@absolutejs/ai 0.3.3 → 0.4.0

package/CHANGELOG.md

@@ -6,6 +6,24 @@ This file is generated by `absolute-changelog` from the entries in
 `changelog/`. Edit an entry, not this file — and add new ones under
 `changelog/unreleased/`.
 
+## 0.4.0 — 2026-10-16
+
+### Added
+
+- **Report a tool call's input while it streams: a `tool_input` chunk, and a `preparing` tools event from streamAIToSSE** (`AIChunk`, `AIToolInputChunk`, `streamAIToSSE`)
+  A tool call with a long input, such as a whole file, takes the model a minute or more to write, and until now nothing arrived meanwhile. The Anthropic provider now yields a `tool_input` chunk when a tool call starts and every 2 KB of its input, with the tool name, the characters received and the top-level string fields already complete (a path, a name; at most 200 characters each, never read from inside other values). With structured events, streamAIToSSE turns it into a `tools` event with status `preparing`. Consumers that do not know the chunk can ignore it.
+
+### Fixed
+
+- **streamAIToSSE sends a turn's text back as one block, not one per streamed delta, so agent loops stop re-writing their prompt cache** (`streamAIToSSE`)
+  streamAIToSSE stored every streamed text delta as its own content block, so an agent loop sent each turn's text back as dozens of blocks. Anthropic looks back only 20 blocks for an earlier cache entry, and a turn with a few sentences of text pushed the previous request's entry out of reach: the next request re-wrote the whole history at the cache-write price. A run of text is now one block, as generateAI and streamAIWithTools already kept it.
+
+## 0.3.4 — 2026-10-15
+
+### Fixed
+
+- **Keep a prompt-cache breakpoint on the previous turn's user message as well as the last message, so an agent loop's next request reads its cached history by exact position instead of re-writing it; a production Studio build re-wrote its whole history on 7 of 48 requests and those writes were about half its cost.** (`anthropic`)
+
 ## 0.3.3 — 2026-10-10
 
 ### Fixed

package/changelog.json

@@ -2,6 +2,43 @@
 	"contract": 1,
 	"name": "@absolutejs/ai",
 	"releases": [
+		{
+			"changes": [
+				{
+					"detail": "streamAIToSSE stored every streamed text delta as its own content block, so an agent loop sent each turn's text back as dozens of blocks. Anthropic looks back only 20 blocks for an earlier cache entry, and a turn with a few sentences of text pushed the previous request's entry out of reach: the next request re-wrote the whole history at the cache-write price. A run of text is now one block, as generateAI and streamAIWithTools already kept it.",
+					"kind": "fixed",
+					"summary": "streamAIToSSE sends a turn's text back as one block, not one per streamed delta, so agent loops stop re-writing their prompt cache",
+					"symbols": [
+						"streamAIToSSE"
+					]
+				},
+				{
+					"detail": "A tool call with a long input, such as a whole file, takes the model a minute or more to write, and until now nothing arrived meanwhile. The Anthropic provider now yields a `tool_input` chunk when a tool call starts and every 2 KB of its input, with the tool name, the characters received and the top-level string fields already complete (a path, a name; at most 200 characters each, never read from inside other values). With structured events, streamAIToSSE turns it into a `tools` event with status `preparing`. Consumers that do not know the chunk can ignore it.",
+					"kind": "added",
+					"summary": "Report a tool call's input while it streams: a `tool_input` chunk, and a `preparing` tools event from streamAIToSSE",
+					"symbols": [
+						"AIChunk",
+						"AIToolInputChunk",
+						"streamAIToSSE"
+					]
+				}
+			],
+			"date": "2026-10-16",
+			"version": "0.4.0"
+		},
+		{
+			"changes": [
+				{
+					"kind": "fixed",
+					"summary": "Keep a prompt-cache breakpoint on the previous turn's user message as well as the last message, so an agent loop's next request reads its cached history by exact position instead of re-writing it; a production Studio build re-wrote its whole history on 7 of 48 requests and those writes were about half its cost.",
+					"symbols": [
+						"anthropic"
+					]
+				}
+			],
+			"date": "2026-10-15",
+			"version": "0.3.4"
+		},
 		{
 			"changes": [
 				{

package/dist/ai/index.js

@@ -999,10 +999,11 @@ var maybeFlushThinking = (chunkState) => {
 };
 var processTextChunk = function* (content, chunkState, renderers, fullResponse, options) {
   maybeFlushThinking(chunkState);
-  chunkState.contentBlocks.push({
-    content,
-    type: "text"
-  });
+  const last = chunkState.contentBlocks.at(-1);
+  if (last?.type === "text")
+    last.content += content;
+  else
+    chunkState.contentBlocks.push({ content, type: "text" });
   yield contentEvent(options, renderers, content, fullResponse + content);
 };
 var processImageChunk = function* (chunk, renderers, options) {
@@ -1056,6 +1057,19 @@ var processChunk = function* (chunk, chunkState, renderers, options, fullRespons
     case "tool_use":
       processToolUseChunk(chunk, chunkState);
       break;
+    case "tool_input":
+      if (options.structuredEvents)
+        yield {
+          data: JSON.stringify({
+            fields: chunk.fields,
+            id: chunk.id,
+            name: chunk.name,
+            receivedChars: chunk.receivedChars,
+            status: "preparing"
+          }),
+          event: "tools"
+        };
+      break;
     case "provider_event":
       maybeFlushThinking(chunkState);
       chunkState.contentBlocks.push({
@@ -6735,9 +6749,15 @@ var buildRequestBody3 = (params, configuredMax, configCaching) => {
   const options = requestOptionsFor(params);
   const messages = params.messages.filter((msg) => msg.role !== "system").map(mapMessage);
   if (caching && messages.length > 1) {
-    const last = messages[messages.length - 1];
+    const lastIndex = messages.length - 1;
+    const last = messages[lastIndex];
     if (last) {
-      messages[messages.length - 1] = cacheLastContentBlock(last);
+      messages[lastIndex] = cacheLastContentBlock(last);
+    }
+    const previousUserIndex = messages.findLastIndex((message, index) => index < lastIndex - 1 && message.role === "user");
+    const previous = messages[previousUserIndex];
+    if (previous) {
+      messages[previousUserIndex] = cacheLastContentBlock(previous);
     }
   }
   const max = typeof params.maxTokens === "number" ? params.maxTokens : configuredMax;
@@ -6872,14 +6892,79 @@ var getNumber = (obj, key) => {
   }
   return 0;
 };
+var TOOL_INPUT_PROGRESS_CHARS = 2048;
+var TOOL_INPUT_FIELD_MAX_CHARS = 200;
+var readJsonString = (json, start) => {
+  for (let index = start + 1;index < json.length; index += 1) {
+    if (json[index] === "\\")
+      index += 1;
+    else if (json[index] === '"')
+      return { end: index + 1, raw: json.slice(start + 1, index) };
+  }
+  return;
+};
+var decodeJsonString = (raw) => {
+  try {
+    const value = JSON.parse(`"${raw}"`);
+    return typeof value === "string" ? value : undefined;
+  } catch {
+    return;
+  }
+};
+var completeStringFields = (partialJson) => {
+  const fields = {};
+  let depth = 0;
+  let key;
+  let index = 0;
+  while (index < partialJson.length) {
+    const character = partialJson[index];
+    if (character === '"') {
+      const text = readJsonString(partialJson, index);
+      if (!text)
+        break;
+      if (depth === 1 && key === undefined) {
+        const after = partialJson.slice(text.end).trimStart();
+        if (after.startsWith(":"))
+          key = text.raw;
+      } else if (depth === 1 && key !== undefined) {
+        const value = decodeJsonString(text.raw);
+        if (value !== undefined && value.length <= TOOL_INPUT_FIELD_MAX_CHARS)
+          fields[key] = value;
+        key = undefined;
+      }
+      index = text.end;
+      continue;
+    }
+    if (character === "{" || character === "[")
+      depth += 1;
+    else if (character === "}" || character === "]")
+      depth -= 1;
+    else if (character === "," && depth === 1)
+      key = undefined;
+    index += 1;
+  }
+  return fields;
+};
+var toolInputProgress = (state) => {
+  state.toolInputReported = state.toolInputJson.length;
+  return {
+    fields: completeStringFields(state.toolInputJson),
+    id: state.currentToolId,
+    name: state.currentToolName,
+    receivedChars: state.toolInputJson.length,
+    type: "tool_input"
+  };
+};
 var handleContentBlockStart = (parsed, state) => {
   const block = getRecord(parsed, "content_block");
   if (block && block.type === "tool_use") {
     state.currentToolId = getString(block, "id");
     state.currentToolName = getString(block, "name");
     state.toolInputJson = "";
+    state.toolInputReported = 0;
     state.isThinkingBlock = false;
     state.currentProviderBlock = undefined;
+    return toolInputProgress(state);
   } else if (block && block.type === "thinking") {
     state.isThinkingBlock = true;
     state.thinkingSignature = "";
@@ -6889,6 +6974,7 @@ var handleContentBlockStart = (parsed, state) => {
     state.currentProviderBlock = block && block.type !== "text" ? { ...block } : undefined;
     state.providerBlockInputJson = "";
   }
+  return;
 };
 var handleContentBlockDelta = (parsed, state) => {
   const delta = getRecord(parsed, "delta");
@@ -6923,6 +7009,8 @@ var handleContentBlockDelta = (parsed, state) => {
       state.providerBlockInputJson += getString(delta, "partial_json");
     } else {
       state.toolInputJson += getString(delta, "partial_json");
+      if (state.toolInputJson.length - state.toolInputReported >= TOOL_INPUT_PROGRESS_CHARS)
+        return toolInputProgress(state);
     }
   }
   if (delta.type === "signature_delta") {
@@ -7042,8 +7130,7 @@ var handleError = (parsed, state) => {
 var processEvent = (eventType, parsed, state) => {
   switch (eventType) {
     case "content_block_start": {
-      handleContentBlockStart(parsed, state);
-      return;
+      return handleContentBlockStart(parsed, state);
     }
     case "content_block_delta": {
       return handleContentBlockDelta(parsed, state);
@@ -7146,6 +7233,7 @@ async function* parseSSEStream4(body, providerName, signal) {
     stopReason: "",
     thinkingSignature: "",
     toolInputJson: "",
+    toolInputReported: 0,
     usage: undefined,
     providerName,
     providerBlockInputJson: ""
@@ -9056,5 +9144,5 @@ export {
   xai
 };
 
-//# debugId=C9BA825E688A09A764756E2164756E21
+//# debugId=4ED2ED351363579D64756E2164756E21
 //# sourceMappingURL=index.js.map