@lowire/loop 0.0.22 → 0.0.24

package/lib/cache.d.ts

@@ -20,4 +20,5 @@ export type ReplayCaches = {
 };
 export declare function cachedComplete(provider: types.Provider, conversation: types.Conversation, caches: ReplayCaches | undefined, options: types.CompletionOptions & {
     secrets?: Record<string, string>;
+    cacheMode?: 'strict' | 'lax';
 }): ReturnType<types.Provider['complete']>;

package/lib/cache.js

@@ -28,7 +28,13 @@ async function cachedComplete(provider, conversation, caches, options) {
 async function cachedCompleteNoSecrets(provider, conversation, caches, options) {
     if (!caches)
         return await provider.complete(conversation, options);
-    const key = calculateSha1(JSON.stringify(conversation));
+    const keyObject = {
+        conversation: options.cacheMode === 'lax' ? { ...conversation, tools: [] } : conversation,
+        maxTokens: options.maxTokens,
+        reasoning: options.reasoning,
+        temperature: options.temperature,
+    };
+    const key = calculateSha1(JSON.stringify(keyObject));
     if (!process.env.LOWIRE_NO_CACHE && caches.input[key]) {
         caches.output[key] = caches.input[key];
         return caches.input[key] ?? caches.output[key];

package/lib/loop.d.ts

@@ -48,6 +48,7 @@ export type LoopOptions = types.CompletionOptions & LoopEvents & {
     maxToolCalls?: number;
     maxToolCallRetries?: number;
     cache?: types.ReplayCache;
+    cacheMode?: 'strict' | 'lax';
     secrets?: Record<string, string>;
     summarize?: boolean;
 };

package/lib/loop.js

@@ -60,11 +60,18 @@ class Loop {
             if (abortController?.signal.aborted)
                 return { status: 'break', usage: totalUsage, turns };
             debug?.('lowire:loop')(`Request`, JSON.stringify({ ...summarizedConversation, tools: `${summarizedConversation.tools.length} tools` }, null, 2));
+            const tokenEstimate = Math.floor(JSON.stringify(summarizedConversation).length / 4);
+            if (budget.tokens !== undefined && tokenEstimate >= budget.tokens)
+                return { status: 'error', error: `Input token estimate ${tokenEstimate} exceeds budget ${budget.tokens}`, usage: totalUsage, turns };
             const { result: assistantMessage, usage } = await (0, cache_1.cachedComplete)(this._provider, summarizedConversation, caches, {
                 ...options,
-                maxTokens: budget.tokens,
+                maxTokens: budget.tokens !== undefined ? budget.tokens - tokenEstimate : undefined,
                 signal: abortController?.signal,
             });
+            if (assistantMessage.stopReason.code === 'error')
+                return { status: 'error', error: assistantMessage.stopReason.message, usage: totalUsage, turns };
+            if (assistantMessage.stopReason.code === 'max_tokens')
+                return { status: 'error', error: `Max tokens exhausted`, usage: totalUsage, turns };
             const intent = assistantMessage.content.filter(part => part.type === 'text').map(part => part.text).join('\n');
             totalUsage.input += usage.input;
             totalUsage.output += usage.output;

package/lib/providers/anthropic.js

@@ -17,11 +17,12 @@
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.Anthropic = void 0;
 const fetchWithTimeout_1 = require("../fetchWithTimeout");
+const types_1 = require("../types");
 class Anthropic {
     name = 'anthropic';
     async complete(conversation, options) {
         const maxTokens = Math.min(options.maxTokens ?? 32_768, 32_768);
-        const response = await create({
+        const { response, error } = await create({
             model: options.model,
             max_tokens: maxTokens,
             temperature: options.temperature,
@@ -33,6 +34,8 @@ class Anthropic {
                 budget_tokens: options.maxTokens ? Math.round(maxTokens / 10) : 1024,
             } : undefined,
         }, options);
+        if (error || !response)
+            return { result: (0, types_1.assistantMessageFromError)(error ?? 'No response from Anthropic API'), usage: (0, types_1.emptyUsage)() };
         const result = toAssistantMessage(response);
         const usage = {
             input: response.usage.input_tokens,
@@ -57,13 +60,14 @@ async function create(createParams, options) {
         signal: options.signal,
         timeout: options.apiTimeout
     });
+    const responseText = await response.text();
+    const responseBody = JSON.parse(responseText);
+    options.debug?.('lowire:anthropic')('Response:', responseText);
     if (!response.ok) {
         options.debug?.('lowire:anthropic')('Response:', response.status);
-        throw new Error(`API error: ${response.status} ${response.statusText} ${await response.text()}`);
+        return { error: `API error: ${response.status} ${response.statusText} ${responseText}` };
     }
-    const responseBody = await response.json();
-    options.debug?.('lowire:anthropic')('Response:', JSON.stringify(responseBody, null, 2));
-    return responseBody;
+    return { response: responseBody };
 }
 function toContentPart(block) {
     if (block.type === 'text') {
@@ -109,9 +113,13 @@ function toAnthropicResultParam(part) {
     throw new Error(`Unsupported content part type: ${part.type}`);
 }
 function toAssistantMessage(message) {
+    const stopReason = { code: 'ok' };
+    if (message.stop_reason === 'max_tokens')
+        stopReason.code = 'max_tokens';
     return {
         role: 'assistant',
         content: message.content.map(toContentPart).filter(Boolean),
+        stopReason,
     };
 }
 function toAnthropicTool(tool) {

package/lib/providers/google.js

@@ -17,11 +17,12 @@
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.Google = void 0;
 const fetchWithTimeout_1 = require("../fetchWithTimeout");
+const types_1 = require("../types");
 class Google {
     name = 'google';
     async complete(conversation, options) {
         const contents = conversation.messages.map(toGeminiContent).flat();
-        const response = await create(options.model ?? 'gemini-2.5-pro', {
+        const { response, error } = await create(options.model ?? 'gemini-2.5-pro', {
             systemInstruction: {
                 role: 'system',
                 parts: [
@@ -30,11 +31,14 @@ class Google {
             },
             contents,
             tools: conversation.tools.length > 0 ? [{ functionDeclarations: conversation.tools.map(toGeminiTool) }] : undefined,
-            generationConfig: { temperature: options.temperature },
+            generationConfig: {
+                temperature: options.temperature,
+                maxOutputTokens: options.maxTokens
+            },
         }, options);
-        const [candidate] = response.candidates ?? [];
-        if (!candidate)
-            throw new Error('No candidates in response');
+        const [candidate] = response?.candidates ?? [];
+        if (error || !response || !candidate)
+            return { result: (0, types_1.assistantMessageFromError)(error ?? 'No response from Google API'), usage: (0, types_1.emptyUsage)() };
         const usage = {
             input: response.usageMetadata?.promptTokenCount ?? 0,
             output: response.usageMetadata?.candidatesTokenCount ?? 0,
@@ -59,11 +63,11 @@ async function create(model, createParams, options) {
     });
     if (!response.ok) {
         options.debug?.('lowire:google')('Response:', response.status);
-        throw new Error(`API error: ${response.status} ${response.statusText} ${await response.text()}`);
+        return { error: `API error: ${response.status} ${response.statusText} ${await response.text()}` };
     }
     const responseBody = await response.json();
     options.debug?.('lowire:google')('Response:', JSON.stringify(responseBody, null, 2));
-    return responseBody;
+    return { response: responseBody };
 }
 function toGeminiTool(tool) {
     return {
@@ -85,9 +89,13 @@ function stripUnsupportedSchemaFields(schema) {
     return cleaned;
 }
 function toAssistantMessage(candidate) {
+    const stopReason = { code: 'ok' };
+    if (candidate.finishReason === 'MAX_TOKENS')
+        stopReason.code = 'max_tokens';
     return {
         role: 'assistant',
-        content: candidate.content.parts.map(toContentPart).filter(Boolean),
+        content: (candidate.content.parts || []).map(toContentPart).filter(Boolean),
+        stopReason,
     };
 }
 function toContentPart(part) {

package/lib/providers/openai.js

@@ -17,6 +17,7 @@
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.OpenAI = void 0;
 const fetchWithTimeout_1 = require("../fetchWithTimeout");
+const types_1 = require("../types");
 class OpenAI {
     name = 'openai';
     async complete(conversation, options) {
@@ -27,7 +28,7 @@ exports.OpenAI = OpenAI;
 async function complete(conversation, options) {
     const inputItems = conversation.messages.map(toResponseInputItems).flat();
     const tools = conversation.tools.map(toOpenAIFunctionTool);
-    const response = await create({
+    const { response, error } = await create({
         model: options.model,
         temperature: options.temperature,
         input: inputItems,
@@ -35,10 +36,22 @@ async function complete(conversation, options) {
         tools: tools.length > 0 ? tools : undefined,
         tool_choice: conversation.tools.length > 0 ? 'auto' : undefined,
         parallel_tool_calls: false,
+        max_output_tokens: options.maxTokens,
         reasoning: toOpenAIReasoning(options.reasoning),
     }, options);
+    if (!response || error)
+        return { result: (0, types_1.assistantMessageFromError)(error ?? 'No response from OpenAI API'), usage: (0, types_1.emptyUsage)() };
     // Parse response output items
-    const result = { role: 'assistant', content: [] };
+    const stopReason = { code: 'ok' };
+    if (response.incomplete_details?.reason === 'max_output_tokens')
+        stopReason.code = 'max_tokens';
+    const result = { role: 'assistant', content: [], stopReason };
+    const usage = {
+        input: response.usage?.input_tokens ?? 0,
+        output: response.usage?.output_tokens ?? 0,
+    };
+    if (stopReason.code !== 'ok')
+        return { result, usage };
     for (const item of response.output) {
         if (item.type === 'message' && item.role === 'assistant') {
             result.openaiId = item.id;
@@ -57,10 +70,6 @@ async function complete(conversation, options) {
             result.content.push(toToolCall(item));
         }
     }
-    const usage = {
-        input: response.usage?.input_tokens ?? 0,
-        output: response.usage?.output_tokens ?? 0,
-    };
     return { result, usage };
 }
 async function create(createParams, options) {
@@ -77,13 +86,18 @@ async function create(createParams, options) {
         signal: options.signal,
         timeout: options.apiTimeout
     });
+    const responseText = await response.text();
+    const responseBody = JSON.parse(responseText);
     if (!response.ok) {
-        options.debug?.('lowire:openai-responses')('Response:', response.status);
-        throw new Error(`API error: ${response.status} ${response.statusText} ${await response.text()}`);
+        try {
+            return { error: responseBody.error.message };
+        }
+        catch {
+            return { error: responseText };
+        }
     }
-    const responseBody = await response.json();
     options.debug?.('lowire:openai-responses')('Response:', JSON.stringify(responseBody, null, 2));
-    return responseBody;
+    return { response: responseBody };
 }
 function toResultContentPart(part) {
     if (part.type === 'text') {

package/lib/providers/openaiCompatible.js

@@ -17,6 +17,7 @@
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.OpenAICompatible = void 0;
 const fetchWithTimeout_1 = require("../fetchWithTimeout");
+const types_1 = require("../types");
 class OpenAICompatible {
     name = 'openai-compatible';
     async complete(conversation, options) {
@@ -32,9 +33,9 @@ async function complete(conversation, options) {
     };
     const openaiMessages = [systemMessage, ...conversation.messages.map(toCompletionsMessages).flat()];
     const openaiTools = conversation.tools.map(t => toCompletionsTool(t));
-    const response = await create({
+    const { response, error } = await create({
         model: options.model,
-        max_tokens: options.maxTokens,
+        max_completion_tokens: options.maxTokens,
         temperature: options.temperature,
         messages: openaiMessages,
         tools: openaiTools,
@@ -42,9 +43,10 @@ async function complete(conversation, options) {
         reasoning_effort: toCompletionsReasoning(options.reasoning),
         parallel_tool_calls: false,
     }, options);
-    if (!response || !response.choices.length)
-        throw new Error('Failed to get response from OpenAI completions');
-    const result = { role: 'assistant', content: [] };
+    if (error || !response)
+        return { result: (0, types_1.assistantMessageFromError)(error?.message ?? 'No response from OpenAI compatible API'), usage: (0, types_1.emptyUsage)() };
+    const result = { role: 'assistant', content: [], stopReason: { code: 'ok' } };
+    const finishReason = response.choices[0]?.finish_reason;
     for (const choice of response.choices) {
         const message = choice.message;
         if (message.content)
@@ -55,6 +57,8 @@ async function complete(conversation, options) {
             result.content.push(toToolCall(entry));
         }
     }
+    if (finishReason === 'length')
+        result.stopReason = { code: 'max_tokens' };
     const usage = {
         input: response.usage?.prompt_tokens ?? 0,
         output: response.usage?.completion_tokens ?? 0,
@@ -75,13 +79,18 @@ async function create(createParams, options) {
         signal: options.signal,
         timeout: options.apiTimeout
     });
+    const responseText = await response.text();
+    const responseBody = JSON.parse(responseText);
     if (!response.ok) {
-        options.debug?.('lowire:openai')('Response:', response.status);
-        throw new Error(`API error: ${response.status} ${response.statusText} ${await response.text()}`);
+        try {
+            return { error: responseBody };
+        }
+        catch {
+            return { error: { type: 'unknown', message: responseText } };
+        }
     }
-    const responseBody = await response.json();
     options.debug?.('lowire:openai')('Response:', JSON.stringify(responseBody, null, 2));
-    return responseBody;
+    return { response: responseBody };
 }
 function toCopilotResultContentPart(part) {
     if (part.type === 'text') {

package/lib/types.d.ts

@@ -38,6 +38,10 @@ export type UserMessage = BaseMessage & {
 export type AssistantMessage = BaseMessage & {
     role: 'assistant';
     content: (TextContentPart | ToolCallContentPart | ThinkingContentPart)[];
+    stopReason: {
+        code: 'ok' | 'max_tokens' | 'error';
+        message?: string;
+    };
     openaiId?: string;
     openaiStatus?: 'completed' | 'incomplete' | 'in_progress';
     toolError?: string;
@@ -116,3 +120,5 @@ export type ReplayCache = Record<string, {
     result: AssistantMessage;
     usage: Usage;
 }>;
+export declare function assistantMessageFromError(error: string): AssistantMessage;
+export declare function emptyUsage(): Usage;

package/lib/types.js

@@ -15,3 +15,18 @@
  * limitations under the License.
  */
 Object.defineProperty(exports, "__esModule", { value: true });
+exports.assistantMessageFromError = assistantMessageFromError;
+exports.emptyUsage = emptyUsage;
+function assistantMessageFromError(error) {
+    return {
+        role: 'assistant',
+        content: [],
+        stopReason: { code: 'error', message: error },
+    };
+}
+function emptyUsage() {
+    return {
+        input: 0,
+        output: 0,
+    };
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lowire/loop",
-  "version": "0.0.22",
+  "version": "0.0.24",
   "description": "Small agentic loop",
   "repository": {
     "type": "git",