@genesislcap/ai-assistant 15.12.0 → 15.13.0

package/dist/chat-driver.cjs

@@ -51,12 +51,14 @@ __export(chat_driver_node_exports, {
   geminiTokenCost: () => geminiTokenCost,
   getMetaEvents: () => getMetaEvents,
   isObservableAIProviderRegistry: () => isObservableAIProviderRegistry,
+  messageUsage: () => messageUsage,
   restoreMachine: () => restoreMachine,
   strictFallbackAgent: () => strictFallbackAgent,
   sumUsage: () => sumUsage,
   totalTokens: () => totalTokens,
   usageRows: () => usageRows,
-  vendorOfModel: () => vendorOfModel
+  vendorOfModel: () => vendorOfModel,
+  withFreshMetaSnapshot: () => withFreshMetaSnapshot
 });
 module.exports = __toCommonJS(chat_driver_node_exports);
 
@@ -3617,6 +3619,24 @@ function addUsage(a, b) {
     outputTokens: a.outputTokens + b.outputTokens
   };
 }
+function messageUsage(m) {
+  if (m.cost == null && m.externalCostUsd == null && m.inputTokens == null && m.outputTokens == null && m.cacheReadTokens == null && m.cacheWriteTokens == null) {
+    return void 0;
+  }
+  const cacheReadTokens = m.cacheReadTokens ?? 0;
+  const cacheWriteTokens = m.cacheWriteTokens ?? 0;
+  return {
+    costUsd: (m.cost ?? 0) + (m.externalCostUsd ?? 0),
+    // The uncached REMAINDER, clamped — the same arithmetic `accumulate` below applies,
+    // for the same reasons. The two are deliberately separate implementations (that walk
+    // mutates one accumulator rather than allocating per message) and must agree;
+    // `sum-usage.test.ts` pins them together.
+    uncachedInputTokens: Math.max(0, (m.inputTokens ?? 0) - cacheReadTokens - cacheWriteTokens),
+    cacheReadTokens,
+    cacheWriteTokens,
+    outputTokens: m.outputTokens ?? 0
+  };
+}
 function sumUsage(messages) {
   const total = emptyUsage();
   accumulate(messages, total);
@@ -4190,7 +4210,14 @@ var ChatDriver = class _ChatDriver extends EventTarget {
     if (resolvedName !== this.lastDispatchedProviderName) {
       this.lastDispatchedProviderName = resolvedName;
       recordMetaEvent(this.sessionKey, "provider.selected", {
+        // `provider` is the registry SLOT (a tier name like 'high'), kept under that key
+        // for compatibility; `model` and `vendor` are what it resolved to. Recording all
+        // three is the difference between "the agent switched to its high tier" and
+        // knowing which model that actually was — a tier can be repointed mid-session,
+        // and a slot name alone cannot distinguish anthropic from gemini.
         provider: resolvedName,
+        model: status.model,
+        vendor: status.provider,
         agent: this.activeAgentName
       });
       this.dispatchEvent(
@@ -4329,6 +4356,10 @@ var ChatDriver = class _ChatDriver extends EventTarget {
    * Push one snapshot to the ring buffer. Called inside `runToolLoop` just
    * before each LLM call — that's the latest point where the prompt, tool
    * surface, and agent state line up with what the model is about to see.
+   *
+   * Returns the pushed object so the caller can back-fill what only the response
+   * knows (`usage`). Mutating it after the fact is safe whether or not the ring
+   * buffer has since evicted it — an evicted snapshot is simply no longer exported.
    */
   recordTurnSnapshot(resolvedSystemPrompt, temperature, toolChoice, tailContext) {
     let agentSnapshot;
@@ -4341,7 +4372,7 @@ var ChatDriver = class _ChatDriver extends EventTarget {
     }
     const turnIndex = String(this.globalTurnIndex);
     this.globalTurnIndex += 1;
-    this.turnSnapshots.push({
+    const snapshot = {
       turnIndex,
       timestamp: (/* @__PURE__ */ new Date()).toISOString(),
       agentName: this.activeAgentName,
@@ -4352,10 +4383,12 @@ var ChatDriver = class _ChatDriver extends EventTarget {
       temperature,
       toolChoice,
       agentSnapshot
-    });
+    };
+    this.turnSnapshots.push(snapshot);
     if (this.turnSnapshots.length > this.maxTurnSnapshots) {
       this.turnSnapshots.shift();
     }
+    return snapshot;
   }
   /**
    * Optional transform applied to conversation history immediately before each LLM request.
@@ -5283,7 +5316,12 @@ Output format (strict):
 ${tailBody}
 </system-reminder>` : void 0;
       const effectiveToolChoice = resolvedToolChoice ?? (this.isSubAgent ? "required" : void 0);
-      this.recordTurnSnapshot(systemPrompt, resolvedTemperature, effectiveToolChoice, tailContext);
+      const turnSnapshot = this.recordTurnSnapshot(
+        systemPrompt,
+        resolvedTemperature,
+        effectiveToolChoice,
+        tailContext
+      );
       const userInputForCall = currentInput;
       const attachmentsForCall = currentAttachments;
       currentInput = "";
@@ -5319,6 +5357,12 @@ ${tailBody}
         fallbacks: this.activeFallbacks
       };
       const activeProvider = await this.resolveProviderForTurn(promptCtx);
+      if (this.lastResolvedProviderName !== void 0) {
+        turnSnapshot.providerName = this.lastResolvedProviderName;
+      }
+      if (this.lastResolvedProvider !== void 0)
+        turnSnapshot.provider = this.lastResolvedProvider;
+      if (this.lastResolvedModel !== void 0) turnSnapshot.model = this.lastResolvedModel;
       let response;
       try {
         response = await activeProvider.chat(historyForCall, userInputForCall, options);
@@ -5436,6 +5480,8 @@ ${tailBody}
       if (this.lastResolvedProviderName !== void 0) {
         response.providerName = this.lastResolvedProviderName;
       }
+      turnSnapshot.usage = messageUsage(response);
+      if (response.model !== void 0) turnSnapshot.model = response.model;
       const isThinkingStep = response.content && response.toolCalls?.length;
       const isEmptyResponse = !response.content?.trim() && !response.toolCalls?.length;
       const isRefusal = response.responseMeta?.finishReason === "refusal";
@@ -5481,7 +5527,14 @@ ${tailBody}
         return this.turnDone(failureReason2);
       } else {
         const { reasoning, ...rest } = response;
-        const displayOnly = { cost: void 0, inputTokens: void 0, outputTokens: void 0 };
+        const displayOnly = {
+          cost: void 0,
+          externalCostUsd: void 0,
+          inputTokens: void 0,
+          outputTokens: void 0,
+          cacheReadTokens: void 0,
+          cacheWriteTokens: void 0
+        };
         if (reasoning) {
           this.appendToHistory({
             ...rest,
@@ -6679,23 +6732,30 @@ function buildTimelineEntries(input) {
 }
 
 // src/state/persistence/diagnostics.ts
+function turnIdentity(entry) {
+  return `turn::${String(entry.turnIndex ?? "")}::${entry.timestamp ?? ""}::${String(entry.agentName ?? "")}`;
+}
 var KIND_RANK = { event: 0, turn: 1, message: 2 };
 var UNKNOWN_KIND_RANK = 99;
 function assembleDebugLog(entries, readme) {
-  const seen = /* @__PURE__ */ new Set();
+  const seenAt = /* @__PURE__ */ new Map();
   let latestMeta;
   const timeline = [];
   for (const entry of entries) {
-    const id = JSON.stringify(entry);
-    if (seen.has(id)) continue;
-    seen.add(id);
     if (entry.kind === "meta-snapshot") {
       if (!latestMeta || (entry.timestamp ?? "") >= (latestMeta.timestamp ?? "")) {
         latestMeta = entry;
       }
-    } else {
-      timeline.push(entry);
+      continue;
     }
+    const id = entry.kind === "turn" ? turnIdentity(entry) : JSON.stringify(entry);
+    const at = seenAt.get(id);
+    if (at !== void 0) {
+      if (entry.usage && !timeline[at].usage) timeline[at] = entry;
+      continue;
+    }
+    seenAt.set(id, timeline.length);
+    timeline.push(entry);
   }
   timeline.sort((a, b) => {
     const ta = a.timestamp ?? "";
@@ -6704,7 +6764,11 @@ function assembleDebugLog(entries, readme) {
     if (ta > tb) return 1;
     return (KIND_RANK[a.kind] ?? UNKNOWN_KIND_RANK) - (KIND_RANK[b.kind] ?? UNKNOWN_KIND_RANK);
   });
-  return { readme, timeline, meta: latestMeta?.meta };
+  const context = latestMeta?.meta?.context;
+  return { readme, sessionUsage: context?.sessionUsage, timeline, meta: latestMeta?.meta };
+}
+function withFreshMetaSnapshot(stored, fresh) {
+  return [...stored.filter((e) => e.kind !== "meta-snapshot"), fresh];
 }
 // Annotate the CommonJS export names for ESM import in node:
 0 && (module.exports = {
@@ -6740,11 +6804,13 @@ function assembleDebugLog(entries, readme) {
   geminiTokenCost,
   getMetaEvents,
   isObservableAIProviderRegistry,
+  messageUsage,
   restoreMachine,
   strictFallbackAgent,
   sumUsage,
   totalTokens,
   usageRows,
-  vendorOfModel
+  vendorOfModel,
+  withFreshMetaSnapshot
 });
 //# sourceMappingURL=chat-driver.cjs.map