@nodaro/shared 3.25.0 → 3.26.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/index.cjs +476 -114
- package/dist/index.cjs.map +1 -1
- package/dist/index.d.cts +160 -53
- package/dist/index.d.ts +160 -53
- package/dist/index.js +467 -115
- package/dist/index.js.map +1 -1
- package/package.json +1 -1
- package/src/__tests__/condition-variables.test.ts +12 -0
- package/src/__tests__/content-recipe-ideas.test.ts +9 -4
- package/src/__tests__/credit-identifiers.test.ts +12 -0
- package/src/__tests__/image-model-roles.test.ts +57 -0
- package/src/__tests__/image-ref-limit.test.ts +1 -0
- package/src/__tests__/llm-models.test.ts +136 -37
- package/src/__tests__/lone-media-url.test.ts +78 -0
- package/src/__tests__/prompt-length-limits.test.ts +2 -0
- package/src/condition-variables.ts +3 -1
- package/src/content-recipe-ideas.ts +5 -2
- package/src/image-model-roles.ts +62 -0
- package/src/index.ts +14 -0
- package/src/llm-models.ts +348 -23
- package/src/lone-media-url.ts +57 -0
- package/src/model-catalog.ts +39 -5
- package/src/model-constants.ts +11 -1
- package/src/node-default-mappings.ts +3 -0
- package/src/settings-input.ts +2 -1
package/src/llm-models.ts
CHANGED
|
@@ -15,7 +15,7 @@
|
|
|
15
15
|
|
|
16
16
|
export type LlmTier = "economy" | "standard" | "premium"
|
|
17
17
|
export type KieApiFormat = "chat-completions" | "messages" | "responses"
|
|
18
|
-
export type LlmVendor = "anthropic" | "google" | "openai" | "xai"
|
|
18
|
+
export type LlmVendor = "anthropic" | "deepseek" | "google" | "moonshot" | "openai" | "xai"
|
|
19
19
|
|
|
20
20
|
export const LLM_REASONING_EFFORTS = ["none", "low", "medium", "high", "xhigh", "max"] as const
|
|
21
21
|
export type LlmReasoningEffort = (typeof LLM_REASONING_EFFORTS)[number]
|
|
@@ -92,6 +92,43 @@ export interface LlmModelDef {
|
|
|
92
92
|
directReasoningEfforts?: readonly LlmReasoningEffort[]
|
|
93
93
|
/** false = model rejects `temperature` (Claude 5-era, GPT-5.6). Absent = accepts. */
|
|
94
94
|
supportsTemperature?: false
|
|
95
|
+
/**
|
|
96
|
+
* false = the model 400s on a FORCED `tool_choice` (`{ type: "tool" }` /
|
|
97
|
+
* `{ type: "any" }`) — Claude Sonnet 5.5 and Opus 5.5 (vendor docs: "not
|
|
98
|
+
* supported for this model"). Absent = forcing is accepted.
|
|
99
|
+
*
|
|
100
|
+
* Only meaningful with `structuredOutputMode: "anthropic-tool"`: such a model
|
|
101
|
+
* is asked with `tool_choice: auto` plus an instruction naming the tool, and
|
|
102
|
+
* the tool carries `strict: true` whenever its schema is expressible in the
|
|
103
|
+
* vendor's strict subset (`anthropicStrictToolSchema` in llm-client). The
|
|
104
|
+
* caller's Zod still validates every answer, so nothing the wire withholds is
|
|
105
|
+
* lost. Declared per model, never matched on a model name.
|
|
106
|
+
*/
|
|
107
|
+
supportsForcedToolChoice?: false
|
|
108
|
+
/**
|
|
109
|
+
* true = the model binds each thinking block to the exact conversation prefix
|
|
110
|
+
* that produced it ("preserved thinking" — Claude Sonnet 5.5 / Opus 5.5):
|
|
111
|
+
* replaying a block after an earlier turn was edited, dropped or rebuilt is a
|
|
112
|
+
* 400 on accounts the vendor enforces. A caller that trims or rebuilds stored
|
|
113
|
+
* history between requests (the Workflow Copilot keeps a token budget by
|
|
114
|
+
* dropping its oldest turns) must strip thinking blocks from the turns it
|
|
115
|
+
* replays. Within one request's own tool loop the history only grows, so
|
|
116
|
+
* blocks produced there stay valid.
|
|
117
|
+
*/
|
|
118
|
+
conversationBoundThinking?: true
|
|
119
|
+
/**
|
|
120
|
+
* true = this model's reasoning effort takes effect ONLY on the vendor's own
|
|
121
|
+
* API: the aggregator accepts the field and silently ignores it. Measured
|
|
122
|
+
* 2026-10-08 for the Claude family on KIE — `low`, `max` and no effort gave
|
|
123
|
+
* the same output length on a reasoning-heavy prompt, and no thinking block
|
|
124
|
+
* ever came back, even on Sonnet 4.6 at `max`.
|
|
125
|
+
*
|
|
126
|
+
* So an effort-bearing call on such a model is SERVED direct, and it bills
|
|
127
|
+
* as Advanced mode does — one rung up (decided 2026-10-08: "effort =
|
|
128
|
+
* Advanced"). {@link llmServesDirect} is the one place that decides it, for
|
|
129
|
+
* the credit identifier and the routing alike.
|
|
130
|
+
*/
|
|
131
|
+
effortRequiresDirect?: true
|
|
95
132
|
/** Claude-only: KIE is the preferred routing, direct Anthropic the fallback. */
|
|
96
133
|
preferKie?: true
|
|
97
134
|
/**
|
|
@@ -306,6 +343,9 @@ export const LLM_MODELS: readonly LlmModelDef[] = [
|
|
|
306
343
|
supportsImages: true,
|
|
307
344
|
maxOutputTokens: 8192,
|
|
308
345
|
directFallbackModel: "claude-haiku-4-5-20251001",
|
|
346
|
+
// KIE-first since 2026-10-08 (a call is priced on the lane it runs on);
|
|
347
|
+
// KIE served it plain and with a tool on its streaming wire that day.
|
|
348
|
+
preferKie: true,
|
|
309
349
|
},
|
|
310
350
|
{
|
|
311
351
|
id: "claude-sonnet-4.6",
|
|
@@ -319,7 +359,12 @@ export const LLM_MODELS: readonly LlmModelDef[] = [
|
|
|
319
359
|
supportsImages: true,
|
|
320
360
|
maxOutputTokens: 16384,
|
|
321
361
|
directFallbackModel: "claude-sonnet-4-6",
|
|
362
|
+
// KIE-first since 2026-10-08 (a call is priced on the lane it runs on);
|
|
363
|
+
// KIE served it plain and with a tool on its streaming wire that day.
|
|
364
|
+
preferKie: true,
|
|
322
365
|
reasoningEfforts: ["low", "medium", "high", "max"],
|
|
366
|
+
// KIE ignores Claude effort (measured 2026-10-08): an effort call runs direct and bills as Advanced.
|
|
367
|
+
effortRequiresDirect: true,
|
|
323
368
|
},
|
|
324
369
|
{
|
|
325
370
|
id: "gpt-5.2",
|
|
@@ -358,14 +403,14 @@ export const LLM_MODELS: readonly LlmModelDef[] = [
|
|
|
358
403
|
// text such as a name badge), which is exactly what video-analysis casts
|
|
359
404
|
// identities from — so adding an entry here would silently move analysis to
|
|
360
405
|
// the shallower lane. Don't. Full comparison lives with the analysis engine.
|
|
361
|
-
//
|
|
362
|
-
//
|
|
363
|
-
//
|
|
364
|
-
// the direct lane
|
|
365
|
-
// control, native media ingestion, and a `responseJsonSchema`
|
|
366
|
-
// `additionalProperties` (KIE's `response_format` silently DROPS
|
|
406
|
+
// KIE-first since 2026-10-08 (decided: a call is priced on the lane it runs
|
|
407
|
+
// on, so the base premium price buys the KIE lane; the direct lane is
|
|
408
|
+
// Advanced mode, billed `premium-direct`). It was the one direct-first
|
|
409
|
+
// Gemini before. What the direct lane buys, now behind Advanced: real
|
|
410
|
+
// `thinkingLevel` control, native media ingestion, and a `responseJsonSchema`
|
|
411
|
+
// that honours `additionalProperties` (KIE's `response_format` silently DROPS
|
|
367
412
|
// record/map-shaped fields — see the z.record rule in backend/CLAUDE.md).
|
|
368
|
-
|
|
413
|
+
// Video-analysis is unaffected: it pins `requireLane: "direct"` itself.
|
|
369
414
|
// Reasons with no thinking param sent on both lanes — the proxied endpoint
|
|
370
415
|
// DEFAULTS to "high" (above), and the direct lane reasons harder still.
|
|
371
416
|
// Floored at its own 16384: its KIE fallback is not known to take more.
|
|
@@ -389,6 +434,8 @@ export const LLM_MODELS: readonly LlmModelDef[] = [
|
|
|
389
434
|
maxOutputTokens: 16384,
|
|
390
435
|
directFallbackModel: "claude-opus-4-7",
|
|
391
436
|
reasoningEfforts: ["low", "medium", "high", "xhigh", "max"],
|
|
437
|
+
// KIE ignores Claude effort (measured 2026-10-08): an effort call runs direct and bills as Advanced.
|
|
438
|
+
effortRequiresDirect: true,
|
|
392
439
|
supportsTemperature: false,
|
|
393
440
|
preferKie: true,
|
|
394
441
|
},
|
|
@@ -538,10 +585,76 @@ export const LLM_MODELS: readonly LlmModelDef[] = [
|
|
|
538
585
|
// why this is per-model and not a lane-wide flag.
|
|
539
586
|
kieCollapseStream: true,
|
|
540
587
|
},
|
|
588
|
+
// ── GPT-6 Luna / Sol / 6.1 Sol ────────────────────────────────────────────
|
|
589
|
+
// Same KIE endpoint and dialect as gpt-6-astra (codex/v1/responses; the doc
|
|
590
|
+
// pages differ only in the model name). Live-probed 2026-10-08, per model:
|
|
591
|
+
// `text.format` json_schema ENFORCED (echo `strict: true`, schema-valid
|
|
592
|
+
// reply), `temperature: 0.2` echoed back as 1 → silently ignored, no
|
|
593
|
+
// reasoning param → echo "medium" (reasons by default), and data-URI vision.
|
|
594
|
+
// Unlike astra, the NON-stream lane served them reliably — luna 10/10 and sol
|
|
595
|
+
// 10/10, 6.1 sol 8/8 at levels other than `none` (incl. 4 long structured
|
|
596
|
+
// calls each, 500–1,500 output tokens) — so none of them is collapsed:
|
|
597
|
+
// `credits_consumed` stays the real provider charge.
|
|
598
|
+
// Their ladder is WIDER than astra's doc-enum one: `none` and `max` were each
|
|
599
|
+
// sent and echoed back by the endpoint (6.1 sol excepted below).
|
|
600
|
+
{
|
|
601
|
+
id: "gpt-6-luna",
|
|
602
|
+
displayName: "GPT-6 Luna",
|
|
603
|
+
desc: "Fastest GPT-6, high-volume workloads",
|
|
604
|
+
tier: "economy",
|
|
605
|
+
kieFormat: "responses",
|
|
606
|
+
kieSlugOrModel: "gpt-6-luna",
|
|
607
|
+
vendor: "openai",
|
|
608
|
+
structuredOutputMode: "responses-json-schema",
|
|
609
|
+
supportsImages: true,
|
|
610
|
+
maxOutputTokens: 16384,
|
|
611
|
+
reasoningEfforts: ["none", "low", "medium", "high", "xhigh", "max"],
|
|
612
|
+
supportsTemperature: false,
|
|
613
|
+
thinkingDefaultOn: true,
|
|
614
|
+
},
|
|
615
|
+
{
|
|
616
|
+
id: "gpt-6-sol",
|
|
617
|
+
displayName: "GPT-6 Sol",
|
|
618
|
+
desc: "Strong GPT-6, deep reasoning at lower cost",
|
|
619
|
+
// Premium by decision (2026-10-08), not by unit cost — "Sol" is the
|
|
620
|
+
// deep-reasoning rung of the family, and gpt-5.6-sol bills premium too.
|
|
621
|
+
tier: "premium",
|
|
622
|
+
kieFormat: "responses",
|
|
623
|
+
kieSlugOrModel: "gpt-6-sol",
|
|
624
|
+
vendor: "openai",
|
|
625
|
+
structuredOutputMode: "responses-json-schema",
|
|
626
|
+
supportsImages: true,
|
|
627
|
+
maxOutputTokens: 16384,
|
|
628
|
+
reasoningEfforts: ["none", "low", "medium", "high", "xhigh", "max"],
|
|
629
|
+
supportsTemperature: false,
|
|
630
|
+
thinkingDefaultOn: true,
|
|
631
|
+
},
|
|
632
|
+
{
|
|
633
|
+
id: "gpt-6.1-sol",
|
|
634
|
+
displayName: "GPT-6.1 Sol",
|
|
635
|
+
// Measured slow: 30–76 s for 500–1,100 output tokens (2026-10-08), roughly
|
|
636
|
+
// twice gpt-6-sol on the same request — say so where the picker shows it.
|
|
637
|
+
desc: "Newest GPT-6 Sol, thorough but slower",
|
|
638
|
+
tier: "premium",
|
|
639
|
+
kieFormat: "responses",
|
|
640
|
+
kieSlugOrModel: "gpt-6-1-sol",
|
|
641
|
+
vendor: "openai",
|
|
642
|
+
structuredOutputMode: "responses-json-schema",
|
|
643
|
+
supportsImages: true,
|
|
644
|
+
maxOutputTokens: 16384,
|
|
645
|
+
// NO `none`: every probe at `reasoning.effort: none` failed (502, a 120 s
|
|
646
|
+
// timeout, 502, 429 — 0/4, 2026-10-08) while every other level answered,
|
|
647
|
+
// and the same request at `none` succeeded on luna and sol. Clamping maps a
|
|
648
|
+
// `none` request to nothing sent (Auto) rather than to a level that fails.
|
|
649
|
+
reasoningEfforts: ["low", "medium", "high", "xhigh", "max"],
|
|
650
|
+
supportsTemperature: false,
|
|
651
|
+
thinkingDefaultOn: true,
|
|
652
|
+
},
|
|
541
653
|
{
|
|
542
654
|
id: "grok-4.6",
|
|
543
655
|
displayName: "Grok 4.6",
|
|
544
|
-
|
|
656
|
+
// Demoted 2026-10-08 when grok-4.7 registered — the flagship copy moved up.
|
|
657
|
+
desc: "Previous Grok, strong reasoning",
|
|
545
658
|
tier: "standard",
|
|
546
659
|
kieFormat: "responses",
|
|
547
660
|
// KIE serves Grok on the responses dialect under its own family path —
|
|
@@ -567,10 +680,35 @@ export const LLM_MODELS: readonly LlmModelDef[] = [
|
|
|
567
680
|
// call needs output headroom, not just xhigh.
|
|
568
681
|
thinkingDefaultOn: true,
|
|
569
682
|
},
|
|
683
|
+
{
|
|
684
|
+
id: "grok-4.7",
|
|
685
|
+
displayName: "Grok 4.7",
|
|
686
|
+
desc: "xAI flagship, strong reasoning",
|
|
687
|
+
tier: "standard",
|
|
688
|
+
kieFormat: "responses",
|
|
689
|
+
// Same grok/v1/responses family path and wire as grok-4.6 (the KIE doc
|
|
690
|
+
// pages differ only in the model name). Live-probed 2026-10-08: non-stream
|
|
691
|
+
// 6/6, `text.format` json_schema enforced, data-URI vision, SSE deltas.
|
|
692
|
+
kieSlugOrModel: "grok-4-7",
|
|
693
|
+
vendor: "xai",
|
|
694
|
+
structuredOutputMode: "responses-json-schema",
|
|
695
|
+
supportsImages: true,
|
|
696
|
+
maxOutputTokens: 16384,
|
|
697
|
+
// 4.6's ladder. `none` is NOT honoured: the endpoint echoed it back as
|
|
698
|
+
// "minimal" and still reasoned (33 of 35 output tokens), so offering it
|
|
699
|
+
// would promise a lever the wire overrides.
|
|
700
|
+
reasoningEfforts: ["low", "medium", "high", "xhigh"],
|
|
701
|
+
// `temperature: 0.2` → echo stayed at the 0.7 default: ignored.
|
|
702
|
+
supportsTemperature: false,
|
|
703
|
+
// "Capital of France" with no reasoning param spent 310 of 311 output
|
|
704
|
+
// tokens reasoning.
|
|
705
|
+
thinkingDefaultOn: true,
|
|
706
|
+
},
|
|
570
707
|
{
|
|
571
708
|
id: "claude-sonnet-5",
|
|
572
709
|
displayName: "Claude Sonnet 5",
|
|
573
|
-
|
|
710
|
+
// Demoted 2026-10-08 when claude-sonnet-5.5 registered.
|
|
711
|
+
desc: "Previous Sonnet, near-Opus quality",
|
|
574
712
|
tier: "standard",
|
|
575
713
|
kieFormat: "messages",
|
|
576
714
|
kieSlugOrModel: "claude-sonnet-5",
|
|
@@ -580,9 +718,43 @@ export const LLM_MODELS: readonly LlmModelDef[] = [
|
|
|
580
718
|
maxOutputTokens: 16384,
|
|
581
719
|
directFallbackModel: "claude-sonnet-5",
|
|
582
720
|
reasoningEfforts: ["low", "medium", "high", "xhigh", "max"],
|
|
721
|
+
// KIE ignores Claude effort (measured 2026-10-08): an effort call runs direct and bills as Advanced.
|
|
722
|
+
effortRequiresDirect: true,
|
|
583
723
|
supportsTemperature: false,
|
|
584
724
|
preferKie: true,
|
|
585
725
|
},
|
|
726
|
+
{
|
|
727
|
+
id: "claude-sonnet-5.5",
|
|
728
|
+
displayName: "Claude Sonnet 5.5",
|
|
729
|
+
desc: "Latest Sonnet, near-Opus quality at Sonnet cost",
|
|
730
|
+
tier: "standard",
|
|
731
|
+
kieFormat: "messages",
|
|
732
|
+
// KIE serves it on the Claude-native messages dialect under its plain id
|
|
733
|
+
// (docs.kie.ai/market/claude/claude-sonnet-5-5.md — the page is
|
|
734
|
+
// claude-opus-5's with the model name swapped). Live-probed 2026-10-08 on
|
|
735
|
+
// the streaming wire: plain, effort, and forced/auto+strict tool calls all
|
|
736
|
+
// answered.
|
|
737
|
+
kieSlugOrModel: "claude-sonnet-5-5",
|
|
738
|
+
vendor: "anthropic",
|
|
739
|
+
structuredOutputMode: "anthropic-tool",
|
|
740
|
+
// Anthropic 400s a forced tool_choice on this model. KIE's proxy happened
|
|
741
|
+
// to accept one in the probe, but the direct lane is where every
|
|
742
|
+
// effort-bearing and every streamed structured call goes.
|
|
743
|
+
supportsForcedToolChoice: false,
|
|
744
|
+
conversationBoundThinking: true,
|
|
745
|
+
supportsImages: true,
|
|
746
|
+
maxOutputTokens: 16384,
|
|
747
|
+
directFallbackModel: "claude-sonnet-5-5",
|
|
748
|
+
reasoningEfforts: ["low", "medium", "high", "xhigh", "max"],
|
|
749
|
+
// KIE ignores Claude effort (measured 2026-10-08): an effort call runs direct and bills as Advanced.
|
|
750
|
+
effortRequiresDirect: true,
|
|
751
|
+
// Anthropic rejects non-default sampling values on this model (KIE's proxy
|
|
752
|
+
// accepted temperature: 0.2 in the probe; the direct lane is the contract).
|
|
753
|
+
supportsTemperature: false,
|
|
754
|
+
preferKie: true,
|
|
755
|
+
// Adaptive thinking runs when no thinking param is sent (vendor default).
|
|
756
|
+
thinkingDefaultOn: true,
|
|
757
|
+
},
|
|
586
758
|
{
|
|
587
759
|
id: "claude-opus-4.8",
|
|
588
760
|
displayName: "Claude Opus 4.8",
|
|
@@ -596,13 +768,16 @@ export const LLM_MODELS: readonly LlmModelDef[] = [
|
|
|
596
768
|
maxOutputTokens: 16384,
|
|
597
769
|
directFallbackModel: "claude-opus-4-8",
|
|
598
770
|
reasoningEfforts: ["low", "medium", "high", "xhigh", "max"],
|
|
771
|
+
// KIE ignores Claude effort (measured 2026-10-08): an effort call runs direct and bills as Advanced.
|
|
772
|
+
effortRequiresDirect: true,
|
|
599
773
|
supportsTemperature: false,
|
|
600
774
|
preferKie: true,
|
|
601
775
|
},
|
|
602
776
|
{
|
|
603
777
|
id: "claude-opus-5",
|
|
604
778
|
displayName: "Claude Opus 5",
|
|
605
|
-
|
|
779
|
+
// Demoted 2026-10-08 when claude-opus-5.5 registered.
|
|
780
|
+
desc: "Previous Opus, deep agentic reasoning",
|
|
606
781
|
tier: "premium",
|
|
607
782
|
kieFormat: "messages",
|
|
608
783
|
// KIE serves Opus 5 on the Claude-native messages dialect under the plain
|
|
@@ -615,12 +790,44 @@ export const LLM_MODELS: readonly LlmModelDef[] = [
|
|
|
615
790
|
maxOutputTokens: 16384,
|
|
616
791
|
directFallbackModel: "claude-opus-5",
|
|
617
792
|
reasoningEfforts: ["low", "medium", "high", "xhigh", "max"],
|
|
793
|
+
// KIE ignores Claude effort (measured 2026-10-08): an effort call runs direct and bills as Advanced.
|
|
794
|
+
effortRequiresDirect: true,
|
|
618
795
|
supportsTemperature: false,
|
|
619
796
|
preferKie: true,
|
|
620
797
|
// Opus 5 reasons with NO thinking param sent (vendor default flipped from
|
|
621
798
|
// Opus 4.8/4.7) — so every call needs output headroom, not just xhigh/max.
|
|
622
799
|
thinkingDefaultOn: true,
|
|
623
800
|
},
|
|
801
|
+
{
|
|
802
|
+
id: "claude-opus-5.5",
|
|
803
|
+
displayName: "Claude Opus 5.5",
|
|
804
|
+
// Positioned as Fable-class (decided 2026-10-08: offered wherever Fable is).
|
|
805
|
+
desc: "Latest Opus, frontier-class reasoning",
|
|
806
|
+
tier: "premium",
|
|
807
|
+
kieFormat: "messages",
|
|
808
|
+
// KIE plain id on the messages dialect (docs.kie.ai/market/claude/
|
|
809
|
+
// claude-opus-5-5 — claude-opus-5's page with the name swapped).
|
|
810
|
+
// Live-probed 2026-10-08 on the streaming wire (plain + effort); the KIE
|
|
811
|
+
// Claude lane was erroring for claude-opus-5 too during the tool probes.
|
|
812
|
+
kieSlugOrModel: "claude-opus-5-5",
|
|
813
|
+
vendor: "anthropic",
|
|
814
|
+
structuredOutputMode: "anthropic-tool",
|
|
815
|
+
// Forced tool_choice is a 400 on this model (vendor docs) — see the field.
|
|
816
|
+
supportsForcedToolChoice: false,
|
|
817
|
+
conversationBoundThinking: true,
|
|
818
|
+
supportsImages: true,
|
|
819
|
+
maxOutputTokens: 16384,
|
|
820
|
+
directFallbackModel: "claude-opus-5-5",
|
|
821
|
+
reasoningEfforts: ["low", "medium", "high", "xhigh", "max"],
|
|
822
|
+
// KIE ignores Claude effort (measured 2026-10-08): an effort call runs direct and bills as Advanced.
|
|
823
|
+
effortRequiresDirect: true,
|
|
824
|
+
supportsTemperature: false,
|
|
825
|
+
preferKie: true,
|
|
826
|
+
// Thinking cannot be turned off on this model at all (`disabled` is a 400
|
|
827
|
+
// at every effort); with no effort sent it runs adaptive at the vendor's
|
|
828
|
+
// `medium` default. Nothing on our wire sends `disabled`.
|
|
829
|
+
thinkingDefaultOn: true,
|
|
830
|
+
},
|
|
624
831
|
{
|
|
625
832
|
id: "claude-fable-5",
|
|
626
833
|
displayName: "Claude Fable 5",
|
|
@@ -634,9 +841,61 @@ export const LLM_MODELS: readonly LlmModelDef[] = [
|
|
|
634
841
|
maxOutputTokens: 16384,
|
|
635
842
|
directFallbackModel: "claude-fable-5",
|
|
636
843
|
reasoningEfforts: ["low", "medium", "high", "xhigh", "max"],
|
|
844
|
+
// KIE ignores Claude effort (measured 2026-10-08): an effort call runs direct and bills as Advanced.
|
|
845
|
+
effortRequiresDirect: true,
|
|
637
846
|
supportsTemperature: false,
|
|
638
847
|
preferKie: true,
|
|
639
848
|
},
|
|
849
|
+
// ── Moonshot / DeepSeek ───────────────────────────────────────────────────
|
|
850
|
+
// Both on KIE's THIRD responses family path, openai/v1/responses (llm-client
|
|
851
|
+
// derives it from `vendor`). KIE labels both pages "Chat only — not adapted
|
|
852
|
+
// for agent use": fine for single (structured) completions, so neither may
|
|
853
|
+
// back a tool-loop feature.
|
|
854
|
+
{
|
|
855
|
+
id: "kimi-k3",
|
|
856
|
+
displayName: "Kimi K3",
|
|
857
|
+
desc: "Moonshot flagship, 1M context, careful reasoning",
|
|
858
|
+
tier: "premium",
|
|
859
|
+
kieFormat: "responses",
|
|
860
|
+
kieSlugOrModel: "kimi-k3",
|
|
861
|
+
vendor: "moonshot",
|
|
862
|
+
// Live-probed 2026-10-08: `text.format` json_schema is ENFORCED — asked to
|
|
863
|
+
// answer "no json here" in plain prose, it returned schema-shaped JSON.
|
|
864
|
+
structuredOutputMode: "responses-json-schema",
|
|
865
|
+
supportsImages: true,
|
|
866
|
+
maxOutputTokens: 16384,
|
|
867
|
+
// KIE's documented enum. The endpoint accepted xhigh/none without error but
|
|
868
|
+
// never echoes an effort back, so nothing beyond the doc is verifiable.
|
|
869
|
+
reasoningEfforts: ["low", "medium", "high"],
|
|
870
|
+
// The doc lists no `temperature` field at all.
|
|
871
|
+
supportsTemperature: false,
|
|
872
|
+
// "Reasons by default" per KIE's doc (its usage reports no reasoning-token
|
|
873
|
+
// split, so this is the doc's word, not a measurement).
|
|
874
|
+
thinkingDefaultOn: true,
|
|
875
|
+
},
|
|
876
|
+
{
|
|
877
|
+
id: "deepseek-v4.1-flash",
|
|
878
|
+
displayName: "DeepSeek V4.1 Flash",
|
|
879
|
+
desc: "Very cheap and fast, solid reasoning",
|
|
880
|
+
tier: "economy",
|
|
881
|
+
kieFormat: "responses",
|
|
882
|
+
kieSlugOrModel: "deepseek-v4-1-flash",
|
|
883
|
+
vendor: "deepseek",
|
|
884
|
+
// Live-probed 2026-10-08: json_schema echoed `strict: true` and enforced.
|
|
885
|
+
structuredOutputMode: "responses-json-schema",
|
|
886
|
+
supportsImages: true,
|
|
887
|
+
// ADVISORY ONLY: KIE documents `max_output_tokens` as "not supported … no
|
|
888
|
+
// effect" on this model (a probe sending it was accepted and ignored), so
|
|
889
|
+
// neither this cap nor the reasoning floor binds its output length, and the
|
|
890
|
+
// `incomplete` cap-stop signal cannot fire from our own cap.
|
|
891
|
+
maxOutputTokens: 16384,
|
|
892
|
+
// The endpoint takes none/minimal/low/medium/high/xhigh/max; each of none,
|
|
893
|
+
// low and xhigh was echoed back. `minimal` is not in our ladder.
|
|
894
|
+
reasoningEfforts: ["none", "low", "medium", "high", "xhigh", "max"],
|
|
895
|
+
// Documented "not supported … no effect".
|
|
896
|
+
supportsTemperature: false,
|
|
897
|
+
thinkingDefaultOn: true,
|
|
898
|
+
},
|
|
640
899
|
] as const
|
|
641
900
|
|
|
642
901
|
export const LLM_MODEL_IDS = LLM_MODELS.map((m) => m.id)
|
|
@@ -658,10 +917,12 @@ export const STRUCTURED_VISION_MODELS = LLM_MODELS.filter(
|
|
|
658
917
|
* ship with its models silently sorted to the end of every menu unlabeled.
|
|
659
918
|
* Alphabetical on purpose: stable, and no vendor-preference fights.
|
|
660
919
|
*/
|
|
661
|
-
export const LLM_VENDOR_ORDER: readonly LlmVendor[] = ["anthropic", "google", "openai", "xai"]
|
|
920
|
+
export const LLM_VENDOR_ORDER: readonly LlmVendor[] = ["anthropic", "deepseek", "google", "moonshot", "openai", "xai"]
|
|
662
921
|
export const LLM_VENDOR_LABELS: Record<LlmVendor, string> = {
|
|
663
922
|
anthropic: "Anthropic",
|
|
923
|
+
deepseek: "DeepSeek",
|
|
664
924
|
google: "Google",
|
|
925
|
+
moonshot: "Moonshot AI",
|
|
665
926
|
openai: "OpenAI",
|
|
666
927
|
xai: "xAI",
|
|
667
928
|
}
|
|
@@ -763,7 +1024,9 @@ export const LLM_FEATURE_DEFAULTS: Record<LlmFeature, string> = {
|
|
|
763
1024
|
// runs once per returned ad — volume, not depth. Must stay an image-capable
|
|
764
1025
|
// structured-output model (STRUCTURED_VISION_MODELS); a registry test pins it.
|
|
765
1026
|
"meta-ads-analysis": "gemini-3.6-flash",
|
|
766
|
-
|
|
1027
|
+
// Moved from claude-opus-5 on 2026-10-08 (decided: Opus 5.5 takes the Opus 5
|
|
1028
|
+
// defaults). Still an image-capable structured model — a registry test pins it.
|
|
1029
|
+
"describe-to-picker": "claude-opus-5.5",
|
|
767
1030
|
"qa-check": "gemini-3.6-flash",
|
|
768
1031
|
"generate-script": "gemini-3.6-flash",
|
|
769
1032
|
"translate": "gemini-3.6-flash",
|
|
@@ -816,11 +1079,21 @@ export const LLM_MODALITY_CAPS: Record<string, { image: boolean; video: boolean;
|
|
|
816
1079
|
"gpt-5.6-sol": { image: true, video: false, audio: false },
|
|
817
1080
|
// KIE's GPT-6 doc lists text + image + file inputs only — no video, no audio.
|
|
818
1081
|
"gpt-6-astra": { image: true, video: false, audio: false },
|
|
1082
|
+
// Same KIE doc as astra: text + image + file inputs only.
|
|
1083
|
+
"gpt-6-luna": { image: true, video: false, audio: false },
|
|
1084
|
+
"gpt-6-sol": { image: true, video: false, audio: false },
|
|
1085
|
+
"gpt-6.1-sol": { image: true, video: false, audio: false },
|
|
819
1086
|
"grok-4.6": { image: true, video: false, audio: false },
|
|
1087
|
+
"grok-4.7": { image: true, video: false, audio: false },
|
|
820
1088
|
"claude-sonnet-5": { image: true, video: false, audio: false },
|
|
1089
|
+
"claude-sonnet-5.5": { image: true, video: false, audio: false },
|
|
821
1090
|
"claude-opus-4.8": { image: true, video: false, audio: false },
|
|
822
1091
|
"claude-opus-5": { image: true, video: false, audio: false },
|
|
1092
|
+
"claude-opus-5.5": { image: true, video: false, audio: false },
|
|
823
1093
|
"claude-fable-5": { image: true, video: false, audio: false },
|
|
1094
|
+
// KIE documents text + image input only for both.
|
|
1095
|
+
"kimi-k3": { image: true, video: false, audio: false },
|
|
1096
|
+
"deepseek-v4.1-flash": { image: true, video: false, audio: false },
|
|
824
1097
|
}
|
|
825
1098
|
|
|
826
1099
|
/** Capability lookup with safe default — unknown models get image-only. */
|
|
@@ -971,13 +1244,59 @@ function bumpTier(tier: LlmTier): LlmTier {
|
|
|
971
1244
|
* rather than on a hand-maintained model list.
|
|
972
1245
|
*/
|
|
973
1246
|
export function supportsAdvancedMode(modelId: string | undefined): boolean {
|
|
974
|
-
|
|
1247
|
+
const model = modelId ? getLlmModel(modelId) : undefined
|
|
1248
|
+
return Boolean(model?.directGeminiModel || model?.directFallbackModel)
|
|
975
1249
|
}
|
|
976
1250
|
|
|
977
1251
|
/** User-facing reason a model can't offer Advanced mode. Single-sourced so the
|
|
978
1252
|
* config panel's disabled hint and the route's 400 say the same thing. */
|
|
979
1253
|
export const ADVANCED_MODE_UNAVAILABLE_REASON =
|
|
980
|
-
"Advanced mode is available on Gemini models — switch the model to enable it."
|
|
1254
|
+
"Advanced mode is available on Gemini and Claude models — switch the model to enable it."
|
|
1255
|
+
|
|
1256
|
+
/**
|
|
1257
|
+
* Is this call served on the vendor's own API — and therefore billed as such?
|
|
1258
|
+
*
|
|
1259
|
+
* The rule (decided 2026-10-08): a call is priced on the lane it runs on.
|
|
1260
|
+
* It runs direct when Advanced mode is on, or when it carries a reasoning
|
|
1261
|
+
* effort on a model whose effort only works there (`effortRequiresDirect` —
|
|
1262
|
+
* the Claude family). Everything else runs on the aggregator at its price,
|
|
1263
|
+
* falling back to the vendor's API only if the aggregator fails (at no extra
|
|
1264
|
+
* charge).
|
|
1265
|
+
*
|
|
1266
|
+
* The SINGLE decision both the credit identifier and the LLM client's routing
|
|
1267
|
+
* read, so what is billed and where it runs cannot disagree. Ignores a stale
|
|
1268
|
+
* `advancedMode` on a model with no direct lane, like the bump always has.
|
|
1269
|
+
*/
|
|
1270
|
+
export function llmServesDirect(
|
|
1271
|
+
modelId: string | undefined,
|
|
1272
|
+
reasoningEffort?: string,
|
|
1273
|
+
advancedMode?: boolean,
|
|
1274
|
+
): boolean {
|
|
1275
|
+
if (!supportsAdvancedMode(modelId)) return false
|
|
1276
|
+
if (advancedMode) return true
|
|
1277
|
+
return Boolean(
|
|
1278
|
+
getLlmModel(modelId!)?.effortRequiresDirect && effectiveReasoningEffort(modelId, reasoningEffort) !== undefined,
|
|
1279
|
+
)
|
|
1280
|
+
}
|
|
1281
|
+
|
|
1282
|
+
/**
|
|
1283
|
+
* The credit rungs an LLM feature bills on, cheapest first. `premium-direct` is
|
|
1284
|
+
* reached only by a premium model served direct (decided 2026-10-08): the bump
|
|
1285
|
+
* for running direct moves one rung on THIS ladder, so a premium model has
|
|
1286
|
+
* somewhere to go. The effort bump keeps the three-rung ladder.
|
|
1287
|
+
*/
|
|
1288
|
+
export const LLM_CREDIT_RUNGS = ["economy", "standard", "premium", "premium-direct"] as const
|
|
1289
|
+
export type LlmCreditRung = (typeof LLM_CREDIT_RUNGS)[number]
|
|
1290
|
+
|
|
1291
|
+
/** The credit identifier for `feature` on one rung — the standard rung is the bare feature id. */
|
|
1292
|
+
export function llmCreditIdForRung(feature: string, rung: LlmCreditRung): string {
|
|
1293
|
+
return rung === "standard" ? feature : `${feature}:${rung}`
|
|
1294
|
+
}
|
|
1295
|
+
|
|
1296
|
+
/** Every credit identifier a tier-priced LLM feature can bill under, in rung order. */
|
|
1297
|
+
export function llmTierCreditIds(feature: string): string[] {
|
|
1298
|
+
return LLM_CREDIT_RUNGS.map((rung) => llmCreditIdForRung(feature, rung))
|
|
1299
|
+
}
|
|
981
1300
|
|
|
982
1301
|
export function buildLlmCreditIdentifier(
|
|
983
1302
|
feature: string,
|
|
@@ -989,14 +1308,20 @@ export function buildLlmCreditIdentifier(
|
|
|
989
1308
|
let tier = getLlmTier(modelId)
|
|
990
1309
|
const eff = effectiveReasoningEffort(modelId, reasoningEffort)
|
|
991
1310
|
if (eff !== undefined && EFFORT_TIER_BUMP.has(eff)) tier = bumpTier(tier)
|
|
992
|
-
//
|
|
993
|
-
//
|
|
994
|
-
//
|
|
995
|
-
//
|
|
996
|
-
//
|
|
997
|
-
|
|
998
|
-
|
|
999
|
-
|
|
1311
|
+
// A call served on the vendor's own API (Advanced mode, or an effort on a
|
|
1312
|
+
// model whose effort only works there — llmServesDirect) bills materially
|
|
1313
|
+
// more per token than the aggregator. It bumps INDEPENDENTLY of the effort
|
|
1314
|
+
// bump — the two are separate cost levers and genuinely stack, so a
|
|
1315
|
+
// max-effort advanced economy call lands at premium — and on the four-rung
|
|
1316
|
+
// ladder, so a premium model served direct lands at `premium-direct`. A stale
|
|
1317
|
+
// flag on a model with no direct lane never inflates a bill.
|
|
1318
|
+
const rung: LlmCreditRung = llmServesDirect(modelId, reasoningEffort, advancedMode) ? bumpToDirectRung(tier) : tier
|
|
1319
|
+
return llmCreditIdForRung(feature, rung)
|
|
1320
|
+
}
|
|
1321
|
+
|
|
1322
|
+
/** One rung up the four-rung ladder — the direct-lane bump. */
|
|
1323
|
+
function bumpToDirectRung(tier: LlmTier): LlmCreditRung {
|
|
1324
|
+
return tier === "premium" ? "premium-direct" : bumpTier(tier)
|
|
1000
1325
|
}
|
|
1001
1326
|
|
|
1002
1327
|
/**
|
|
@@ -0,0 +1,57 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* The ONE rule both engines use when a text-shaped wire reaches a social post:
|
|
3
|
+
* a value that is exactly one media LINK — and nothing else — carries the
|
|
4
|
+
* medium, not the caption. An Extract Field over a saved record's cover posts
|
|
5
|
+
* the picture; a page link, a sentence with a link in it, a JSON value or a
|
|
6
|
+
* format no platform posts stays the caption (decided 2026-10-08).
|
|
7
|
+
*
|
|
8
|
+
* The file name is read from the parsed URL's PATH, anchored at its end: a host
|
|
9
|
+
* that merely contains a file ending (`movistar.es`, `webmd.com`, `gifs.com`)
|
|
10
|
+
* or a file name in the query (`?img=a.png`) is a page link. Only the formats
|
|
11
|
+
* the platforms accept count — avif / heic / svg / bmp are not pictures to
|
|
12
|
+
* Telegram, X or Instagram and would fail the post.
|
|
13
|
+
*/
|
|
14
|
+
import { AUDIO_PRODUCER_TYPES, DYNAMIC_PRODUCER_TYPES, IMAGE_PRODUCER_TYPES, VIDEO_PRODUCER_TYPES } from "./producer-types.js"
|
|
15
|
+
import { ENTITY_NODE_KINDS } from "./entity-node-fields.js"
|
|
16
|
+
|
|
17
|
+
export type LoneMediaUrlKind = "image" | "video"
|
|
18
|
+
|
|
19
|
+
const IMAGE_FILE_RE = /\.(png|jpe?g|gif|webp)$/i
|
|
20
|
+
const VIDEO_FILE_RE = /\.(mp4|mov|webm|m4v)$/i
|
|
21
|
+
|
|
22
|
+
/** Which medium a text value is when its WHOLE value is one media link; `null`
|
|
23
|
+
* for anything else (empty, several words, a comma-joined pair of links, a
|
|
24
|
+
* page link, a file ending in the host or the query, an unsupported format). */
|
|
25
|
+
export function loneMediaUrlKind(value: string): LoneMediaUrlKind | null {
|
|
26
|
+
const s = value.trim()
|
|
27
|
+
if (s.length === 0 || /[\s,]/.test(s) || !/^https?:\/\//i.test(s)) return null
|
|
28
|
+
let pathname: string
|
|
29
|
+
try {
|
|
30
|
+
pathname = new URL(s).pathname
|
|
31
|
+
} catch {
|
|
32
|
+
return null
|
|
33
|
+
}
|
|
34
|
+
if (IMAGE_FILE_RE.test(pathname)) return "image"
|
|
35
|
+
if (VIDEO_FILE_RE.test(pathname)) return "video"
|
|
36
|
+
return null
|
|
37
|
+
}
|
|
38
|
+
|
|
39
|
+
/** A trigger whose ONE wire carries a message's text AND its medium (the
|
|
40
|
+
* caption and the photo of an incoming Telegram message): both engines route
|
|
41
|
+
* the two together from the message itself, so the wire's single value is
|
|
42
|
+
* never read as "one media link". */
|
|
43
|
+
const MESSAGE_TRIGGER_TYPES: ReadonlySet<string> = new Set(["telegram-trigger", "telegram-account-trigger"])
|
|
44
|
+
|
|
45
|
+
/** A source whose wire a social post reads as TEXT: not a declared image, video
|
|
46
|
+
* or audio producer, not a dynamic (multi-typed) one such as List — whose typed
|
|
47
|
+
* columns route by column type — not a message trigger, and not an entity
|
|
48
|
+
* whose picture is a reference. Derived from the producer sets, so a new text
|
|
49
|
+
* node is covered and a new media node (which must join its producer set to
|
|
50
|
+
* connect at all) is excluded, with nothing to remember. */
|
|
51
|
+
export function isLoneMediaLinkSource(nodeType: string | null | undefined): boolean {
|
|
52
|
+
if (!nodeType) return false
|
|
53
|
+
if (IMAGE_PRODUCER_TYPES.has(nodeType) || VIDEO_PRODUCER_TYPES.has(nodeType) || AUDIO_PRODUCER_TYPES.has(nodeType)) return false
|
|
54
|
+
if (DYNAMIC_PRODUCER_TYPES.has(nodeType) || MESSAGE_TRIGGER_TYPES.has(nodeType)) return false
|
|
55
|
+
if ((ENTITY_NODE_KINDS as readonly string[]).includes(nodeType) || nodeType === "face") return false
|
|
56
|
+
return true
|
|
57
|
+
}
|