@absolutejs/ai 0.2.4 → 0.2.6

package/CHANGELOG.md

@@ -6,6 +6,18 @@ 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/`.
 
+## 0.2.6 — 2026-09-20
+
+### Fixed
+
+- **Count supported Anthropic caller input when hosted tools cannot be pre-counted, preserving full generation requests and exposing honest count scope** (`anthropic`, `inspectAIInput`, `AIInputCapacity`, `AIProviderConfig`, `createAIProviderRouter`)
+
+## 0.2.5 — 2026-09-20
+
+### Fixed
+
+- **Allow SSE tool loops to reread changed files without silently completing; retain turn, time and token limits** (`streamAIToSSE`)
+
 ## 0.2.4 — 2026-09-20
 
 ### Added

package/README.md

@@ -388,6 +388,17 @@ Unknown models/providers fail with `AIInputError` (`capacity_unavailable`) rathe
 than guessing. Custom providers can implement the optional `inputCapacity`
 capability. Existing `stream` calls remain unchanged.
 
+Anthropic cannot pre-count hosted web search, web fetch, code execution or tool
+search definitions. The adapter omits only those definitions (and dependent tool
+choice) from the counting request, preserving system text, messages and client
+tools. `inspectAIInput(...).inputTokenScope` is `"caller-input"` in this case,
+and `"request"` otherwise; provider routers preserve the scope. `fits` describes
+the counted scope, not a guarantee that future hosted results fit. Generation
+still sends the complete hosted-tool configuration, and Anthropic validates the
+full request and reports actual usage. No guessed overhead or paid probe is used.
+Custom providers can expose `inputCapacity.countScope(params)` without changing
+the numeric `countTokens` contract. See the [Anthropic counting limitations](https://platform.claude.com/docs/en/build-with-claude/token-counting).
+
 Automatic preparation supports text-only history. It rejects oversized tool or
 multimodal histories rather than flattening them. Handle preparation errors in
 the UI while preserving the original input; expose retry without clearing the

package/changelog.json

@@ -2,6 +2,36 @@
 	"contract": 1,
 	"name": "@absolutejs/ai",
 	"releases": [
+		{
+			"changes": [
+				{
+					"kind": "fixed",
+					"summary": "Count supported Anthropic caller input when hosted tools cannot be pre-counted, preserving full generation requests and exposing honest count scope",
+					"symbols": [
+						"anthropic",
+						"inspectAIInput",
+						"AIInputCapacity",
+						"AIProviderConfig",
+						"createAIProviderRouter"
+					]
+				}
+			],
+			"date": "2026-09-20",
+			"version": "0.2.6"
+		},
+		{
+			"changes": [
+				{
+					"kind": "fixed",
+					"summary": "Allow SSE tool loops to reread changed files without silently completing; retain turn, time and token limits",
+					"symbols": [
+						"streamAIToSSE"
+					]
+				}
+			],
+			"date": "2026-09-20",
+			"version": "0.2.5"
+		},
 		{
 			"changes": [
 				{

package/dist/ai/index.js

@@ -40,6 +40,7 @@ var inspectAIInput = async (provider, params) => {
   return {
     limits,
     inputTokens,
+    inputTokenScope: provider.inputCapacity.countScope?.(params) ?? "request",
     outputTokens,
     availableInputTokens,
     fits: outputTokens <= maxOutputTokens && inputTokens <= availableInputTokens
@@ -2886,6 +2887,35 @@ var anthropic = (config2) => {
       signal: params.signal ? AbortSignal.any([params.signal, AbortSignal.timeout(15000)]) : AbortSignal.timeout(15000)
     }));
   };
+  const countingRequest = (params) => {
+    const built = buildRequestBody3(params, configuredMax, promptCaching);
+    const body = config2.transformRequestBody ? config2.transformRequestBody(built, params) : built;
+    const counted = {};
+    for (const key of [
+      "model",
+      "messages",
+      "system",
+      "tools",
+      "tool_choice",
+      "thinking"
+    ])
+      if (body[key] !== undefined)
+        counted[key] = body[key];
+    const tools = Array.isArray(body.tools) ? body.tools : [];
+    const hosted = tools.filter((tool) => isRecord4(tool) && typeof tool.type === "string" && /^(web_search_|web_fetch_|code_execution_|bash_code_execution_|text_editor_code_execution_|tool_search_tool_)/u.test(tool.type));
+    const scope = hosted.length > 0 || body.mcp_servers !== undefined ? "caller-input" : "request";
+    if (hosted.length > 0) {
+      const clientTools = tools.filter((tool) => !hosted.includes(tool));
+      if (clientTools.length)
+        counted.tools = clientTools;
+      else
+        delete counted.tools;
+      const choice = counted.tool_choice;
+      if (!clientTools.length || isRecord4(choice) && choice.type === "tool" && hosted.some((tool) => isRecord4(tool) && tool.name === choice.name))
+        delete counted.tool_choice;
+    }
+    return { counted, scope };
+  };
   const inputCapacity = {
     outputTokens: (params) => {
       const built = buildRequestBody3(params, configuredMax, promptCaching);
@@ -2901,20 +2931,9 @@ var anthropic = (config2) => {
         maxOutputTokens: positiveTokenLimit(data.max_tokens)
       };
     }, config2.tokenSource || typeof config2.headers === "function" ? 0 : 300000),
+    countScope: (params) => countingRequest(params).scope,
     countTokens: async (params) => {
-      const built = buildRequestBody3(params, configuredMax, promptCaching);
-      const body = config2.transformRequestBody ? config2.transformRequestBody(built, params) : built;
-      const counted = {};
-      for (const key of [
-        "model",
-        "messages",
-        "system",
-        "tools",
-        "tool_choice",
-        "thinking"
-      ])
-        if (body[key] !== undefined)
-          counted[key] = body[key];
+      const { counted } = countingRequest(params);
       return inputTokenCount((await capacityRequest(params, "/v1/messages/count_tokens", counted)).input_tokens);
     }
   };
@@ -4892,7 +4911,6 @@ var executeTool2 = async (options, toolName, toolInput) => {
     return `Error: ${err instanceof Error ? err.message : String(err)}`;
   }
 };
-var serializeToolCall2 = (name, input) => `${name}:${JSON.stringify(input)}`;
 var contentEvent = (options, renderers, delta, full) => options.structuredEvents ? { data: JSON.stringify({ delta, full }), event: "content" } : { data: renderers.chunk(delta, full), event: "content" };
 var thinkingEvent = (options, renderers, text) => options.structuredEvents ? { data: JSON.stringify({ text }), event: "thinking" } : { data: renderers.thinking(text), event: "thinking" };
 var imageEvent = (options, renderers, data, format, revisedPrompt) => options.structuredEvents ? { data: JSON.stringify({ data, format, revisedPrompt }), event: "images" } : { data: renderers.image(data, format, revisedPrompt), event: "images" };
@@ -5074,7 +5092,6 @@ var executeToolCalls = async function* (pendingToolCalls, options, renderers, tu
       tool_use_id: toolCall.id,
       type: "tool_result"
     });
-    turnState.executedToolKeys.add(serializeToolCall2(toolCall.name, toolCall.input));
   }
   return toolResultBlocks;
 };
@@ -5089,12 +5106,7 @@ var consumeStream2 = async function* (stream, chunkState, renderers, options, tu
     turnState.fullResponse = prevResponse + chunk.content;
   }
 };
-var shouldStopToolLoop = (chunkState, turnState, signal) => {
-  if (chunkState.pendingToolCalls.length === 0 || signal.aborted) {
-    return true;
-  }
-  return chunkState.pendingToolCalls.every((toolCall) => turnState.executedToolKeys.has(serializeToolCall2(toolCall.name, toolCall.input)));
-};
+var shouldStopToolLoop = (chunkState, signal) => chunkState.pendingToolCalls.length === 0 || signal.aborted;
 var processTurn = async function* (chunkState, options, renderers, turnState) {
   turnState.currentMessages.push({
     content: chunkState.contentBlocks,
@@ -5110,7 +5122,6 @@ var streamTurns = async function* (options, renderers, messages, signal, startTi
   const turnState = {
     allToolsHtml: "",
     currentMessages: [...messages],
-    executedToolKeys: new Set,
     fullResponse: "",
     turn: 0
   };
@@ -5172,7 +5183,7 @@ var streamTurns = async function* (options, renderers, messages, signal, startTi
         yield stoppedEvent(options, renderers, "max_duration_ms", `Stopped: time budget reached (${Math.round((Date.now() - startTime) / 1000)}s over ` + `${turnState.turn} turns). Narrow the request or raise maxDurationMs.`);
         return;
       }
-      if (shouldStopToolLoop(chunkState, turnState, signal)) {
+      if (shouldStopToolLoop(chunkState, signal)) {
         if (signal.aborted) {
           finishReason = "aborted";
           yield stoppedEvent(options, renderers, "aborted", "Aborted by caller.");
@@ -8773,6 +8784,7 @@ var createAIProviderRouter = (routes) => {
       outputTokens: (params) => capacity(params).outputTokens?.(params),
       isContextError: (error) => [...new Set(providers.values())].some((provider) => provider.inputCapacity?.isContextError?.(error) === true),
       getLimits: async (params) => capacity(params).getLimits(params),
+      countScope: (params) => capacity(params).countScope?.(params) ?? "request",
       countTokens: async (params) => capacity(params).countTokens(params)
     },
     stream: (params) => resolve(params).stream(params)
@@ -8888,5 +8900,5 @@ export {
   xai
 };
 
-//# debugId=C24827A491DCDA4A64756E2164756E21
+//# debugId=33E15B35E923B1E064756E2164756E21
 //# sourceMappingURL=index.js.map