@nodaro/shared 2.22.0 → 2.24.0

package/dist/index.js

@@ -1981,11 +1981,11 @@ var VIDEO_MODELS = {
     description: "Legacy fast-tier analysis model (pre-2026-07). Kept so stored raw-model configs keep running and keep pricing under their own identifier; new fast-tier runs use the current fast model.",
     useCases: ["video-analysis", "shot-list", "fast"],
     pricing: [
-      { identifier: "video-analysis:gemini-3-flash", credits: 848, note: "10-min ceiling (no duration given)" },
+      { identifier: "video-analysis:gemini-3-flash", credits: 849, note: "10-min ceiling (no duration given)" },
       { identifier: "video-analysis:gemini-3-flash:60s", credits: 181 },
-      { identifier: "video-analysis:gemini-3-flash:180s", credits: 185 },
-      { identifier: "video-analysis:gemini-3-flash:360s", credits: 515 },
-      { identifier: "video-analysis:gemini-3-flash:600s", credits: 848, note: "10-min ceiling" }
+      { identifier: "video-analysis:gemini-3-flash:180s", credits: 186 },
+      { identifier: "video-analysis:gemini-3-flash:360s", credits: 516 },
+      { identifier: "video-analysis:gemini-3-flash:600s", credits: 849, note: "10-min ceiling" }
     ]
   },
   "gemini-3.6-flash-video-analysis": {
@@ -1998,11 +1998,11 @@ var VIDEO_MODELS = {
     description: "Analyze a video into a structured shot list (scenes, camera, audio) \u2014 fast, economy tier. Billed per duration bucket.",
     useCases: ["video-analysis", "shot-list", "fast"],
     pricing: [
-      { identifier: "video-analysis:gemini-3.6-flash", credits: 992, note: "10-min ceiling (no duration given)" },
-      { identifier: "video-analysis:gemini-3.6-flash:60s", credits: 204 },
+      { identifier: "video-analysis:gemini-3.6-flash", credits: 995, note: "10-min ceiling (no duration given)" },
+      { identifier: "video-analysis:gemini-3.6-flash:60s", credits: 205 },
       { identifier: "video-analysis:gemini-3.6-flash:180s", credits: 219 },
-      { identifier: "video-analysis:gemini-3.6-flash:360s", credits: 602 },
-      { identifier: "video-analysis:gemini-3.6-flash:600s", credits: 992, note: "10-min ceiling" }
+      { identifier: "video-analysis:gemini-3.6-flash:360s", credits: 603 },
+      { identifier: "video-analysis:gemini-3.6-flash:600s", credits: 995, note: "10-min ceiling" }
     ]
   },
   "gemini-3.1-pro-video-analysis": {
@@ -2015,11 +2015,11 @@ var VIDEO_MODELS = {
     description: "Analyze a video into a structured shot list (scenes, camera, audio) \u2014 higher-fidelity, default tier. Billed per duration bucket.",
     useCases: ["video-analysis", "shot-list", "cinematic"],
     pricing: [
-      { identifier: "video-analysis:gemini-3.1-pro", credits: 1056, note: "10-min ceiling (no duration given)" },
-      { identifier: "video-analysis:gemini-3.1-pro:60s", credits: 216 },
-      { identifier: "video-analysis:gemini-3.1-pro:180s", credits: 232 },
-      { identifier: "video-analysis:gemini-3.1-pro:360s", credits: 640 },
-      { identifier: "video-analysis:gemini-3.1-pro:600s", credits: 1056, note: "10-min ceiling" }
+      { identifier: "video-analysis:gemini-3.1-pro", credits: 1059, note: "10-min ceiling (no duration given)" },
+      { identifier: "video-analysis:gemini-3.1-pro:60s", credits: 217 },
+      { identifier: "video-analysis:gemini-3.1-pro:180s", credits: 233 },
+      { identifier: "video-analysis:gemini-3.1-pro:360s", credits: 642 },
+      { identifier: "video-analysis:gemini-3.1-pro:600s", credits: 1059, note: "10-min ceiling" }
     ]
   },
   // Both mixed tiers are variants of the same advanced multi-engine analysis
@@ -2035,11 +2035,11 @@ var VIDEO_MODELS = {
     description: "Our most advanced analysis tier \u2014 multiple analysis engines combined into one result for maximum completeness and accuracy. Billed per duration bucket.",
     useCases: ["video-analysis", "shot-list", "premium", "most-complete"],
     pricing: [
-      { identifier: "video-analysis:mixed", credits: 1177, note: "10-min ceiling (no duration given)" },
+      { identifier: "video-analysis:mixed", credits: 1181, note: "10-min ceiling (no duration given)" },
       { identifier: "video-analysis:mixed:60s", credits: 270 },
-      { identifier: "video-analysis:mixed:180s", credits: 291 },
-      { identifier: "video-analysis:mixed:360s", credits: 729 },
-      { identifier: "video-analysis:mixed:600s", credits: 1177, note: "10-min ceiling" }
+      { identifier: "video-analysis:mixed:180s", credits: 292 },
+      { identifier: "video-analysis:mixed:360s", credits: 731 },
+      { identifier: "video-analysis:mixed:600s", credits: 1181, note: "10-min ceiling" }
     ]
   },
   // SMART — the accuracy tier. Since the 2026-08-03 hybrid re-plan it is a
@@ -2057,11 +2057,11 @@ var VIDEO_MODELS = {
     description: "Highest-accuracy analysis \u2014 a hybrid pass that blends a native skeleton read with several donor analysis rolls, then always refines the merged result to find shot boundaries and identify the cast by appearance. Best choice when the shot list will drive regeneration. Billed per duration bucket.",
     useCases: ["video-analysis", "shot-list", "premium", "most-accurate"],
     pricing: [
-      { identifier: "video-analysis:smart", credits: 2076, note: "10-min ceiling (no duration given)" },
-      { identifier: "video-analysis:smart:60s", credits: 413 },
-      { identifier: "video-analysis:smart:180s", credits: 503 },
-      { identifier: "video-analysis:smart:360s", credits: 1267 },
-      { identifier: "video-analysis:smart:600s", credits: 2076, note: "10-min ceiling" }
+      { identifier: "video-analysis:smart", credits: 2081, note: "10-min ceiling (no duration given)" },
+      { identifier: "video-analysis:smart:60s", credits: 414 },
+      { identifier: "video-analysis:smart:180s", credits: 504 },
+      { identifier: "video-analysis:smart:360s", credits: 1270 },
+      { identifier: "video-analysis:smart:600s", credits: 2081, note: "10-min ceiling" }
     ]
   },
   // ── AI Audit (`video-audit`) ── a SEPARATE node from video-analysis (own
@@ -2085,11 +2085,11 @@ var VIDEO_MODELS = {
     description: "Re-watches a clip against a wired analysis, applies video-verified corrections under guards, and returns a disclosed report of what changed. Billed per duration bucket.",
     useCases: ["video-audit", "qa", "shot-list", "verification"],
     pricing: [
-      { identifier: "video-audit", credits: 1073, note: "10-min ceiling (no duration given)" },
+      { identifier: "video-audit", credits: 1075, note: "10-min ceiling (no duration given)" },
       { identifier: "video-audit:60s", credits: 215 },
-      { identifier: "video-audit:180s", credits: 290 },
-      { identifier: "video-audit:360s", credits: 663 },
-      { identifier: "video-audit:600s", credits: 1073, note: "10-min ceiling" }
+      { identifier: "video-audit:180s", credits: 291 },
+      { identifier: "video-audit:360s", credits: 664 },
+      { identifier: "video-audit:600s", credits: 1075, note: "10-min ceiling" }
     ]
   },
   "video-audit-auto": {
@@ -2101,11 +2101,11 @@ var VIDEO_MODELS = {
     description: "Same video-verified audit as AI Audit, but auto-runs a fast analysis first when none is wired in, then applies corrections under guards and returns a disclosed report. Billed per duration bucket.",
     useCases: ["video-audit", "qa", "shot-list", "verification", "auto-analysis"],
     pricing: [
-      { identifier: "video-audit:auto", credits: 1921, note: "10-min ceiling (no duration given)" },
+      { identifier: "video-audit:auto", credits: 1924, note: "10-min ceiling (no duration given)" },
       { identifier: "video-audit:auto:60s", credits: 396 },
-      { identifier: "video-audit:auto:180s", credits: 475 },
-      { identifier: "video-audit:auto:360s", credits: 1178 },
-      { identifier: "video-audit:auto:600s", credits: 1921, note: "10-min ceiling" }
+      { identifier: "video-audit:auto:180s", credits: 477 },
+      { identifier: "video-audit:auto:360s", credits: 1180 },
+      { identifier: "video-audit:auto:600s", credits: 1924, note: "10-min ceiling" }
     ]
   }
 };
@@ -5424,7 +5424,9 @@ var LLM_MODELS = [
   {
     id: "gemini-3.6-flash",
     displayName: "Gemini 3.6 Flash",
-    desc: "Latest fast Gemini, sharper reasoning",
+    // Demoted 2026-09-06 alongside 3.7: superlatives belong to the CURRENT top
+    // model of a family only, and 3.8 now holds that slot.
+    desc: "Fast Gemini, sharper reasoning",
     tier: "economy",
     kieFormat: "chat-completions",
     // KIE serves Gemini 3.6 Flash on the OpenAI-compatible dialect under this
@@ -5454,7 +5456,11 @@ var LLM_MODELS = [
   {
     id: "gemini-3.7-flash",
     displayName: "Gemini 3.7 Flash",
-    desc: "Newest fast Gemini, agentic-tuned",
+    // Demoted 2026-09-06 when 3.8 registered below: superlatives belong to the
+    // CURRENT top model of a family only (same rule the Opus entries follow) —
+    // `desc` renders in every picker, so leaving "Newest" on the older flash
+    // steers the A/B's traffic backwards.
+    desc: "Fast Gemini, agentic-tuned",
     tier: "economy",
     kieFormat: "chat-completions",
     // KIE serves it on the OpenAI-compatible dialect under this slug
@@ -5475,6 +5481,46 @@ var LLM_MODELS = [
     directReasoningEfforts: ["none", "low", "medium", "high"],
     directGeminiModel: "gemini-3.7-flash"
   },
+  {
+    id: "gemini-3.8-flash",
+    displayName: "Gemini 3.8 Flash",
+    // Inherits "Newest" from 3.7 (demoted above). Deliberately does NOT name
+    // agentic video: that capability exists only on the direct Google lane
+    // (Interactions API, not wired here) and the modality caps below withhold
+    // video — a picker desc must not promise a lever this entry cannot expose.
+    desc: "Newest fast Gemini, agentic-tuned",
+    tier: "economy",
+    kieFormat: "chat-completions",
+    // KIE serves it on the OpenAI-compatible dialect under this slug
+    // (docs.kie.ai/market/gemini/gemini-3-8-flash-openai.md) — identical
+    // chat-completions path shape to gemini-3.7-flash / gemini-3.6-flash,
+    // different slug prefix.
+    kieSlugOrModel: "gemini-3-8-flash-openai",
+    vendor: "google",
+    // Live-verified 2026-09-06 on the KIE lane: `response_format: json_schema`
+    // is ENFORCED (the reply came back exact-schema valid, credits_consumed
+    // present) — not merely accepted-and-ignored.
+    structuredOutputMode: "kie-response-format",
+    supportsImages: true,
+    // 16384 here, NOT the 8192 its 3.6 / 3.7 siblings sit at. Those two stay at
+    // the KIE-safe intersection because nobody measured their endpoints past
+    // it; 3.8's WAS measured — live-verified 2026-09-06, `max_tokens: 20000`
+    // was honored for 14,892 completion tokens with finish_reason "stop" (no
+    // truncation), so the KIE lane is not the binding constraint. 16384 keeps
+    // it level with the rest of the registry's ceiling rather than at the cap.
+    maxOutputTokens: 16384,
+    // KIE's 3.8 endpoint enumerates reasoning_effort low | high (its doc enum),
+    // and usage carries completion_tokens_details.reasoning_tokens — same
+    // KIE-safe intersection as 3.6 / 3.7, because ONE field feeds both lanes
+    // and this model is KIE-first.
+    reasoningEfforts: ["low", "high"],
+    // Google's own API accepts the full minimal→high ladder here (`none` maps
+    // to `minimal`), same as 3.7. Advanced mode is what unlocks it.
+    directReasoningEfforts: ["none", "low", "medium", "high"],
+    // KIE-first (no `preferDirect`) — 3.7's posture exactly: the cheap lane
+    // serves the A/B, direct is Advanced mode + the reliability fallback.
+    directGeminiModel: "gemini-3.8-flash"
+  },
   {
     id: "claude-haiku-4.5",
     displayName: "Claude Haiku 4.5",
@@ -5571,7 +5617,9 @@ var LLM_MODELS = [
   {
     id: "gpt-5.4",
     displayName: "GPT-5.4",
-    desc: "Latest GPT, premium quality",
+    // Demoted 2026-09-06: two GPT generations have shipped above it, so
+    // "Latest" steered the picker at the oldest premium GPT in the registry.
+    desc: "Older GPT, premium quality",
     tier: "premium",
     kieFormat: "responses",
     kieSlugOrModel: "gpt-5-4",
@@ -5626,7 +5674,11 @@ var LLM_MODELS = [
   {
     id: "gpt-5.6-sol",
     displayName: "GPT-5.6 Sol",
-    desc: "Flagship GPT-5.6, deepest reasoning",
+    // Demoted 2026-09-06 when gpt-6-astra registered below — "Flagship /
+    // deepest" is the current top GPT's copy only. Worded to stay distinct
+    // from gpt-5.5's "Previous flagship GPT, deep reasoning": all three read
+    // premium on the invoice, so a stale superlative mis-steers silently.
+    desc: "Previous flagship GPT-5.6, deep reasoning",
     tier: "premium",
     kieFormat: "responses",
     kieSlugOrModel: "gpt-5-6-sol",
@@ -5637,6 +5689,61 @@ var LLM_MODELS = [
     reasoningEfforts: ["none", "low", "medium", "high", "xhigh", "max"],
     supportsTemperature: false
   },
+  {
+    id: "gpt-6-astra",
+    displayName: "GPT-6 Astra",
+    desc: "Flagship GPT-6, deepest reasoning",
+    tier: "premium",
+    kieFormat: "responses",
+    // KIE serves GPT-6 on the responses dialect under the OpenAI family path —
+    // codex/v1/responses, which llm-client DERIVES from `vendor` (the same
+    // derivation that keeps Grok on grok/v1/responses). So this field is the
+    // body `model` only, not a path segment. Live-verified 2026-09-06 on the
+    // streaming path (array `input`, SSE deltas), which is the ONLY path we
+    // serve it on — see `kieCollapseStream` below. The non-stream reply does
+    // carry `credits_consumed` (and cached_tokens/cache_write_tokens in usage)
+    // on the ~1 call in 3 that returns one, but we no longer take that reply,
+    // so provider cost here is the rate-table estimate.
+    // KIE's codex/v1/responses injects its own Codex-agent `instructions`
+    // prompt ABOVE the caller's `developer` message — observed in the
+    // 2026-07-14 live probe of that endpoint, and unchanged for GPT-6. Budget
+    // for it when a developer message has to dominate.
+    kieSlugOrModel: "gpt-6-astra",
+    vendor: "openai",
+    // Live-verified 2026-09-06: `text.format: json_schema` is passed through
+    // and ENFORCED — the server echoed `strict: true` and returned
+    // schema-valid JSON.
+    structuredOutputMode: "responses-json-schema",
+    // text + image + file inputs per KIE's doc. No video/audio — see the
+    // image-only LLM_MODALITY_CAPS row below.
+    supportsImages: true,
+    maxOutputTokens: 16384,
+    // KIE's documented enum for this endpoint. No `none` on purpose: the
+    // endpoint reasons unconditionally (see thinkingDefaultOn below), so a
+    // "none" level would be a lie the wire silently overrides.
+    reasoningEfforts: ["low", "medium", "high", "xhigh"],
+    // Live-probed 2026-09-06: `temperature: 0.2` was sent and the request echo
+    // stayed at 1.0 — silently IGNORED, so never send it. Same treatment as the
+    // GPT-5.6 family and grok-4.6.
+    supportsTemperature: false,
+    // With NO reasoning param sent, the server echoed effort "medium" — it
+    // reasons by default (2026-09-06), so reasoning tokens share `max_tokens`
+    // on EVERY call, not just effort-bearing ones. Consumers must floor output
+    // headroom off this flag or a premium answer truncates into a paid-for
+    // empty reply.
+    thinkingDefaultOn: true,
+    // Measured 2026-09-06, 12 identical requests (developer + user message,
+    // reasoning.effort low, text.format json_schema): `stream: false`
+    // succeeded 2/6 in 4–5 s and 500'd 4/6 with
+    // {"error":{"type":"server_error"}} after 34, 34, 35 and 64 s, while the
+    // same body with `stream: true` succeeded 5/6 in the same 4–5 s. An
+    // earlier non-stream probe timed out at 90 s with 0 bytes, and a
+    // non-stream call WITHOUT a schema 500'd after 65 s — the schema is not
+    // the trigger, the non-stream lane is. The very same endpoint serves
+    // gpt-5.4/5.5/5.6 non-stream reliably (live-verified 2026-07-14), which is
+    // why this is per-model and not a lane-wide flag.
+    kieCollapseStream: true
+  },
   {
     id: "grok-4.6",
     displayName: "Grok 4.6",
@@ -5791,6 +5898,15 @@ var LLM_MODALITY_CAPS = {
   // deferred while the smart-family A/B routes this model internally (#747).
   // Flip these two flags ONLY together with that VA-side decision.
   "gemini-3.7-flash": { image: true, video: false, audio: false },
+  // gemini-3.8-flash is IMAGE-ONLY by DECISION, not omission — 3.7's rationale
+  // exactly: full video+audio caps would auto-enroll it in
+  // VIDEO_ANALYSIS_LLM_MODELS (derived below) and force a video-analysis tier +
+  // pricing decision that stays deferred until the 3.8-vs-3.7 analysis A/B
+  // concludes. Google's own 3.8 lane DOES understand video (agentic video, via
+  // the Interactions API we don't wire) — which is precisely why withholding it
+  // has to be a decision rather than a gap. Flip these two flags ONLY together
+  // with that VA-side decision.
+  "gemini-3.8-flash": { image: true, video: false, audio: false },
   "gemini-3.1-pro": { image: true, video: true, audio: true },
   "claude-haiku-4.5": { image: true, video: false, audio: false },
   "claude-sonnet-4.6": { image: true, video: false, audio: false },
@@ -5801,6 +5917,8 @@ var LLM_MODALITY_CAPS = {
   "gpt-5.6-luna": { image: true, video: false, audio: false },
   "gpt-5.6-terra": { image: true, video: false, audio: false },
   "gpt-5.6-sol": { image: true, video: false, audio: false },
+  // KIE's GPT-6 doc lists text + image + file inputs only — no video, no audio.
+  "gpt-6-astra": { image: true, video: false, audio: false },
   "grok-4.6": { image: true, video: false, audio: false },
   "claude-sonnet-5": { image: true, video: false, audio: false },
   "claude-opus-4.8": { image: true, video: false, audio: false },
@@ -14287,36 +14405,36 @@ var VIDEO_ANALYSIS_WINDOW = { LEN: WINDOW_LEN, STRIDE: WINDOW_STRIDE, OVERLAP: W
 var VIDEO_ANALYSIS_BUCKET_CREDITS = {
   // Legacy fast-tier model (pre-2026-07) — kept for stored raw-id configs.
   "video-analysis:gemini-3-flash:60s": 181,
-  "video-analysis:gemini-3-flash:180s": 185,
-  "video-analysis:gemini-3-flash:360s": 515,
-  "video-analysis:gemini-3-flash:600s": 848,
+  "video-analysis:gemini-3-flash:180s": 186,
+  "video-analysis:gemini-3-flash:360s": 516,
+  "video-analysis:gemini-3-flash:600s": 849,
   // Current fast tier — regenerated from the private formula for its backing
   // model; higher than the legacy fast schedule but still ≤ pro per bucket.
-  "video-analysis:gemini-3.6-flash:60s": 204,
+  "video-analysis:gemini-3.6-flash:60s": 205,
   "video-analysis:gemini-3.6-flash:180s": 219,
-  "video-analysis:gemini-3.6-flash:360s": 602,
-  "video-analysis:gemini-3.6-flash:600s": 992,
-  "video-analysis:gemini-3.1-pro:60s": 216,
-  "video-analysis:gemini-3.1-pro:180s": 232,
-  "video-analysis:gemini-3.1-pro:360s": 640,
-  "video-analysis:gemini-3.1-pro:600s": 1056,
+  "video-analysis:gemini-3.6-flash:360s": 603,
+  "video-analysis:gemini-3.6-flash:600s": 995,
+  "video-analysis:gemini-3.1-pro:60s": 217,
+  "video-analysis:gemini-3.1-pro:180s": 233,
+  "video-analysis:gemini-3.1-pro:360s": 642,
+  "video-analysis:gemini-3.1-pro:600s": 1059,
   // Mixed tiers (`mixed` + `mixed-fast`) share ONE credit family — they are
   // variants of the same engine plan (plan internals live in the private
   // analysis plugin). Admin-tunable via model_pricing like every other row.
   "video-analysis:mixed:60s": 270,
-  "video-analysis:mixed:180s": 291,
-  "video-analysis:mixed:360s": 729,
-  "video-analysis:mixed:600s": 1177,
+  "video-analysis:mixed:180s": 292,
+  "video-analysis:mixed:360s": 731,
+  "video-analysis:mixed:600s": 1181,
   // SMART — the accuracy tier, and since the 2026-08-03 re-plan a multi-roll
   // plan like the others, always refined (`selectionMode` does not apply
   // here — smart always refines; it never offers a cheaper "choose" path).
   // Priced above the economy tiers because it genuinely costs more to run;
   // the only tier whose accuracy is validated against a hand-counted edit
   // list, re-validated at the current plan before this schedule shipped.
-  "video-analysis:smart:60s": 413,
-  "video-analysis:smart:180s": 503,
-  "video-analysis:smart:360s": 1267,
-  "video-analysis:smart:600s": 2076
+  "video-analysis:smart:60s": 414,
+  "video-analysis:smart:180s": 504,
+  "video-analysis:smart:360s": 1270,
+  "video-analysis:smart:600s": 2081
 };
 function videoAnalysisCreditSegment(modelOrSentinel) {
   return modelOrSentinel === "mixed-fast" ? "mixed" : modelOrSentinel;
@@ -14338,13 +14456,13 @@ function videoAnalysisNumWindows(bucketSec) {
 }
 var VIDEO_AUDIT_BUCKET_CREDITS = {
   "video-audit:60s": 215,
-  "video-audit:180s": 290,
-  "video-audit:360s": 663,
-  "video-audit:600s": 1073,
+  "video-audit:180s": 291,
+  "video-audit:360s": 664,
+  "video-audit:600s": 1075,
   "video-audit:auto:60s": 396,
-  "video-audit:auto:180s": 475,
-  "video-audit:auto:360s": 1178,
-  "video-audit:auto:600s": 1921
+  "video-audit:auto:180s": 477,
+  "video-audit:auto:360s": 1180,
+  "video-audit:auto:600s": 1924
 };
 function buildVideoAuditCreditId(args) {
   const { analysisProvided, durationSec } = args;