framework-agents 0.1.5 → 0.1.6

package/dist/{chunk-G3CMR4MD.js → chunk-JUOOPH4Z.js}

@@ -100,6 +100,16 @@ var ToolRejection = class extends Error {
     this.feedback = feedback;
   }
 };
+var RETRYABLE = /* @__PURE__ */ Symbol.for("framework.retryable");
+function stampRetryable(err, retryable) {
+  err[RETRYABLE] = retryable;
+  return err;
+}
+function getRetryable(err) {
+  if (err instanceof AgentError) return err.retryable;
+  if (err instanceof Error) return err[RETRYABLE];
+  return void 0;
+}
 var HitlInterrupt = class extends Error {
   interruptId;
   tool;
@@ -461,6 +471,27 @@ async function runWithBudget(fn, budget = {}) {
 }
 
 // src/agent/index.ts
+var EventQueue = class {
+  items = [];
+  wakes = [];
+  closed = false;
+  push(item) {
+    this.items.push(item);
+    this.wakes.shift()?.();
+  }
+  close() {
+    this.closed = true;
+    for (const w of this.wakes.splice(0)) w();
+  }
+  async next() {
+    for (; ; ) {
+      const item = this.items.shift();
+      if (item !== void 0) return { done: false, value: item };
+      if (this.closed) return { done: true };
+      await new Promise((resolve2) => this.wakes.push(resolve2));
+    }
+  }
+};
 var DEFAULT_SYSTEM = "You are a helpful agent. Use tools when needed. Be concise.";
 var DEFAULT_MAX_STEPS = 6;
 var toolJsonSchemas = /* @__PURE__ */ new WeakMap();
@@ -798,27 +829,82 @@ var Agent = class _Agent {
           ...strategy === "provider" && format ? { responseFormat: { name: responseFormatName(format), schema: responseFormatJsonSchema(format) } } : {}
         };
         for (const mw of middleware) req = await mw.beforeModel?.(req, ctxFor(step)) ?? req;
-        const pending = [];
-        const callModel = (r) => model.complete({
-          ...r,
-          // Propagate the run's abort into the in-flight provider request.
-          ...signal ? { signal } : {},
-          onToken: (d) => {
-            events?.onToken?.(d);
-            pending.push({ type: "token", delta: d });
-          },
-          onThinking: (d) => {
-            events?.onThinking?.(d);
-            pending.push({ type: "thinking", delta: d });
+        const needsBatching = middleware.some((mw) => mw.wrapModel || mw.afterModel);
+        const buffered = [];
+        const channel = new EventQueue();
+        const emitDelta = (type, d) => {
+          const event = { type, delta: d };
+          if (needsBatching) buffered.push(event);
+          else channel.push(event);
+        };
+        const callModel = (r) => {
+          const start = buffered.length;
+          let failed = false;
+          try {
+            return model.complete({
+              ...r,
+              // Propagate the run's abort into the in-flight provider request.
+              ...signal ? { signal } : {},
+              onToken: (d) => {
+                events?.onToken?.(d);
+                if (failed) return;
+                emitDelta("token", d);
+              },
+              onThinking: (d) => {
+                events?.onThinking?.(d);
+                if (failed) return;
+                emitDelta("thinking", d);
+              }
+            }).catch((err) => {
+              failed = true;
+              buffered.splice(start);
+              throw err;
+            });
+          } catch (err) {
+            failed = true;
+            buffered.splice(start);
+            throw err;
           }
-        });
+        };
         const modelChain = middleware.filter((mw) => mw.wrapModel).reduceRight(
           (next, mw) => (r) => mw.wrapModel(r, next, ctxFor(step)),
           callModel
         );
-        let completion = await modelChain(req);
-        for (const e of pending.splice(0)) {
-          if (e.delta !== "") yield e;
+        let deltasEmittedThisCall = false;
+        let completion;
+        const pendingCompletion = modelChain(req).then(
+          (c) => {
+            channel.close();
+            return c;
+          },
+          (e) => {
+            channel.close();
+            throw e;
+          }
+        );
+        pendingCompletion.catch(() => {
+        });
+        try {
+          if (needsBatching) {
+            completion = await pendingCompletion;
+            for (const e of buffered.splice(0)) {
+              if (e.delta !== "") yield e;
+            }
+          } else {
+            for (; ; ) {
+              const next = await channel.next();
+              if (next.done) break;
+              deltasEmittedThisCall = true;
+              if (next.value.delta !== "") yield next.value;
+            }
+            completion = await pendingCompletion;
+          }
+        } catch (err) {
+          buffered.splice(0);
+          if (deltasEmittedThisCall && getRetryable(err) === void 0) {
+            stampRetryable(err, false);
+          }
+          throw err;
         }
         for (const mw of [...middleware].reverse()) completion = await mw.afterModel?.(completion, req, ctxFor(step)) ?? completion;
         usage.inputTokens += completion.usage.inputTokens;

package/dist/cli.js

@@ -9,7 +9,7 @@ import {
   imagesOf,
   todoListMiddleware,
   wireTextOf
-} from "./chunk-G3CMR4MD.js";
+} from "./chunk-JUOOPH4Z.js";
 
 // src/serve/cli.ts
 import { resolve as resolve3 } from "path";
@@ -184,8 +184,8 @@ var OpenAICompatibleProvider = class {
         }
       } : {},
       messages: options.messages.map(toWireMessage),
-      ...options.tools.length ? {
-        tools: options.tools.map((t) => ({
+      ...options.tools?.length ? {
+        tools: (options.tools ?? []).map((t) => ({
           type: "function",
           function: {
             name: t.name,
@@ -252,91 +252,107 @@ var OpenAICompatibleProvider = class {
   /** SSE: `data: {...}` lines; delta.content -> onToken, delta.reasoning_content -> onThinking. */
   async readStream(res, options) {
     if (!res.body) throw new Error("Streaming response has no body");
-    const reader = res.body.getReader();
-    const decoder = new TextDecoder();
-    let text = "";
-    let thinking = "";
-    let inputTokens = 0;
-    let outputTokens = 0;
-    let reasoningTokens;
-    let finishReason;
-    const toolArgBuffers = /* @__PURE__ */ new Map();
-    let buffer = "";
-    const flush = (chunk) => {
-      buffer += chunk;
-      const lines = buffer.split("\n");
-      buffer = lines.pop() ?? "";
-      for (const line of lines) {
-        const t = line.trim();
-        if (!t.startsWith("data:")) continue;
-        const payload = t.slice(5).trim();
-        if (payload === "[DONE]") continue;
-        let evt;
-        try {
-          evt = JSON.parse(payload);
-        } catch {
-          continue;
-        }
-        if (evt.choices?.[0]?.finish_reason) finishReason = evt.choices[0].finish_reason ?? void 0;
-        if (evt.usage) {
-          if (evt.usage.prompt_tokens != null) inputTokens = evt.usage.prompt_tokens;
-          if (evt.usage.completion_tokens != null) outputTokens = evt.usage.completion_tokens;
-          const rt = reasoningOf(evt.usage);
-          if (rt != null) reasoningTokens = rt;
-          continue;
-        }
-        const delta = evt.choices?.[0]?.delta;
-        if (!delta) continue;
-        const think = thinkingOf(delta);
-        if (think) {
-          thinking += think;
-          options.onThinking?.(think);
-        }
-        if (delta.content) {
-          text += delta.content;
-          options.onToken?.(delta.content);
-        }
-        for (const [i, tc] of (delta.tool_calls ?? []).entries()) {
-          const key = tc.index ?? i;
-          const slot = toolArgBuffers.get(key) ?? { id: "", name: "", args: "" };
-          if (tc.id) slot.id = tc.id;
-          if (tc.function?.name) slot.name = tc.function.name;
-          if (tc.function?.arguments) slot.args += tc.function.arguments;
-          toolArgBuffers.set(key, slot);
+    return this._readSse(res.body, options).catch((err) => {
+      if (err instanceof AgentError || err instanceof Error && err.name === "AbortError") throw err;
+      throw new AgentError("MODEL_ERROR", `Stream broke mid-response: ${err}`, false);
+    });
+  }
+  async _readSse(body, options) {
+    let delivered = false;
+    try {
+      const reader = body.getReader();
+      const decoder = new TextDecoder();
+      let text = "";
+      let thinking = "";
+      let inputTokens = 0;
+      let outputTokens = 0;
+      let reasoningTokens;
+      let finishReason;
+      const toolArgBuffers = /* @__PURE__ */ new Map();
+      let buffer = "";
+      const flush = (chunk) => {
+        buffer += chunk;
+        const lines = buffer.split("\n");
+        buffer = lines.pop() ?? "";
+        for (const line of lines) {
+          const t = line.trim();
+          if (!t.startsWith("data:")) continue;
+          const payload = t.slice(5).trim();
+          if (payload === "[DONE]") continue;
+          let evt;
+          try {
+            evt = JSON.parse(payload);
+          } catch {
+            continue;
+          }
+          if (evt.choices?.[0]?.finish_reason) finishReason = evt.choices[0].finish_reason ?? void 0;
+          if (evt.usage) {
+            if (evt.usage.prompt_tokens != null) inputTokens = evt.usage.prompt_tokens;
+            if (evt.usage.completion_tokens != null) outputTokens = evt.usage.completion_tokens;
+            const rt = reasoningOf(evt.usage);
+            if (rt != null) reasoningTokens = rt;
+            continue;
+          }
+          const delta = evt.choices?.[0]?.delta;
+          if (!delta) continue;
+          const think = thinkingOf(delta);
+          if (think) {
+            thinking += think;
+            delivered = true;
+            options.onThinking?.(think);
+          }
+          if (delta.content) {
+            text += delta.content;
+            delivered = true;
+            options.onToken?.(delta.content);
+          }
+          for (const [i, tc] of (delta.tool_calls ?? []).entries()) {
+            const key = tc.index ?? i;
+            delivered = true;
+            const slot = toolArgBuffers.get(key) ?? { id: "", name: "", args: "" };
+            if (tc.id) slot.id = tc.id;
+            if (tc.function?.name) slot.name = tc.function.name;
+            if (tc.function?.arguments) slot.args += tc.function.arguments;
+            toolArgBuffers.set(key, slot);
+          }
         }
+      };
+      for (; ; ) {
+        const { done, value } = await reader.read();
+        if (done) break;
+        flush(decoder.decode(value, { stream: true }));
       }
-    };
-    for (; ; ) {
-      const { done, value } = await reader.read();
-      if (done) break;
-      flush(decoder.decode(value, { stream: true }));
-    }
-    flush(decoder.decode());
-    if (!text && !toolArgBuffers.size && thinking && finishReason === "stop") {
-      throw new AgentError(
-        "MODEL_ERROR",
-        "Stream ended at finish_reason=stop having put the entire answer in the reasoning channel (no content, no tool calls, thinking present). The reasoning/content split lost the reply \u2014 retry or raise maxTokens.",
-        true
-      );
-    }
-    if (!text && !toolArgBuffers.size && finishReason === "length") {
-      throw new AgentError(
-        "MODEL_ERROR",
-        `Stream ended at finish_reason=length with no content (${outputTokens} tokens burned). Raise maxTokens.`,
-        false
-      );
-    }
-    return {
-      text,
-      ...thinking ? { thinking } : {},
-      ...finishReason ? { finishReason } : {},
-      toolCalls: [...toolArgBuffers.values()].filter((t) => t.name).map((t) => ({ id: t.id || `call_${++callSeq}`, name: t.name, args: safeJson(t.args) })),
-      usage: {
-        inputTokens,
-        outputTokens,
-        ...reasoningTokens != null ? { reasoningTokens } : {}
+      flush(decoder.decode());
+      if (!text && !toolArgBuffers.size && thinking && finishReason === "stop") {
+        throw new AgentError(
+          "MODEL_ERROR",
+          "Stream ended at finish_reason=stop having put the entire answer in the reasoning channel (no content, no tool calls, thinking present). The reasoning/content split lost the reply \u2014 retry or raise maxTokens.",
+          true
+        );
       }
-    };
+      if (!text && !toolArgBuffers.size && finishReason === "length") {
+        throw new AgentError(
+          "MODEL_ERROR",
+          `Stream ended at finish_reason=length with no content (${outputTokens} tokens burned). Raise maxTokens.`,
+          false
+        );
+      }
+      return {
+        text,
+        ...thinking ? { thinking } : {},
+        ...finishReason ? { finishReason } : {},
+        toolCalls: [...toolArgBuffers.values()].map((t) => ({ id: t.id || `call_${++callSeq}`, name: t.name, args: safeJson(t.args) })),
+        usage: {
+          inputTokens,
+          outputTokens,
+          ...reasoningTokens != null ? { reasoningTokens } : {}
+        }
+      };
+    } catch (err) {
+      if (err instanceof Error && err.name === "AbortError") throw err;
+      if (err instanceof AgentError) throw err;
+      throw new AgentError("MODEL_ERROR", `Stream broke mid-response${delivered ? " (deltas already delivered)" : ""}: ${err}`, !delivered);
+    }
   }
 };
 var callSeq = 0;
@@ -431,8 +447,8 @@ var AnthropicProvider = class {
       ...thinkingBudget ? { thinking: { type: "enabled", budget_tokens: thinkingBudget } } : {},
       ...stream ? { stream: true } : {},
       messages,
-      ...options.tools.length ? {
-        tools: options.tools.map((t) => ({
+      ...options.tools?.length ? {
+        tools: (options.tools ?? []).map((t) => ({
           name: t.name,
           description: t.description,
           input_schema: t.parametersSchema ?? { type: "object" }

package/dist/index.d.ts

@@ -625,6 +625,7 @@ declare class OpenAICompatibleProvider implements ModelProvider {
     private readJson;
     /** SSE: `data: {...}` lines; delta.content -> onToken, delta.reasoning_content -> onThinking. */
     private readStream;
+    private _readSse;
 }
 /** Build a provider from env config (TogetherAI preferred when key present). */
 declare function providerFromEnv(active: {
@@ -768,7 +769,7 @@ interface ResponseFormatWire {
 interface CompleteOptions {
     messages: ChatMessage[];
     /** Tool metadata; `parametersSchema` (JSON Schema) lets models call with real args — omit only for zero-arg tools. */
-    tools: {
+    tools?: {
         name: string;
         description: string;
         parametersSchema?: Record<string, unknown>;

package/dist/index.js

@@ -71,7 +71,7 @@ import {
   usageCost,
   user,
   wireTextOf
-} from "./chunk-2XXGDP7B.js";
+} from "./chunk-AJI2QSJE.js";
 
 // src/config/index.ts
 import { existsSync } from "fs";
@@ -864,8 +864,8 @@ var OpenAICompatibleProvider = class {
         }
       } : {},
       messages: options.messages.map(toWireMessage),
-      ...options.tools.length ? {
-        tools: options.tools.map((t) => ({
+      ...options.tools?.length ? {
+        tools: (options.tools ?? []).map((t) => ({
           type: "function",
           function: {
             name: t.name,
@@ -932,91 +932,107 @@ var OpenAICompatibleProvider = class {
   /** SSE: `data: {...}` lines; delta.content -> onToken, delta.reasoning_content -> onThinking. */
   async readStream(res, options) {
     if (!res.body) throw new Error("Streaming response has no body");
-    const reader = res.body.getReader();
-    const decoder = new TextDecoder();
-    let text = "";
-    let thinking = "";
-    let inputTokens = 0;
-    let outputTokens = 0;
-    let reasoningTokens;
-    let finishReason;
-    const toolArgBuffers = /* @__PURE__ */ new Map();
-    let buffer = "";
-    const flush = (chunk) => {
-      buffer += chunk;
-      const lines = buffer.split("\n");
-      buffer = lines.pop() ?? "";
-      for (const line of lines) {
-        const t = line.trim();
-        if (!t.startsWith("data:")) continue;
-        const payload = t.slice(5).trim();
-        if (payload === "[DONE]") continue;
-        let evt;
-        try {
-          evt = JSON.parse(payload);
-        } catch {
-          continue;
-        }
-        if (evt.choices?.[0]?.finish_reason) finishReason = evt.choices[0].finish_reason ?? void 0;
-        if (evt.usage) {
-          if (evt.usage.prompt_tokens != null) inputTokens = evt.usage.prompt_tokens;
-          if (evt.usage.completion_tokens != null) outputTokens = evt.usage.completion_tokens;
-          const rt = reasoningOf(evt.usage);
-          if (rt != null) reasoningTokens = rt;
-          continue;
-        }
-        const delta = evt.choices?.[0]?.delta;
-        if (!delta) continue;
-        const think = thinkingOf(delta);
-        if (think) {
-          thinking += think;
-          options.onThinking?.(think);
-        }
-        if (delta.content) {
-          text += delta.content;
-          options.onToken?.(delta.content);
-        }
-        for (const [i, tc] of (delta.tool_calls ?? []).entries()) {
-          const key = tc.index ?? i;
-          const slot = toolArgBuffers.get(key) ?? { id: "", name: "", args: "" };
-          if (tc.id) slot.id = tc.id;
-          if (tc.function?.name) slot.name = tc.function.name;
-          if (tc.function?.arguments) slot.args += tc.function.arguments;
-          toolArgBuffers.set(key, slot);
+    return this._readSse(res.body, options).catch((err) => {
+      if (err instanceof AgentError || err instanceof Error && err.name === "AbortError") throw err;
+      throw new AgentError("MODEL_ERROR", `Stream broke mid-response: ${err}`, false);
+    });
+  }
+  async _readSse(body, options) {
+    let delivered = false;
+    try {
+      const reader = body.getReader();
+      const decoder = new TextDecoder();
+      let text = "";
+      let thinking = "";
+      let inputTokens = 0;
+      let outputTokens = 0;
+      let reasoningTokens;
+      let finishReason;
+      const toolArgBuffers = /* @__PURE__ */ new Map();
+      let buffer = "";
+      const flush = (chunk) => {
+        buffer += chunk;
+        const lines = buffer.split("\n");
+        buffer = lines.pop() ?? "";
+        for (const line of lines) {
+          const t = line.trim();
+          if (!t.startsWith("data:")) continue;
+          const payload = t.slice(5).trim();
+          if (payload === "[DONE]") continue;
+          let evt;
+          try {
+            evt = JSON.parse(payload);
+          } catch {
+            continue;
+          }
+          if (evt.choices?.[0]?.finish_reason) finishReason = evt.choices[0].finish_reason ?? void 0;
+          if (evt.usage) {
+            if (evt.usage.prompt_tokens != null) inputTokens = evt.usage.prompt_tokens;
+            if (evt.usage.completion_tokens != null) outputTokens = evt.usage.completion_tokens;
+            const rt = reasoningOf(evt.usage);
+            if (rt != null) reasoningTokens = rt;
+            continue;
+          }
+          const delta = evt.choices?.[0]?.delta;
+          if (!delta) continue;
+          const think = thinkingOf(delta);
+          if (think) {
+            thinking += think;
+            delivered = true;
+            options.onThinking?.(think);
+          }
+          if (delta.content) {
+            text += delta.content;
+            delivered = true;
+            options.onToken?.(delta.content);
+          }
+          for (const [i, tc] of (delta.tool_calls ?? []).entries()) {
+            const key = tc.index ?? i;
+            delivered = true;
+            const slot = toolArgBuffers.get(key) ?? { id: "", name: "", args: "" };
+            if (tc.id) slot.id = tc.id;
+            if (tc.function?.name) slot.name = tc.function.name;
+            if (tc.function?.arguments) slot.args += tc.function.arguments;
+            toolArgBuffers.set(key, slot);
+          }
         }
+      };
+      for (; ; ) {
+        const { done, value } = await reader.read();
+        if (done) break;
+        flush(decoder.decode(value, { stream: true }));
       }
-    };
-    for (; ; ) {
-      const { done, value } = await reader.read();
-      if (done) break;
-      flush(decoder.decode(value, { stream: true }));
-    }
-    flush(decoder.decode());
-    if (!text && !toolArgBuffers.size && thinking && finishReason === "stop") {
-      throw new AgentError(
-        "MODEL_ERROR",
-        "Stream ended at finish_reason=stop having put the entire answer in the reasoning channel (no content, no tool calls, thinking present). The reasoning/content split lost the reply \u2014 retry or raise maxTokens.",
-        true
-      );
-    }
-    if (!text && !toolArgBuffers.size && finishReason === "length") {
-      throw new AgentError(
-        "MODEL_ERROR",
-        `Stream ended at finish_reason=length with no content (${outputTokens} tokens burned). Raise maxTokens.`,
-        false
-      );
-    }
-    return {
-      text,
-      ...thinking ? { thinking } : {},
-      ...finishReason ? { finishReason } : {},
-      toolCalls: [...toolArgBuffers.values()].filter((t) => t.name).map((t) => ({ id: t.id || `call_${++callSeq}`, name: t.name, args: safeJson(t.args) })),
-      usage: {
-        inputTokens,
-        outputTokens,
-        ...reasoningTokens != null ? { reasoningTokens } : {}
+      flush(decoder.decode());
+      if (!text && !toolArgBuffers.size && thinking && finishReason === "stop") {
+        throw new AgentError(
+          "MODEL_ERROR",
+          "Stream ended at finish_reason=stop having put the entire answer in the reasoning channel (no content, no tool calls, thinking present). The reasoning/content split lost the reply \u2014 retry or raise maxTokens.",
+          true
+        );
       }
-    };
+      if (!text && !toolArgBuffers.size && finishReason === "length") {
+        throw new AgentError(
+          "MODEL_ERROR",
+          `Stream ended at finish_reason=length with no content (${outputTokens} tokens burned). Raise maxTokens.`,
+          false
+        );
+      }
+      return {
+        text,
+        ...thinking ? { thinking } : {},
+        ...finishReason ? { finishReason } : {},
+        toolCalls: [...toolArgBuffers.values()].map((t) => ({ id: t.id || `call_${++callSeq}`, name: t.name, args: safeJson(t.args) })),
+        usage: {
+          inputTokens,
+          outputTokens,
+          ...reasoningTokens != null ? { reasoningTokens } : {}
+        }
+      };
+    } catch (err) {
+      if (err instanceof Error && err.name === "AbortError") throw err;
+      if (err instanceof AgentError) throw err;
+      throw new AgentError("MODEL_ERROR", `Stream broke mid-response${delivered ? " (deltas already delivered)" : ""}: ${err}`, !delivered);
+    }
   }
 };
 var callSeq = 0;
@@ -1114,8 +1130,8 @@ var AnthropicProvider = class {
       ...thinkingBudget ? { thinking: { type: "enabled", budget_tokens: thinkingBudget } } : {},
       ...stream ? { stream: true } : {},
       messages,
-      ...options.tools.length ? {
-        tools: options.tools.map((t) => ({
+      ...options.tools?.length ? {
+        tools: (options.tools ?? []).map((t) => ({
           name: t.name,
           description: t.description,
           input_schema: t.parametersSchema ?? { type: "object" }