@firstlovecenter/ai-chat 0.7.0 → 0.8.0

package/dist/server/index.d.cts

@@ -55,6 +55,14 @@ type NormalizedToolResult = {
 type NormalizedMessage = {
     role: 'user';
     text: string;
+    /**
+     * Cache hint: when true, the producing route is asking the provider
+     * to mark this message's content with a cache breakpoint so the
+     * full prefix becomes cacheable on a subsequent request. Anthropic
+     * applies `cache_control: ephemeral`; Vertex Gemini ignores the
+     * flag (its prefix cache works automatically).
+     */
+    cached?: boolean;
 } | {
     role: 'assistant';
     /** Free text the model emitted (zero-or-more text blocks joined as-is). */
@@ -69,6 +77,8 @@ type NormalizedMessage = {
      * thought_signature`. Other adapters can ignore.
      */
     providerData?: unknown;
+    /** See `user.cached`. */
+    cached?: boolean;
 } | {
     role: 'tool';
     results: NormalizedToolResult[];
@@ -155,6 +165,16 @@ type AgentInput<S = unknown> = {
     systemBlocks: SystemBlock[];
     /** Constructed ToolProvider — caller resolves the right one via toolProviders[id].createProvider({...}). */
     provider: ToolProvider;
+    /**
+     * Conversation history to seed the prompt with, in chronological order.
+     * Hosts pass this to give the model memory across turns in a chat session
+     * (so a follow-up like "summarize that" resolves the antecedent). The
+     * route handler is responsible for fetching prior `chat_messages` and
+     * normalising them; see `historyToNormalizedMessages` in `./history`.
+     * Tool-call provenance is intentionally not replayed — assistant turns
+     * here should be plain text only.
+     */
+    priorMessages?: NormalizedMessage[];
     /** Optional caps. Default both. */
     maxToolTurns?: number;
     maxOutputTokens?: number;

package/dist/server/index.d.ts

@@ -55,6 +55,14 @@ type NormalizedToolResult = {
 type NormalizedMessage = {
     role: 'user';
     text: string;
+    /**
+     * Cache hint: when true, the producing route is asking the provider
+     * to mark this message's content with a cache breakpoint so the
+     * full prefix becomes cacheable on a subsequent request. Anthropic
+     * applies `cache_control: ephemeral`; Vertex Gemini ignores the
+     * flag (its prefix cache works automatically).
+     */
+    cached?: boolean;
 } | {
     role: 'assistant';
     /** Free text the model emitted (zero-or-more text blocks joined as-is). */
@@ -69,6 +77,8 @@ type NormalizedMessage = {
      * thought_signature`. Other adapters can ignore.
      */
     providerData?: unknown;
+    /** See `user.cached`. */
+    cached?: boolean;
 } | {
     role: 'tool';
     results: NormalizedToolResult[];
@@ -155,6 +165,16 @@ type AgentInput<S = unknown> = {
     systemBlocks: SystemBlock[];
     /** Constructed ToolProvider — caller resolves the right one via toolProviders[id].createProvider({...}). */
     provider: ToolProvider;
+    /**
+     * Conversation history to seed the prompt with, in chronological order.
+     * Hosts pass this to give the model memory across turns in a chat session
+     * (so a follow-up like "summarize that" resolves the antecedent). The
+     * route handler is responsible for fetching prior `chat_messages` and
+     * normalising them; see `historyToNormalizedMessages` in `./history`.
+     * Tool-call provenance is intentionally not replayed — assistant turns
+     * here should be plain text only.
+     */
+    priorMessages?: NormalizedMessage[];
     /** Optional caps. Default both. */
     maxToolTurns?: number;
     maxOutputTokens?: number;

package/dist/server/index.js

@@ -23,7 +23,10 @@ async function runAgent(input) {
   const maxOutputTokens = input.maxOutputTokens ?? DEFAULT_MAX_OUTPUT_TOKENS;
   const transcript = [];
   transcript.push({ kind: "user", text: input.question });
-  const messages = [{ role: "user", text: input.question }];
+  const messages = [
+    ...input.priorMessages ?? [],
+    { role: "user", text: input.question }
+  ];
   const system = input.systemBlocks;
   const toolSchemas = Object.values(input.tools).map((t) => t.schema);
   let toolCallCount = 0;
@@ -216,11 +219,28 @@ function toAnthropicMessages(messages) {
   const out = [];
   for (const msg of messages) {
     if (msg.role === "user") {
-      out.push({ role: "user", content: msg.text });
+      if (msg.cached) {
+        out.push({
+          role: "user",
+          content: [
+            {
+              type: "text",
+              text: msg.text,
+              cache_control: { type: "ephemeral" }
+            }
+          ]
+        });
+      } else {
+        out.push({ role: "user", content: msg.text });
+      }
     } else if (msg.role === "assistant") {
       const blocks = [];
       if (msg.text) {
-        blocks.push({ type: "text", text: msg.text });
+        const textBlock = { type: "text", text: msg.text };
+        if (msg.cached && msg.toolCalls.length === 0) {
+          textBlock.cache_control = { type: "ephemeral" };
+        }
+        blocks.push(textBlock);
       }
       for (const tc of msg.toolCalls) {
         blocks.push({
@@ -572,6 +592,87 @@ var toolProviders = [
 function getToolProvider(id) {
   return toolProviders.find((p) => p.id === id);
 }
+
+// src/server/history.ts
+var DEFAULT_MAX_HISTORY_PAIRS = 20;
+var DEFAULT_MAX_TEXT_CHARS = 4e3;
+function historyToNormalizedMessages(rows, opts = {}) {
+  const maxPairs = opts.maxPairs ?? DEFAULT_MAX_HISTORY_PAIRS;
+  const maxTextChars = opts.maxTextChars ?? DEFAULT_MAX_TEXT_CHARS;
+  const pairs = [];
+  let i = 0;
+  while (i < rows.length) {
+    const row = rows[i];
+    if (row.role !== "user" || !row.question) {
+      i += 1;
+      continue;
+    }
+    const next = rows[i + 1];
+    if (next?.role !== "assistant") {
+      i += 1;
+      continue;
+    }
+    const assistantText = truncate(
+      assistantMessageToText(next),
+      maxTextChars
+    );
+    if (assistantText) {
+      pairs.push([
+        { role: "user", text: truncate(row.question, maxTextChars) },
+        { role: "assistant", text: assistantText, toolCalls: [] }
+      ]);
+    }
+    i += 2;
+  }
+  const kept = maxPairs > 0 ? pairs.slice(-maxPairs) : pairs;
+  return kept.flat();
+}
+function truncate(text, max) {
+  if (max <= 0 || text.length <= max) return text;
+  return text.slice(0, max);
+}
+function assistantMessageToText(row) {
+  if (row.errorJson) return "";
+  const proseText = proseToText(row.prose);
+  if (proseText) return proseText;
+  const blockText = blocksToText(row.blocks);
+  return blockText;
+}
+function proseToText(prose) {
+  if (!prose || typeof prose !== "object") return "";
+  const entries = Object.entries(prose).map(([k, v]) => [Number(k), typeof v === "string" ? v : ""]).filter(([k, v]) => Number.isFinite(k) && v.length > 0).sort(([a], [b]) => a - b);
+  return entries.map(([, v]) => v).join("\n\n").trim();
+}
+function blocksToText(blocks) {
+  if (!Array.isArray(blocks)) return "";
+  const parts = [];
+  for (const raw of blocks) {
+    if (!raw || typeof raw !== "object") continue;
+    const b = raw;
+    switch (b.kind) {
+      case "paragraph_brief": {
+        const facts = (b.key_facts ?? []).filter((f) => f && f.trim());
+        if (b.topic) parts.push(b.topic);
+        if (facts.length) parts.push(facts.join("\n"));
+        break;
+      }
+      case "list": {
+        const items = (b.items ?? []).filter((s) => s && s.trim());
+        if (b.title) parts.push(b.title);
+        if (items.length) parts.push(items.map((s) => `- ${s}`).join("\n"));
+        break;
+      }
+      case "chart":
+      case "table":
+        if (b.title) parts.push(`[${b.title}]`);
+        break;
+      case "callout":
+        if (b.text) parts.push(b.text);
+        break;
+    }
+  }
+  return parts.join("\n\n").trim();
+}
 function vertexHost2(location) {
   return location === "global" ? "aiplatform.googleapis.com" : `${location}-aiplatform.googleapis.com`;
 }
@@ -914,12 +1015,23 @@ function createAgentCustomRoutes(ctx) {
       const aiSettings = await persistence.getAiSettings();
       const effectiveProjectId = aiSettings.gcpProjectId ?? vertex.projectId;
       let chatSessionId;
+      let priorMessages = [];
       if (incomingChatSessionId !== null) {
         const owned = await persistence.getSession(incomingChatSessionId, userId);
         if (!owned) {
           return jsonError(404, "NOT_FOUND", "Chat session not found.");
         }
         chatSessionId = owned.id;
+        const stored = await persistence.listMessagesForSession(chatSessionId, userId);
+        priorMessages = historyToNormalizedMessages(stored);
+        if (priorMessages.length > 0) {
+          const last = priorMessages[priorMessages.length - 1];
+          if (last.role === "assistant" && last.toolCalls.length === 0) {
+            priorMessages[priorMessages.length - 1] = { ...last, cached: true };
+          } else if (last.role === "user") {
+            priorMessages[priorMessages.length - 1] = { ...last, cached: true };
+          }
+        }
       } else {
         const created = await persistence.createSession({
           userId,
@@ -1007,6 +1119,7 @@ data: ${JSON.stringify(data)}
             send("meta", { chatSessionId, scopeLabel });
             const agentResult = await runAgent({
               question,
+              priorMessages,
               ctx: toolContext,
               tools: tools.tools,
               systemBlocks,
@@ -1225,7 +1338,17 @@ function createAgentVercelRoutes(ctx) {
         if (short) return short;
       }
       const body = await req.json().catch(() => null);
-      const question = typeof body?.question === "string" ? body.question.trim() : "";
+      let question = typeof body?.question === "string" ? body.question.trim() : "";
+      if (!question && Array.isArray(body?.messages)) {
+        const msgs = body.messages;
+        for (let i = msgs.length - 1; i >= 0; i -= 1) {
+          const m = msgs[i];
+          if (m && m.role === "user" && typeof m.content === "string") {
+            question = m.content.trim();
+            break;
+          }
+        }
+      }
       if (!question) {
         return jsonError2(
           400,
@@ -1240,6 +1363,7 @@ function createAgentVercelRoutes(ctx) {
       const aiSettings = await persistence.getAiSettings();
       const effectiveProjectId = aiSettings.gcpProjectId ?? vertex.projectId;
       let chatSessionId;
+      let priorMessages = [];
       if (incomingChatSessionId !== null) {
         const owned = await persistence.getSession(
           incomingChatSessionId,
@@ -1249,6 +1373,11 @@ function createAgentVercelRoutes(ctx) {
           return jsonError2(404, "NOT_FOUND", "Chat session not found.");
         }
         chatSessionId = owned.id;
+        const stored = await persistence.listMessagesForSession(
+          chatSessionId,
+          userId
+        );
+        priorMessages = historyToNormalizedMessages(stored);
       } else {
         const created = await persistence.createSession({
           userId,
@@ -1318,10 +1447,16 @@ function createAgentVercelRoutes(ctx) {
           location: aiSettings.gcpLocation,
           googleAuthOptions: {}
         })(vertex.modelIds.gemini);
+        const priorCoreMessages = priorMessages.filter(
+          (m) => m.role === "user" || m.role === "assistant"
+        ).map((m) => ({ role: m.role, content: m.text }));
         const result = streamText({
           model,
           system,
-          messages: [{ role: "user", content: question }],
+          messages: [
+            ...priorCoreMessages,
+            { role: "user", content: question }
+          ],
           tools: vercelTools,
           maxSteps: 12,
           maxTokens: aiSettings.maxOutputTokens,