@vincemakes/kiso-runtime 0.37.0 → 0.38.0

package/dist/provider/metadata.js

@@ -18,6 +18,32 @@
  * Lives under runtime/internal — the curated root surface (44 names)
  * does not move; core is untouched.
  */
+/**
+ * REG-1 — the OFF-PEAK rates, read from the vendor's pricing table on
+ * 2026-09-17. The table carries TWO sets: off-peak and peak, the peak
+ * rates exactly double, and peak is 01:00-04:00 and 06:00-10:00 UTC
+ * Mon-Fri. `ModelPricing` has one set of rates and no period, so the
+ * OFF-PEAK numbers are the row's and the doubling is stated here rather
+ * than by inventing a field the type does not have. A bill computed from
+ * this row during peak hours is half the truth, and that is the honest
+ * shape of a one-rate type against a two-rate vendor.
+ */
+const DEEPSEEK_FLASH_PRICING = {
+    inputPerM: 0.15,
+    outputPerM: 0.6,
+    cacheReadPerM: 0.003,
+    cacheWritePerM: 0,
+    asOf: "2026-09-17",
+    source: "https://api-docs.deepseek.com/quick_start/pricing/",
+};
+const DEEPSEEK_PRO_PRICING = {
+    inputPerM: 0.66,
+    outputPerM: 1.98,
+    cacheReadPerM: 0.022,
+    cacheWritePerM: 0,
+    asOf: "2026-09-17",
+    source: "https://api-docs.deepseek.com/quick_start/pricing/",
+};
 const DEEPSEEK_PRICING = {
     // The E2-frozen rates (pricing table v1, freeze date 2026-08-13),
     // re-homed here with their provenance made explicit. The caveat
@@ -185,8 +211,28 @@ const ENTRIES = [
         // stays null until read from the live billing page and dated.
         model: "deepseek-v4-flash",
         providerId: "deepseek",
+        // REG-1: RETIRED, and the vendor says so in footnote (1) of the
+        // pricing table: "The legacy names deepseek-v4-flash and
+        // deepseek-v4-flash-vision-exp are still accepted, but the
+        // corresponding models have been retired, their requests are served
+        // by the DeepSeek-V4.1-Flash model and billed at the Flash price."
+        // Still accepted is not still supported: a config naming it works
+        // today on borrowed time, which is why the row stays (removing it
+        // would break configs that run) and is marked (so the product can
+        // say the word it has the evidence for).
+        deprecated: { asOf: "2026-09-17", source: "https://api-docs.deepseek.com/quick_start/pricing/" },
         endpoint: "https://api.deepseek.com",
-        capabilities: { contextWindow: null, maxOutputTokens: null, promptCaching: "automatic", reasoning: {
+        // REG-1: the window and max output are READ from the vendor's pricing
+        // table — "CONTEXT LENGTH 1M", "MAX OUTPUT MAXIMUM: 384K". The date
+        // and page live in `capabilitiesAsOf` / `capabilitiesSource`, which
+        // exist for exactly this; the reasoning block carries its own pair
+        // and already cited the same page. A dated vendor
+        // statement is this registry's contract and every other row lives on
+        // it; holding DeepSeek to a stricter "nobody measured it" standard
+        // left the window null, and a null window sends the CLI to its
+        // 200,000 fallback — so the product compacted a 1M model at 100,000
+        // tokens, a tenth of its capacity, while sounding principled.
+        capabilities: { contextWindow: 1_000_000, maxOutputTokens: 384_000, promptCaching: "automatic", reasoning: {
                 emitsThinkingStream: true,
                 thinking: { modes: ["enabled", "disabled"], default: "enabled" },
                 // `none` added 2026-09-14 from a live probe. The three natives were
@@ -214,9 +260,9 @@ const ENTRIES = [
                 asOf: "2026-08-26",
                 source: "https://api-docs.deepseek.com/guides/thinking_mode",
             }, inputModalities: null },
-        capabilitiesAsOf: "2026-08-26",
-        capabilitiesSource: "https://api-docs.deepseek.com/guides/thinking_mode",
-        pricing: null,
+        capabilitiesAsOf: "2026-09-17",
+        capabilitiesSource: "https://api-docs.deepseek.com/quick_start/pricing/",
+        pricing: DEEPSEEK_FLASH_PRICING,
     },
     {
         // Astra F5: THE VENDOR'S CURRENT RECOMMENDED NAME, registered so that
@@ -241,7 +287,17 @@ const ENTRIES = [
         model: "deepseek-flash",
         providerId: "deepseek",
         endpoint: "https://api.deepseek.com",
-        capabilities: { contextWindow: null, maxOutputTokens: null, promptCaching: "automatic", reasoning: {
+        // REG-1: the window and max output are READ from the vendor's pricing
+        // table — "CONTEXT LENGTH 1M", "MAX OUTPUT MAXIMUM: 384K". The date
+        // and page live in `capabilitiesAsOf` / `capabilitiesSource`, which
+        // exist for exactly this; the reasoning block carries its own pair
+        // and already cited the same page. A dated vendor
+        // statement is this registry's contract and every other row lives on
+        // it; holding DeepSeek to a stricter "nobody measured it" standard
+        // left the window null, and a null window sends the CLI to its
+        // 200,000 fallback — so the product compacted a 1M model at 100,000
+        // tokens, a tenth of its capacity, while sounding principled.
+        capabilities: { contextWindow: 1_000_000, maxOutputTokens: 384_000, promptCaching: "automatic", reasoning: {
                 emitsThinkingStream: true,
                 thinking: { modes: ["enabled", "disabled"], default: "enabled" },
                 // `none` added 2026-09-14 from a live probe. The three natives were
@@ -269,15 +325,25 @@ const ENTRIES = [
                 asOf: "2026-09-12",
                 source: "https://api-docs.deepseek.com/quick_start/pricing/",
             }, inputModalities: null },
-        capabilitiesAsOf: "2026-09-12",
+        capabilitiesAsOf: "2026-09-17",
         capabilitiesSource: "https://api-docs.deepseek.com/quick_start/pricing/",
-        pricing: null,
+        pricing: DEEPSEEK_FLASH_PRICING,
     },
     {
         model: "deepseek-v4-pro",
         providerId: "deepseek",
         endpoint: "https://api.deepseek.com",
-        capabilities: { contextWindow: null, maxOutputTokens: null, promptCaching: "automatic", reasoning: {
+        // REG-1: the window and max output are READ from the vendor's pricing
+        // table — "CONTEXT LENGTH 1M", "MAX OUTPUT MAXIMUM: 384K". The date
+        // and page live in `capabilitiesAsOf` / `capabilitiesSource`, which
+        // exist for exactly this; the reasoning block carries its own pair
+        // and already cited the same page. A dated vendor
+        // statement is this registry's contract and every other row lives on
+        // it; holding DeepSeek to a stricter "nobody measured it" standard
+        // left the window null, and a null window sends the CLI to its
+        // 200,000 fallback — so the product compacted a 1M model at 100,000
+        // tokens, a tenth of its capacity, while sounding principled.
+        capabilities: { contextWindow: 1_000_000, maxOutputTokens: 384_000, promptCaching: "automatic", reasoning: {
                 emitsThinkingStream: true,
                 thinking: { modes: ["enabled", "disabled"], default: "enabled" },
                 // `none` added 2026-09-14 from a live probe. The three natives were
@@ -305,9 +371,9 @@ const ENTRIES = [
                 asOf: "2026-08-26",
                 source: "https://api-docs.deepseek.com/guides/thinking_mode",
             }, inputModalities: null },
-        capabilitiesAsOf: "2026-08-26",
-        capabilitiesSource: "https://api-docs.deepseek.com/guides/thinking_mode",
-        pricing: null,
+        capabilitiesAsOf: "2026-09-17",
+        capabilitiesSource: "https://api-docs.deepseek.com/quick_start/pricing/",
+        pricing: DEEPSEEK_PRO_PRICING,
     },
     {
         model: "gpt-4o",

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@vincemakes/kiso-runtime",
-	"version": "0.37.0",
+	"version": "0.38.0",
 	"description": "kiso runtime \u2014 durable multi-turn agent sessions: AgentDefinition, AgentRuntime, AgentSession, Run, append-only JSONL store.",
 	"type": "module",
 	"license": "MIT",
@@ -25,12 +25,12 @@
 		"test": "vitest run"
 	},
 	"dependencies": {
-		"@vincemakes/kiso-core": "0.37.0"
+		"@vincemakes/kiso-core": "0.38.0"
 	},
 	"peerDependencies": {
-		"@vincemakes/kiso-provider-anthropic": "0.37.0",
-		"@vincemakes/kiso-provider-openai": "0.37.0",
-		"@vincemakes/kiso-provider-openai-responses": "0.37.0"
+		"@vincemakes/kiso-provider-anthropic": "0.38.0",
+		"@vincemakes/kiso-provider-openai": "0.38.0",
+		"@vincemakes/kiso-provider-openai-responses": "0.38.0"
 	},
 	"peerDependenciesMeta": {
 		"@vincemakes/kiso-provider-anthropic": {
@@ -44,7 +44,7 @@
 		}
 	},
 	"devDependencies": {
-		"@vincemakes/kiso-evals": "0.37.0",
+		"@vincemakes/kiso-evals": "0.38.0",
 		"@types/node": "^26.1.2",
 		"typescript": "^5.7.2",
 		"vitest": "^3.0.0"