@genesislcap/ai-assistant 15.22.1 → 15.23.0

package/dist/chat-driver.mjs

@@ -1134,6 +1134,7 @@ var SUPPORTED_GEMINI_MODEL_IDS = [
   "gemini-2.5-pro",
   "gemini-2.5-flash",
   "gemini-2.5-flash-lite",
+  "gemini-3.8-flash",
   "gemini-3.5-flash",
   "gemini-3.1-flash-lite",
   "gemini-3.1-pro-preview"
@@ -1236,6 +1237,22 @@ var GEMINI_PRICING = {
     rate: { promptPerMillion: 0.25, candidatePerMillion: 1.5 }
   },
   "gemini-3.5-flash": { kind: "flat", rate: { promptPerMillion: 1.5, candidatePerMillion: 9 } },
+  // INTRODUCTORY RATE, and the only dated entry in this file. Google publishes
+  // $0.75 / $3.75 per MTok **through 2026-12-31**, rising to $1.50 / $7.50 on 2027-01-01 —
+  // so today this model is CHEAPER than gemini-3.5-flash above and in January it is not.
+  //
+  // The listed figure is the one in force, not the one coming. That is the lesson the
+  // `claude-sonnet-5` entry records from the other direction: charging a scheduled future
+  // rate early over-billed every request by 50% for as long as nobody noticed. A rate that
+  // has not taken effect is not a price, and `GeminiPricing` deliberately has no time
+  // dimension to express one — flip these two numbers on 2027-01-01 instead.
+  //
+  // That flip is ENFORCED, not remembered: `token-cost.test.ts` carries a dated assertion that
+  // turns the build red on 2027-01-01 and names the three-line edit (these two rates, plus the
+  // now-false `< gemini-3.5-flash` comparison in that same test — both models are $1.50 input
+  // from that date). There is no ticket to lose.
+  // https://ai.google.dev/gemini-api/docs/pricing
+  "gemini-3.8-flash": { kind: "flat", rate: { promptPerMillion: 0.75, candidatePerMillion: 3.75 } },
   "gemini-3.1-pro-preview": {
     kind: "tiered",
     standard: { promptPerMillion: 2, candidatePerMillion: 12 },
@@ -2722,6 +2739,7 @@ var GEMINI_CONTEXT_LIMITS = {
   "gemini-2.5-pro": 1048576,
   "gemini-2.5-flash": 1048576,
   "gemini-2.5-flash-lite": 1048576,
+  "gemini-3.8-flash": 1048576,
   "gemini-3.5-flash": 1048576,
   "gemini-3.1-flash-lite": 1048576,
   "gemini-3.1-pro-preview": 1048576
@@ -2746,6 +2764,10 @@ var GEMINI_MODEL_WARNINGS = {
   "gemini-2.5-flash": "GeminiTransport: using gemini-2.5-flash \u2014 higher cost than flash-lite; use for harder reasoning or agent tasks.",
   "gemini-2.5-pro": "GeminiTransport: using gemini-2.5-pro \u2014 significantly higher, prompt-size-tiered cost; reserve for tasks where flash reliability is insufficient.",
   "gemini-3.5-flash": "GeminiTransport: using gemini-3.5-flash \u2014 frontier model, materially higher cost than flash-lite; use for harder reasoning or agent tasks.",
+  // Warns despite currently costing LESS than gemini-3.5-flash, because that is temporary: the
+  // rate doubles on 2027-01-01 (see GEMINI_PRICING). A caller that picks this tier on today's
+  // price and never revisits the choice gets twice the bill in January without changing a line.
+  "gemini-3.8-flash": "GeminiTransport: using gemini-3.8-flash \u2014 frontier model, materially higher cost than flash-lite. Its $0.75/$3.75 per-MTok rate is introductory and DOUBLES to $1.50/$7.50 on 2027-01-01; re-check this tier choice before then.",
   "gemini-3.1-pro-preview": "GeminiTransport: using gemini-3.1-pro-preview \u2014 PREVIEW model (no stability guarantee, may be withdrawn) with significantly higher, prompt-size-tiered cost; reserve for tasks where flash reliability is insufficient."
 };
 var GEMINI_THINKING_DISABLEABLE = [
@@ -2753,15 +2775,32 @@ var GEMINI_THINKING_DISABLEABLE = [
   "gemini-2.5-flash-lite"
 ];
 var GEMINI_MEDIA_RESOLUTION_TIERS = [
+  "gemini-3.8-flash",
   "gemini-3.5-flash",
   "gemini-3.1-flash-lite",
   "gemini-3.1-pro-preview"
 ];
-var GEMINI_THINKING_LEVEL_TIERS = [
-  "gemini-3.5-flash",
-  "gemini-3.1-flash-lite",
-  "gemini-3.1-pro-preview"
-];
+var GEMINI_LEVEL_LADDER = ["minimal", "low", "medium", "high"];
+var GEMINI_THINKING_LEVELS = {
+  "gemini-3.8-flash": ["low", "medium", "high"],
+  "gemini-3.5-flash": ["minimal", "low", "medium", "high"],
+  "gemini-3.1-flash-lite": ["minimal", "low", "medium", "high"],
+  "gemini-3.1-pro-preview": ["low", "medium", "high"]
+};
+function clampThinkingLevel(supported, level) {
+  if (!(supported === null || supported === void 0 ? void 0 : supported.length) || supported.includes(level)) {
+    return level;
+  }
+  const target = GEMINI_LEVEL_LADDER.indexOf(level);
+  const distance = (candidate) => Math.abs(GEMINI_LEVEL_LADDER.indexOf(candidate) - target);
+  let best = supported[0];
+  for (const candidate of supported) {
+    if (distance(candidate) <= distance(best)) {
+      best = candidate;
+    }
+  }
+  return best;
+}
 var GEMINI_THINKING_LEVEL = {
   minimal: "minimal",
   low: "low",
@@ -2769,25 +2808,16 @@ var GEMINI_THINKING_LEVEL = {
   high: "high",
   max: "high"
 };
-var GEMINI_COARSE_LEVEL_TIERS = [
-  "gemini-3.5-flash",
-  "gemini-3.1-flash-lite",
-  "gemini-3.1-pro-preview"
-];
-var GEMINI_COARSE_LEVEL = {
-  minimal: "low",
-  low: "low",
-  medium: "high",
-  high: "high"
-};
 var isGradedLevel = (p) => p !== void 0 && p !== "auto" && p !== "off";
 function geminiThinkingConfig(model, policy) {
-  if (isGradedLevel(policy) && GEMINI_THINKING_LEVEL_TIERS.includes(model)) {
-    const level = GEMINI_THINKING_LEVEL[policy];
-    return {
-      includeThoughts: true,
-      thinkingLevel: GEMINI_COARSE_LEVEL_TIERS.includes(model) ? GEMINI_COARSE_LEVEL[level] : level
-    };
+  if (isGradedLevel(policy)) {
+    const accepted = GEMINI_THINKING_LEVELS[model];
+    if (accepted) {
+      return {
+        includeThoughts: true,
+        thinkingLevel: clampThinkingLevel(accepted, GEMINI_THINKING_LEVEL[policy])
+      };
+    }
   }
   if (policy === "off" && GEMINI_THINKING_DISABLEABLE.includes(model)) {
     return { includeThoughts: false, thinkingBudget: 0 };
@@ -2828,7 +2858,7 @@ var GeminiTransport = class _GeminiTransport {
   /**
    * Warn once when a requested policy cannot be honoured on this model — a *cost* decision the
    * caller made that the wire silently reversed, so they hear about it, but only once, not per
-   * turn. Two shapes reach here (the Anthropic twin warns on the same two):
+   * turn. Three shapes reach here (the Anthropic twin warns on the first two):
    *
    * - `'off'` on a tier that can't disable thinking — only the flash tiers accept a zero thinking
    *   budget; everything else keeps paying for reasoning it was told to stop. (`'auto'` is NOT this
@@ -2838,6 +2868,12 @@ var GeminiTransport = class _GeminiTransport {
    *   not the Gemini 3 `thinkingLevel` enum, so the level is dropped and the turn runs at the
    *   default posture. Without this, a caller asking for `high` on `gemini-2.5-pro` and quietly
    *   getting the default is left to wonder why.
+   * - A graded depth whose DIALECT a Gemini 3 tier speaks but whose particular value it does not
+   *   accept — `minimal` on `gemini-3.8-flash` or `gemini-3.1-pro-preview`, the two tiers that
+   *   reject it outright. Nothing is dropped here; `clampThinkingLevel` MOVES the level onto the
+   *   nearest one the tier takes, and since ties resolve upward the move is usually to something
+   *   DEARER than the caller chose. A cost decision reversed on the wire in the expensive
+   *   direction is exactly what this warning exists for — without it the only trace is the bill.
    */
   warnIfThinkingUnclampable(policy) {
     if (this.warnedThinkingClamped) {
@@ -2848,9 +2884,19 @@ var GeminiTransport = class _GeminiTransport {
       logger.warn(`GeminiTransport: thinkingPolicy 'off' ignored \u2014 ${this.model} runs its default thinking posture and takes no budget we can safely set. Reasoning tokens are still billed at the candidate rate; use a flash tier if you need them gone.`);
       return;
     }
-    if (isGradedLevel(policy) && !GEMINI_THINKING_LEVEL_TIERS.includes(this.model)) {
-      this.warnedThinkingClamped = true;
-      logger.warn(`GeminiTransport: thinkingPolicy '${policy}' ignored \u2014 ${this.model} takes a numeric thinking budget, not the graded 'thinkingLevel' the Gemini 3 tiers use, so this turn runs at its default reasoning posture. Use a Gemini 3 tier to set reasoning depth.`);
+    if (isGradedLevel(policy)) {
+      const accepted = GEMINI_THINKING_LEVELS[this.model];
+      if (!accepted) {
+        this.warnedThinkingClamped = true;
+        logger.warn(`GeminiTransport: thinkingPolicy '${policy}' ignored \u2014 ${this.model} takes a numeric thinking budget, not the graded 'thinkingLevel' the Gemini 3 tiers use, so this turn runs at its default reasoning posture. Use a Gemini 3 tier to set reasoning depth.`);
+        return;
+      }
+      const requested = GEMINI_THINKING_LEVEL[policy];
+      const sent = clampThinkingLevel(accepted, requested);
+      if (sent !== requested) {
+        this.warnedThinkingClamped = true;
+        logger.warn(`GeminiTransport: thinkingPolicy '${policy}' sent as thinkingLevel '${sent}' \u2014 ${this.model} does not accept '${requested}' (it takes ${accepted.join("/")}). Reasoning bills at the candidate rate, so an upward clamp costs more than you asked for.`);
+      }
     }
   }
   constructor(config = {}) {