@genesislcap/ai-assistant 15.4.1 → 15.6.0

package/dist/chat-driver.mjs

@@ -1518,6 +1518,8 @@ ${att.content}` });
     var _a, _b, _c, _d, _e, _f, _g, _h;
     let inputTokens;
     let outputTokens;
+    let cacheReadTokens;
+    let cacheWriteTokens;
     let cost;
     if (response.usage) {
       const uncachedInput = (_a = response.usage.input_tokens) !== null && _a !== void 0 ? _a : 0;
@@ -1527,6 +1529,8 @@ ${att.content}` });
       const cacheCreation = (_d = response.usage.cache_creation_input_tokens) !== null && _d !== void 0 ? _d : breakdown ? ((_e = breakdown.ephemeral_5m_input_tokens) !== null && _e !== void 0 ? _e : 0) + cacheCreation1h : 0;
       cost = this.logTokenUsage(uncachedInput, (_f = response.usage.output_tokens) !== null && _f !== void 0 ? _f : 0, cacheRead, cacheCreation, cacheCreation1h);
       inputTokens = uncachedInput + cacheRead + cacheCreation;
+      cacheReadTokens = cacheRead;
+      cacheWriteTokens = cacheCreation;
       if (response.usage.output_tokens != null) {
         outputTokens = response.usage.output_tokens;
       }
@@ -1557,6 +1561,10 @@ ${att.content}` });
       base.inputTokens = inputTokens;
     if (outputTokens != null)
       base.outputTokens = outputTokens;
+    if (cacheReadTokens != null)
+      base.cacheReadTokens = cacheReadTokens;
+    if (cacheWriteTokens != null)
+      base.cacheWriteTokens = cacheWriteTokens;
     if (cost != null)
       base.cost = cost;
     if (response.stop_reason === "max_tokens") {
@@ -2083,28 +2091,45 @@ ${att.content}`
     }
     return contents;
   }
+  /**
+   * Map Gemini's `usageMetadata` onto the provider-agnostic usage fields, pricing
+   * the call as a side effect (`logTokenUsage`). Returns an empty object when the
+   * response carries no usage — every field is optional, so an absent count stays
+   * `undefined` rather than being reported as a measured 0.
+   *
+   * Split out of `fromGeminiResponse` because it is self-contained bookkeeping
+   * that reads better alone — and because folding it inline pushed that method
+   * past the complexity ceiling.
+   */
+  usageFromGemini(usageMetadata) {
+    var _a, _b, _c, _d, _e, _f;
+    if (!usageMetadata)
+      return {};
+    const usage = usageMetadata;
+    const cost = this.logTokenUsage((_a = usage.promptTokenCount) !== null && _a !== void 0 ? _a : 0, (_b = usage.candidatesTokenCount) !== null && _b !== void 0 ? _b : 0, (_c = usage.thoughtsTokenCount) !== null && _c !== void 0 ? _c : 0, (_d = usage.cachedContentTokenCount) !== null && _d !== void 0 ? _d : 0);
+    return {
+      inputTokens: usage.promptTokenCount,
+      // Thinking tokens are generated output, billed at the candidate rate, but
+      // reported in a field disjoint from `candidatesTokenCount`. Fold them into the
+      // provider-agnostic `outputTokens` so it reflects the true generated total —
+      // matching Anthropic, whose `output_tokens` already includes them. Left
+      // `undefined` only when the provider reported neither count.
+      outputTokens: usage.candidatesTokenCount != null || usage.thoughtsTokenCount != null ? ((_e = usage.candidatesTokenCount) !== null && _e !== void 0 ? _e : 0) + ((_f = usage.thoughtsTokenCount) !== null && _f !== void 0 ? _f : 0) : void 0,
+      thoughtsTokens: usage.thoughtsTokenCount,
+      // Gemini reports cache HITS only — implicit caching bills no separate write, so
+      // `cacheWriteTokens` is left unset rather than 0 (there is no write bucket to
+      // measure, which is not the same as measuring 0). Already counted inside
+      // `promptTokenCount`, so this breaks that total down rather than adding to it.
+      cacheReadTokens: usage.cachedContentTokenCount,
+      cost
+    };
+  }
   fromGeminiResponse(response, offeredToolNames = /* @__PURE__ */ new Set()) {
-    var _a, _b, _c, _d, _e, _f, _g, _h, _j, _k;
-    let inputTokens;
-    let outputTokens;
-    let thoughtsTokens;
-    let cost;
-    if (response.usageMetadata) {
-      const usage = response.usageMetadata;
-      cost = this.logTokenUsage((_a = usage.promptTokenCount) !== null && _a !== void 0 ? _a : 0, (_b = usage.candidatesTokenCount) !== null && _b !== void 0 ? _b : 0, (_c = usage.thoughtsTokenCount) !== null && _c !== void 0 ? _c : 0, (_d = usage.cachedContentTokenCount) !== null && _d !== void 0 ? _d : 0);
-      if (usage.promptTokenCount != null) {
-        inputTokens = usage.promptTokenCount;
-      }
-      if (usage.thoughtsTokenCount != null) {
-        thoughtsTokens = usage.thoughtsTokenCount;
-      }
-      if (usage.candidatesTokenCount != null || usage.thoughtsTokenCount != null) {
-        outputTokens = ((_e = usage.candidatesTokenCount) !== null && _e !== void 0 ? _e : 0) + ((_f = usage.thoughtsTokenCount) !== null && _f !== void 0 ? _f : 0);
-      }
-    }
+    var _a, _b, _c, _d;
+    const { inputTokens, outputTokens, thoughtsTokens, cacheReadTokens, cost } = this.usageFromGemini(response.usageMetadata);
     const candidates = response === null || response === void 0 ? void 0 : response.candidates;
     const firstCandidate = candidates === null || candidates === void 0 ? void 0 : candidates[0];
-    const parts = (_h = (_g = firstCandidate === null || firstCandidate === void 0 ? void 0 : firstCandidate.content) === null || _g === void 0 ? void 0 : _g.parts) !== null && _h !== void 0 ? _h : [];
+    const parts = (_b = (_a = firstCandidate === null || firstCandidate === void 0 ? void 0 : firstCandidate.content) === null || _a === void 0 ? void 0 : _a.parts) !== null && _b !== void 0 ? _b : [];
     const toolCalls = [];
     const thoughtParts = [];
     const textParts = [];
@@ -2117,7 +2142,7 @@ ${att.content}`
         toolCalls.push({
           id: crypto.randomUUID(),
           name: part.functionCall.name,
-          args: (_j = part.functionCall.args) !== null && _j !== void 0 ? _j : {},
+          args: (_c = part.functionCall.args) !== null && _c !== void 0 ? _c : {},
           providerMetadata: { [GEMINI_PROVIDER_KEY]: state }
         });
       } else if (part.thought && part.text) {
@@ -2151,9 +2176,11 @@ ${att.content}`
       base.inputTokens = inputTokens;
     if (outputTokens != null)
       base.outputTokens = outputTokens;
+    if (cacheReadTokens != null)
+      base.cacheReadTokens = cacheReadTokens;
     if (cost != null)
       base.cost = cost;
-    const blockReason = (_k = response.promptFeedback) === null || _k === void 0 ? void 0 : _k.blockReason;
+    const blockReason = (_d = response.promptFeedback) === null || _d === void 0 ? void 0 : _d.blockReason;
     if (finishReason != null || thoughtsTokens != null || blockReason != null) {
       base.responseMeta = {
         finishReason,
@@ -2769,6 +2796,9 @@ function estimateTokensSaved(origLen, stubLen) {
 function applyCondensation(history, policies, ctx, onCondensed) {
   if (policies.size === 0) return history;
   const triggersOf = (entry) => Array.isArray(entry.policy.on) ? entry.policy.on : [entry.policy.on];
+  const batchCalls = Math.max(1, Math.floor(ctx.batchCalls ?? 1));
+  const asOf = batchCalls === 1 ? ctx.modelCall : Math.floor(ctx.modelCall / batchCalls) * batchCalls;
+  const withinBatch = (entry) => batchCalls === 1 || entry.iteration <= asOf;
   const latestByKey = /* @__PURE__ */ new Map();
   const nameById = /* @__PURE__ */ new Map();
   for (const msg of history) {
@@ -2776,7 +2806,7 @@ function applyCondensation(history, policies, ctx, onCondensed) {
     for (const tc of msg.toolCalls) {
       nameById.set(tc.id, tc.name);
       const entry = policies.get(tc.id);
-      if (!entry) continue;
+      if (!entry || !withinBatch(entry)) continue;
       for (const t of triggersOf(entry)) {
         if (t.kind === "superseded") latestByKey.set(t.by, tc.id);
       }
@@ -2788,7 +2818,7 @@ function applyCondensation(history, policies, ctx, onCondensed) {
       // past the last allowed view: turns:1 → seen once, then gone. Monotonic
       // clock, so this holds across turn boundaries too.
       case "age":
-        return ctx.modelCall - entry.iteration > trig.turns;
+        return asOf - entry.iteration > trig.turns;
       case "turnEnd":
         return ctx.turn > entry.turn;
       case "agentEnd":
@@ -2802,7 +2832,7 @@ function applyCondensation(history, policies, ctx, onCondensed) {
         return latestByKey.get(trig.by) !== toolCallId;
     }
   };
-  const firstFired = (toolCallId, entry) => triggersOf(entry).find((t) => triggerFires(toolCallId, entry, t));
+  const firstFired = (toolCallId, entry) => withinBatch(entry) ? triggersOf(entry).find((t) => triggerFires(toolCallId, entry, t)) : void 0;
   return history.map((msg) => {
     if (msg.toolCalls?.length) {
       let changed = false;
@@ -2950,6 +2980,46 @@ The above summarizes earlier parts of this conversation that were compacted. Con
   });
 }
 
+// src/utils/sum-usage.ts
+function emptyUsage() {
+  return {
+    costUsd: 0,
+    uncachedInputTokens: 0,
+    cacheReadTokens: 0,
+    cacheWriteTokens: 0,
+    outputTokens: 0
+  };
+}
+function sumUsage(messages) {
+  const total = emptyUsage();
+  accumulate(messages, total);
+  return total;
+}
+function addInto(into, from) {
+  into.costUsd += from.costUsd;
+  into.uncachedInputTokens += from.uncachedInputTokens;
+  into.cacheReadTokens += from.cacheReadTokens;
+  into.cacheWriteTokens += from.cacheWriteTokens;
+  into.outputTokens += from.outputTokens;
+}
+function accumulate(messages, into) {
+  for (const m of messages) {
+    if (m.cost != null) into.costUsd += m.cost;
+    if (m.externalCostUsd != null) into.costUsd += m.externalCostUsd;
+    const cacheRead = m.cacheReadTokens ?? 0;
+    const cacheWrite = m.cacheWriteTokens ?? 0;
+    into.cacheReadTokens += cacheRead;
+    into.cacheWriteTokens += cacheWrite;
+    into.uncachedInputTokens += Math.max(0, (m.inputTokens ?? 0) - cacheRead - cacheWrite);
+    into.outputTokens += m.outputTokens ?? 0;
+    const rolled = m.compaction?.rolledUpUsage;
+    if (rolled) addInto(into, rolled);
+    for (const tc of m.toolCalls ?? []) {
+      if (tc.subAgentTrace) accumulate(tc.subAgentTrace, into);
+    }
+  }
+}
+
 // src/utils/tool-fold.ts
 var TOOL_FOLD_SYMBOL = Symbol("toolFold");
 
@@ -3143,11 +3213,13 @@ var ChatDriver = class _ChatDriver extends EventTarget {
       primerHistory,
       maxToolIterations = DEFAULT_MAX_TOOL_ITERATIONS,
       maxFoldOperations = DEFAULT_MAX_FOLD_OPERATIONS,
+      condenseBatchCalls = 1,
       maxTurnSnapshots = DEFAULT_MAX_TURN_SNAPSHOTS,
       sessionKey = "",
       activityBus = NOOP_ACTIVITY_BUS
     } = config;
     this.maxToolIterations = maxToolIterations;
+    this.condenseBatchCalls = condenseBatchCalls;
     this.sessionKey = sessionKey;
     this.activityBus = activityBus;
     if (typeof toolHandlers === "function") {
@@ -3697,7 +3769,12 @@ Output format (strict):
         coveredThroughTimestamp: toCompact[toCompact.length - 1]?.timestamp,
         tokensBefore,
         model: status?.model,
-        createdAt
+        createdAt,
+        // Bank what the replaced turns spent. `toCompact` is dropped from history
+        // below, taking its per-message cost/token fields with it — so a host total
+        // re-derived from the transcript would otherwise fall by exactly this much
+        // on every compaction. Summarising a turn does not refund what it cost.
+        rolledUpUsage: sumUsage(toCompact)
       }
     };
     this.loadHistory([summaryMessage, ...tail]);
@@ -4381,7 +4458,8 @@ Output format (strict):
           activationEnded: (activation) => activation < this.currentActivation,
           // A call's phase has ended once the app declared a later phase epoch
           // current via `endPhase()`.
-          phaseEnded: (phaseEpoch) => phaseEpoch < this.currentPhaseEpoch
+          phaseEnded: (phaseEpoch) => phaseEpoch < this.currentPhaseEpoch,
+          batchCalls: this.condenseBatchCalls
         },
         (detail) => recordMetaEvent(this.sessionKey, "context.condensed", detail)
       );