247
|
-
|
|
1248
|
-
|
|
1249
|
-
|
|
1250
|
-
|
|
1251
|
-
|
|
1252
|
-
|
|
1253
|
-
|
|
1254
|
-
|
|
1255
|
-
|
|
1256
|
-
|
|
1257
|
-
|
|
1258
|
-
|
|
1259
|
-
|
|
1260
|
-
|
|
1261
|
-
|
|
1262
|
-
|
|
1263
|
-
|
|
1264
|
-
|
|
1265
|
-
|
|
1266
|
-
|
|
1267
|
-
|
|
1268
|
-
|
|
1269
|
-
|
|
1270
|
-
|
|
1271
|
-
|
|
1272
|
-
|
|
1273
|
-
|
|
1274
|
-
|
|
1275
|
-
|
|
1276
|
-
|
|
1277
|
-
|
|
1278
|
-
|
|
1279
|
-
|
|
1280
|
-
|
|
1281
|
-
|
|
1282
|
-
|
|
1283
|
-
|
|
1284
|
-
|
|
1285
|
-
|
|
1286
|
-
|
|
1287
|
-
|
|
1288
|
-
|
|
1289
|
-
|
|
1290
|
-
|
|
1291
|
-
|
|
1292
|
-
|
|
1293
|
-
|
|
1294
|
-
|
|
1295
|
-
|
|
1296
|
-
|
|
1297
|
-
|
|
1298
|
-
|
|
1299
|
-
|
|
1300
|
-
|
|
1301
|
-
|
|
1302
|
-
|
|
1303
|
-
|
|
1304
|
-
|
|
1305
|
-
|
|
1306
|
-
|
|
1307
|
-
|
|
1308
|
-
|
|
1309
|
-
|
|
1310
|
-
|
|
1311
|
-
|
|
1312
|
-
|
|
1313
|
-
|
|
1314
|
-
|
|
1315
|
-
|
|
1316
|
-
|
|
1317
|
-
|
|
1318
|
-
|
|
1319
|
-
|
|
1320
|
-
|
|
1321
|
-
|
|
1322
|
-
|
|
1323
|
-
|
|
1324
|
-
|
|
1325
|
-
|
|
1326
|
-
|
|
1327
|
-
|
|
1328
|
-
|
|
1329
|
-
|
|
1330
|
-
|
|
1331
|
-
|
|
1332
|
-
|
|
1333
|
-
|
|
1334
|
-
|
|
1335
|
-
|
|
1336
|
-
|
|
1337
|
-
|
|
1338
|
-
|
|
1339
|
-
|
|
1340
|
-
|
|
1341
|
-
|
|
1342
|
-
|
|
1343
|
-
|
|
1344
|
-
|
|
1345
|
-
|
|
1346
|
-
|
|
1347
|
-
|
|
1348
|
-
|
|
1349
|
-
|
|
1350
|
-
|
|
1351
|
-
|
|
1352
|
-
|
|
1353
|
-
|
|
1354
|
-
|
|
1355
|
-
|
|
1356
|
-
|
|
1357
|
-
|
|
1358
|
-
|
|
1359
|
-
|
|
1360
|
-
|
|
1361
|
-
|
|
1362
|
-
|
|
1363
|
-
|
|
1364
|
-
|
|
1365
|
-
|
|
1366
|
-
|
|
1367
|
-
|
|
1368
|
-
|
|
1369
|
-
|
|
1370
|
-
|
|
1371
|
-
|
|
1372
|
-
|
|
1373
|
-
|
|
1374
|
-
|
|
1375
|
-
|
|
1376
|
-
|
|
1377
|
-
|
|
1378
|
-
|
|
1379
|
-
|
|
1380
|
-
|
|
1381
|
-
|
|
1382
|
-
|
|
1383
|
-
|
|
1384
|
-
|
|
1385
|
-
|
|
1386
|
-
|
|
1387
|
-
|
|
1388
|
-
|
|
1389
|
-
|
|
1390
|
-
|
|
1391
|
-
|
|
1392
|
-
|
|
1393
|
-
|
|
1394
|
-
|
|
1395
|
-
|
|
1396
|
-
|
|
1397
|
-
|
|
1398
|
-
|
|
1399
|
-
|
|
1400
|
-
|
|
1401
|
-
|
|
1402
|
-
|
|
1403
|
-
|
|
1404
|
-
|
|
1405
|
-
|
|
1406
|
-
|
|
1407
|
-
|
|
1408
|
-
|
|
1409
|
-
|
|
1410
|
-
|
|
1411
|
-
|
|
1412
|
-
|
|
1413
|
-
|
|
1414
|
-
|
|
1415
|
-
|
|
1416
|
-
|
|
1417
|
-
|
|
1418
|
-
|
|
1419
|
-
|
|
1420
|
-
|
|
1421
|
-
|
|
1422
|
-
|
|
1423
|
-
|
|
1424
|
-
|
|
1425
|
-
|
|
1426
|
-
|
|
1427
|
-
|
|
1428
|
-
|
|
1429
|
-
|
|
1430
|
-
|
|
1431
|
-
|
|
1432
|
-
|
|
1433
|
-
|
|
1434
|
-
|
|
1435
|
-
|
|
1436
|
-
|
|
1437
|
-
|
|
1438
|
-
|
|
1439
|
-
|
|
1440
|
-
|
|
1441
|
-
|
|
1442
|
-
|
|
1443
|
-
|
|
1444
|
-
|
|
1445
|
-
|
|
1446
|
-
|
|
1447
|
-
|
|
1448
|
-
|
|
1449
|
-
|
|
1450
|
-
|
|
1451
|
-
|
|
1452
|
-
|
|
1453
|
-
|
|
1454
|
-
|
|
1455
|
-
|
|
1456
|
-
|
|
1457
|
-
|
|
1458
|
-
|
|
1459
|
-
|
|
1460
|
-
|
|
1461
|
-
|
|
1462
|
-
|
|
1463
|
-
|
|
1464
|
-
|
|
1465
|
-
|
|
1466
|
-
|
|
1467
|
-
|
|
1468
|
-
|
|
1469
|
-
|
|
1470
|
-
|
|
1471
|
-
|
|
1472
|
-
|
|
1473
|
-
|
|
1474
|
-
|
|
1475
|
-
|
|
1476
|
-
|
|
1477
|
-
|
|
1478
|
-
|
|
1479
|
-
|
|
1480
|
-
|
|
1481
|
-
|
|
1482
|
-
|
|
1483
|
-
|
|
1484
|
-
|
|
1485
|
-
|
|
1486
|
-
|
|
1487
|
-
|
|
1488
|
-
|
|
1489
|
-
|
|
1490
|
-
|
|
1491
|
-
|
|
1492
|
-
|
|
1493
|
-
|
|
1494
|
-
|
|
1495
|
-
|
|
1496
|
-
|
|
1497
|
-
|
|
1498
|
-
|
|
1499
|
-
|
|
1500
|
-
|
|
1501
|
-
|
|
1502
|
-
|
|
1503
|
-
|
|
1504
|
-
|
|
1505
|
-
|
|
1506
|
-
|
|
1507
|
-
|
|
1508
|
-
|
|
1509
|
-
|
|
1510
|
-
|
|
1511
|
-
|
|
1512
|
-
|
|
1513
|
-
|
|
1514
|
-
|
|
1515
|
-
|
|
1516
|
-
|
|
1517
|
-
|
|
1518
|
-
|
|
1519
|
-
|
|
1520
|
-
|
|
1521
|
-
|
|
1522
|
-
|
|
1523
|
-
|
|
1524
|
-
|
|
1525
|
-
|
|
1526
|
-
|
|
1527
|
-
|
|
1528
|
-
|
|
1529
|
-
|
|
1530
|
-
|
|
1531
|
-
|
|
1532
|
-
|
|
1533
|
-
|
|
1534
|
-
|
|
1535
|
-
|
|
1536
|
-
|
|
1537
|
-
|
|
1538
|
-
|
|
1539
|
-
|
|
1540
|
-
|
|
1541
|
-
|
|
1542
|
-
|
|
1543
|
-
|
|
1544
|
-
|
|
1545
|
-
|
|
1546
|
-
|
|
1547
|
-
|
|
1548
|
-
|
|
1549
|
-
|
|
1550
|
-
|
|
1551
|
-
|
|
1552
|
-
|
|
1553
|
-
|
|
1554
|
-
|
|
1555
|
-
|
|
1556
|
-
|
|
1557
|
-
|
|
1558
|
-
|
|
1559
|
-
|
|
1
|
+
import base64
|
|
2
|
+
import binascii
|
|
3
|
+
import json
|
|
4
|
+
import logging
|
|
5
|
+
import math
|
|
6
|
+
import os
|
|
7
|
+
from dataclasses import dataclass, field
|
|
8
|
+
from typing import Optional
|
|
9
|
+
from uuid import UUID, uuid4
|
|
10
|
+
|
|
11
|
+
import config
|
|
12
|
+
import credentials
|
|
13
|
+
from credentials import (
|
|
14
|
+
load_provider_data,
|
|
15
|
+
no_providers_message,
|
|
16
|
+
unknown_provider_message,
|
|
17
|
+
)
|
|
18
|
+
|
|
19
|
+
# ROADMAP_v2 §33 W7 (#135). The three provider SDKs are NOT imported here.
|
|
20
|
+
#
|
|
21
|
+
# core.loop imports this module and main.py imports core.loop, so a module
|
|
22
|
+
# -scope `from openai import OpenAI` is paid by every invocation of either
|
|
23
|
+
# shell -- `--help`, `--summary`, a mistyped flag, the thread picker --
|
|
24
|
+
# before argparse runs. Measured on Windows: 5.6s to print --help, of
|
|
25
|
+
# which 3.5s was the three SDKs, and no run uses more than one of them.
|
|
26
|
+
#
|
|
27
|
+
# Each import therefore sits at the point of use, INSIDE the branch that
|
|
28
|
+
# needs it rather than at the top of the enclosing function: a
|
|
29
|
+
# GOOGLE-only name imported at the top of `_messages_for_provider` would
|
|
30
|
+
# be paid on every Anthropic turn, which is most of them, and would leave
|
|
31
|
+
# the defect in place while looking fixed.
|
|
32
|
+
#
|
|
33
|
+
# Deferred imports are otherwise not this file's idiom. The trade is
|
|
34
|
+
# stated here rather than left to be re-derived: `import` is cached after
|
|
35
|
+
# the first call, so the cost is a dict lookup per call on the hot path
|
|
36
|
+
# and seconds saved on every launch that never makes one.
|
|
37
|
+
|
|
38
|
+
logger = logging.getLogger(__name__)
|
|
39
|
+
|
|
40
|
+
# Caches initialized clients so repeated calls for the same provider don't
|
|
41
|
+
# rebuild the client (and re-read the credentials file) every time.
|
|
42
|
+
_client_cache: dict[str, object] = {}
|
|
43
|
+
|
|
44
|
+
|
|
45
|
+
def api_initialization(provider_name: str):
|
|
46
|
+
if provider_name in _client_cache:
|
|
47
|
+
return _client_cache[provider_name]
|
|
48
|
+
|
|
49
|
+
provider_data = load_provider_data()
|
|
50
|
+
|
|
51
|
+
# #24. The file's absence and the provider's absence are different
|
|
52
|
+
# problems and now say so: the first names the path that was checked
|
|
53
|
+
# and its remedy, the second names what IS configured, like /model's
|
|
54
|
+
# refusal. Still ValueError either way -- callers (and tests) catch
|
|
55
|
+
# the type, not the text. The constant is read THROUGH the module,
|
|
56
|
+
# not imported: a from-import would freeze the value at this module's
|
|
57
|
+
# import time and a test's monkeypatch of credentials.LLM_PROVIDERS_FILE
|
|
58
|
+
# would not be seen here.
|
|
59
|
+
if not provider_data and not os.path.exists(
|
|
60
|
+
credentials.LLM_PROVIDERS_FILE):
|
|
61
|
+
raise ValueError(no_providers_message())
|
|
62
|
+
if provider_name not in provider_data:
|
|
63
|
+
raise ValueError(unknown_provider_message(provider_name, provider_data))
|
|
64
|
+
|
|
65
|
+
entry = provider_data[provider_name]
|
|
66
|
+
|
|
67
|
+
# ROADMAP_v3 §83. Routed by NAME, before the flag: the plan is not
|
|
68
|
+
# v1-compatible and must never fall to the Chat Completions branch.
|
|
69
|
+
if provider_name == "CHATGPT":
|
|
70
|
+
client = _chatgpt_client()
|
|
71
|
+
elif entry.get("is_v1_compatible", False):
|
|
72
|
+
from openai import OpenAI
|
|
73
|
+
client = OpenAI(base_url=entry["API_URL"], api_key=entry["API_KEY"])
|
|
74
|
+
elif provider_name == "GOOGLE":
|
|
75
|
+
from google import genai
|
|
76
|
+
client = genai.Client(api_key=entry["API_KEY"])
|
|
77
|
+
elif provider_name == "ANTHROPIC":
|
|
78
|
+
# Batch 234. Refused HERE, before a client exists, because the SDK
|
|
79
|
+
# no longer says what is wrong: 2.x sends the key as `Bearer <key>`,
|
|
80
|
+
# httpx2 refuses `Bearer ` as an illegal header before any request,
|
|
81
|
+
# and the SDK wraps that in APIConnectionError -- which
|
|
82
|
+
# provider_errors retries as a network blip. 0.116.0 refused the
|
|
83
|
+
# empty key itself, so no configuration that worked stops working.
|
|
84
|
+
# Nothing is cached, so adding the key takes effect on the next call.
|
|
85
|
+
if not entry.get("API_KEY"):
|
|
86
|
+
raise ValueError(credentials.empty_key_refusal(provider_name))
|
|
87
|
+
from anthropic import Anthropic
|
|
88
|
+
client = Anthropic(api_key=entry["API_KEY"])
|
|
89
|
+
# Add further proprietary API formats here
|
|
90
|
+
else:
|
|
91
|
+
raise ValueError(f"Provider {provider_name} is not currently supported")
|
|
92
|
+
|
|
93
|
+
_client_cache[provider_name] = client
|
|
94
|
+
return client
|
|
95
|
+
|
|
96
|
+
|
|
97
|
+
def _chatgpt_client():
|
|
98
|
+
"""The openai SDK, set up for the ChatGPT plan (ROADMAP_v3 §83, CG21).
|
|
99
|
+
|
|
100
|
+
`api_key` is a FUNCTION, `chatgpt_auth.access_token`: the SDK calls it
|
|
101
|
+
before every request (measured 25 of 25 in slice 2a), so the hourly
|
|
102
|
+
refresh needs no other plumbing and a cached client never holds a
|
|
103
|
+
stale token. `max_retries=0`, so CG10 decides every retry and the SDK
|
|
104
|
+
decides none: its own would resend a usage-limit 429.
|
|
105
|
+
|
|
106
|
+
The address is `chatgpt_auth.RESOURCE` and never the entry's `API_URL`.
|
|
107
|
+
The token is minted for that resource alone, and a hand-edited entry
|
|
108
|
+
must not be able to send it anywhere else.
|
|
109
|
+
|
|
110
|
+
Refused on a hosted deployment (CG1), as the sign-in already is, so a
|
|
111
|
+
`providers.json` carried onto one cannot reach the plan either.
|
|
112
|
+
"""
|
|
113
|
+
from security import hosted
|
|
114
|
+
|
|
115
|
+
if hosted.is_hosted():
|
|
116
|
+
raise ValueError(
|
|
117
|
+
"The ChatGPT plan answers on a local install only (ROADMAP_v3 "
|
|
118
|
+
"§83, CG1): OpenAI offers it to open-source apps running on your "
|
|
119
|
+
"own machine.")
|
|
120
|
+
from openai import OpenAI
|
|
121
|
+
|
|
122
|
+
from core import chatgpt_auth
|
|
123
|
+
|
|
124
|
+
return OpenAI(base_url=chatgpt_auth.RESOURCE,
|
|
125
|
+
api_key=chatgpt_auth.access_token, max_retries=0)
|
|
126
|
+
|
|
127
|
+
|
|
128
|
+
def forget_client(provider_name: str) -> None:
|
|
129
|
+
"""Drop the cached client for `provider_name`, so the next call builds a
|
|
130
|
+
new one from what `providers.json` (or the session) says NOW.
|
|
131
|
+
|
|
132
|
+
`/connect` is the caller (ROADMAP_v3 §79, HD5): the cache above is keyed
|
|
133
|
+
by provider name for the life of the process, so a key added or replaced
|
|
134
|
+
mid-session was otherwise ignored until a restart, with every call still
|
|
135
|
+
going out under the old one.
|
|
136
|
+
"""
|
|
137
|
+
_client_cache.pop(provider_name, None)
|
|
138
|
+
|
|
139
|
+
|
|
140
|
+
def accepts_images(provider_name: str) -> bool:
|
|
141
|
+
"""ROADMAP_v3 §81 (IM4). Whether an image in a tool result is DELIVERED
|
|
142
|
+
to this provider, as opposed to replaced by a note saying it was not.
|
|
143
|
+
|
|
144
|
+
Opt-in per provider through `accepts_images` in providers.json, on
|
|
145
|
+
`echoes_reasoning`'s terms: whether a provider takes an image depends on
|
|
146
|
+
the endpoint and often the model, cannot be discovered, and must not be
|
|
147
|
+
guessed, so it lives beside the endpoint and defaults to the
|
|
148
|
+
conservative answer. The test is `is True`, never truthiness:
|
|
149
|
+
`bool("false")` is True, and a quoted flag is a hand edit away (A29).
|
|
150
|
+
|
|
151
|
+
GOOGLE is honoured like the others since batch 173 (IM9). Until then it
|
|
152
|
+
was never delivered to and a flag on it was warned about: the old pin's
|
|
153
|
+
FunctionResponse had no field an image could go in, and the wire that
|
|
154
|
+
does work -- a harness turn after the batch -- had not been measured.
|
|
155
|
+
"""
|
|
156
|
+
entry = load_provider_data().get(provider_name, {})
|
|
157
|
+
return isinstance(entry, dict) and entry.get("accepts_images") is True
|
|
158
|
+
|
|
159
|
+
|
|
160
|
+
# ============================================================================
|
|
161
|
+
# ---- Normalized response shape ---------------------------------------------
|
|
162
|
+
# ============================================================================
|
|
163
|
+
|
|
164
|
+
@dataclass
|
|
165
|
+
class ToolCallRequest:
|
|
166
|
+
id: str
|
|
167
|
+
name: str
|
|
168
|
+
input: dict
|
|
169
|
+
# §33 W2 (#37). Set when the provider's `arguments` string could not
|
|
170
|
+
# be read as a JSON object -- most often because the response was cut
|
|
171
|
+
# off mid-arguments by the token limit. TRANSPORT ONLY: it describes
|
|
172
|
+
# one wire read, not the thread, so `add_assistant_message` persists
|
|
173
|
+
# {id, name, input} exactly as before and this never reaches storage.
|
|
174
|
+
#
|
|
175
|
+
# The call is still constructed rather than dropped. The model made a
|
|
176
|
+
# call and is owed an answer about it; a silently missing tool_result
|
|
177
|
+
# is M4's pairing defect, and `input={}` with no error would run the
|
|
178
|
+
# tool with arguments nobody sent.
|
|
179
|
+
parse_error: Optional[str] = None
|
|
180
|
+
|
|
181
|
+
|
|
182
|
+
@dataclass
|
|
183
|
+
class ModelResponse:
|
|
184
|
+
text: str
|
|
185
|
+
tool_calls: list[ToolCallRequest] = field(default_factory=list)
|
|
186
|
+
usage: dict = field(default_factory=lambda: {"input_tokens": 0, "output_tokens": 0})
|
|
187
|
+
stop_reason: str = "complete" # set by RunAgentLoop._run(), not here -- see core/loop.py
|
|
188
|
+
thread_id: Optional[UUID] = None # set by the entry points (stream_deep_research_mode,
|
|
189
|
+
# continue_conversation) so callers can persist/retry
|
|
190
|
+
# against the same thread. call_model() itself NEVER
|
|
191
|
+
# sets this -- it has no thread concept.
|
|
192
|
+
# §25: calls that proceeded on a PRE-FLIGHT grant rather than a live
|
|
193
|
+
# approval. Set by _run() (like stop_reason), never by call_model().
|
|
194
|
+
#
|
|
195
|
+
# Authorization moved up-front, so accountability moves after: this is
|
|
196
|
+
# the record of what a single launch-time yes actually authorised,
|
|
197
|
+
# written into the run artifacts for a pipeline nobody was watching.
|
|
198
|
+
# Empty for every run that granted nothing, which is every run today.
|
|
199
|
+
granted_calls: list = field(default_factory=list)
|
|
200
|
+
# ROADMAP_v2 §21: compaction notices raised during this run, each a
|
|
201
|
+
# {"kind", "text"} dict. Set by _run() like stop_reason and
|
|
202
|
+
# granted_calls, never by call_model().
|
|
203
|
+
#
|
|
204
|
+
# Carried here as well as yielded as a LoopEvent because
|
|
205
|
+
# run_to_completion() discards non-final events: the TUI watches the
|
|
206
|
+
# generator, but the CLI and the pipeline drain it, so a notice with
|
|
207
|
+
# only the event route would be invisible in two shells out of three.
|
|
208
|
+
notices: list = field(default_factory=list)
|
|
209
|
+
# #4 (batch 27): the two per-turn figures the loop computes, set by
|
|
210
|
+
# _run() like stop_reason -- never by call_model(), which has no turn
|
|
211
|
+
# concept. UNPRICED APPROXIMATIONS, deliberately: raw provider counts,
|
|
212
|
+
# no cache-read/write pricing and no currency anywhere. Precise spend
|
|
213
|
+
# lives in the provider's dashboard; these exist for RELATIVE reasoning
|
|
214
|
+
# ("is this turn ballooning the context?"), not billing truth.
|
|
215
|
+
#
|
|
216
|
+
# turn_billed_tokens: cumulative input+output over this user-turn so
|
|
217
|
+
# far -- the spend meter's running total (the prompt is re-sent every
|
|
218
|
+
# step, so this grows quadratically in a tool-using turn, which is
|
|
219
|
+
# correct: that is what the provider bills).
|
|
220
|
+
# turn_new_tokens: what THIS TURN added to the thread -- output plus
|
|
221
|
+
# positive input deltas, the first call's inherited prompt excluded
|
|
222
|
+
# and mid-turn compaction shrinks clamped away. This is the figure
|
|
223
|
+
# anything reasoning about SIZE reads; the spend meter must never be
|
|
224
|
+
# read that way again (TECHNICAL_DEBT item 9, closed batch 27).
|
|
225
|
+
# turn_output_tokens: what the MODEL WROTE this turn -- output tokens
|
|
226
|
+
# and nothing else, summed across every step. Batch 61, and it is a
|
|
227
|
+
# THIRD instrument rather than a convenience: the two above are a
|
|
228
|
+
# SPEND meter and a SIZE meter, and a generation RATE can be derived
|
|
229
|
+
# from neither. turn_billed_tokens counts the prompt again on every
|
|
230
|
+
# step; turn_new_tokens adds the positive input deltas, so a
|
|
231
|
+
# tool-using turn's tool results land in it and it overstates what was
|
|
232
|
+
# generated. Dividing either by a duration produces a number that
|
|
233
|
+
# looks like tok/s and is not, which is exactly the misreading
|
|
234
|
+
# TECHNICAL_DEBT item 9 was closed for.
|
|
235
|
+
turn_billed_tokens: Optional[int] = None
|
|
236
|
+
turn_new_tokens: Optional[int] = None
|
|
237
|
+
turn_output_tokens: Optional[int] = None
|
|
238
|
+
# Batch 44 (§44). The turn's reasoning, as the PROVIDER'S OWN BLOCKS
|
|
239
|
+
# rather than as the text §38 streams for display:
|
|
240
|
+
#
|
|
241
|
+
# {"provider": ..., "model": ..., "blocks": [ ...native... ]}
|
|
242
|
+
#
|
|
243
|
+
# §38's O1 made thinking display-only -- never persisted, never back
|
|
244
|
+
# on the wire -- and sidestepped "does a provider need its own
|
|
245
|
+
# thinking returned next turn?" by never returning it. That answer
|
|
246
|
+
# stopped being free once it was clear the model had never seen its
|
|
247
|
+
# own reasoning at ALL: not across a restart, and not between two
|
|
248
|
+
# turns of one live session, so a turn re-derives what it already
|
|
249
|
+
# worked out. O1 is amended here.
|
|
250
|
+
#
|
|
251
|
+
# BLOCKS, NOT THE STREAMED TEXT, and that is the whole reason this is
|
|
252
|
+
# a field rather than an accumulation in core/loop.py. An Anthropic
|
|
253
|
+
# thinking block carries a `signature`; the model reuses its own state
|
|
254
|
+
# through the block, not through the prose, so a re-serialised copy of
|
|
255
|
+
# the deltas would be unsigned and rejected. The blocks are read off
|
|
256
|
+
# get_final_message(), which already has them in hand.
|
|
257
|
+
#
|
|
258
|
+
# `provider` and `model` ride along because the replay rule is
|
|
259
|
+
# per-model: blocks go back UNCHANGED to the model that produced them
|
|
260
|
+
# and are dropped for anything else.
|
|
261
|
+
thinking: Optional[dict] = None
|
|
262
|
+
|
|
263
|
+
|
|
264
|
+
def _tool_arguments(raw_args: str, tool_name: str,
|
|
265
|
+
finish_reason: Optional[str] = None) -> tuple:
|
|
266
|
+
"""(input, parse_error) for one provider `arguments` string.
|
|
267
|
+
|
|
268
|
+
The ONLY place a wire string is parsed as JSON. Google and Anthropic
|
|
269
|
+
hand back dicts, so this is the OpenAI-compatible branch's -- which is
|
|
270
|
+
the branch a local model server takes, and truncated or malformed tool
|
|
271
|
+
arguments are likeliest there -- and, since ROADMAP_v3 §83, CHATGPT's,
|
|
272
|
+
whose Responses function calls carry `arguments` as a string too.
|
|
273
|
+
|
|
274
|
+
§33 W1 (#37): a failure is RETURNED, not raised. `dispatch()` turning
|
|
275
|
+
handler exceptions into {"error": ...} is the whole reason a raising
|
|
276
|
+
TOOL cannot kill a run; nothing played that role for a raising
|
|
277
|
+
TRANSLATION one layer up, so a response cut mid-arguments took a
|
|
278
|
+
JSONDecodeError out of this generator, out of _run(), and into the
|
|
279
|
+
pipeline's `except Exception` -- recording status='failed' on a
|
|
280
|
+
finished ten-pass run.
|
|
281
|
+
|
|
282
|
+
§33 W4: `finish_reason` is what separates "the provider emitted bad
|
|
283
|
+
JSON" from "the answer was cut off", which is the likeliest cause by
|
|
284
|
+
far and the only one the model can act on.
|
|
285
|
+
"""
|
|
286
|
+
# W9: no arguments is not an error. `get_time` declares properties=[]
|
|
287
|
+
# and `pin` has no required parameters, so this is a live path -- and
|
|
288
|
+
# json.loads("") raises. Dropping this fallback was green on the whole
|
|
289
|
+
# suite when audit unit 3 ran mutation P10.
|
|
290
|
+
if not raw_args:
|
|
291
|
+
return {}, None
|
|
292
|
+
|
|
293
|
+
try:
|
|
294
|
+
parsed = json.loads(raw_args)
|
|
295
|
+
except json.JSONDecodeError as exc:
|
|
296
|
+
if finish_reason == "length":
|
|
297
|
+
return {}, (
|
|
298
|
+
f"The arguments for {tool_name} were cut off by the response "
|
|
299
|
+
f"token limit, so the call could not be made. Reissue it with "
|
|
300
|
+
f"shorter arguments."
|
|
301
|
+
)
|
|
302
|
+
return {}, (
|
|
303
|
+
f"The arguments for {tool_name} were not valid JSON ({exc.msg}), "
|
|
304
|
+
f"so the call could not be made. Reissue it with valid JSON "
|
|
305
|
+
f"arguments."
|
|
306
|
+
)
|
|
307
|
+
|
|
308
|
+
# Valid JSON that is not an object -- "null", "[1,2]", "3". Rarer than
|
|
309
|
+
# a truncation and caught in the same place because the consequence is
|
|
310
|
+
# the same: dispatch() would be handed something that is not params.
|
|
311
|
+
if not isinstance(parsed, dict):
|
|
312
|
+
return {}, (
|
|
313
|
+
f"The arguments for {tool_name} parsed as "
|
|
314
|
+
f"{type(parsed).__name__}, not a JSON object, so the call could "
|
|
315
|
+
f"not be made. Reissue it with a JSON object."
|
|
316
|
+
)
|
|
317
|
+
|
|
318
|
+
return parsed, None
|
|
319
|
+
|
|
320
|
+
|
|
321
|
+
@dataclass
|
|
322
|
+
class StreamToken:
|
|
323
|
+
"""One token-level event yielded by call_model_stream(). Exactly one
|
|
324
|
+
field is populated per yield: text_delta for incremental text,
|
|
325
|
+
thinking_delta for incremental REASONING text, or final_response for
|
|
326
|
+
the terminal ModelResponse (with accumulated tool calls and usage).
|
|
327
|
+
|
|
328
|
+
§38 O1: thinking is DISPLAY-ONLY. It never joins the accumulated
|
|
329
|
+
`text`, is never persisted, and is never sent back on the wire -- so
|
|
330
|
+
the ModelResponse each branch produces is byte-identical to what it
|
|
331
|
+
produced before thinking was captured, and D20's persisted turn and
|
|
332
|
+
M4's fold boundary keep exactly the bytes they have today.
|
|
333
|
+
|
|
334
|
+
Which providers actually populate it (§38 O2): Anthropic, and every
|
|
335
|
+
OpenAI-compatible provider that streams `reasoning_content` or
|
|
336
|
+
`reasoning` (DeepSeek, Qwen, Z.AI, OpenRouter). It stays empty on
|
|
337
|
+
OPENAI -- its Chat Completions endpoint returns no reasoning text at
|
|
338
|
+
all -- and on GOOGLE, whose request deliberately does not ask for
|
|
339
|
+
thought summaries. Absence is not a bug there; see O2.
|
|
340
|
+
"""
|
|
341
|
+
text_delta: Optional[str] = None
|
|
342
|
+
thinking_delta: Optional[str] = None
|
|
343
|
+
final_response: Optional[ModelResponse] = None
|
|
344
|
+
|
|
345
|
+
|
|
346
|
+
# ============================================================================
|
|
347
|
+
# ---- Token usage, which every branch reports and no branch can trust -------
|
|
348
|
+
# ============================================================================
|
|
349
|
+
|
|
350
|
+
def _token_count(usage_obj, field: str) -> Optional[int]:
|
|
351
|
+
"""One token count off a provider's usage object, or None for "this
|
|
352
|
+
provider did not report that number".
|
|
353
|
+
|
|
354
|
+
`_record_window`'s rule applied to the other nullable figure these SDKs
|
|
355
|
+
carry -- "present" is not "usable". A bool is an int in Python and is
|
|
356
|
+
not a token count; a negative count is not one either.
|
|
357
|
+
|
|
358
|
+
WHY THIS IS NOT A `getattr(obj, field, 0)` DEFAULT, which is the
|
|
359
|
+
softened form D21's own note forbids: the attribute EXISTS on all three
|
|
360
|
+
pinned SDKs and holds None, so the default never fires and the None
|
|
361
|
+
travels into `core/loop.py`'s `usage_in + usage_out`. Measured per SDK
|
|
362
|
+
(batch 147), because the three reasons are different and only one of
|
|
363
|
+
them is visible in a type annotation. test_sdk_conformance.py holds each
|
|
364
|
+
on the pin; the versions below are batch 234's:
|
|
365
|
+
|
|
366
|
+
google-genai 2.32.0 every field of (as on 1.0.0 before it)
|
|
367
|
+
GenerateContentResponseUsageMetadata is
|
|
368
|
+
Optional[int] = None, and a bare instance is
|
|
369
|
+
TRUTHY -- so `if usage_meta:` passes and then
|
|
370
|
+
every read is None.
|
|
371
|
+
anthropic 2.1.0 Usage.input_tokens and CompletionUsage.
|
|
372
|
+
openai 3.36.0 prompt_tokens are annotated `int` and REQUIRED,
|
|
373
|
+
which says what the server promises rather than
|
|
374
|
+
what the object holds: both SDKs parse wire JSON
|
|
375
|
+
through pydantic's construct(), which skips
|
|
376
|
+
validation. `construct_type({"output_tokens":
|
|
377
|
+
7}, Usage)` yields input_tokens=None, and
|
|
378
|
+
anthropic's own stream accumulator guards
|
|
379
|
+
`input_tokens` with `is not None` against that
|
|
380
|
+
same `int` annotation -- while assigning
|
|
381
|
+
`output_tokens` unconditionally.
|
|
382
|
+
|
|
383
|
+
So an `int` annotation is not a runtime guarantee anywhere here, which
|
|
384
|
+
is why ONE reader serves all three branches rather than the Google one
|
|
385
|
+
that the defect was found in.
|
|
386
|
+
"""
|
|
387
|
+
if usage_obj is None:
|
|
388
|
+
return None
|
|
389
|
+
value = getattr(usage_obj, field, None)
|
|
390
|
+
if isinstance(value, bool) or not isinstance(value, int):
|
|
391
|
+
return None
|
|
392
|
+
return value if value >= 0 else None
|
|
393
|
+
|
|
394
|
+
|
|
395
|
+
def _stream_usage(provider_name: str, counts: dict,
|
|
396
|
+
supports_stream_usage: bool) -> dict:
|
|
397
|
+
"""D21's usage-or-raise for all three streaming branches, and the one
|
|
398
|
+
place that turns "not reported" into a number.
|
|
399
|
+
|
|
400
|
+
RETURNS A TOTAL SHAPE: both values are ints, always. That is what makes
|
|
401
|
+
a belt in `core/loop.py` unnecessary rather than merely unfashionable --
|
|
402
|
+
the producer is where this belongs (AGENTS.md), and a consumer-side
|
|
403
|
+
guard would leave the next consumer to write its own.
|
|
404
|
+
|
|
405
|
+
D21 says a stream must not silently degrade token accounting, and the
|
|
406
|
+
guard that enforced it compared `input == 0 and output == 0`. A
|
|
407
|
+
PARTIAL report defeats that comparison without reporting anything
|
|
408
|
+
usable: `None != 0`, and once coerced, an unreported output count is
|
|
409
|
+
indistinguishable from a genuine zero. §64's A20 widens the rule to
|
|
410
|
+
match what D21 says rather than what it happened to compare: under the
|
|
411
|
+
flag, a count that is not a number raises, because the flag is the
|
|
412
|
+
per-provider statement that this provider DOES report usage and a
|
|
413
|
+
missing number is that promise broken.
|
|
414
|
+
|
|
415
|
+
Two messages, deliberately. "Completed without usage data" is kept
|
|
416
|
+
verbatim for the case it has always named -- nothing reported at all,
|
|
417
|
+
or an explicit zero for both -- so the existing guards keep their
|
|
418
|
+
meaning. The partial case gets its own line naming which count is
|
|
419
|
+
missing, because that is the one a reader cannot diagnose from the
|
|
420
|
+
other.
|
|
421
|
+
|
|
422
|
+
With the flag FALSE the coercion is all that happens, and 0 is the
|
|
423
|
+
right answer there: the flag's whole purpose is to let a provider that
|
|
424
|
+
genuinely omits usage run anyway.
|
|
425
|
+
"""
|
|
426
|
+
reported_in = counts["input_tokens"]
|
|
427
|
+
reported_out = counts["output_tokens"]
|
|
428
|
+
if supports_stream_usage:
|
|
429
|
+
if (reported_in is None and reported_out is None) or (
|
|
430
|
+
reported_in == 0 and reported_out == 0):
|
|
431
|
+
raise RuntimeError(
|
|
432
|
+
f"Streaming call to {provider_name} completed without usage "
|
|
433
|
+
f"data despite supports_stream_usage=True. The stream may "
|
|
434
|
+
f"have been interrupted before the final usage chunk."
|
|
435
|
+
)
|
|
436
|
+
missing = [name for name, value in
|
|
437
|
+
(("input_tokens", reported_in),
|
|
438
|
+
("output_tokens", reported_out)) if value is None]
|
|
439
|
+
if missing:
|
|
440
|
+
raise RuntimeError(
|
|
441
|
+
f"Streaming call to {provider_name} reported "
|
|
442
|
+
f"{' and '.join(missing)} as absent or not a number despite "
|
|
443
|
+
f"supports_stream_usage=True, so this step's token "
|
|
444
|
+
f"accounting would be short by it."
|
|
445
|
+
)
|
|
446
|
+
return {"input_tokens": reported_in or 0,
|
|
447
|
+
"output_tokens": reported_out or 0}
|
|
448
|
+
|
|
449
|
+
|
|
450
|
+
# ============================================================================
|
|
451
|
+
# ---- Tool schema translation ------------------------------------------------
|
|
452
|
+
# ============================================================================
|
|
453
|
+
|
|
454
|
+
#: ROADMAP_v3 §82 (GS4). The `Schema` fields the Gemini API is known to take:
|
|
455
|
+
#: the 22 google-genai 1.0.0 had, which batch 132 and §9 verified on the
|
|
456
|
+
#: wire. The SDK serves Vertex AI as well, and 2.29.0 added three fields for
|
|
457
|
+
#: it -- `additional_properties`, which the Gemini API refused when measured
|
|
458
|
+
#: in batch 172 (400, `Unknown name "additional_properties"`), `defs` and
|
|
459
|
+
#: `ref`. One refused key fails every GOOGLE call, because the declarations
|
|
460
|
+
#: travel together, so a new SDK field waits to be measured.
|
|
461
|
+
_GOOGLE_SCHEMA_VERIFIED = frozenset({
|
|
462
|
+
"any_of", "default", "description", "enum", "example", "format",
|
|
463
|
+
"items", "max_items", "max_length", "max_properties", "maximum",
|
|
464
|
+
"min_items", "min_length", "min_properties", "minimum",
|
|
465
|
+
"nullable", "pattern", "properties", "property_ordering",
|
|
466
|
+
"required", "title", "type",
|
|
467
|
+
})
|
|
468
|
+
|
|
469
|
+
|
|
470
|
+
def _google_schema_keys(genai_types) -> set:
|
|
471
|
+
"""Every key the pinned `types.Schema` will accept AND the Gemini API is
|
|
472
|
+
known to take, read off the SDK.
|
|
473
|
+
|
|
474
|
+
Field names plus their camelCase aliases, because `Schema` sets
|
|
475
|
+
`populate_by_name` and JSON Schema spells things `anyOf`/`maxItems`
|
|
476
|
+
while the model declares `any_of`/`max_items`. Read off the SDK so that
|
|
477
|
+
a field the installed version lacks is never sent -- and read at CALL
|
|
478
|
+
time, not import time, because the root `conftest.py` puts a fake
|
|
479
|
+
`google.genai` in `sys.modules` before collection and an import-time
|
|
480
|
+
constant would bind the fake's fields for the whole test session.
|
|
481
|
+
|
|
482
|
+
It used to take EVERY field, on the reasoning that a version which adds
|
|
483
|
+
one should carry it with no change here. The 2.29.0 pin disproved that
|
|
484
|
+
(§82): a field on the type is not a field the Gemini API accepts. So the
|
|
485
|
+
SDK's set is narrowed to `_GOOGLE_SCHEMA_VERIFIED`.
|
|
486
|
+
"""
|
|
487
|
+
keys = set()
|
|
488
|
+
# NOT `field`: this module imports dataclasses.field at the top, and
|
|
489
|
+
# shadowing it here is what ruff F402 flagged in batch 141 -- found only
|
|
490
|
+
# then, because the lint job is not on main and so has never run.
|
|
491
|
+
for name, model_field in genai_types.Schema.model_fields.items():
|
|
492
|
+
if name not in _GOOGLE_SCHEMA_VERIFIED:
|
|
493
|
+
continue
|
|
494
|
+
keys.add(name)
|
|
495
|
+
alias = getattr(model_field, "alias", None)
|
|
496
|
+
if alias:
|
|
497
|
+
keys.add(alias)
|
|
498
|
+
return keys
|
|
499
|
+
|
|
500
|
+
|
|
501
|
+
def _reduce_for_google(node, accepted: set):
|
|
502
|
+
"""JSON Schema -> what the pinned `types.Schema` can actually carry.
|
|
503
|
+
|
|
504
|
+
`Schema.model_config` is `extra="forbid"`, so a key it has no field for
|
|
505
|
+
is a `ValidationError` rather than something ignored. The translation
|
|
506
|
+
used to hand `input_schema` over untouched, and because it is one list
|
|
507
|
+
comprehension over every schema, ONE refusal took the whole advertised
|
|
508
|
+
set with it: no model call on GOOGLE could be built at all. This
|
|
509
|
+
repository's own `probability_stats` was the entry that did it --
|
|
510
|
+
pydantic renders its `Optional[dict[str, float]]` as
|
|
511
|
+
`{"anyOf": [{"additionalProperties": {...}, "type": "object"},
|
|
512
|
+
{"type": "null"}]}`, and `additionalProperties` is not a field.
|
|
513
|
+
|
|
514
|
+
THE REDUCTION IS HERE AND NOT IN THE TOOL, because the schemas are
|
|
515
|
+
generated by `model_json_schema()`: fixing the one tool leaves the next
|
|
516
|
+
`Optional[dict]` to find this again, which is the shape D24 settled for
|
|
517
|
+
permissions. Two rules:
|
|
518
|
+
|
|
519
|
+
* drop any key `Schema` has no field or alias for;
|
|
520
|
+
* flatten `anyOf: [X, {"type": "null"}]` to X plus `nullable`, which
|
|
521
|
+
is how this SDK spells an optional value. Dropping the null member
|
|
522
|
+
is not cosmetic -- the pinned SDK coerces `{"type": "null"}` with a
|
|
523
|
+
`UserWarning: null is not a valid Type` and carries on, so without
|
|
524
|
+
the flattening a nullable field reaches the wire describing a type
|
|
525
|
+
that does not exist.
|
|
526
|
+
|
|
527
|
+
`docs/ARCHITECTURE.md` 4.7 said the raw dict "auto-converts to
|
|
528
|
+
`types.Schema`". It does not; it validates, and refuses.
|
|
529
|
+
|
|
530
|
+
Since batch 172 (§82) the pin's `Schema` HAS `additional_properties`
|
|
531
|
+
(2.29.0 then, 2.32.0 since batch 234, the same 25 fields) -- and the
|
|
532
|
+
Gemini API refuses it, so it is still dropped: `accepted` is the
|
|
533
|
+
verified set, not the SDK's.
|
|
534
|
+
"""
|
|
535
|
+
if isinstance(node, list):
|
|
536
|
+
return [_reduce_for_google(item, accepted) for item in node]
|
|
537
|
+
if not isinstance(node, dict):
|
|
538
|
+
return node
|
|
539
|
+
|
|
540
|
+
members = node.get("anyOf")
|
|
541
|
+
if isinstance(members, list):
|
|
542
|
+
live = [m for m in members
|
|
543
|
+
if not (isinstance(m, dict) and m.get("type") == "null")]
|
|
544
|
+
nullable = len(live) != len(members)
|
|
545
|
+
siblings = {k: v for k, v in node.items() if k != "anyOf"}
|
|
546
|
+
if len(live) == 1:
|
|
547
|
+
# The Optional[X] case: X itself, carrying anything that sat
|
|
548
|
+
# beside the union (a description, a default) and the flag.
|
|
549
|
+
merged = dict(_reduce_for_google(live[0], accepted))
|
|
550
|
+
merged.update(_reduce_for_google(siblings, accepted))
|
|
551
|
+
if nullable:
|
|
552
|
+
merged["nullable"] = True
|
|
553
|
+
return merged
|
|
554
|
+
out = _reduce_for_google(siblings, accepted)
|
|
555
|
+
out["anyOf"] = [_reduce_for_google(m, accepted) for m in live]
|
|
556
|
+
if nullable:
|
|
557
|
+
out["nullable"] = True
|
|
558
|
+
return out
|
|
559
|
+
|
|
560
|
+
out = {}
|
|
561
|
+
for key, value in node.items():
|
|
562
|
+
if key not in accepted:
|
|
563
|
+
continue
|
|
564
|
+
if key == "properties" and isinstance(value, dict):
|
|
565
|
+
out[key] = {name: _reduce_for_google(sub, accepted)
|
|
566
|
+
for name, sub in value.items()}
|
|
567
|
+
elif key in ("items", "anyOf"):
|
|
568
|
+
out[key] = _reduce_for_google(value, accepted)
|
|
569
|
+
else:
|
|
570
|
+
out[key] = value
|
|
571
|
+
return out
|
|
572
|
+
|
|
573
|
+
|
|
574
|
+
def _tools_for_provider(provider_name: str, tool_schemas: list[dict]):
|
|
575
|
+
"""
|
|
576
|
+
tool_schemas arrive in Anthropic's native shape (name/description/
|
|
577
|
+
input_schema), since that's what every tool's TOOL_SCHEMA already
|
|
578
|
+
produces. Translate into whatever shape the target provider expects.
|
|
579
|
+
"""
|
|
580
|
+
if not tool_schemas:
|
|
581
|
+
return None
|
|
582
|
+
|
|
583
|
+
if provider_name == "ANTHROPIC":
|
|
584
|
+
return tool_schemas
|
|
585
|
+
|
|
586
|
+
if provider_name == "GOOGLE":
|
|
587
|
+
from google.genai import types as genai_types
|
|
588
|
+
accepted = _google_schema_keys(genai_types)
|
|
589
|
+
return [genai_types.Tool(
|
|
590
|
+
function_declarations=[
|
|
591
|
+
genai_types.FunctionDeclaration(
|
|
592
|
+
name=s["name"],
|
|
593
|
+
description=s["description"],
|
|
594
|
+
parameters=_reduce_for_google(s["input_schema"], accepted),
|
|
595
|
+
)
|
|
596
|
+
for s in tool_schemas
|
|
597
|
+
]
|
|
598
|
+
)]
|
|
599
|
+
|
|
600
|
+
if provider_name == "CHATGPT":
|
|
601
|
+
# ROADMAP_v3 §83. The Responses API's FLAT function shape, and no
|
|
602
|
+
# `strict`. Slice 2a measured it: the registry's fifteen schemas,
|
|
603
|
+
# none of which meets strict mode's rules, completed a call without
|
|
604
|
+
# the key -- and cost fewer input tokens than with `strict: false`.
|
|
605
|
+
# Never a hosted tool (CG3): only what the harness itself runs.
|
|
606
|
+
return [
|
|
607
|
+
{
|
|
608
|
+
"type": "function",
|
|
609
|
+
"name": s["name"],
|
|
610
|
+
"description": s["description"],
|
|
611
|
+
"parameters": s["input_schema"],
|
|
612
|
+
}
|
|
613
|
+
for s in tool_schemas
|
|
614
|
+
]
|
|
615
|
+
|
|
616
|
+
# OpenAI-compatible providers
|
|
617
|
+
return [
|
|
618
|
+
{
|
|
619
|
+
"type": "function",
|
|
620
|
+
"function": {
|
|
621
|
+
"name": s["name"],
|
|
622
|
+
"description": s["description"],
|
|
623
|
+
"parameters": s["input_schema"],
|
|
624
|
+
},
|
|
625
|
+
}
|
|
626
|
+
for s in tool_schemas
|
|
627
|
+
]
|
|
628
|
+
|
|
629
|
+
|
|
630
|
+
# ============================================================================
|
|
631
|
+
# ---- Message history translation -------------------------------------------
|
|
632
|
+
# ============================================================================
|
|
633
|
+
# core/memory.py stores history in a provider-neutral shape (see that
|
|
634
|
+
# file's docstring). This is the counterpart to _tools_for_provider above,
|
|
635
|
+
# applied to conversation history instead of tool schemas -- the ONLY
|
|
636
|
+
# place either provider's specific message wire format should exist.
|
|
637
|
+
|
|
638
|
+
def _is_tool_result_message(msg: dict) -> bool:
|
|
639
|
+
return isinstance(msg.get("content"), list) and all(
|
|
640
|
+
isinstance(b, dict) and b.get("type") == "tool_result" for b in msg["content"]
|
|
641
|
+
)
|
|
642
|
+
|
|
643
|
+
|
|
644
|
+
def _is_empty_assistant_turn(msg: dict) -> bool:
|
|
645
|
+
"""An assistant turn carrying neither text nor tool calls (#13, #36).
|
|
646
|
+
|
|
647
|
+
Produced when the model returns nothing at all -- a safety filter, a
|
|
648
|
+
truncation, an empty pass response. `core/loop.py` persists it because
|
|
649
|
+
D20 requires the assistant turn to be written before any stop-condition
|
|
650
|
+
branching, so it reaches translation on every resume of that thread.
|
|
651
|
+
|
|
652
|
+
It is meaningless on every provider, and each branch used to mangle it
|
|
653
|
+
differently from the same neutral input:
|
|
654
|
+
|
|
655
|
+
OPENAI: {"role": "assistant", "content": null} <- #13
|
|
656
|
+
ANTHROPIC: {"role": "assistant", "content": []} <- #36
|
|
657
|
+
GOOGLE: skipped entirely <- guarded
|
|
658
|
+
|
|
659
|
+
Only Google was guarded, and only because Google was the provider being
|
|
660
|
+
debugged when it surfaced. Dropping it above the branch split is
|
|
661
|
+
AGENTS.md's *fix at the producer, not the consumer*: all three branches
|
|
662
|
+
become correct at once, and so does the next one.
|
|
663
|
+
|
|
664
|
+
Requiring BOTH halves to be empty is load-bearing. An empty text WITH
|
|
665
|
+
tool calls is legitimate on the OpenAI schema -- a tool-call-only turn
|
|
666
|
+
-- and `test_messages_openai_assistant_with_null_text` covers it, so a
|
|
667
|
+
guard keyed on text alone breaks a valid case.
|
|
668
|
+
"""
|
|
669
|
+
return (msg.get("role") == "assistant"
|
|
670
|
+
and not msg.get("text")
|
|
671
|
+
and not msg.get("tool_calls"))
|
|
672
|
+
|
|
673
|
+
|
|
674
|
+
def _echoed_blocks(provider_name: str, model: Optional[str],
|
|
675
|
+
msg: dict) -> list:
|
|
676
|
+
"""This turn's stored thinking blocks, if they may go back (§44).
|
|
677
|
+
|
|
678
|
+
THE GATE IS THE PAIR, and it is the whole safety argument. A stored
|
|
679
|
+
record names the provider and model that produced it; blocks go back
|
|
680
|
+
UNCHANGED to that model and to nothing else. An Anthropic block carries
|
|
681
|
+
a signature the model verifies against its own state, so handing it to
|
|
682
|
+
a different model is at best ignored and at worst a 400 -- and after a
|
|
683
|
+
/model switch the thread is full of blocks the new model never wrote.
|
|
684
|
+
Dropping them is always safe: the harness then does exactly what it did
|
|
685
|
+
before this batch.
|
|
686
|
+
|
|
687
|
+
Empty for every message written before §44, which is what keeps the
|
|
688
|
+
translation byte-identical for every existing thread.
|
|
689
|
+
"""
|
|
690
|
+
record = msg.get("thinking")
|
|
691
|
+
if not isinstance(record, dict) or not model:
|
|
692
|
+
return []
|
|
693
|
+
if record.get("provider") != provider_name or record.get("model") != model:
|
|
694
|
+
return []
|
|
695
|
+
return [b for b in record.get("blocks") or [] if isinstance(b, dict)]
|
|
696
|
+
|
|
697
|
+
|
|
698
|
+
def _chatgpt_reasoning(model: Optional[str], msg: dict) -> list:
|
|
699
|
+
"""The reasoning items a CHATGPT turn goes back with (ROADMAP_v3 §83,
|
|
700
|
+
CG18, CG23): exactly as the SDK parsed them, `None` fields left out,
|
|
701
|
+
and only to the model that wrote them -- §44's pair, through
|
|
702
|
+
`_echoed_blocks`.
|
|
703
|
+
|
|
704
|
+
EVERY turn this model wrote, the ones that ended in calls included
|
|
705
|
+
(CG23, the owner's choice). Slice 2a measured the item accepted under
|
|
706
|
+
`store: false`, with its id and without, before a user message; inside a
|
|
707
|
+
tool loop it was never sent, because the model did not reason before a
|
|
708
|
+
trivial call. The item type is the same and so is the list it sits in,
|
|
709
|
+
and the tool loop is where a reasoning model most needs what it already
|
|
710
|
+
worked out. A refusal would fail that step with a 400 naming the input,
|
|
711
|
+
which is the owner's live check to see."""
|
|
712
|
+
return [b for b in _echoed_blocks("CHATGPT", model, msg)
|
|
713
|
+
if b.get("type") == "reasoning"]
|
|
714
|
+
|
|
715
|
+
|
|
716
|
+
#: ROADMAP_v3 §82 (GS3). Google's stand-in for a thought signature, for a
|
|
717
|
+
#: function call the target model did not write. Gemini 3 refuses a request
|
|
718
|
+
#: whose current turn holds an unsigned call, and a thread that switched
|
|
719
|
+
#: provider or model mid-turn has exactly that. Measured accepted on
|
|
720
|
+
#: gemini-3.1-flash-lite (batch 172); a real signature is always preferred.
|
|
721
|
+
_GOOGLE_FOREIGN_SIGNATURE = b"skip_thought_signature_validator"
|
|
722
|
+
|
|
723
|
+
|
|
724
|
+
def _google_signature_block(call_id: str, signature: Optional[bytes]) -> dict:
|
|
725
|
+
"""One function call's thought signature, as §44's record holds it
|
|
726
|
+
(§82, GS2). One block per call, INCLUDING a call that carried none --
|
|
727
|
+
Gemini 3 signs only the first of a parallel batch, and a model that
|
|
728
|
+
does not think signs nothing -- so "this model wrote the turn and
|
|
729
|
+
attached nothing" stays distinguishable from "nothing is known". Base64
|
|
730
|
+
in the record, because the record is JSON; and no `text` or `thinking`
|
|
731
|
+
key, so replay renders nothing for it."""
|
|
732
|
+
return {"type": "thought_signature", "call_id": call_id,
|
|
733
|
+
"signature": (base64.b64encode(signature).decode("ascii")
|
|
734
|
+
if signature else None)}
|
|
735
|
+
|
|
736
|
+
|
|
737
|
+
def _google_signatures(model: Optional[str], msg: dict) -> Optional[dict]:
|
|
738
|
+
"""{call_id: signature bytes or None} for a turn THIS model wrote, or
|
|
739
|
+
None when it wrote nothing here (§82, GS2/GS3).
|
|
740
|
+
|
|
741
|
+
The gate is §44's pair, read through `_echoed_blocks`: a signature goes
|
|
742
|
+
back unchanged to the model that produced it and to nothing else. A
|
|
743
|
+
stored value that will not decode costs that call its signature and
|
|
744
|
+
gets the placeholder instead, `_decode_thinking`'s rule: an unreadable
|
|
745
|
+
entry costs the entry, never the request."""
|
|
746
|
+
blocks = [b for b in _echoed_blocks("GOOGLE", model, msg)
|
|
747
|
+
if b.get("type") == "thought_signature"]
|
|
748
|
+
if not blocks:
|
|
749
|
+
return None
|
|
750
|
+
signatures: dict = {}
|
|
751
|
+
for block in blocks:
|
|
752
|
+
value = block.get("signature")
|
|
753
|
+
if value is None:
|
|
754
|
+
signatures[block.get("call_id")] = None
|
|
755
|
+
continue
|
|
756
|
+
try:
|
|
757
|
+
signatures[block.get("call_id")] = base64.b64decode(
|
|
758
|
+
value, validate=True)
|
|
759
|
+
except (binascii.Error, TypeError, ValueError):
|
|
760
|
+
signatures[block.get("call_id")] = _GOOGLE_FOREIGN_SIGNATURE
|
|
761
|
+
return signatures
|
|
762
|
+
|
|
763
|
+
|
|
764
|
+
#: ROADMAP_v3 §81 (IM10). The image types GOOGLE is sent: Gemini's documented
|
|
765
|
+
#: list (PNG, JPEG, WEBP, HEIC, HEIF) met with what `view_image` attaches.
|
|
766
|
+
#: GIF is left out, the owner's choice. Two Gemini 3 models read one correctly
|
|
767
|
+
#: when measured, but it is not documented, and an image the endpoint refused
|
|
768
|
+
#: would stay in the thread and fail every request after it.
|
|
769
|
+
_GOOGLE_IMAGE_TYPES = frozenset({"image/png", "image/jpeg", "image/webp"})
|
|
770
|
+
|
|
771
|
+
#: What a tool message says on the two wires that deliver images after the
|
|
772
|
+
#: batch, so the model meets the images knowing whose they are (IM4, IM9).
|
|
773
|
+
_IMAGES_FOLLOW = ("[The attached image(s) follow these tool results, in a "
|
|
774
|
+
"harness message.]")
|
|
775
|
+
|
|
776
|
+
|
|
777
|
+
def _image_note(images: list,
|
|
778
|
+
reason: str = "this provider is not set up to receive images"
|
|
779
|
+
) -> str:
|
|
780
|
+
"""What a tool result's images become on a wire that does not receive
|
|
781
|
+
them (ROADMAP_v3 §81, IM5): a sentence saying so, never the bytes and
|
|
782
|
+
never silence. One rule for every way to get here -- a provider not
|
|
783
|
+
marked `accepts_images`, a thread that switched to one after the image
|
|
784
|
+
was attached, and since batch 173 an image GOOGLE is not sent (IM10). A
|
|
785
|
+
model told nothing would answer from a result that reads as complete,
|
|
786
|
+
which is A32's lesson about MCP's placeholder."""
|
|
787
|
+
kinds = ", ".join(sorted({image["media_type"] for image in images}))
|
|
788
|
+
return (f"[{len(images)} image(s) attached to this result ({kinds}) were "
|
|
789
|
+
f"not sent: {reason}.]")
|
|
790
|
+
|
|
791
|
+
|
|
792
|
+
def _with_image_note(msg: dict) -> str:
|
|
793
|
+
"""A tool row's text, with the IM5 note appended when it carries images
|
|
794
|
+
this wire will not receive."""
|
|
795
|
+
images = msg.get("images")
|
|
796
|
+
if not images:
|
|
797
|
+
return msg["content"]
|
|
798
|
+
return f"{msg['content']}\n{_image_note(images)}"
|
|
799
|
+
|
|
800
|
+
|
|
801
|
+
def _openai_image_turn(pending: list) -> dict:
|
|
802
|
+
"""The user-role message that carries a tool batch's images on the
|
|
803
|
+
OpenAI-compatible wire (IM4).
|
|
804
|
+
|
|
805
|
+
A tool message there holds TEXT PARTS ONLY -- the installed SDK types its
|
|
806
|
+
content as text parts, and the API rejects an image part in it -- so an
|
|
807
|
+
image a tool returned can only travel in a user message. It goes AFTER
|
|
808
|
+
the whole batch: every tool message must directly follow the assistant
|
|
809
|
+
turn that called it, so one placed between two of them would break the
|
|
810
|
+
pairing. Built here and never persisted, like the compaction summary
|
|
811
|
+
(M8): the archive keeps the tool row with its images, and each request
|
|
812
|
+
derives this message from it again. It says it is the harness speaking,
|
|
813
|
+
so the model does not read the images as something the user sent.
|
|
814
|
+
"""
|
|
815
|
+
parts: list = []
|
|
816
|
+
for tool_call_id, images in pending:
|
|
817
|
+
parts.append({"type": "text", "text": (
|
|
818
|
+
f"[Harness: the image(s) attached to the result of tool call "
|
|
819
|
+
f"{tool_call_id}. A tool message on this API carries text only, "
|
|
820
|
+
f"so they are delivered here.]")})
|
|
821
|
+
for image in images:
|
|
822
|
+
parts.append({"type": "image_url", "image_url": {
|
|
823
|
+
"url": f"data:{image['media_type']};base64,{image['data']}"}})
|
|
824
|
+
return {"role": "user", "content": parts}
|
|
825
|
+
|
|
826
|
+
|
|
827
|
+
def _google_tool_images(msg: dict, images: bool) -> tuple:
|
|
828
|
+
"""(result text, [(media_type, bytes)]) for one tool row on GOOGLE
|
|
829
|
+
(ROADMAP_v3 §81, IM9/IM10).
|
|
830
|
+
|
|
831
|
+
The bytes go out DECODED: `Blob.data` is bytes on the pinned type, and
|
|
832
|
+
a base64 string handed to it would be sent as the text of the base64.
|
|
833
|
+
Each image that is not sent says why in the result: the provider is not
|
|
834
|
+
marked (IM5), the type is not one Google is sent (IM10), or the stored
|
|
835
|
+
copy will not decode -- `_decode_thinking`'s rule, so one unreadable
|
|
836
|
+
image costs that image, never the request."""
|
|
837
|
+
attached = msg.get("images") or []
|
|
838
|
+
if not attached:
|
|
839
|
+
return msg["content"], []
|
|
840
|
+
if not images:
|
|
841
|
+
return _with_image_note(msg), []
|
|
842
|
+
sent, refused, unreadable = [], [], []
|
|
843
|
+
for image in attached:
|
|
844
|
+
if image["media_type"] not in _GOOGLE_IMAGE_TYPES:
|
|
845
|
+
refused.append(image)
|
|
846
|
+
continue
|
|
847
|
+
try:
|
|
848
|
+
sent.append((image["media_type"],
|
|
849
|
+
base64.b64decode(image["data"], validate=True)))
|
|
850
|
+
except (binascii.Error, TypeError, ValueError):
|
|
851
|
+
unreadable.append(image)
|
|
852
|
+
text = msg["content"]
|
|
853
|
+
if refused:
|
|
854
|
+
kinds = ", ".join(sorted({image["media_type"] for image in refused}))
|
|
855
|
+
text += "\n" + _image_note(refused, f"this provider is not sent {kinds}")
|
|
856
|
+
if unreadable:
|
|
857
|
+
text += "\n" + _image_note(unreadable,
|
|
858
|
+
"the stored copy could not be read")
|
|
859
|
+
if sent:
|
|
860
|
+
text += "\n" + _IMAGES_FOLLOW
|
|
861
|
+
return text, sent
|
|
862
|
+
|
|
863
|
+
|
|
864
|
+
def _google_image_turn(pending: list, genai_types):
|
|
865
|
+
"""The user-role Content that carries a tool batch's images on GOOGLE
|
|
866
|
+
(ROADMAP_v3 §81, IM9).
|
|
867
|
+
|
|
868
|
+
A FunctionResponse cannot carry them on the models a new account can
|
|
869
|
+
call: Gemini 2.5 answers "multimodal function responses are not
|
|
870
|
+
supported", and only Gemini 3 takes `FunctionResponse.parts`. A harness
|
|
871
|
+
turn after the WHOLE batch can, measured on two Gemini 3 models, and it
|
|
872
|
+
is the shape Google's own ADK uses for an artifact a tool loads. After
|
|
873
|
+
the batch, so the function responses still answer the call turn
|
|
874
|
+
directly; derived per request and never persisted, `_openai_image_turn`'s
|
|
875
|
+
rule; and it says it is the harness speaking."""
|
|
876
|
+
parts: list = []
|
|
877
|
+
for tool_call_id, images in pending:
|
|
878
|
+
parts.append(genai_types.Part(text=(
|
|
879
|
+
f"[Harness: the image(s) attached to the result of tool call "
|
|
880
|
+
f"{tool_call_id}. A function response on this API carries no "
|
|
881
|
+
f"image, so they are delivered here.]")))
|
|
882
|
+
for media_type, data in images:
|
|
883
|
+
parts.append(genai_types.Part(inline_data=genai_types.Blob(
|
|
884
|
+
mime_type=media_type, data=data)))
|
|
885
|
+
return genai_types.Content(role="user", parts=parts)
|
|
886
|
+
|
|
887
|
+
|
|
888
|
+
def _messages_for_provider(provider_name: str, neutral_messages: list[dict],
|
|
889
|
+
model: Optional[str] = None,
|
|
890
|
+
echo_reasoning: bool = False,
|
|
891
|
+
images: bool = False) -> list:
|
|
892
|
+
"""The wire shape for one provider.
|
|
893
|
+
|
|
894
|
+
`model` and `echo_reasoning` are §44's: they decide whether a stored
|
|
895
|
+
thinking record may travel. Both DEFAULT to the pre-§44 behaviour --
|
|
896
|
+
no model means no echo, and the OpenAI-compatible side is opt-in per
|
|
897
|
+
provider -- so a caller that does not pass them gets exactly the
|
|
898
|
+
translation this function produced before. On GOOGLE `model` also
|
|
899
|
+
decides whether a function call's thought signature goes back (§82),
|
|
900
|
+
and the echo is not opt-in there: Gemini 3 refuses a turn without one.
|
|
901
|
+
|
|
902
|
+
`images` is ROADMAP_v3 §81's (IM4): whether a tool row's images are
|
|
903
|
+
delivered to this provider, or replaced by `_image_note`. It defaults
|
|
904
|
+
to the note, so a caller that does not pass it sends no bytes. GOOGLE
|
|
905
|
+
reads it too since batch 173 (IM9); it ignored it while that wire was
|
|
906
|
+
deferred.
|
|
907
|
+
|
|
908
|
+
CHATGPT (ROADMAP_v3 §83) gets the Responses API's `input` items, each
|
|
909
|
+
shape measured on the plan route in slice 2a. Its reasoning goes back
|
|
910
|
+
on §44's pair rule like Anthropic's, and is not opt-in: OpenAI's own
|
|
911
|
+
stateless pattern hands it back (CG18, CG23).
|
|
912
|
+
"""
|
|
913
|
+
# One rule, above the split (#13, #36). Note this can leave two
|
|
914
|
+
# consecutive user turns -- which is the shape Google's branch has
|
|
915
|
+
# always produced here, and which
|
|
916
|
+
# `test_messages_google_empty_assistant_turn_skipped` has always
|
|
917
|
+
# asserted.
|
|
918
|
+
neutral_messages = [m for m in neutral_messages
|
|
919
|
+
if not _is_empty_assistant_turn(m)]
|
|
920
|
+
|
|
921
|
+
if provider_name == "CHATGPT":
|
|
922
|
+
# Slice 2a, item by item: an earlier reply as a plain assistant
|
|
923
|
+
# message, two user turns in a row, and a call WITHOUT its `fc_`
|
|
924
|
+
# item id -- under any call id, another provider's included, which
|
|
925
|
+
# is what a thread holds after a /model switch. So nothing beyond
|
|
926
|
+
# the neutral {id, name, input} has to be kept for a call.
|
|
927
|
+
translated: list = []
|
|
928
|
+
for msg in neutral_messages:
|
|
929
|
+
role = msg["role"]
|
|
930
|
+
if role == "user":
|
|
931
|
+
translated.append({"role": "user", "content": msg["content"]})
|
|
932
|
+
elif role == "assistant":
|
|
933
|
+
# The reasoning FIRST: the order the model emitted it in.
|
|
934
|
+
translated.extend(_chatgpt_reasoning(model, msg))
|
|
935
|
+
if msg.get("text"):
|
|
936
|
+
translated.append(
|
|
937
|
+
{"role": "assistant", "content": msg["text"]})
|
|
938
|
+
for tc in msg.get("tool_calls", []):
|
|
939
|
+
translated.append({
|
|
940
|
+
"type": "function_call", "call_id": tc["id"],
|
|
941
|
+
"name": tc["name"],
|
|
942
|
+
"arguments": json.dumps(tc["input"]),
|
|
943
|
+
})
|
|
944
|
+
elif role == "tool":
|
|
945
|
+
# CG7: no image is delivered on this wire. One swatch was
|
|
946
|
+
# misread in slice 0, and delivering one is a measurement and
|
|
947
|
+
# a translation still to come, so a result's images become
|
|
948
|
+
# IM5's note whatever `accepts_images` says.
|
|
949
|
+
translated.append({
|
|
950
|
+
"type": "function_call_output",
|
|
951
|
+
"call_id": msg["tool_call_id"],
|
|
952
|
+
"output": _with_image_note(msg),
|
|
953
|
+
})
|
|
954
|
+
return translated
|
|
955
|
+
|
|
956
|
+
if provider_name == "GOOGLE":
|
|
957
|
+
from google.genai import types as genai_types
|
|
958
|
+
# Build a {tool_call_id: function_name} lookup from assistant
|
|
959
|
+
# messages so we can populate FunctionResponse.name (required by
|
|
960
|
+
# Google but absent from the neutral tool-result shape).
|
|
961
|
+
call_id_to_name: dict[str, str] = {}
|
|
962
|
+
for msg in neutral_messages:
|
|
963
|
+
if msg["role"] == "assistant":
|
|
964
|
+
for tc in msg.get("tool_calls", []):
|
|
965
|
+
call_id_to_name[tc["id"]] = tc["name"]
|
|
966
|
+
|
|
967
|
+
translated = []
|
|
968
|
+
# §81 (IM9). (tool_call_id, [(media_type, bytes)]) for the tool
|
|
969
|
+
# batch being translated, sent as one harness turn when the batch
|
|
970
|
+
# ends -- see _google_image_turn.
|
|
971
|
+
pending_images: list = []
|
|
972
|
+
for msg in neutral_messages:
|
|
973
|
+
role = msg["role"]
|
|
974
|
+
|
|
975
|
+
if role != "tool" and pending_images:
|
|
976
|
+
translated.append(
|
|
977
|
+
_google_image_turn(pending_images, genai_types))
|
|
978
|
+
pending_images = []
|
|
979
|
+
|
|
980
|
+
if role == "user":
|
|
981
|
+
translated.append(genai_types.Content(
|
|
982
|
+
role="user",
|
|
983
|
+
parts=[genai_types.Part(text=msg["content"])],
|
|
984
|
+
))
|
|
985
|
+
|
|
986
|
+
elif role == "assistant":
|
|
987
|
+
parts = []
|
|
988
|
+
if msg.get("text"):
|
|
989
|
+
parts.append(genai_types.Part(text=msg["text"]))
|
|
990
|
+
# §82 (GS2/GS3). Each call goes back with exactly what this
|
|
991
|
+
# model attached to it. A turn it did not write gets the
|
|
992
|
+
# placeholder on its FIRST call -- the one Gemini 3 checks --
|
|
993
|
+
# and nothing on the rest, which is how it signs its own.
|
|
994
|
+
own = _google_signatures(model, msg)
|
|
995
|
+
for index, tc in enumerate(msg.get("tool_calls", [])):
|
|
996
|
+
fallback = _GOOGLE_FOREIGN_SIGNATURE if index == 0 else None
|
|
997
|
+
parts.append(genai_types.Part(
|
|
998
|
+
function_call=genai_types.FunctionCall(
|
|
999
|
+
name=tc["name"],
|
|
1000
|
+
args=tc["input"],
|
|
1001
|
+
id=tc["id"],
|
|
1002
|
+
),
|
|
1003
|
+
thought_signature=(fallback if own is None
|
|
1004
|
+
else own.get(tc["id"], fallback)),
|
|
1005
|
+
))
|
|
1006
|
+
# Google rejects Content with empty parts. The turn that
|
|
1007
|
+
# produced them is now dropped above the branch split, so
|
|
1008
|
+
# this branch no longer needs its own guard -- one rule
|
|
1009
|
+
# rather than two that can disagree (#36).
|
|
1010
|
+
translated.append(genai_types.Content(role="model", parts=parts))
|
|
1011
|
+
|
|
1012
|
+
elif role == "tool":
|
|
1013
|
+
name = call_id_to_name.get(msg["tool_call_id"])
|
|
1014
|
+
if name is None:
|
|
1015
|
+
raise ValueError(
|
|
1016
|
+
f"Tool result references tool_call_id={msg['tool_call_id']!r} "
|
|
1017
|
+
f"but no assistant message in the history made that call"
|
|
1018
|
+
)
|
|
1019
|
+
# §81 (IM9/IM10). The images wait for the harness turn after
|
|
1020
|
+
# the batch; the result says they follow, or why not.
|
|
1021
|
+
result, sent = _google_tool_images(msg, images)
|
|
1022
|
+
if sent:
|
|
1023
|
+
pending_images.append((msg["tool_call_id"], sent))
|
|
1024
|
+
part = genai_types.Part(
|
|
1025
|
+
function_response=genai_types.FunctionResponse(
|
|
1026
|
+
name=name,
|
|
1027
|
+
response={"result": result},
|
|
1028
|
+
)
|
|
1029
|
+
)
|
|
1030
|
+
# Batch consecutive tool results into a single user
|
|
1031
|
+
# Content — Google requires strict user/model alternation.
|
|
1032
|
+
# The harness image turn comes only after the batch, so it
|
|
1033
|
+
# never sits between two of them.
|
|
1034
|
+
if (translated
|
|
1035
|
+
and translated[-1].role == "user"
|
|
1036
|
+
and all(p.function_response is not None for p in translated[-1].parts)):
|
|
1037
|
+
translated[-1].parts.append(part)
|
|
1038
|
+
else:
|
|
1039
|
+
translated.append(genai_types.Content(role="user", parts=[part]))
|
|
1040
|
+
|
|
1041
|
+
if pending_images:
|
|
1042
|
+
translated.append(_google_image_turn(pending_images, genai_types))
|
|
1043
|
+
return translated
|
|
1044
|
+
|
|
1045
|
+
if provider_name == "ANTHROPIC":
|
|
1046
|
+
translated: list[dict] = []
|
|
1047
|
+
for msg in neutral_messages:
|
|
1048
|
+
role = msg["role"]
|
|
1049
|
+
|
|
1050
|
+
if role == "user":
|
|
1051
|
+
translated.append({"role": "user", "content": msg["content"]})
|
|
1052
|
+
|
|
1053
|
+
elif role == "assistant":
|
|
1054
|
+
# §44. FIRST in the content array, because that is the
|
|
1055
|
+
# order the model emitted them in and the order the API
|
|
1056
|
+
# documents for the replay -- a signed block after the
|
|
1057
|
+
# text it preceded is not the same message.
|
|
1058
|
+
content = list(_echoed_blocks(provider_name, model, msg))
|
|
1059
|
+
if msg.get("text"):
|
|
1060
|
+
content.append({"type": "text", "text": msg["text"]})
|
|
1061
|
+
for tc in msg.get("tool_calls", []):
|
|
1062
|
+
content.append({
|
|
1063
|
+
"type": "tool_use", "id": tc["id"], "name": tc["name"], "input": tc["input"],
|
|
1064
|
+
})
|
|
1065
|
+
translated.append({"role": "assistant", "content": content})
|
|
1066
|
+
|
|
1067
|
+
elif role == "tool":
|
|
1068
|
+
# Anthropic wants every tool result from one turn batched
|
|
1069
|
+
# into a SINGLE user message with multiple tool_result
|
|
1070
|
+
# blocks, not one message per result -- merge consecutive
|
|
1071
|
+
# tool-role entries into the same message rather than
|
|
1072
|
+
# appending a new one each time.
|
|
1073
|
+
#
|
|
1074
|
+
# §81 (IM4). An image travels INSIDE its tool_result here --
|
|
1075
|
+
# the one wire of the three whose tool result takes an image
|
|
1076
|
+
# block -- after the text, as a content array. The text
|
|
1077
|
+
# block is omitted when empty, because the API rejects an
|
|
1078
|
+
# empty text block.
|
|
1079
|
+
content = msg["content"]
|
|
1080
|
+
if msg.get("images"):
|
|
1081
|
+
if images:
|
|
1082
|
+
content = (
|
|
1083
|
+
[{"type": "text", "text": msg["content"]}]
|
|
1084
|
+
if msg["content"] else [])
|
|
1085
|
+
content += [{
|
|
1086
|
+
"type": "image",
|
|
1087
|
+
"source": {"type": "base64",
|
|
1088
|
+
"media_type": image["media_type"],
|
|
1089
|
+
"data": image["data"]},
|
|
1090
|
+
} for image in msg["images"]]
|
|
1091
|
+
else:
|
|
1092
|
+
content = _with_image_note(msg)
|
|
1093
|
+
block = {
|
|
1094
|
+
"type": "tool_result",
|
|
1095
|
+
"tool_use_id": msg["tool_call_id"],
|
|
1096
|
+
"content": content,
|
|
1097
|
+
}
|
|
1098
|
+
if translated and translated[-1]["role"] == "user" and _is_tool_result_message(translated[-1]):
|
|
1099
|
+
translated[-1]["content"].append(block)
|
|
1100
|
+
else:
|
|
1101
|
+
translated.append({"role": "user", "content": [block]})
|
|
1102
|
+
|
|
1103
|
+
return translated
|
|
1104
|
+
|
|
1105
|
+
# OpenAI-compatible: close to 1:1, each tool result is its own message
|
|
1106
|
+
translated = []
|
|
1107
|
+
# §81 (IM4). (tool_call_id, images) for the tool batch being translated,
|
|
1108
|
+
# sent as one user message when the batch ends -- see _openai_image_turn.
|
|
1109
|
+
pending_images: list = []
|
|
1110
|
+
for msg in neutral_messages:
|
|
1111
|
+
role = msg["role"]
|
|
1112
|
+
|
|
1113
|
+
if role != "tool" and pending_images:
|
|
1114
|
+
translated.append(_openai_image_turn(pending_images))
|
|
1115
|
+
pending_images = []
|
|
1116
|
+
|
|
1117
|
+
if role == "user":
|
|
1118
|
+
translated.append({"role": "user", "content": msg["content"]})
|
|
1119
|
+
|
|
1120
|
+
elif role == "assistant":
|
|
1121
|
+
entry: dict = {"role": "assistant", "content": msg.get("text") or None}
|
|
1122
|
+
# §44, and OPT-IN PER PROVIDER rather than on by default.
|
|
1123
|
+
# There is no standard for sending reasoning back on this wire
|
|
1124
|
+
# the way there is for receiving it: DeepSeek documents that
|
|
1125
|
+
# reasoning_content must not be supplied on input and 400s,
|
|
1126
|
+
# while OpenRouter accepts it for some models and ignores it
|
|
1127
|
+
# for others. Off means today's behaviour exactly, so a
|
|
1128
|
+
# provider that has never been tried cannot be broken by this.
|
|
1129
|
+
if echo_reasoning:
|
|
1130
|
+
echoed = "\n\n".join(
|
|
1131
|
+
b.get("text", "") for b in _echoed_blocks(
|
|
1132
|
+
provider_name, model, msg)
|
|
1133
|
+
if b.get("text"))
|
|
1134
|
+
if echoed:
|
|
1135
|
+
entry["reasoning_content"] = echoed
|
|
1136
|
+
if msg.get("tool_calls"):
|
|
1137
|
+
entry["tool_calls"] = [
|
|
1138
|
+
{
|
|
1139
|
+
"id": tc["id"],
|
|
1140
|
+
"type": "function",
|
|
1141
|
+
"function": {"name": tc["name"], "arguments": json.dumps(tc["input"])},
|
|
1142
|
+
}
|
|
1143
|
+
for tc in msg["tool_calls"]
|
|
1144
|
+
]
|
|
1145
|
+
translated.append(entry)
|
|
1146
|
+
|
|
1147
|
+
elif role == "tool":
|
|
1148
|
+
content = msg["content"]
|
|
1149
|
+
if msg.get("images"):
|
|
1150
|
+
if images:
|
|
1151
|
+
pending_images.append(
|
|
1152
|
+
(msg["tool_call_id"], msg["images"]))
|
|
1153
|
+
content = f"{content}\n{_IMAGES_FOLLOW}"
|
|
1154
|
+
else:
|
|
1155
|
+
content = _with_image_note(msg)
|
|
1156
|
+
translated.append({
|
|
1157
|
+
"role": "tool", "tool_call_id": msg["tool_call_id"], "content": content,
|
|
1158
|
+
})
|
|
1159
|
+
|
|
1160
|
+
if pending_images:
|
|
1161
|
+
translated.append(_openai_image_turn(pending_images))
|
|
1162
|
+
return translated
|
|
1163
|
+
|
|
1164
|
+
|
|
1165
|
+
# ============================================================================
|
|
1166
|
+
# ---- The actual call --------------------------------------------------------
|
|
1167
|
+
# ============================================================================
|
|
1168
|
+
|
|
1169
|
+
# ============================================================================
|
|
1170
|
+
# ---- Sampling + reasoning-effort translation (ROADMAP_v2 §16) ---------------
|
|
1171
|
+
# ============================================================================
|
|
1172
|
+
#
|
|
1173
|
+
# Both helpers below are the same "translate at the boundary, once" case that
|
|
1174
|
+
# _tools_for_provider / _messages_for_provider exist for: the three providers
|
|
1175
|
+
# express these concepts differently enough that no caller should have to know.
|
|
1176
|
+
|
|
1177
|
+
# (provider_name, model) -> supported effort levels. Populated lazily; the
|
|
1178
|
+
# Anthropic branch queries the API once per process rather than per call.
|
|
1179
|
+
_effort_levels_cache: dict[tuple[str, str], list[str]] = {}
|
|
1180
|
+
|
|
1181
|
+
# (provider_name, model) -> context window in tokens, or None for "asked,
|
|
1182
|
+
# and this provider does not report one". Same lazy, once-per-process
|
|
1183
|
+
# posture as the effort cache above; see context_window_for() for why None
|
|
1184
|
+
# is cached and a raised query is not.
|
|
1185
|
+
_context_window_cache: dict[tuple[str, str], Optional[int]] = {}
|
|
1186
|
+
|
|
1187
|
+
# (provider_name, model) -> how many times the query has RAISED.
|
|
1188
|
+
#
|
|
1189
|
+
# The middle ground between #35's two bad options. Caching a failure makes
|
|
1190
|
+
# one transient error suppress the query for the life of the process;
|
|
1191
|
+
# never caching one means a provider with no /models endpoint at all --
|
|
1192
|
+
# Grok, Fireworks, Z.AI -- is asked again on every turn, which across a
|
|
1193
|
+
# ten-pass research run is ten failing HTTP calls to learn the same thing.
|
|
1194
|
+
# So a failure is retried, and after this many it settles on the table and
|
|
1195
|
+
# says so once. Transient blips still recover; a permanent 404 costs three
|
|
1196
|
+
# calls per process, not one per turn.
|
|
1197
|
+
_context_window_failures: dict[tuple[str, str], int] = {}
|
|
1198
|
+
MAX_CONTEXT_WINDOW_QUERY_ATTEMPTS = 3
|
|
1199
|
+
|
|
1200
|
+
#: ROADMAP_v3 §83 (CG22). The plan route's whole effort vocabulary, read off
|
|
1201
|
+
#: slice 0's `invalid_value` refusal of `ultra`. The model list names a level
|
|
1202
|
+
#: the route refuses (`ultra`), so a level reaches /effort only when the list
|
|
1203
|
+
#: names it AND this does. Slice 2a found the rule exact on both models whose
|
|
1204
|
+
#: refusal lists their levels.
|
|
1205
|
+
CHATGPT_EFFORT_VOCABULARY = ("none", "minimal", "low", "medium", "high",
|
|
1206
|
+
"xhigh", "max")
|
|
1207
|
+
#: CG24, the owner's choice. Six models accepted `none` in slice 2a and two
|
|
1208
|
+
#: refused it; no list entry names it, and a table of the six would go stale
|
|
1209
|
+
#: with a roster that changed between two runs a day apart.
|
|
1210
|
+
CHATGPT_NEVER_OFFERED = frozenset({"none"})
|
|
1211
|
+
|
|
1212
|
+
|
|
1213
|
+
def _chatgpt_listing(client) -> list:
|
|
1214
|
+
"""The plan's model list, every entry (ROADMAP_v3 §83).
|
|
1215
|
+
|
|
1216
|
+
`client.get("/models", cast_to=object)`, because the plan's list is
|
|
1217
|
+
`{"models": [...]}` keyed by `slug`, which the SDK's `models.list()` --
|
|
1218
|
+
built for `{"data": [...]}` -- cannot read; slice 2a read it this way.
|
|
1219
|
+
Raises when the list cannot be read, which each caller counts as a
|
|
1220
|
+
failed query.
|
|
1221
|
+
"""
|
|
1222
|
+
body = client.get("/models", cast_to=object)
|
|
1223
|
+
entries = body.get("models") if isinstance(body, dict) else None
|
|
1224
|
+
if not isinstance(entries, list):
|
|
1225
|
+
raise ValueError(
|
|
1226
|
+
"the ChatGPT plan's model list carried no `models` array")
|
|
1227
|
+
return [entry for entry in entries if isinstance(entry, dict)]
|
|
1228
|
+
|
|
1229
|
+
|
|
1230
|
+
def _chatgpt_entry(entries: list, model: str) -> Optional[dict]:
|
|
1231
|
+
"""`model`'s entry, or None when the list does not name it. Any entry
|
|
1232
|
+
counts, a `visibility: hide` one included: a hidden model works when
|
|
1233
|
+
named, and its levels and window are its own."""
|
|
1234
|
+
return next((e for e in entries if e.get("slug") == model), None)
|
|
1235
|
+
|
|
1236
|
+
|
|
1237
|
+
def _chatgpt_unlisted(model: str, entries: list) -> None:
|
|
1238
|
+
"""Say that the plan does not list `model`, and what it does list.
|
|
1239
|
+
|
|
1240
|
+
Batch 189, the owner's live check: a typo (`gpt-6-luba`) fell through to
|
|
1241
|
+
the static table's warning, which advised adding a config entry and
|
|
1242
|
+
named `ultra` -- a level the CHATGPT filter removes after that warning
|
|
1243
|
+
is written. On the plan an unlisted name is likeliest misspelt, and the
|
|
1244
|
+
list that would show it was already in hand.
|
|
1245
|
+
"""
|
|
1246
|
+
listed = [e.get("slug") for e in entries
|
|
1247
|
+
if e.get("visibility") == "list" and e.get("slug")]
|
|
1248
|
+
logger.warning(
|
|
1249
|
+
"%r is not in your ChatGPT plan's model list, so the next turn will "
|
|
1250
|
+
"probably be refused. The plan lists: %s.", model,
|
|
1251
|
+
", ".join(listed) or "no models")
|
|
1252
|
+
|
|
1253
|
+
|
|
1254
|
+
def _sampling_kwargs(provider_name: str, model: str, temperature,
|
|
1255
|
+
effort: Optional[str] = None) -> dict:
|
|
1256
|
+
"""Sampling parameters for this call, or {} when the model rejects them.
|
|
1257
|
+
|
|
1258
|
+
Current Anthropic models removed temperature/top_p/top_k outright (any
|
|
1259
|
+
value returns a 400) and Sonnet 5 rejects non-default values. Sending
|
|
1260
|
+
one anyway is not a degraded request, it is a failed one -- which is how
|
|
1261
|
+
ROADMAP §10's ensemble mode came to be built, documented as working, and
|
|
1262
|
+
incapable of running against this harness's own default model.
|
|
1263
|
+
|
|
1264
|
+
TWO MORE, MEASURED FROM THE PROVIDERS' OWN DOCUMENTATION (ROADMAP_v3
|
|
1265
|
+
§87, #6). OpenAI's rule for its reasoning models is "When reasoning
|
|
1266
|
+
effort is not `none`, remove `temperature`, `top_p`, and
|
|
1267
|
+
`top_logprobs`", and a model that reasons refuses one with a 400 -- so
|
|
1268
|
+
`effort` decides it on OPENAI, the value `_thinking_for_provider` is
|
|
1269
|
+
about to send, with no model list to go stale. And Google's Gemini 3.6
|
|
1270
|
+
Flash and 3.5 Flash-Lite IGNORE a temperature, as Google says every
|
|
1271
|
+
later Gemini will: no error, no effect, which is the one failure here
|
|
1272
|
+
worse than a 400. Those are named in the set below like the rest.
|
|
1273
|
+
|
|
1274
|
+
ON ANTHROPIC, EVERY MODEL, SINCE BATCH 234 (A82, amended). anthropic 2.x
|
|
1275
|
+
sends API version 2026-10-21, which refuses `temperature`, `top_p` and
|
|
1276
|
+
`top_k` with a 400 on every model, even at the default; and 1.0 took
|
|
1277
|
+
them out of `messages.stream()`, so passing one is a TypeError before
|
|
1278
|
+
anything is sent (measured). The set's Claude names stay for Claude
|
|
1279
|
+
reached through an OpenAI-compatible gateway, where they are still the
|
|
1280
|
+
only thing that can drop it.
|
|
1281
|
+
|
|
1282
|
+
Dropping the parameter is logged at WARNING rather than done silently.
|
|
1283
|
+
|
|
1284
|
+
THIS FUNCTION NOW HAS NO CALLER THAT DEPENDS ON IT, and the warning is
|
|
1285
|
+
what makes that safe. Ensemble mode used to be the one, and
|
|
1286
|
+
orchestrator.py carried a guard refusing to run it on a model in the set
|
|
1287
|
+
below. §10's revisit replaced sampling variation with a roster of
|
|
1288
|
+
different models, so the guard is gone and nothing in the harness passes
|
|
1289
|
+
a temperature any more.
|
|
1290
|
+
|
|
1291
|
+
Kept rather than removed: `temperature` still threads through
|
|
1292
|
+
core/loop.py, this is still the correct boundary translation for it, and
|
|
1293
|
+
the next caller to want sampling variation needs to be told rather than
|
|
1294
|
+
silently ignored. If this warning ever fires, something new is relying on
|
|
1295
|
+
sampling variation without knowing it cannot have it -- which is exactly
|
|
1296
|
+
how §10 came to be built, documented as working, and incapable of running
|
|
1297
|
+
against this harness's own default model.
|
|
1298
|
+
"""
|
|
1299
|
+
if temperature is None:
|
|
1300
|
+
return {}
|
|
1301
|
+
if provider_name == "CHATGPT":
|
|
1302
|
+
# ROADMAP_v3 §83: the plan route answers a temperature with 400
|
|
1303
|
+
# "Unsupported parameter" (slice 0), on every model.
|
|
1304
|
+
logger.warning(
|
|
1305
|
+
"Dropping temperature=%s: the ChatGPT plan's route refuses "
|
|
1306
|
+
"sampling parameters.", temperature)
|
|
1307
|
+
return {}
|
|
1308
|
+
if provider_name == "ANTHROPIC":
|
|
1309
|
+
logger.warning(
|
|
1310
|
+
"Dropping temperature=%s: Anthropic's API version 2026-10-21 "
|
|
1311
|
+
"refuses sampling parameters on every model.", temperature)
|
|
1312
|
+
return {}
|
|
1313
|
+
if provider_name == "OPENAI" and effort and effort != "none":
|
|
1314
|
+
# §87. Sent no effort at all, a reasoning model still reasons at
|
|
1315
|
+
# its own default and still refuses: that stays the loud 400 an
|
|
1316
|
+
# unlisted model has always been, rather than a guess here.
|
|
1317
|
+
logger.warning(
|
|
1318
|
+
"Dropping temperature=%s: OpenAI refuses sampling parameters "
|
|
1319
|
+
"while a model reasons (reasoning effort %r).", temperature,
|
|
1320
|
+
effort)
|
|
1321
|
+
return {}
|
|
1322
|
+
if model in config.MODELS_REJECTING_SAMPLING_PARAMS:
|
|
1323
|
+
logger.warning(
|
|
1324
|
+
"Dropping temperature=%s: model %r rejects sampling parameters. "
|
|
1325
|
+
"Steer via prompting or reasoning effort instead.",
|
|
1326
|
+
temperature, model,
|
|
1327
|
+
)
|
|
1328
|
+
return {}
|
|
1329
|
+
return {"temperature": temperature}
|
|
1330
|
+
|
|
1331
|
+
|
|
1332
|
+
def effort_for(client, provider_name: str, model: str,
|
|
1333
|
+
requested: Optional[str]) -> Optional[str]:
|
|
1334
|
+
"""The effort level to actually SEND to (provider_name, model), or None.
|
|
1335
|
+
|
|
1336
|
+
One validator for every send site, because nothing else checked a
|
|
1337
|
+
level against the model that would receive it. Three ways a wrong one
|
|
1338
|
+
used to reach the wire, all of them 400ing every turn until the user
|
|
1339
|
+
intervened by hand:
|
|
1340
|
+
|
|
1341
|
+
* a persisted settings.json `tui.effort` was trusted as-is;
|
|
1342
|
+
* the TUI's current level was inherited by an agent that declares
|
|
1343
|
+
its OWN model/provider, so a level validated for one model was
|
|
1344
|
+
sent to a different one;
|
|
1345
|
+
* nothing revalidated after a /model or agent switch.
|
|
1346
|
+
|
|
1347
|
+
`effort=None` sends nothing on every provider and is the one
|
|
1348
|
+
universally safe value, so dropping is always available as an answer.
|
|
1349
|
+
|
|
1350
|
+
On a FAILED lookup the requested level is passed through unverified
|
|
1351
|
+
with a warning rather than dropped: a transient network blip must not
|
|
1352
|
+
silently downgrade a deliberate choice. A genuinely wrong level still
|
|
1353
|
+
fails at the provider, which is the honest outcome.
|
|
1354
|
+
"""
|
|
1355
|
+
if not requested:
|
|
1356
|
+
return None
|
|
1357
|
+
levels, authoritative = _effort_levels(client, provider_name, model)
|
|
1358
|
+
if requested in levels:
|
|
1359
|
+
return requested
|
|
1360
|
+
if not authoritative:
|
|
1361
|
+
logger.warning(
|
|
1362
|
+
"Could not verify effort %r for %s/%s; sending it unverified.",
|
|
1363
|
+
requested, provider_name, model)
|
|
1364
|
+
return requested
|
|
1365
|
+
logger.warning(
|
|
1366
|
+
"Dropping effort %r: %s/%s accepts %s. Sending no effort parameter.",
|
|
1367
|
+
requested, provider_name, model, levels or "no effort control")
|
|
1368
|
+
return None
|
|
1369
|
+
|
|
1370
|
+
|
|
1371
|
+
def effort_levels_for_model(client, provider_name: str, model: str) -> list[str]:
|
|
1372
|
+
"""Reasoning-effort levels this model accepts, most-to-least capable last.
|
|
1373
|
+
|
|
1374
|
+
Anthropic is queried, not tabulated: its Models API reports per-model
|
|
1375
|
+
effort support (`capabilities.effort.<level>.supported`), so a newly
|
|
1376
|
+
released Anthropic model needs no manual entry here. Every other provider
|
|
1377
|
+
is tabulated because there is nothing to ask -- no OpenAI-compatible
|
|
1378
|
+
provider in providers.json exposes a capability endpoint, and Google
|
|
1379
|
+
expresses depth as an integer budget rather than a level set at all.
|
|
1380
|
+
|
|
1381
|
+
Falling back logs at WARNING, matching §21's MODEL_CONTEXT_WINDOWS
|
|
1382
|
+
posture: a static table is honest about being incomplete, but it should
|
|
1383
|
+
say so out loud when it is the thing answering.
|
|
1384
|
+
|
|
1385
|
+
Returns [] when the model supports no effort control, which callers must
|
|
1386
|
+
treat as "offer no picker and send no effort parameter".
|
|
1387
|
+
"""
|
|
1388
|
+
return _effort_levels(client, provider_name, model)[0]
|
|
1389
|
+
|
|
1390
|
+
|
|
1391
|
+
def effort_choice(levels, requested: str, current: Optional[str],
|
|
1392
|
+
model: str) -> tuple:
|
|
1393
|
+
"""`/effort`'s decision in either shell: `(level, sentence, refused)`.
|
|
1394
|
+
|
|
1395
|
+
`levels` is what `effort_levels_for_model` answered for the session's
|
|
1396
|
+
model, `requested` the command's argument and `current` the level in
|
|
1397
|
+
force. The level comes back unchanged for a report or a refusal, None
|
|
1398
|
+
for `auto`, and the requested one when the model accepts it.
|
|
1399
|
+
|
|
1400
|
+
ONE COPY (ROADMAP_v3 §89): the TUI answers on a worker's message and
|
|
1401
|
+
the CLI inline, and what each may accept is the same decision. Moved
|
|
1402
|
+
out of the TUI's `_apply_effort_command`.
|
|
1403
|
+
"""
|
|
1404
|
+
if not levels:
|
|
1405
|
+
return current, f"{model} exposes no reasoning-effort control.", False
|
|
1406
|
+
available = ", ".join(levels)
|
|
1407
|
+
if not requested:
|
|
1408
|
+
return (current,
|
|
1409
|
+
f"Effort: {current or 'auto'}. Available: {available} "
|
|
1410
|
+
f"(or 'auto').", False)
|
|
1411
|
+
if requested == "auto":
|
|
1412
|
+
return None, "Effort set to auto.", False
|
|
1413
|
+
if requested in levels:
|
|
1414
|
+
return requested, f"Effort set to {requested}.", False
|
|
1415
|
+
return (current,
|
|
1416
|
+
f"{model} does not support effort {requested!r}. "
|
|
1417
|
+
f"Available: {available}.", True)
|
|
1418
|
+
|
|
1419
|
+
|
|
1420
|
+
def _effort_levels(client, provider_name: str, model: str) -> tuple:
|
|
1421
|
+
"""(levels, authoritative).
|
|
1422
|
+
|
|
1423
|
+
`authoritative` is False when the answer came from the static table
|
|
1424
|
+
after the capability QUERY failed, as opposed to because no query was
|
|
1425
|
+
ever possible for that provider. Only effort_for() needs the
|
|
1426
|
+
distinction -- an unverified answer must not be used to drop a level
|
|
1427
|
+
the user deliberately chose.
|
|
1428
|
+
"""
|
|
1429
|
+
key = (provider_name, model)
|
|
1430
|
+
if key in _effort_levels_cache:
|
|
1431
|
+
return _effort_levels_cache[key], True
|
|
1432
|
+
|
|
1433
|
+
levels: Optional[list[str]] = None
|
|
1434
|
+
query_failed = False
|
|
1435
|
+
if provider_name == "GOOGLE" and not _google_supports_thinking_budget():
|
|
1436
|
+
# Reporting no levels is the honest answer, not a silent [] : on the
|
|
1437
|
+
# pinned SDK there is no field to put a budget in, so a picker
|
|
1438
|
+
# offering levels would be offering something unsendable.
|
|
1439
|
+
_effort_levels_cache[key] = []
|
|
1440
|
+
return [], True
|
|
1441
|
+
if provider_name == "ANTHROPIC":
|
|
1442
|
+
try:
|
|
1443
|
+
# ATTRIBUTES, not subscripts (#35). This read used to be
|
|
1444
|
+
# `capabilities["effort"]` with `effort.get(name)` and
|
|
1445
|
+
# `isinstance(..., dict)`. On the exactly-pinned SDK those are
|
|
1446
|
+
# pydantic models -- ModelCapabilities, EffortCapability,
|
|
1447
|
+
# CapabilitySupport -- supporting neither [] nor .get(), so the
|
|
1448
|
+
# TypeError was caught below, query_failed was set, and the
|
|
1449
|
+
# query FELL BACK EVERY SINGLE TIME. Not "failed sometimes":
|
|
1450
|
+
# it could not succeed. All three dict spellings were wrong,
|
|
1451
|
+
# so correcting the subscript alone would have failed one line
|
|
1452
|
+
# further down.
|
|
1453
|
+
#
|
|
1454
|
+
# No dict/attribute dual path on purpose. A defensive default
|
|
1455
|
+
# is only safe where absence is impossible, requirements.txt
|
|
1456
|
+
# pins anthropic exactly, and test_sdk_conformance.py exists
|
|
1457
|
+
# to go red when the pinned shape moves -- which is a better
|
|
1458
|
+
# signal than a fallback that silently keeps guessing.
|
|
1459
|
+
info = client.models.retrieve(model)
|
|
1460
|
+
# ONE retrieve, TWO readings (§21's revisit). The context
|
|
1461
|
+
# window is on this SAME ModelInfo as `max_input_tokens`, so
|
|
1462
|
+
# recording it here IS the entire Anthropic half of the
|
|
1463
|
+
# window query: no second call, no second failure mode, and
|
|
1464
|
+
# it lands on the path every send already crosses
|
|
1465
|
+
# (core/loop.py calls effort_for before the wire). §21
|
|
1466
|
+
# rejected a window query on the belief that there was
|
|
1467
|
+
# nothing to ask; for this provider the answer was already
|
|
1468
|
+
# in hand and being dropped. See context_window_for below.
|
|
1469
|
+
_record_window(provider_name, model,
|
|
1470
|
+
getattr(info, "max_input_tokens", None))
|
|
1471
|
+
capabilities = info.capabilities
|
|
1472
|
+
effort = capabilities.effort
|
|
1473
|
+
levels = [
|
|
1474
|
+
name for name in ("low", "medium", "high", "xhigh", "max")
|
|
1475
|
+
# getattr, because `xhigh` is Optional on the pinned type
|
|
1476
|
+
# and a model that does not offer a level reports None
|
|
1477
|
+
# rather than supported=False.
|
|
1478
|
+
if getattr(getattr(effort, name, None), "supported", False)
|
|
1479
|
+
]
|
|
1480
|
+
except Exception as e:
|
|
1481
|
+
# Network failure, an SDK too old to expose capabilities, or a
|
|
1482
|
+
# model the endpoint doesn't know. Fall through to the table.
|
|
1483
|
+
query_failed = True
|
|
1484
|
+
logger.warning(
|
|
1485
|
+
"Could not read effort capabilities for %r from the Anthropic "
|
|
1486
|
+
"Models API (%s); falling back to the static table.", model, e,
|
|
1487
|
+
)
|
|
1488
|
+
if provider_name == "CHATGPT":
|
|
1489
|
+
try:
|
|
1490
|
+
entries = _chatgpt_listing(client)
|
|
1491
|
+
except Exception as e:
|
|
1492
|
+
query_failed = True
|
|
1493
|
+
logger.warning(
|
|
1494
|
+
"Could not read effort levels for %r from the ChatGPT plan's "
|
|
1495
|
+
"model list (%s); falling back to the static table.", model, e)
|
|
1496
|
+
else:
|
|
1497
|
+
entry = _chatgpt_entry(entries, model)
|
|
1498
|
+
if entry is not None:
|
|
1499
|
+
# ONE read, TWO readings, as Anthropic's retrieve: the entry
|
|
1500
|
+
# carries the window too.
|
|
1501
|
+
_record_window(provider_name, model,
|
|
1502
|
+
entry.get("context_window"))
|
|
1503
|
+
levels = [lv.get("effort") if isinstance(lv, dict) else lv
|
|
1504
|
+
for lv in entry.get("supported_reasoning_levels")
|
|
1505
|
+
or []]
|
|
1506
|
+
else:
|
|
1507
|
+
# The list answered and does not name the model: said in
|
|
1508
|
+
# the plan's own terms, and the default set is filtered
|
|
1509
|
+
# below like any other, without the static table's advice.
|
|
1510
|
+
_chatgpt_unlisted(model, entries)
|
|
1511
|
+
levels = list(config.DEFAULT_EFFORT_LEVELS)
|
|
1512
|
+
|
|
1513
|
+
if levels is None:
|
|
1514
|
+
levels = config.MODEL_EFFORT_LEVELS.get(model)
|
|
1515
|
+
if levels is None:
|
|
1516
|
+
logger.warning(
|
|
1517
|
+
"No effort levels known for %r on %s; assuming %s. Add an entry "
|
|
1518
|
+
"to config.MODEL_EFFORT_LEVELS if this is wrong.",
|
|
1519
|
+
model, provider_name, config.DEFAULT_EFFORT_LEVELS,
|
|
1520
|
+
)
|
|
1521
|
+
levels = list(config.DEFAULT_EFFORT_LEVELS)
|
|
1522
|
+
|
|
1523
|
+
if provider_name == "CHATGPT":
|
|
1524
|
+
# CG22 and CG24, applied to the static table's answer as well: its
|
|
1525
|
+
# default set ends in `ultra`, which the route refuses.
|
|
1526
|
+
levels = [lv for lv in levels if lv in CHATGPT_EFFORT_VOCABULARY
|
|
1527
|
+
and lv not in CHATGPT_NEVER_OFFERED]
|
|
1528
|
+
|
|
1529
|
+
if query_failed:
|
|
1530
|
+
# Do NOT cache a failure-derived guess. Caching it made one
|
|
1531
|
+
# transient error permanently suppress the query that is this
|
|
1532
|
+
# function's entire reason to exist: for the whole process life
|
|
1533
|
+
# the picker would offer a guessed level set, and every selection
|
|
1534
|
+
# from it could 400 with no path back short of a restart.
|
|
1535
|
+
return levels, False
|
|
1536
|
+
|
|
1537
|
+
_effort_levels_cache[key] = levels
|
|
1538
|
+
return levels, True
|
|
1539
|
+
|
|
1540
|
+
|
|
1541
|
+
# The field names a provider might report a context window under, in the
|
|
1542
|
+
# order they are tried. ONE reader covers every OpenAI-compatible provider
|
|
1543
|
+
# rather than one adapter each, because openai.types.Model is declared
|
|
1544
|
+
# `extra="allow"` -- a field the SDK does not model survives on the parsed
|
|
1545
|
+
# object and reaches `.model_extra`. Verified against the pinned openai
|
|
1546
|
+
# (2.45.0 when written, 3.36.0 since batch 234) in test_sdk_conformance.py,
|
|
1547
|
+
# because that permissiveness is the whole reason this is one function.
|
|
1548
|
+
#
|
|
1549
|
+
# context_window Groq
|
|
1550
|
+
# context_length Together, OpenRouter, Cohere (native /v1 only)
|
|
1551
|
+
# max_context_length Mistral
|
|
1552
|
+
# max_input_tokens Anthropic's spelling, accepted here so a gateway
|
|
1553
|
+
# proxying Anthropic through an OpenAI-shaped API
|
|
1554
|
+
# is not silently unreadable
|
|
1555
|
+
#
|
|
1556
|
+
# OpenAI, DeepSeek and Perplexity report NONE of these -- they answer with
|
|
1557
|
+
# id/created/object/owned_by and nothing else. That is not a failure, it is
|
|
1558
|
+
# the answer, and it is why config.MODEL_CONTEXT_WINDOWS still exists.
|
|
1559
|
+
_WINDOW_FIELD_ALIASES = (
|
|
1560
|
+
"context_window",
|
|
1561
|
+
"context_length",
|
|
1562
|
+
"max_context_length",
|
|
1563
|
+
"max_input_tokens",
|
|
1564
|
+
)
|
|
1565
|
+
|
|
1566
|
+
#: Returned by _model_info when a COMPLETE listing did not contain the model.
|
|
1567
|
+
#: Distinct from a raise: the provider answered, so there is nothing to retry.
|
|
1568
|
+
_NOT_LISTED = object()
|
|
1569
|
+
|
|
1570
|
+
|
|
1571
|
+
def _window_from_model_object(info) -> Optional[int]:
|
|
1572
|
+
"""The context window off one provider model object, or None.
|
|
1573
|
+
|
|
1574
|
+
Declared fields first, then model_extra -- the same object may carry
|
|
1575
|
+
either depending on whether the installed SDK models the field.
|
|
1576
|
+
ATTRIBUTES not subscripts, for #35's reason: these are pydantic objects
|
|
1577
|
+
on every pinned SDK here.
|
|
1578
|
+
|
|
1579
|
+
`top_provider.context_length` is the one NESTED spelling, and it is
|
|
1580
|
+
OpenRouter's: the flat `context_length` is the model's maximum across
|
|
1581
|
+
every backend, while top_provider carries the routed one. Read second,
|
|
1582
|
+
so the flat value still wins where both exist -- this is a fallback for
|
|
1583
|
+
entries that carry only the nested form, not a preference for the
|
|
1584
|
+
narrower number.
|
|
1585
|
+
"""
|
|
1586
|
+
extra = getattr(info, "model_extra", None) or {}
|
|
1587
|
+
for name in _WINDOW_FIELD_ALIASES:
|
|
1588
|
+
value = getattr(info, name, None)
|
|
1589
|
+
if value is None:
|
|
1590
|
+
value = extra.get(name)
|
|
1591
|
+
if value is not None:
|
|
1592
|
+
return value
|
|
1593
|
+
top = extra.get("top_provider")
|
|
1594
|
+
if isinstance(top, dict):
|
|
1595
|
+
return top.get("context_length")
|
|
1596
|
+
return None
|
|
1597
|
+
|
|
1598
|
+
|
|
1599
|
+
def _model_info(client, model):
|
|
1600
|
+
"""The provider's object for one model, via retrieve or via the listing.
|
|
1601
|
+
|
|
1602
|
+
THE ROUTE WAS THE BUG, NOT THE READER (batch 44). This asked
|
|
1603
|
+
`client.models.retrieve(model)` alone, which is `GET {base}/models/{id}`
|
|
1604
|
+
-- and OpenRouter implements `GET /api/v1/models` plus
|
|
1605
|
+
`/models/:author/:slug/endpoints` and no plain per-id retrieve, so an id
|
|
1606
|
+
containing a slash (`minimax/minimax-m3`) turns the path into a route
|
|
1607
|
+
that does not exist. 404, every time, on a provider whose listing has
|
|
1608
|
+
carried `context_length` all along and which _WINDOW_FIELD_ALIASES'
|
|
1609
|
+
own comment names. NVIDIA's endpoint has the same shape, and so does
|
|
1610
|
+
every other provider whose ids are `vendor/model`.
|
|
1611
|
+
|
|
1612
|
+
Retrieve stays FIRST: it is one call and it is the right one for Groq,
|
|
1613
|
+
Together and Mistral, and a listing can be hundreds of entries. The
|
|
1614
|
+
fallback costs one extra call per provider per process, on the path
|
|
1615
|
+
that was previously three failures and a permanent give-up.
|
|
1616
|
+
|
|
1617
|
+
Three outcomes, deliberately distinct:
|
|
1618
|
+
an object something answered
|
|
1619
|
+
_NOT_LISTED a complete listing came back without this model, which
|
|
1620
|
+
is an ANSWER ("this provider has nothing to say about
|
|
1621
|
+
it") and must not be retried like a network blip
|
|
1622
|
+
raises neither route could answer; the caller counts it
|
|
1623
|
+
"""
|
|
1624
|
+
try:
|
|
1625
|
+
return client.models.retrieve(model)
|
|
1626
|
+
except Exception as retrieve_error:
|
|
1627
|
+
listing = getattr(getattr(client, "models", None), "list", None)
|
|
1628
|
+
if listing is None:
|
|
1629
|
+
raise
|
|
1630
|
+
try:
|
|
1631
|
+
entries = list(listing())
|
|
1632
|
+
except Exception:
|
|
1633
|
+
# The listing is not a second opinion about the retrieve; it is
|
|
1634
|
+
# the same question by another route. Report the FIRST failure,
|
|
1635
|
+
# which is the one that names the endpoint the caller asked for.
|
|
1636
|
+
raise retrieve_error from None
|
|
1637
|
+
for entry in entries:
|
|
1638
|
+
if getattr(entry, "id", None) == model:
|
|
1639
|
+
return entry
|
|
1640
|
+
if not entries:
|
|
1641
|
+
# An empty listing answered nothing at all; report the failure
|
|
1642
|
+
# that names the endpoint the caller actually asked for.
|
|
1643
|
+
raise retrieve_error from None
|
|
1644
|
+
return _NOT_LISTED
|
|
1645
|
+
|
|
1646
|
+
|
|
1647
|
+
def _record_window(provider_name: str, model: str, value) -> Optional[int]:
|
|
1648
|
+
"""Cache a queried window, coercing anything unusable to None.
|
|
1649
|
+
|
|
1650
|
+
A POSITIVE int or nothing. The Anthropic Models API types
|
|
1651
|
+
`max_input_tokens` as nullable and its own documented example carries
|
|
1652
|
+
a placeholder 0, so "present" is not "usable" -- and a 0 reaching
|
|
1653
|
+
compaction would make `context_limit - backstop` clamp to the floor
|
|
1654
|
+
and compact continuously. Anything that is not a positive int is
|
|
1655
|
+
recorded as None, which means "asked, got no usable answer" and sends
|
|
1656
|
+
the caller to the static table.
|
|
1657
|
+
"""
|
|
1658
|
+
window = value if isinstance(value, int) and not isinstance(value, bool) else None
|
|
1659
|
+
if window is not None and window <= 0:
|
|
1660
|
+
window = None
|
|
1661
|
+
_context_window_cache[(provider_name, model)] = window
|
|
1662
|
+
return window
|
|
1663
|
+
|
|
1664
|
+
|
|
1665
|
+
def known_context_window(provider_name: str, model: str) -> Optional[int]:
|
|
1666
|
+
"""The queried window for (provider, model), or None if none is known.
|
|
1667
|
+
|
|
1668
|
+
CACHE ONLY -- this never makes a call, which is what lets
|
|
1669
|
+
core/compaction.py ask the question without holding a client or
|
|
1670
|
+
importing an SDK. The cache is primed by context_window_for() on the
|
|
1671
|
+
path core/loop.py already crosses before every send, so by the time a
|
|
1672
|
+
compaction check runs the answer is there if the provider had one.
|
|
1673
|
+
"""
|
|
1674
|
+
return _context_window_cache.get((provider_name, model))
|
|
1675
|
+
|
|
1676
|
+
|
|
1677
|
+
def context_window_for(client, provider_name: str, model: str) -> tuple:
|
|
1678
|
+
"""(window_or_None, authoritative). Queries the provider once per process.
|
|
1679
|
+
|
|
1680
|
+
§21 rejected this as needing "fifteen adapters" for a roster where no
|
|
1681
|
+
uniform endpoint existed. Re-measured against the PINNED SDKs, that
|
|
1682
|
+
premise did not hold, and the shape below is what it actually costs:
|
|
1683
|
+
|
|
1684
|
+
ANTHROPIC already answered. _effort_levels retrieves the ModelInfo
|
|
1685
|
+
to read capabilities.effort and max_input_tokens rides
|
|
1686
|
+
along on it, so this branch is usually a cache hit and
|
|
1687
|
+
makes no call at all.
|
|
1688
|
+
GOOGLE models.get(...).input_token_limit, one field on the
|
|
1689
|
+
pinned google-genai type.
|
|
1690
|
+
CHATGPT the plan's model list, `context_window` on the model's
|
|
1691
|
+
entry (ROADMAP_v3 §83) -- not the larger
|
|
1692
|
+
`max_context_window`. Usually a cache hit too: _effort_levels
|
|
1693
|
+
records it off the same read.
|
|
1694
|
+
v1-compat ONE alias sniff over model_extra, not thirteen adapters,
|
|
1695
|
+
because openai.types.Model allows extra fields -- over
|
|
1696
|
+
whichever of models.retrieve() and models.list() answers
|
|
1697
|
+
(see _model_info: providers whose ids carry a slash have
|
|
1698
|
+
no per-id route, so retrieve 404s on them).
|
|
1699
|
+
|
|
1700
|
+
§21's CONCLUSION survives intact and is why the table remains: OpenAI,
|
|
1701
|
+
DeepSeek and Perplexity report nothing, so a fallback constant is
|
|
1702
|
+
still needed at the end of it.
|
|
1703
|
+
|
|
1704
|
+
THE FAILURE/SILENCE SPLIT IS THE POINT, and it is _effort_levels'
|
|
1705
|
+
query_failed rule (#35's lesson) applied to a second query. A call
|
|
1706
|
+
that RAISED is warned and NOT cached: a transient network error must
|
|
1707
|
+
not permanently suppress the query for the life of the process. A
|
|
1708
|
+
call that SUCCEEDED and reported no window IS cached as None, because
|
|
1709
|
+
that is a real and stable answer, and re-asking it on every
|
|
1710
|
+
compaction evaluation would spend a request per step to learn the
|
|
1711
|
+
same thing.
|
|
1712
|
+
"""
|
|
1713
|
+
key = (provider_name, model)
|
|
1714
|
+
if key in _context_window_cache:
|
|
1715
|
+
return _context_window_cache[key], True
|
|
1716
|
+
|
|
1717
|
+
try:
|
|
1718
|
+
if provider_name == "GOOGLE":
|
|
1719
|
+
info = client.models.get(model=model)
|
|
1720
|
+
return _record_window(
|
|
1721
|
+
provider_name, model,
|
|
1722
|
+
getattr(info, "input_token_limit", None)), True
|
|
1723
|
+
if provider_name == "CHATGPT":
|
|
1724
|
+
entry = _chatgpt_entry(_chatgpt_listing(client), model)
|
|
1725
|
+
return _record_window(
|
|
1726
|
+
provider_name, model,
|
|
1727
|
+
entry.get("context_window") if entry else None), True
|
|
1728
|
+
info = _model_info(client, model)
|
|
1729
|
+
if info is _NOT_LISTED:
|
|
1730
|
+
# Asked and answered by a listing that did not name it.
|
|
1731
|
+
return _record_window(provider_name, model, None), True
|
|
1732
|
+
window = _window_from_model_object(info)
|
|
1733
|
+
if window is not None:
|
|
1734
|
+
return _record_window(provider_name, model, window), True
|
|
1735
|
+
# Asked and answered: this provider reports no window.
|
|
1736
|
+
return _record_window(provider_name, model, None), True
|
|
1737
|
+
except Exception as e:
|
|
1738
|
+
# Network failure, a provider with no models endpoint at all, or a
|
|
1739
|
+
# model the endpoint does not know. Not cached as an ANSWER -- see
|
|
1740
|
+
# the docstring -- but counted, so a permanent failure stops being
|
|
1741
|
+
# re-asked once it has proved it is not a blip.
|
|
1742
|
+
attempts = _context_window_failures.get(key, 0) + 1
|
|
1743
|
+
_context_window_failures[key] = attempts
|
|
1744
|
+
giving_up = attempts >= MAX_CONTEXT_WINDOW_QUERY_ATTEMPTS
|
|
1745
|
+
logger.warning(
|
|
1746
|
+
"Could not read a context window for %r from %s (%s); "
|
|
1747
|
+
"config.MODEL_CONTEXT_WINDOWS will answer instead.%s",
|
|
1748
|
+
model, provider_name, e,
|
|
1749
|
+
" Not asking again this session." if giving_up else "")
|
|
1750
|
+
if giving_up:
|
|
1751
|
+
_record_window(provider_name, model, None)
|
|
1752
|
+
return None, False
|
|
1753
|
+
|
|
1754
|
+
|
|
1755
|
+
_google_budget_support: Optional[bool] = None
|
|
1756
|
+
|
|
1757
|
+
|
|
1758
|
+
def _google_supports_thinking_budget() -> bool:
|
|
1759
|
+
"""Whether the INSTALLED google-genai can express a thinking budget.
|
|
1760
|
+
|
|
1761
|
+
Checked rather than assumed, and the check earned its keep: the old pin,
|
|
1762
|
+
google-genai 1.0.0, shipped a ThinkingConfig whose only field was
|
|
1763
|
+
`include_thoughts`, so ThinkingConfig(thinking_budget=N) raised a
|
|
1764
|
+
TypeError and reasoning effort was unavailable on Google.
|
|
1765
|
+
|
|
1766
|
+
The pin moved in batch 172 (ROADMAP_v3 §82, GS1) -- for thought
|
|
1767
|
+
signatures, not for effort -- and 2.29.0 carries the field, so this now
|
|
1768
|
+
answers True and Google effort is on. The check stays: it is still the
|
|
1769
|
+
dependency's real shape that decides, D22's rule, and a downgrade would
|
|
1770
|
+
turn effort off again rather than send a field the SDK cannot carry.
|
|
1771
|
+
"""
|
|
1772
|
+
from google.genai import types as genai_types
|
|
1773
|
+
|
|
1774
|
+
global _google_budget_support
|
|
1775
|
+
if _google_budget_support is None:
|
|
1776
|
+
fields = getattr(genai_types.ThinkingConfig, "model_fields", {})
|
|
1777
|
+
_google_budget_support = "thinking_budget" in fields
|
|
1778
|
+
if not _google_budget_support:
|
|
1779
|
+
logger.info(
|
|
1780
|
+
"Installed google-genai has no ThinkingConfig.thinking_budget; "
|
|
1781
|
+
"reasoning effort is unavailable on GOOGLE until the pin moves."
|
|
1782
|
+
)
|
|
1783
|
+
return _google_budget_support
|
|
1784
|
+
|
|
1785
|
+
|
|
1786
|
+
def _thinking_record(provider_name: str, model: str,
|
|
1787
|
+
blocks: list) -> Optional[dict]:
|
|
1788
|
+
"""The turn's reasoning in the shape ModelResponse.thinking carries.
|
|
1789
|
+
|
|
1790
|
+
None when the turn produced none, so "no reasoning" and "a provider
|
|
1791
|
+
that reports none" stay one thing and no row is written for either.
|
|
1792
|
+
"""
|
|
1793
|
+
if not blocks:
|
|
1794
|
+
return None
|
|
1795
|
+
return {"provider": provider_name, "model": model, "blocks": blocks}
|
|
1796
|
+
|
|
1797
|
+
|
|
1798
|
+
def _thinking_for_provider(provider_name: str, effort: Optional[str]) -> dict:
|
|
1799
|
+
"""Reasoning-effort kwargs for this provider, or {} when no effort is set.
|
|
1800
|
+
|
|
1801
|
+
Effort is opt-in: `effort=None` sends nothing at all and every provider
|
|
1802
|
+
behaves exactly as it did before §16. That matters because the parameter
|
|
1803
|
+
is not universally safe -- `reasoning_effort` is rejected by
|
|
1804
|
+
OpenAI-compatible NON-reasoning models, and `thinking: {"type": "adaptive"}`
|
|
1805
|
+
is rejected by pre-4.6 Anthropic models. Callers are expected to offer
|
|
1806
|
+
only the levels effort_levels_for_model() reported, and to offer nothing
|
|
1807
|
+
when it reported none.
|
|
1808
|
+
|
|
1809
|
+
Returned keys are merged into the provider's own kwargs by the caller;
|
|
1810
|
+
the Google entry is nested inside GenerateContentConfig rather than sent
|
|
1811
|
+
top-level, matching how that provider takes every other setting.
|
|
1812
|
+
|
|
1813
|
+
CHATGPT is the one provider that sends something with no effort set
|
|
1814
|
+
(ROADMAP_v3 §83, CG18). Every request asks for a reasoning summary and
|
|
1815
|
+
for the encrypted reasoning: both were accepted with and without an
|
|
1816
|
+
effort in slice 2a, the encrypted reasoning arrives unasked anyway, and
|
|
1817
|
+
asking says what this code relies on rather than leaning on a default.
|
|
1818
|
+
"auto" leaves the summary to the server, which wrote one on one of the
|
|
1819
|
+
two requests that asked.
|
|
1820
|
+
"""
|
|
1821
|
+
if provider_name == "CHATGPT":
|
|
1822
|
+
reasoning = {"summary": "auto"}
|
|
1823
|
+
if effort:
|
|
1824
|
+
reasoning["effort"] = effort
|
|
1825
|
+
return {"reasoning": reasoning,
|
|
1826
|
+
"include": ["reasoning.encrypted_content"]}
|
|
1827
|
+
if not effort:
|
|
1828
|
+
return {}
|
|
1829
|
+
if provider_name == "ANTHROPIC":
|
|
1830
|
+
# Adaptive thinking + an effort level is the current-model pairing;
|
|
1831
|
+
# `display` is EXPLICIT because its default changed under us. On
|
|
1832
|
+
# Fable 5, Opus 5, Opus 4.8/4.7 and Sonnet 5 it defaults to
|
|
1833
|
+
# "omitted", which still streams thinking blocks but with EMPTY
|
|
1834
|
+
# text -- so §38's whole feature rendered nothing at all on
|
|
1835
|
+
# Anthropic: ev.thinking was "", core/loop.py's `if
|
|
1836
|
+
# token.thinking_delta` never fired, and neither the inline block
|
|
1837
|
+
# nor ThinkingIndicator ever appeared. It looked like it worked
|
|
1838
|
+
# because the OpenAI-compatible branch reads reasoning_content and
|
|
1839
|
+
# is unaffected. The raw chain of thought is never exposed on any
|
|
1840
|
+
# model; "summarized" is the readable form, and it is what the
|
|
1841
|
+
# blocks captured below carry back to the model that wrote them.
|
|
1842
|
+
# budget_tokens was removed and returns a 400.
|
|
1843
|
+
return {
|
|
1844
|
+
"thinking": {"type": "adaptive", "display": "summarized"},
|
|
1845
|
+
"output_config": {"effort": effort},
|
|
1846
|
+
}
|
|
1847
|
+
if provider_name == "GOOGLE":
|
|
1848
|
+
from google.genai import types as genai_types
|
|
1849
|
+
budget = config.GOOGLE_THINKING_BUDGETS.get(effort)
|
|
1850
|
+
if budget is None or not _google_supports_thinking_budget():
|
|
1851
|
+
return {}
|
|
1852
|
+
return {"thinking_config": genai_types.ThinkingConfig(thinking_budget=budget)}
|
|
1853
|
+
return {"reasoning_effort": effort}
|
|
1854
|
+
|
|
1855
|
+
|
|
1856
|
+
# ============================================================================
|
|
1857
|
+
# ---- The ChatGPT plan's Responses path (ROADMAP_v3 §83, slice 2) -----------
|
|
1858
|
+
# ============================================================================
|
|
1859
|
+
#
|
|
1860
|
+
# The one route the plan answers on is the Responses API (CG3), carried by the
|
|
1861
|
+
# openai SDK (CG21). Every shape below was measured on the owner's account in
|
|
1862
|
+
# slices 0 and 2a, except CG23's reasoning inside a tool loop, which is said
|
|
1863
|
+
# where it is sent (_chatgpt_reasoning).
|
|
1864
|
+
|
|
1865
|
+
class ChatGPTPlanError(RuntimeError):
|
|
1866
|
+
"""A plan request OpenAI refused for a reason another attempt cannot fix
|
|
1867
|
+
(CG10). The message is the person's to read.
|
|
1868
|
+
|
|
1869
|
+
ITS OWN TYPE, WITH NO `status_code`, and that is what makes it permanent
|
|
1870
|
+
to `core/provider_errors.py`, which reads a failure by its class names
|
|
1871
|
+
and its `status_code` and `body` and finds none of them here. So the
|
|
1872
|
+
loop's generic 429 retry never sees a usage limit, which CG10 says is
|
|
1873
|
+
never retried.
|
|
1874
|
+
"""
|
|
1875
|
+
|
|
1876
|
+
|
|
1877
|
+
class ChatGPTUsageLimit(ChatGPTPlanError):
|
|
1878
|
+
"""The plan's usage limit. Its message LEADS with Manage usage (CG8)."""
|
|
1879
|
+
|
|
1880
|
+
|
|
1881
|
+
class ChatGPTNotEligible(ChatGPTPlanError):
|
|
1882
|
+
"""This account cannot share its plan, or its sign-in was kept without
|
|
1883
|
+
the plan scope (CG6)."""
|
|
1884
|
+
|
|
1885
|
+
|
|
1886
|
+
class ChatGPTUnsupported(ChatGPTPlanError):
|
|
1887
|
+
"""A field the plan does not support. The message names it, and the
|
|
1888
|
+
same request is not sent again (CG10)."""
|
|
1889
|
+
|
|
1890
|
+
|
|
1891
|
+
#: §83's adopted default: an ineligible account is told the whole story once
|
|
1892
|
+
#: per process and briefly after that.
|
|
1893
|
+
_ineligible_explained = False
|
|
1894
|
+
|
|
1895
|
+
|
|
1896
|
+
def _wire_dict(obj) -> dict:
|
|
1897
|
+
"""An SDK object as the plain dict the wire spelled, `None` fields left
|
|
1898
|
+
out: what goes into §44's record and back unchanged, the shape slice 2a
|
|
1899
|
+
handed back."""
|
|
1900
|
+
if isinstance(obj, dict):
|
|
1901
|
+
return {k: v for k, v in obj.items() if v is not None}
|
|
1902
|
+
dump = getattr(obj, "model_dump", None)
|
|
1903
|
+
if callable(dump):
|
|
1904
|
+
return dump(exclude_none=True, mode="json")
|
|
1905
|
+
return {}
|
|
1906
|
+
|
|
1907
|
+
|
|
1908
|
+
def _chatgpt_fields(body) -> tuple:
|
|
1909
|
+
"""(code, param, message) off an error body, whichever of OpenAI's two
|
|
1910
|
+
shapes it came in: the standard error object, or `{"detail": ...}`."""
|
|
1911
|
+
error = body.get("error") if isinstance(body, dict) and isinstance(
|
|
1912
|
+
body.get("error"), dict) else body
|
|
1913
|
+
if not isinstance(error, dict):
|
|
1914
|
+
return None, None, None
|
|
1915
|
+
return (error.get("code"), error.get("param"),
|
|
1916
|
+
error.get("message") or error.get("detail"))
|
|
1917
|
+
|
|
1918
|
+
|
|
1919
|
+
def _chatgpt_refusal(body) -> Optional[Exception]:
|
|
1920
|
+
"""CG10's exception for a refusal its code names, or None.
|
|
1921
|
+
|
|
1922
|
+
None hands the failure back to `core/provider_errors.py`, as for every
|
|
1923
|
+
other provider: a 503, another 429 or a dropped connection is retried
|
|
1924
|
+
with the loop's bounded backoff, and anything else fails the turn.
|
|
1925
|
+
"""
|
|
1926
|
+
global _ineligible_explained
|
|
1927
|
+
from core import chatgpt_auth
|
|
1928
|
+
|
|
1929
|
+
code, param, message = _chatgpt_fields(body)
|
|
1930
|
+
said = (f" OpenAI said: {' '.join(str(message).split())[:300]}"
|
|
1931
|
+
if message else "")
|
|
1932
|
+
if code == "subscription_sharing_usage_limit_exceeded":
|
|
1933
|
+
return ChatGPTUsageLimit(
|
|
1934
|
+
f"Manage usage: {chatgpt_auth.USAGE_URL} -- this ChatGPT plan "
|
|
1935
|
+
f"has reached its usage limit, so the request was not sent again. "
|
|
1936
|
+
f"Your next message tries once more.{said}")
|
|
1937
|
+
if code == "subscription_sharing_user_not_eligible":
|
|
1938
|
+
if _ineligible_explained:
|
|
1939
|
+
return ChatGPTNotEligible(
|
|
1940
|
+
"This ChatGPT account cannot share its plan (said above).")
|
|
1941
|
+
_ineligible_explained = True
|
|
1942
|
+
return ChatGPTNotEligible(
|
|
1943
|
+
f"This ChatGPT account cannot share its plan with an app, so "
|
|
1944
|
+
f"nothing was sent again and no new sign-in was asked for. Switch "
|
|
1945
|
+
f"with /model, or sign out with /disconnect chatgpt.{said}")
|
|
1946
|
+
if code == "subscription_sharing_unsupported_capability":
|
|
1947
|
+
return ChatGPTUnsupported(
|
|
1948
|
+
f"The ChatGPT plan does not support "
|
|
1949
|
+
f"{param or 'a field this request sent'}, so the request was not "
|
|
1950
|
+
f"sent again.{said}")
|
|
1951
|
+
return None
|
|
1952
|
+
|
|
1953
|
+
|
|
1954
|
+
def _chatgpt_open(client, request: dict):
|
|
1955
|
+
"""The stream, opened with CG10's one forced refresh on a 401.
|
|
1956
|
+
|
|
1957
|
+
The SDK called `chatgpt_auth.access_token` for the token it sent and
|
|
1958
|
+
holds that token as `client.api_key`. `replacing=` refreshes only while
|
|
1959
|
+
the file still holds it, so two requests refused together spend one
|
|
1960
|
+
refresh between them. A refresh refused as unusable raises `SignedOut`
|
|
1961
|
+
with the session cleared; a second 401 after a good refresh says to sign
|
|
1962
|
+
in again and keeps the tokens, because only an unusable refresh clears
|
|
1963
|
+
them (CG10). A refresh that could not be made at all is a
|
|
1964
|
+
`ConnectionError`, which the loop retries like any dropped connection.
|
|
1965
|
+
|
|
1966
|
+
THE ONE RESEND THIS FILE MAKES, and it is not a retry in AGENTS.md's
|
|
1967
|
+
sense, which this file does not own: nothing was wrong with the request
|
|
1968
|
+
or the weather, the credential was stale, and the second attempt is made
|
|
1969
|
+
with a different one. Every retry for a failure the provider might not
|
|
1970
|
+
repeat is still `core/loop.py`'s, asked of `core/provider_errors.py`.
|
|
1971
|
+
"""
|
|
1972
|
+
from core import chatgpt_auth
|
|
1973
|
+
|
|
1974
|
+
for attempt in (1, 2):
|
|
1975
|
+
try:
|
|
1976
|
+
return client.responses.create(**request)
|
|
1977
|
+
except chatgpt_auth.Transient as exc:
|
|
1978
|
+
raise ConnectionError(str(exc)) from exc
|
|
1979
|
+
except chatgpt_auth.ChatGPTAuthError:
|
|
1980
|
+
raise
|
|
1981
|
+
except Exception as exc:
|
|
1982
|
+
refusal = _chatgpt_refusal(getattr(exc, "body", None))
|
|
1983
|
+
if refusal is not None:
|
|
1984
|
+
raise refusal from exc
|
|
1985
|
+
if getattr(exc, "status_code", None) != 401:
|
|
1986
|
+
raise
|
|
1987
|
+
if attempt == 2:
|
|
1988
|
+
raise ChatGPTPlanError(
|
|
1989
|
+
"OpenAI refused the ChatGPT plan's sign-in again after a "
|
|
1990
|
+
"refresh. /connect chatgpt signs in again; the stored "
|
|
1991
|
+
"tokens are kept.") from exc
|
|
1992
|
+
try:
|
|
1993
|
+
chatgpt_auth.access_token(
|
|
1994
|
+
replacing=getattr(client, "api_key", None) or None)
|
|
1995
|
+
except chatgpt_auth.Transient as again:
|
|
1996
|
+
raise ConnectionError(str(again)) from again
|
|
1997
|
+
raise AssertionError("unreachable: the second attempt returns or raises")
|
|
1998
|
+
|
|
1999
|
+
|
|
2000
|
+
def _chatgpt_stream_failure(source) -> Exception:
|
|
2001
|
+
"""An `error` event or a `response.failed` as an exception. Neither has
|
|
2002
|
+
been seen: nothing failed in slice 2a.
|
|
2003
|
+
|
|
2004
|
+
The openai SDK raises only on an event whose top-level `error` key is an
|
|
2005
|
+
object, which neither of these has, so they arrive as events and this
|
|
2006
|
+
reads them. CG10's type when the code names one; otherwise the rule
|
|
2007
|
+
`core/provider_errors.py` applies to a stream that failed after it
|
|
2008
|
+
opened, on every provider.
|
|
2009
|
+
"""
|
|
2010
|
+
from core import provider_errors
|
|
2011
|
+
|
|
2012
|
+
body = _wire_dict(source)
|
|
2013
|
+
refusal = _chatgpt_refusal(body)
|
|
2014
|
+
if refusal is not None:
|
|
2015
|
+
return refusal
|
|
2016
|
+
code, _, message = _chatgpt_fields(body)
|
|
2017
|
+
text = (f"The ChatGPT plan's response failed ({code or 'no code'}): "
|
|
2018
|
+
f"{message or 'no message'}")
|
|
2019
|
+
if provider_errors.open_stream_error_is_transient(body):
|
|
2020
|
+
return ConnectionError(text)
|
|
2021
|
+
return ChatGPTPlanError(text)
|
|
2022
|
+
|
|
2023
|
+
|
|
2024
|
+
def _chatgpt_stream(client, model: str, items: list, system_prompt: str,
|
|
2025
|
+
tools, reasoning: dict, supports_stream_usage: bool):
|
|
2026
|
+
"""The CHATGPT branch of `call_model_stream`.
|
|
2027
|
+
|
|
2028
|
+
DISPATCH IS ON `event.type` ALONE. The SDK parses an event it does not
|
|
2029
|
+
know as whichever class its union tries first: slice 2a's
|
|
2030
|
+
`response.metadata` came out as a `ResponseAudioDeltaEvent` whose `type`
|
|
2031
|
+
still said `response.metadata`. A class check would have read it as
|
|
2032
|
+
audio; the string is the one thing that is right.
|
|
2033
|
+
|
|
2034
|
+
Never sent: `temperature` (refused, slice 0), `previous_response_id`
|
|
2035
|
+
(the whole history goes in `input`), `service_tier`, a hosted tool, and
|
|
2036
|
+
the Codex instructions the model list carries (CG3).
|
|
2037
|
+
"""
|
|
2038
|
+
from core import chatgpt_auth
|
|
2039
|
+
|
|
2040
|
+
state = chatgpt_auth.status()
|
|
2041
|
+
if state.signed_in and not state.plan_usage:
|
|
2042
|
+
raise ChatGPTNotEligible(
|
|
2043
|
+
"This ChatGPT sign-in did not share the plan: OpenAI did not "
|
|
2044
|
+
"grant chatgpt.tokens.use.direct, so nothing is sent on it (CG6). "
|
|
2045
|
+
"/connect chatgpt signs in again.")
|
|
2046
|
+
request = dict(model=model, instructions=system_prompt, input=items,
|
|
2047
|
+
store=False, stream=True,
|
|
2048
|
+
max_output_tokens=config.MAX_TOKENS)
|
|
2049
|
+
if tools:
|
|
2050
|
+
request["tools"] = tools
|
|
2051
|
+
request.update(reasoning)
|
|
2052
|
+
|
|
2053
|
+
stream = _chatgpt_open(client, request)
|
|
2054
|
+
text_parts: list = []
|
|
2055
|
+
pending_calls: list = []
|
|
2056
|
+
reasoning_items: list = []
|
|
2057
|
+
# None rather than 0, for the Google branch's reason (D21).
|
|
2058
|
+
usage = {"input_tokens": None, "output_tokens": None}
|
|
2059
|
+
finish_reason = None
|
|
2060
|
+
summary_started = False
|
|
2061
|
+
try:
|
|
2062
|
+
for event in stream:
|
|
2063
|
+
kind = getattr(event, "type", None)
|
|
2064
|
+
if kind == "response.output_text.delta":
|
|
2065
|
+
delta = getattr(event, "delta", None)
|
|
2066
|
+
if isinstance(delta, str) and delta:
|
|
2067
|
+
text_parts.append(delta)
|
|
2068
|
+
yield StreamToken(text_delta=delta)
|
|
2069
|
+
elif kind == "response.reasoning_summary_part.added":
|
|
2070
|
+
# §38: a summary in several parts reads as paragraphs.
|
|
2071
|
+
if summary_started:
|
|
2072
|
+
yield StreamToken(thinking_delta="\n\n")
|
|
2073
|
+
summary_started = True
|
|
2074
|
+
elif kind == "response.reasoning_summary_text.delta":
|
|
2075
|
+
delta = getattr(event, "delta", None)
|
|
2076
|
+
if isinstance(delta, str) and delta:
|
|
2077
|
+
yield StreamToken(thinking_delta=delta)
|
|
2078
|
+
elif kind == "response.output_item.done":
|
|
2079
|
+
item = getattr(event, "item", None)
|
|
2080
|
+
item_type = getattr(item, "type", None)
|
|
2081
|
+
if item_type == "function_call":
|
|
2082
|
+
pending_calls.append(item)
|
|
2083
|
+
elif item_type == "reasoning":
|
|
2084
|
+
reasoning_items.append(_wire_dict(item))
|
|
2085
|
+
elif kind in ("response.completed", "response.incomplete"):
|
|
2086
|
+
response = getattr(event, "response", None)
|
|
2087
|
+
usage_obj = getattr(response, "usage", None)
|
|
2088
|
+
for key in ("input_tokens", "output_tokens"):
|
|
2089
|
+
reported = _token_count(usage_obj, key)
|
|
2090
|
+
if reported is not None:
|
|
2091
|
+
usage[key] = reported
|
|
2092
|
+
details = getattr(response, "incomplete_details", None)
|
|
2093
|
+
if getattr(details, "reason", None) == "max_output_tokens":
|
|
2094
|
+
finish_reason = "length"
|
|
2095
|
+
elif kind == "response.failed":
|
|
2096
|
+
raise _chatgpt_stream_failure(
|
|
2097
|
+
getattr(getattr(event, "response", None), "error", None))
|
|
2098
|
+
elif kind == "error":
|
|
2099
|
+
raise _chatgpt_stream_failure(event)
|
|
2100
|
+
except Exception as exc:
|
|
2101
|
+
# An error the SDK raised mid-stream may still carry a code CG10
|
|
2102
|
+
# names; anything else, ours included, goes up as it is.
|
|
2103
|
+
refusal = _chatgpt_refusal(getattr(exc, "body", None))
|
|
2104
|
+
if refusal is None:
|
|
2105
|
+
raise
|
|
2106
|
+
raise refusal from exc
|
|
2107
|
+
finally:
|
|
2108
|
+
close = getattr(stream, "close", None)
|
|
2109
|
+
if callable(close):
|
|
2110
|
+
close()
|
|
2111
|
+
|
|
2112
|
+
usage = _stream_usage("CHATGPT", usage, supports_stream_usage)
|
|
2113
|
+
calls = []
|
|
2114
|
+
for item in pending_calls:
|
|
2115
|
+
name = getattr(item, "name", None) or ""
|
|
2116
|
+
parsed, parse_error = _tool_arguments(
|
|
2117
|
+
getattr(item, "arguments", None) or "", name, finish_reason)
|
|
2118
|
+
if parse_error is not None:
|
|
2119
|
+
logger.warning(
|
|
2120
|
+
"%s: %s (finish_reason=%r)", name, parse_error, finish_reason)
|
|
2121
|
+
calls.append(ToolCallRequest(
|
|
2122
|
+
# The id rule the other branches keep (batch 76).
|
|
2123
|
+
id=getattr(item, "call_id", None) or str(uuid4()),
|
|
2124
|
+
name=name, input=parsed, parse_error=parse_error))
|
|
2125
|
+
yield StreamToken(final_response=ModelResponse(
|
|
2126
|
+
text="".join(text_parts), tool_calls=calls, usage=usage,
|
|
2127
|
+
thinking=_thinking_record("CHATGPT", model, reasoning_items),
|
|
2128
|
+
))
|
|
2129
|
+
|
|
2130
|
+
|
|
2131
|
+
# ============================================================================
|
|
2132
|
+
# ---- Streaming call + helpers -----------------------------------------------
|
|
2133
|
+
# ============================================================================
|
|
2134
|
+
|
|
2135
|
+
def call_model_stream(
|
|
2136
|
+
client,
|
|
2137
|
+
provider_name: str,
|
|
2138
|
+
model: str,
|
|
2139
|
+
messages: list[dict],
|
|
2140
|
+
system_prompt: str,
|
|
2141
|
+
tool_schemas: list[dict],
|
|
2142
|
+
temperature: Optional[float] = None,
|
|
2143
|
+
effort: Optional[str] = None,
|
|
2144
|
+
):
|
|
2145
|
+
"""The one model call, for every provider. Yields StreamToken
|
|
2146
|
+
events: text_delta for incremental text, thinking_delta for
|
|
2147
|
+
incremental reasoning text where the provider sends any (§38 O2),
|
|
2148
|
+
final_response for the terminal ModelResponse with accumulated tool
|
|
2149
|
+
calls and usage.
|
|
2150
|
+
|
|
2151
|
+
Four provider implementations — Anthropic typed events, Google
|
|
2152
|
+
chunked candidates, OpenAI-compatible index-accumulated tool-call
|
|
2153
|
+
fragments, and since ROADMAP_v3 §83 the ChatGPT plan's Responses events
|
|
2154
|
+
(_chatgpt_stream) — which must agree with each other about the
|
|
2155
|
+
ModelResponse they produce for the same inputs.
|
|
2156
|
+
|
|
2157
|
+
There used to be a non-streaming `call_model()` beside this, and this
|
|
2158
|
+
docstring used to promise the two were identical. They were not: it
|
|
2159
|
+
set ModelResponse.raw and no streaming branch did, which is what a
|
|
2160
|
+
second implementation nobody runs accumulates (#38). It is gone, and
|
|
2161
|
+
with it `raw`; the only thing left to keep in step is the three
|
|
2162
|
+
branches below.
|
|
2163
|
+
|
|
2164
|
+
D21 (usage-or-raise): if the provider is marked as supporting stream
|
|
2165
|
+
usage but the stream completes with zero usage, raise rather than
|
|
2166
|
+
silently reporting zero (which disables budget enforcement and §21's
|
|
2167
|
+
compaction trigger).
|
|
2168
|
+
"""
|
|
2169
|
+
# Read once so every branch applies the same D21 usage-or-raise rule --
|
|
2170
|
+
# all three of them, since audit #40. This comment used to claim that in
|
|
2171
|
+
# its first sentence and take it back in the second ("the flag only gates
|
|
2172
|
+
# the Google and OpenAI-compatible branches"), which left the one
|
|
2173
|
+
# provider whose example config ships the flag `true` as the one provider
|
|
2174
|
+
# where setting it did nothing.
|
|
2175
|
+
provider_data = load_provider_data()
|
|
2176
|
+
supports_stream_usage = provider_data.get(provider_name, {}).get(
|
|
2177
|
+
"supports_stream_usage", False
|
|
2178
|
+
)
|
|
2179
|
+
# §44. Whether this OpenAI-compatible provider accepts its own
|
|
2180
|
+
# reasoning back on the next request. supports_stream_usage's
|
|
2181
|
+
# precedent: a per-provider fact that cannot be discovered and must not
|
|
2182
|
+
# be guessed, so it lives beside the endpoint it describes and defaults
|
|
2183
|
+
# to the conservative answer.
|
|
2184
|
+
echoes_reasoning = provider_data.get(provider_name, {}).get(
|
|
2185
|
+
"echoes_reasoning", False
|
|
2186
|
+
)
|
|
2187
|
+
|
|
2188
|
+
translated_messages = _messages_for_provider(
|
|
2189
|
+
provider_name, messages, model, echoes_reasoning,
|
|
2190
|
+
images=accepts_images(provider_name))
|
|
2191
|
+
sampling = _sampling_kwargs(provider_name, model, temperature, effort)
|
|
2192
|
+
thinking = _thinking_for_provider(provider_name, effort)
|
|
2193
|
+
|
|
2194
|
+
if provider_name == "CHATGPT":
|
|
2195
|
+
# ROADMAP_v3 §83. `sampling` is always {} here: _sampling_kwargs
|
|
2196
|
+
# drops a temperature for this provider, which the route refuses.
|
|
2197
|
+
yield from _chatgpt_stream(
|
|
2198
|
+
client, model, translated_messages, system_prompt,
|
|
2199
|
+
_tools_for_provider(provider_name, tool_schemas), thinking,
|
|
2200
|
+
supports_stream_usage)
|
|
2201
|
+
return
|
|
2202
|
+
|
|
2203
|
+
if provider_name == "ANTHROPIC":
|
|
2204
|
+
stream_kwargs = dict(
|
|
2205
|
+
model=model,
|
|
2206
|
+
max_tokens=config.MAX_TOKENS,
|
|
2207
|
+
system=system_prompt,
|
|
2208
|
+
messages=translated_messages,
|
|
2209
|
+
tools=_tools_for_provider(provider_name, tool_schemas),
|
|
2210
|
+
)
|
|
2211
|
+
stream_kwargs.update(sampling)
|
|
2212
|
+
stream_kwargs.update(thinking)
|
|
2213
|
+
|
|
2214
|
+
with client.messages.stream(**stream_kwargs) as stream:
|
|
2215
|
+
# §38 O3. Iterating the STREAM rather than stream.text_stream,
|
|
2216
|
+
# because text_stream yields text deltas only and there is no
|
|
2217
|
+
# way to have both: `text_stream` is an instance attribute
|
|
2218
|
+
# built from the same underlying iterator in MessageStream's
|
|
2219
|
+
# __init__, so consuming one consumes the other.
|
|
2220
|
+
#
|
|
2221
|
+
# The SDK emits its own synthetic events beside the raw ones --
|
|
2222
|
+
# TextEvent(type="text", text=<delta>) and
|
|
2223
|
+
# ThinkingEvent(type="thinking", thinking=<delta>) -- which is
|
|
2224
|
+
# why this reads two attribute names rather than unpacking
|
|
2225
|
+
# content_block_delta by hand. Verified against the pinned
|
|
2226
|
+
# anthropic (0.116.0, and 2.1.0 since batch 234) rather than
|
|
2227
|
+
# assumed (D22), and pinned in tests/test_sdk_conformance.py.
|
|
2228
|
+
#
|
|
2229
|
+
# Every other event kind is ignored on purpose: `text` is the
|
|
2230
|
+
# accumulated answer read off get_final_message() below, not
|
|
2231
|
+
# off these deltas, so nothing here can change what this
|
|
2232
|
+
# branch returns.
|
|
2233
|
+
for ev in stream:
|
|
2234
|
+
kind = getattr(ev, "type", None)
|
|
2235
|
+
if kind == "text":
|
|
2236
|
+
yield StreamToken(text_delta=ev.text)
|
|
2237
|
+
elif kind == "thinking":
|
|
2238
|
+
yield StreamToken(thinking_delta=ev.thinking)
|
|
2239
|
+
final_msg = stream.get_final_message()
|
|
2240
|
+
text = "\n".join(
|
|
2241
|
+
b.text for b in final_msg.content if b.type == "text"
|
|
2242
|
+
)
|
|
2243
|
+
calls = [
|
|
2244
|
+
ToolCallRequest(id=b.id, name=b.name, input=b.input)
|
|
2245
|
+
for b in final_msg.content if b.type == "tool_use"
|
|
2246
|
+
]
|
|
2247
|
+
# model_dump() rather than a hand-built dict: what goes back on
|
|
2248
|
+
# the wire has to be what came off it, field for field, or the
|
|
2249
|
+
# signature does not verify. redacted_thinking is included for
|
|
2250
|
+
# the same reason -- it is opaque to us and meaningful to the
|
|
2251
|
+
# model, so dropping it would silently break a turn that had
|
|
2252
|
+
# one.
|
|
2253
|
+
thinking_blocks = [
|
|
2254
|
+
b.model_dump() for b in final_msg.content
|
|
2255
|
+
if b.type in ("thinking", "redacted_thinking")
|
|
2256
|
+
]
|
|
2257
|
+
usage_obj = getattr(final_msg, "usage", None)
|
|
2258
|
+
usage = {
|
|
2259
|
+
"input_tokens": _token_count(usage_obj, "input_tokens"),
|
|
2260
|
+
"output_tokens": _token_count(usage_obj, "output_tokens"),
|
|
2261
|
+
}
|
|
2262
|
+
|
|
2263
|
+
# D21 parity with the other two branches (audit #40). This branch
|
|
2264
|
+
# used to be exempt on the reasoning that get_final_message() always
|
|
2265
|
+
# carries usage -- which is very likely right for anthropic==0.116.0,
|
|
2266
|
+
# and is exactly the kind of assumption D21 exists because nobody can
|
|
2267
|
+
# verify from inside. The getattr defaults that used to be above were
|
|
2268
|
+
# the softened form D21's own note forbids, so on the DEFAULT
|
|
2269
|
+
# provider an SDK field rename would have zeroed the budget silently,
|
|
2270
|
+
# disabling both the budget stop condition and §21's compaction
|
|
2271
|
+
# trigger. That is the mcp v1->v2 failure mode, and #35 is this
|
|
2272
|
+
# file's live example of it.
|
|
2273
|
+
#
|
|
2274
|
+
# Honouring the flag rather than always raising keeps it meaning ONE
|
|
2275
|
+
# thing on all three providers, which is the coherence #40 is about;
|
|
2276
|
+
# providers.json.example ships it true for ANTHROPIC, so the default
|
|
2277
|
+
# posture is the checked one, and a proxy that genuinely omits usage
|
|
2278
|
+
# can still opt out instead of being unable to run.
|
|
2279
|
+
#
|
|
2280
|
+
# Batch 147: the rule itself, and the coercion that makes `usage`
|
|
2281
|
+
# two ints, are in `_stream_usage` -- one copy for three branches
|
|
2282
|
+
# that carried three.
|
|
2283
|
+
usage = _stream_usage(provider_name, usage, supports_stream_usage)
|
|
2284
|
+
|
|
2285
|
+
yield StreamToken(final_response=ModelResponse(
|
|
2286
|
+
text=text, tool_calls=calls, usage=usage,
|
|
2287
|
+
thinking=_thinking_record(provider_name, model, thinking_blocks),
|
|
2288
|
+
))
|
|
2289
|
+
return
|
|
2290
|
+
|
|
2291
|
+
if provider_name == "GOOGLE":
|
|
2292
|
+
from google.genai import types as genai_types
|
|
2293
|
+
genai_config_kwargs = dict(
|
|
2294
|
+
system_instruction=system_prompt,
|
|
2295
|
+
tools=_tools_for_provider(provider_name, tool_schemas),
|
|
2296
|
+
max_output_tokens=config.MAX_TOKENS,
|
|
2297
|
+
)
|
|
2298
|
+
genai_config_kwargs.update(sampling)
|
|
2299
|
+
genai_config_kwargs.update(thinking)
|
|
2300
|
+
|
|
2301
|
+
text_parts = []
|
|
2302
|
+
calls = []
|
|
2303
|
+
# §82 (GS2). One block per function call, read off the part that
|
|
2304
|
+
# carried the call: Gemini 3 refuses the next request unless each
|
|
2305
|
+
# call goes back with the signature it came with.
|
|
2306
|
+
signatures = []
|
|
2307
|
+
# None rather than 0: "not reported" has to stay distinguishable
|
|
2308
|
+
# from "reported as zero" until D21 has looked at it, or the
|
|
2309
|
+
# partial-report case the widened rule exists for arrives here
|
|
2310
|
+
# already coerced and indistinguishable (batch 147).
|
|
2311
|
+
usage = {"input_tokens": None, "output_tokens": None}
|
|
2312
|
+
|
|
2313
|
+
for chunk in client.models.generate_content_stream(
|
|
2314
|
+
model=model,
|
|
2315
|
+
contents=translated_messages,
|
|
2316
|
+
config=genai_types.GenerateContentConfig(**genai_config_kwargs),
|
|
2317
|
+
):
|
|
2318
|
+
candidate = chunk.candidates[0] if chunk.candidates else None
|
|
2319
|
+
content = candidate.content if candidate else None
|
|
2320
|
+
# `or []`: Content.parts is Optional on the pinned type, and a
|
|
2321
|
+
# chunk can carry a content with no parts at all.
|
|
2322
|
+
parts = (content.parts if content else None) or []
|
|
2323
|
+
for p in parts:
|
|
2324
|
+
if p.text is not None:
|
|
2325
|
+
text_parts.append(p.text)
|
|
2326
|
+
yield StreamToken(text_delta=p.text)
|
|
2327
|
+
if p.function_call is not None:
|
|
2328
|
+
fc = p.function_call
|
|
2329
|
+
call_id = fc.id or str(uuid4())
|
|
2330
|
+
calls.append(ToolCallRequest(
|
|
2331
|
+
id=call_id,
|
|
2332
|
+
name=fc.name,
|
|
2333
|
+
input=fc.args or {},
|
|
2334
|
+
))
|
|
2335
|
+
# getattr: the field is on every pinned Part
|
|
2336
|
+
# (test_sdk_conformance asks the real one), but not on
|
|
2337
|
+
# the bare namespaces the stream tests build parts from.
|
|
2338
|
+
signatures.append(_google_signature_block(
|
|
2339
|
+
call_id, getattr(p, "thought_signature", None)))
|
|
2340
|
+
# PER FIELD, and only when the field was reported. The
|
|
2341
|
+
# guard used to be `if usage_meta:` -- which passes for a bare
|
|
2342
|
+
# instance, because every field is Optional and the object is
|
|
2343
|
+
# truthy anyway -- and then overwrote both counts with
|
|
2344
|
+
# whatever it found, None included. A chunk that carries
|
|
2345
|
+
# usage_metadata without a count now leaves an earlier chunk's
|
|
2346
|
+
# figure alone: reported beats unreported whatever the order,
|
|
2347
|
+
# which is the only rule that does not depend on which chunk
|
|
2348
|
+
# the provider chose to put the totals on.
|
|
2349
|
+
usage_meta = getattr(chunk, "usage_metadata", None)
|
|
2350
|
+
if usage_meta is not None:
|
|
2351
|
+
for key, attribute in (
|
|
2352
|
+
("input_tokens", "prompt_token_count"),
|
|
2353
|
+
("output_tokens", "candidates_token_count")):
|
|
2354
|
+
reported = _token_count(usage_meta, attribute)
|
|
2355
|
+
if reported is not None:
|
|
2356
|
+
usage[key] = reported
|
|
2357
|
+
|
|
2358
|
+
# D21, in the one place all three branches read it (batch 147).
|
|
2359
|
+
# This is the branch the nullable-field defect was found in: the
|
|
2360
|
+
# pinned google-genai declares every usage field Optional[int], so
|
|
2361
|
+
# `usage` could leave here holding None and the raise could not
|
|
2362
|
+
# fire -- `None != 0`.
|
|
2363
|
+
usage = _stream_usage(provider_name, usage, supports_stream_usage)
|
|
2364
|
+
|
|
2365
|
+
yield StreamToken(final_response=ModelResponse(
|
|
2366
|
+
text="".join(text_parts), tool_calls=calls, usage=usage,
|
|
2367
|
+
thinking=_thinking_record(provider_name, model, signatures),
|
|
2368
|
+
))
|
|
2369
|
+
return
|
|
2370
|
+
|
|
2371
|
+
# OpenAI-compatible path (supports_stream_usage already read above)
|
|
2372
|
+
full_messages = [{"role": "system", "content": system_prompt}] + translated_messages
|
|
2373
|
+
openai_kwargs = dict(
|
|
2374
|
+
model=model,
|
|
2375
|
+
messages=full_messages,
|
|
2376
|
+
max_completion_tokens=config.MAX_TOKENS,
|
|
2377
|
+
tools=_tools_for_provider(provider_name, tool_schemas),
|
|
2378
|
+
stream=True,
|
|
2379
|
+
)
|
|
2380
|
+
openai_kwargs.update(sampling)
|
|
2381
|
+
openai_kwargs.update(thinking)
|
|
2382
|
+
if supports_stream_usage:
|
|
2383
|
+
openai_kwargs["stream_options"] = {"include_usage": True}
|
|
2384
|
+
|
|
2385
|
+
text_parts = []
|
|
2386
|
+
reasoning_parts = []
|
|
2387
|
+
tool_fragments: dict[int, dict] = {}
|
|
2388
|
+
# None, for the reason the Google branch states: D21 cannot tell an
|
|
2389
|
+
# unreported count from a zero one once it has been coerced.
|
|
2390
|
+
usage = {"input_tokens": None, "output_tokens": None}
|
|
2391
|
+
|
|
2392
|
+
finish_reason = None
|
|
2393
|
+
for chunk in client.chat.completions.create(**openai_kwargs):
|
|
2394
|
+
if chunk.choices:
|
|
2395
|
+
# W4 (#37): free -- chunk.choices[0] is already read here. The
|
|
2396
|
+
# provider sets this on the last content chunk, and `or
|
|
2397
|
+
# finish_reason` keeps it once seen because the usage chunk
|
|
2398
|
+
# that follows carries none. getattr because a chunk is only
|
|
2399
|
+
# required to have `delta`.
|
|
2400
|
+
choice = chunk.choices[0]
|
|
2401
|
+
finish_reason = getattr(choice, "finish_reason", None) or finish_reason
|
|
2402
|
+
delta = choice.delta
|
|
2403
|
+
if delta.content:
|
|
2404
|
+
text_parts.append(delta.content)
|
|
2405
|
+
yield StreamToken(text_delta=delta.content)
|
|
2406
|
+
# §38 O3. Reasoning text, where the provider sends any. There
|
|
2407
|
+
# is no standard field: DeepSeek, Qwen and Z.AI use
|
|
2408
|
+
# `reasoning_content`, OpenRouter and Groq use `reasoning`,
|
|
2409
|
+
# and OpenAI's own endpoint sends neither. openai's ChoiceDelta
|
|
2410
|
+
# is extra="allow" (pinned in test_sdk_conformance.py), so an
|
|
2411
|
+
# unknown-to-the-SDK field arrives as a real attribute.
|
|
2412
|
+
#
|
|
2413
|
+
# NEVER appended to text_parts: reasoning is not the answer,
|
|
2414
|
+
# and folding it in would change every ModelResponse this
|
|
2415
|
+
# branch returns and put reasoning into the thread as though
|
|
2416
|
+
# the model had said it. It is accumulated SEPARATELY instead
|
|
2417
|
+
# (batch 44), into ModelResponse.thinking, where it is kept as
|
|
2418
|
+
# this provider's own field name rather than as prose -- an
|
|
2419
|
+
# echo back on the wire has to spell it the way the provider
|
|
2420
|
+
# spelled it.
|
|
2421
|
+
reasoning = (getattr(delta, "reasoning_content", None)
|
|
2422
|
+
or getattr(delta, "reasoning", None))
|
|
2423
|
+
if isinstance(reasoning, str) and reasoning:
|
|
2424
|
+
reasoning_parts.append(reasoning)
|
|
2425
|
+
yield StreamToken(thinking_delta=reasoning)
|
|
2426
|
+
if delta.tool_calls:
|
|
2427
|
+
for tc_delta in delta.tool_calls:
|
|
2428
|
+
idx = tc_delta.index
|
|
2429
|
+
if idx not in tool_fragments:
|
|
2430
|
+
tool_fragments[idx] = {
|
|
2431
|
+
"id": "", "name_parts": [], "arguments_parts": [],
|
|
2432
|
+
}
|
|
2433
|
+
if tc_delta.id:
|
|
2434
|
+
tool_fragments[idx]["id"] = tc_delta.id
|
|
2435
|
+
# Accumulate name like arguments: some v1-compatible
|
|
2436
|
+
# providers split function.name across deltas, so a
|
|
2437
|
+
# last-write-wins assignment would corrupt the name.
|
|
2438
|
+
if tc_delta.function and tc_delta.function.name:
|
|
2439
|
+
tool_fragments[idx]["name_parts"].append(
|
|
2440
|
+
tc_delta.function.name
|
|
2441
|
+
)
|
|
2442
|
+
if tc_delta.function and tc_delta.function.arguments:
|
|
2443
|
+
tool_fragments[idx]["arguments_parts"].append(
|
|
2444
|
+
tc_delta.function.arguments
|
|
2445
|
+
)
|
|
2446
|
+
# Per field and only when reported -- the Google branch's shape,
|
|
2447
|
+
# for its reason, and because `CompletionUsage` is parsed by the
|
|
2448
|
+
# SDK's construct() and so can hold None against an `int`
|
|
2449
|
+
# annotation (batch 147).
|
|
2450
|
+
usage_obj = getattr(chunk, "usage", None)
|
|
2451
|
+
if usage_obj is not None:
|
|
2452
|
+
for key, attribute in (("input_tokens", "prompt_tokens"),
|
|
2453
|
+
("output_tokens", "completion_tokens")):
|
|
2454
|
+
reported = _token_count(usage_obj, attribute)
|
|
2455
|
+
if reported is not None:
|
|
2456
|
+
usage[key] = reported
|
|
2457
|
+
|
|
2458
|
+
# D21: if the provider is marked as supporting stream usage but we
|
|
2459
|
+
# got no usable usage, the stream was likely interrupted before the
|
|
2460
|
+
# final usage chunk arrived. Raise rather than silently reporting zero.
|
|
2461
|
+
# One copy for three branches since batch 147.
|
|
2462
|
+
usage = _stream_usage(provider_name, usage, supports_stream_usage)
|
|
2463
|
+
|
|
2464
|
+
calls = []
|
|
2465
|
+
for idx in sorted(tool_fragments):
|
|
2466
|
+
frag = tool_fragments[idx]
|
|
2467
|
+
name = "".join(frag["name_parts"])
|
|
2468
|
+
parsed, parse_error = _tool_arguments(
|
|
2469
|
+
"".join(frag["arguments_parts"]), name, finish_reason)
|
|
2470
|
+
if parse_error is not None:
|
|
2471
|
+
logger.warning(
|
|
2472
|
+
"%s: %s (finish_reason=%r)", name, parse_error, finish_reason)
|
|
2473
|
+
calls.append(ToolCallRequest(
|
|
2474
|
+
# `or str(uuid4())` for the Google branch's reason, which this
|
|
2475
|
+
# one was missing (batch 76). The fragment's id starts as ""
|
|
2476
|
+
# and is filled only if some delta carries one, so a provider
|
|
2477
|
+
# that streams parallel calls without ids hands back two
|
|
2478
|
+
# requests that are EQUAL on the field everything downstream
|
|
2479
|
+
# pairs by: the batch's own bookkeeping, `_spawn_threads` in
|
|
2480
|
+
# the TUI, and `parent_call_id` in storage. The generated id
|
|
2481
|
+
# is what goes back on the wire too, on BOTH halves of the
|
|
2482
|
+
# pair -- `add_assistant_message` persists the call and
|
|
2483
|
+
# `add_tool_result` answers it, so the two agree by
|
|
2484
|
+
# construction, which is all M4/D20 pairing needs. That is
|
|
2485
|
+
# also why "" was not the safer answer: a provider that
|
|
2486
|
+
# validates `tool_call_id` at all rejects an empty one.
|
|
2487
|
+
id=frag["id"] or str(uuid4()),
|
|
2488
|
+
name=name, input=parsed, parse_error=parse_error,
|
|
2489
|
+
))
|
|
2490
|
+
|
|
2491
|
+
# ONE block, not one per delta. There is no block structure on this
|
|
2492
|
+
# wire -- reasoning arrives as a flat string of deltas -- so the record
|
|
2493
|
+
# is the concatenation under the field name the request would use to
|
|
2494
|
+
# send it back. `type` names the shape for _messages_for_provider;
|
|
2495
|
+
# nothing about it is the provider's, which is why it is spelled after
|
|
2496
|
+
# the field rather than after Anthropic's "thinking".
|
|
2497
|
+
reasoning_blocks = ([{"type": "reasoning_content",
|
|
2498
|
+
"text": "".join(reasoning_parts)}]
|
|
2499
|
+
if reasoning_parts else [])
|
|
2500
|
+
|
|
2501
|
+
yield StreamToken(final_response=ModelResponse(
|
|
2502
|
+
text="".join(text_parts), tool_calls=calls, usage=usage,
|
|
2503
|
+
thinking=_thinking_record(provider_name, model, reasoning_blocks),
|
|
2504
|
+
))
|
|
2505
|
+
|
|
2506
|
+
|
|
2507
|
+
# ===========================================================================
|
|
2508
|
+
# ---- Embeddings (ROADMAP_v2 §45, SQ2/SQ7) ---------------------------------
|
|
2509
|
+
# ===========================================================================
|
|
2510
|
+
#
|
|
2511
|
+
# §33's rule is that this file is where a model call lives, and an
|
|
2512
|
+
# embedding is a model call. It shares `api_initialization` and its client
|
|
2513
|
+
# cache with the chat path, so an embedder on a provider already in use
|
|
2514
|
+
# costs no second client.
|
|
2515
|
+
#
|
|
2516
|
+
# WHAT THIS DELIBERATELY DOES NOT SHARE. No streaming, no tool
|
|
2517
|
+
# translation, no `effort`, no context-window check, and no entry in
|
|
2518
|
+
# `ModelResponse`: none of those mean anything for a call that takes
|
|
2519
|
+
# strings and returns vectors. `EmbeddingResult` is its own small shape
|
|
2520
|
+
# for the same reason ToolCallRequest is -- widening ModelResponse with
|
|
2521
|
+
# three fields that are null on every chat turn is how a response object
|
|
2522
|
+
# stops describing anything.
|
|
2523
|
+
|
|
2524
|
+
|
|
2525
|
+
class EmbeddingError(RuntimeError):
|
|
2526
|
+
"""An embedding call that could not produce usable vectors.
|
|
2527
|
+
|
|
2528
|
+
ITS OWN TYPE, so §45's scoring stage can catch exactly this and fall
|
|
2529
|
+
back to the model's own similarity (SQ10) without also swallowing a
|
|
2530
|
+
programming error from inside its own chunker. The message names the
|
|
2531
|
+
provider and the model, because the two most common causes -- a
|
|
2532
|
+
provider with no embeddings endpoint, and a chat model slug passed to
|
|
2533
|
+
one -- are both fixed by changing one of those two words.
|
|
2534
|
+
"""
|
|
2535
|
+
|
|
2536
|
+
|
|
2537
|
+
@dataclass
|
|
2538
|
+
class EmbeddingResult:
|
|
2539
|
+
"""Vectors for the texts that were sent, in the order they were sent.
|
|
2540
|
+
|
|
2541
|
+
`usage` mirrors ModelResponse's shape but carries input tokens only:
|
|
2542
|
+
an embedding call produces no output tokens, and reporting a zero
|
|
2543
|
+
under a key that means something elsewhere would put embedding spend
|
|
2544
|
+
into a total that is documented as billed input+output of model calls.
|
|
2545
|
+
§45 counts it separately and says so.
|
|
2546
|
+
"""
|
|
2547
|
+
|
|
2548
|
+
vectors: list = field(default_factory=list)
|
|
2549
|
+
model: str = ""
|
|
2550
|
+
provider_name: str = ""
|
|
2551
|
+
usage: dict = field(default_factory=lambda: {"input_tokens": 0})
|
|
2552
|
+
|
|
2553
|
+
@property
|
|
2554
|
+
def dimension(self) -> int:
|
|
2555
|
+
return len(self.vectors[0]) if self.vectors else 0
|
|
2556
|
+
|
|
2557
|
+
|
|
2558
|
+
# Providers with no embeddings endpoint at all, where the failure is worth
|
|
2559
|
+
# naming rather than surfacing as a 404 from inside a research run.
|
|
2560
|
+
#
|
|
2561
|
+
# ANTHROPIC IS NOT AN OVERSIGHT. There is no Anthropic embeddings API; the
|
|
2562
|
+
# recommended route is a third-party embedder, which this harness reaches
|
|
2563
|
+
# as its own provider entry. OPENROUTER proxies chat completions only.
|
|
2564
|
+
# CHATGPT is the plan, which answers on the Responses API and nothing else
|
|
2565
|
+
# (ROADMAP_v3 §83, CG3).
|
|
2566
|
+
#
|
|
2567
|
+
# INCOMPLETE ON PURPOSE, and safe when incomplete -- the same posture as
|
|
2568
|
+
# MODELS_REJECTING_SAMPLING_PARAMS. An unlisted provider that turns out to
|
|
2569
|
+
# have no endpoint fails loudly on the probe `/embedder` runs at selection
|
|
2570
|
+
# time, which is a measurement rather than a guess and is why there is no
|
|
2571
|
+
# `supports_embeddings` table to keep in step with the world.
|
|
2572
|
+
PROVIDERS_WITHOUT_EMBEDDINGS = frozenset({"ANTHROPIC", "OPENROUTER", "CHATGPT"})
|
|
2573
|
+
|
|
2574
|
+
|
|
2575
|
+
def embed_texts(provider_name: str, model: str, texts,
|
|
2576
|
+
*, prefix: str = "", task_type: str = "") -> EmbeddingResult:
|
|
2577
|
+
"""Embed `texts` with `model` on `provider_name`.
|
|
2578
|
+
|
|
2579
|
+
`prefix` is prepended to every text in this call. Asymmetric embedders
|
|
2580
|
+
-- e5's `"query: "` / `"passage: "`, BGE's instruction prefix -- score
|
|
2581
|
+
materially worse without it and produce no error at all, so it is
|
|
2582
|
+
configuration rather than something inferred from a model slug.
|
|
2583
|
+
`task_type` is Google's equivalent (`RETRIEVAL_QUERY` /
|
|
2584
|
+
`RETRIEVAL_DOCUMENT`) and is ignored elsewhere.
|
|
2585
|
+
|
|
2586
|
+
Raises EmbeddingError, never a provider exception: the caller's whole
|
|
2587
|
+
contract (SQ10) is to degrade to a documented fallback, and it cannot
|
|
2588
|
+
do that against an exception type that differs per SDK.
|
|
2589
|
+
"""
|
|
2590
|
+
items = [f"{prefix}{text}" for text in texts]
|
|
2591
|
+
if not items:
|
|
2592
|
+
return EmbeddingResult(model=model, provider_name=provider_name)
|
|
2593
|
+
|
|
2594
|
+
if provider_name in PROVIDERS_WITHOUT_EMBEDDINGS:
|
|
2595
|
+
raise EmbeddingError(
|
|
2596
|
+
f"{provider_name} has no embeddings endpoint. Configure an "
|
|
2597
|
+
f"embedder on a provider that does -- /embedder <PROVIDER> "
|
|
2598
|
+
f"<model> -- or leave it unset to score similarity with the "
|
|
2599
|
+
f"grounding model itself."
|
|
2600
|
+
)
|
|
2601
|
+
|
|
2602
|
+
try:
|
|
2603
|
+
client = api_initialization(provider_name)
|
|
2604
|
+
except ValueError as exc:
|
|
2605
|
+
raise EmbeddingError(str(exc)) from exc
|
|
2606
|
+
|
|
2607
|
+
try:
|
|
2608
|
+
if provider_name == "GOOGLE":
|
|
2609
|
+
vectors, tokens = _embed_google(client, model, items, task_type)
|
|
2610
|
+
else:
|
|
2611
|
+
vectors, tokens = _embed_openai_compatible(client, model, items)
|
|
2612
|
+
except EmbeddingError:
|
|
2613
|
+
raise
|
|
2614
|
+
except Exception as exc:
|
|
2615
|
+
raise EmbeddingError(
|
|
2616
|
+
f"{provider_name} embeddings call failed for model {model!r}: "
|
|
2617
|
+
f"{type(exc).__name__}: {exc}"
|
|
2618
|
+
) from exc
|
|
2619
|
+
|
|
2620
|
+
_check_vectors(provider_name, model, vectors, len(items))
|
|
2621
|
+
return EmbeddingResult(
|
|
2622
|
+
vectors=vectors, model=model, provider_name=provider_name,
|
|
2623
|
+
usage={"input_tokens": tokens},
|
|
2624
|
+
)
|
|
2625
|
+
|
|
2626
|
+
|
|
2627
|
+
def _embed_openai_compatible(client, model, items):
|
|
2628
|
+
"""`/v1/embeddings`, which is what every is_v1_compatible provider
|
|
2629
|
+
exposes if it exposes embeddings at all.
|
|
2630
|
+
|
|
2631
|
+
SORTED BY INDEX rather than trusted in arrival order. The response
|
|
2632
|
+
carries an explicit `index` per item precisely because the order is
|
|
2633
|
+
not guaranteed, and a silently permuted batch pairs every claim with
|
|
2634
|
+
another claim's vector -- a wrong number with no symptom, which is the
|
|
2635
|
+
failure this whole section exists to stop producing.
|
|
2636
|
+
"""
|
|
2637
|
+
response = client.embeddings.create(model=model, input=items)
|
|
2638
|
+
data = sorted(response.data, key=lambda item: getattr(item, "index", 0))
|
|
2639
|
+
vectors = [list(item.embedding) for item in data]
|
|
2640
|
+
usage = getattr(response, "usage", None)
|
|
2641
|
+
return vectors, int(getattr(usage, "prompt_tokens", 0) or 0)
|
|
2642
|
+
|
|
2643
|
+
|
|
2644
|
+
def _embed_google(client, model, items, task_type):
|
|
2645
|
+
"""google-genai's `embed_content`, which takes the batch as
|
|
2646
|
+
`contents` and returns `embeddings[].values`.
|
|
2647
|
+
|
|
2648
|
+
The config object is built ONLY when a task type was given, and
|
|
2649
|
+
through getattr rather than a direct import of the types module: a
|
|
2650
|
+
pinned SDK that does not carry EmbedContentConfig would otherwise turn
|
|
2651
|
+
an optional tuning parameter into an import error on every Google
|
|
2652
|
+
embedding call. D22's rule -- check the dependency's real shape, do
|
|
2653
|
+
not assume it.
|
|
2654
|
+
"""
|
|
2655
|
+
kwargs = {"model": model, "contents": items}
|
|
2656
|
+
if task_type:
|
|
2657
|
+
from google.genai import types as genai_types
|
|
2658
|
+
config_cls = getattr(genai_types, "EmbedContentConfig", None)
|
|
2659
|
+
if config_cls is not None:
|
|
2660
|
+
kwargs["config"] = config_cls(task_type=task_type)
|
|
2661
|
+
else:
|
|
2662
|
+
logger.warning(
|
|
2663
|
+
"Ignoring embedder task_type=%r: the installed google-genai "
|
|
2664
|
+
"has no EmbedContentConfig.", task_type)
|
|
2665
|
+
|
|
2666
|
+
response = client.models.embed_content(**kwargs)
|
|
2667
|
+
vectors = [list(item.values) for item in (response.embeddings or [])]
|
|
2668
|
+
# google-genai reports embedding token usage inconsistently across
|
|
2669
|
+
# releases; absent is reported as zero rather than guessed at.
|
|
2670
|
+
metadata = getattr(response, "metadata", None)
|
|
2671
|
+
return vectors, int(getattr(metadata, "billable_character_count", 0) or 0)
|
|
2672
|
+
|
|
2673
|
+
|
|
2674
|
+
def _check_vectors(provider_name, model, vectors, expected) -> None:
|
|
2675
|
+
"""Four ways a successful response is still unusable.
|
|
2676
|
+
|
|
2677
|
+
The first three produce a WRONG NUMBER rather than an error downstream:
|
|
2678
|
+
a short batch pairs texts with the wrong vectors, a ragged batch makes
|
|
2679
|
+
cosine compare different spaces, and an empty vector makes every
|
|
2680
|
+
cosine zero -- which reads as "no source supports this claim" and is
|
|
2681
|
+
indistinguishable from a real finding.
|
|
2682
|
+
|
|
2683
|
+
The fourth (batch 219, #218) is a component that is not a finite
|
|
2684
|
+
number. A string or None raised TypeError in `l2_normalize`, outside
|
|
2685
|
+
every handler, and flipped a run that had already grounded and scored
|
|
2686
|
+
to `status='failed'`; a NaN or an infinity raised nothing and scored
|
|
2687
|
+
NaN, the wrong-number class again. As an EmbeddingError it takes the
|
|
2688
|
+
path every other unusable response takes: the scorer gives up for the
|
|
2689
|
+
run and the stage falls back, with the cause in the trace.
|
|
2690
|
+
"""
|
|
2691
|
+
if len(vectors) != expected:
|
|
2692
|
+
raise EmbeddingError(
|
|
2693
|
+
f"{provider_name} returned {len(vectors)} embedding(s) for "
|
|
2694
|
+
f"{expected} input(s) with model {model!r}.")
|
|
2695
|
+
if not all(isinstance(v, (list, tuple))
|
|
2696
|
+
and all(_is_finite_number(c) for c in v) for v in vectors):
|
|
2697
|
+
raise EmbeddingError(
|
|
2698
|
+
f"{provider_name} returned an embedding whose components are not "
|
|
2699
|
+
f"all finite numbers for model {model!r}.")
|
|
2700
|
+
dimensions = {len(v) for v in vectors}
|
|
2701
|
+
if len(dimensions) > 1:
|
|
2702
|
+
raise EmbeddingError(
|
|
2703
|
+
f"{provider_name} returned embeddings of differing dimensions "
|
|
2704
|
+
f"{sorted(dimensions)} for model {model!r}.")
|
|
2705
|
+
if dimensions == {0}:
|
|
2706
|
+
raise EmbeddingError(
|
|
2707
|
+
f"{provider_name} returned empty embeddings for model "
|
|
2708
|
+
f"{model!r}.")
|
|
2709
|
+
|
|
2710
|
+
|
|
2711
|
+
def _is_finite_number(component) -> bool:
|
|
2712
|
+
"""An int or a float, not a bool, and not NaN or an infinity."""
|
|
2713
|
+
return (isinstance(component, (int, float))
|
|
2714
|
+
and not isinstance(component, bool)
|
|
2715
|
+
and math.isfinite(component))
|