llmz 0.1.5 → 0.2.0

package/dist/{chunk-TIZWOIDF.js → chunk-33AB6SVX.js}

@@ -2,7 +2,7 @@ import {
   DualModePrompt,
   extractType,
   inspect
-} from "./chunk-2O3UTPVH.js";
+} from "./chunk-E5NHK3FS.js";
 import {
   Tool
 } from "./chunk-IRCB4FRM.js";
@@ -14,10 +14,14 @@ import {
 } from "./chunk-XVVYNOTX.js";
 import {
   assertValidComponent
-} from "./chunk-3T5IU7OU.js";
+} from "./chunk-KWQTYXAT.js";
 import {
+  stripTruncationTags,
   wrapContent
-} from "./chunk-L74UZONX.js";
+} from "./chunk-HM5H556L.js";
+import {
+  getTokenizer
+} from "./chunk-YX27XM67.js";
 import {
   cloneDeep_default,
   isPlainObject_default,
@@ -777,6 +781,33 @@ var ExecutionResult = class {
   get iterations() {
     return this.context.iterations ?? [];
   }
+  /**
+   * Gets the total token usage of the execution, aggregated across all iterations.
+   *
+   * `input`/`output` are the provider-reported token counts of every LLM call made
+   * during the execution. Per-iteration detail (including the measured context size
+   * breakdown by prompt part) is available on each iteration's `tokens` property.
+   *
+   * @example
+   * ```typescript
+   * const result = await execute({ ... })
+   * console.log(`Execution used ${result.tokens.total} tokens`)
+   *
+   * for (const iteration of result.iterations) {
+   *   console.log(iteration.id, iteration.tokens?.context)
+   * }
+   * ```
+   */
+  get tokens() {
+    var _a, _b;
+    let input = 0;
+    let output = 0;
+    for (const iteration of this.iterations) {
+      input += ((_a = iteration.tokens) == null ? void 0 : _a.input) ?? 0;
+      output += ((_b = iteration.tokens) == null ? void 0 : _b.output) ?? 0;
+    }
+    return { input, output, total: input + output };
+  }
 };
 var SuccessExecutionResult = class extends ExecutionResult {
   result;
@@ -959,8 +990,18 @@ var Iteration = class {
   id;
   messages;
   code;
+  /** Messages (`■send` blocks) parsed from the assistant response, in order. */
+  sends;
+  /** The `■next` exit parsed from the assistant response, if any. */
+  next;
   traces;
   variables;
+  /**
+   * Token usage of this iteration's LLM call. The `context` breakdown is measured
+   * when the prompt is assembled; `input`/`output` are filled in once the LLM call
+   * completes, from the provider-reported usage.
+   */
+  tokens;
   started_ts;
   ended_ts;
   status;
@@ -994,7 +1035,7 @@ var Iteration = class {
     return this._parameters.reasoningEffort;
   }
   get exits() {
-    const exits = [...this._parameters.exits, ThinkExit];
+    const exits = [...this._parameters.exits];
     if (this.isChatEnabled) {
       exits.push(ListenExit);
     }
@@ -1048,7 +1089,7 @@ var Iteration = class {
     return null;
   }
   get isChatEnabled() {
-    return this._parameters.tools.find((x) => x.name.toLowerCase() === "message") !== void 0;
+    return this._parameters.components.length > 0;
   }
   constructor(props) {
     this.id = props.id;
@@ -1082,6 +1123,7 @@ var Iteration = class {
       status: this.status,
       mutations: [...this._mutations.values()],
       llm: this.llm,
+      tokens: this.tokens,
       transcript: [...this._parameters.transcript],
       tools: this._parameters.tools.map((tool) => tool.toJSON()),
       objects: this._parameters.objects.map((obj) => obj.toJSON()),
@@ -1107,6 +1149,16 @@ var Context = class {
   timeout = 6e4;
   // Default timeout of 60 seconds
   loop;
+  /**
+   * Optional cap on the model's context window. The effective limit is
+   * `min(maxTokens, model's max input tokens)`.
+   */
+  maxTokens;
+  /**
+   * Maximum time to wait for the first streamed token, in milliseconds,
+   * before the cognitive service falls back to the next model/provider.
+   */
+  maxTimeToFirstToken;
   metadata;
   snapshot;
   iteration = 0;
@@ -1121,13 +1173,15 @@ var Context = class {
       );
     }
     const parameters = await this._refreshIterationParameters();
-    const messages = await this._getIterationMessages(parameters);
+    const { messages, parts } = await this._getIterationMessages(parameters);
+    const contextTokens = this._measureContextTokens(messages, parts);
     const iteration = new Iteration({
       id: `${this.id}_${this.iterations.length + 1}`,
       variables: this._getIterationVariables(),
       parameters,
       messages
     });
+    iteration.tokens = { input: 0, output: 0, total: 0, context: contextTokens };
     this.iterations.push(iteration);
     this.iteration = this.iterations.length;
     this.snapshot = void 0;
@@ -1155,72 +1209,80 @@ var Context = class {
     }
     return variables;
   }
+  /**
+   * Measures the token size of each part of the prompt (pre-truncation).
+   * The named parts come from the system prompt; everything else in the system
+   * message is `framework`. Non-system messages count towards `iterations`,
+   * except the very first user message of a fresh execution (task recap),
+   * which is prompt scaffolding.
+   */
+  _measureContextTokens(messages, parts) {
+    const tokenizer = getTokenizer();
+    const countText = (text) => (text == null ? void 0 : text.length) ? tokenizer.count(stripTruncationTags(text)) : 0;
+    const countMessage = (message) => {
+      if (typeof message.content === "string") {
+        return countText(message.content);
+      }
+      if (Array.isArray(message.content)) {
+        return message.content.reduce((acc, part) => acc + (part.type === "text" ? countText(part.text) : 0), 0);
+      }
+      return 0;
+    };
+    const instructions = countText(parts.instructions);
+    const tools = countText(parts.tools);
+    const transcript = countText(parts.transcript);
+    const protocol = countText(parts.protocol);
+    const systemTokens = messages.filter((x) => x.role === "system").reduce((acc, x) => acc + countMessage(x), 0);
+    const otherTokens = messages.filter((x) => x.role !== "system").reduce((acc, x) => acc + countMessage(x), 0);
+    const isFirstIteration = this.iterations.length === 0 && !this.snapshot;
+    const framework = Math.max(0, systemTokens - (instructions + tools + transcript + protocol));
+    const iterations = isFirstIteration ? 0 : otherTokens;
+    return {
+      total: framework + instructions + tools + transcript + protocol + iterations + (isFirstIteration ? otherTokens : 0),
+      framework: framework + (isFirstIteration ? otherTokens : 0),
+      instructions,
+      tools,
+      transcript,
+      protocol,
+      iterations
+    };
+  }
   async _getIterationMessages(parameters) {
     var _a, _b, _c, _d, _e, _f;
     const lastIteration = this.iterations.at(-1);
+    const promptProps = {
+      globalTools: parameters.tools,
+      objects: parameters.objects,
+      instructions: parameters.instructions,
+      transcript: parameters.transcript,
+      // ListenExit is protocol-level in chat mode: it must be documented alongside user-defined exits
+      exits: parameters.components.length ? [...parameters.exits, ListenExit] : parameters.exits,
+      components: parameters.components
+    };
+    const { message: systemMessage, parts } = await this.version.getSystemMessage(promptProps);
+    const withParts = (messages) => ({ messages, parts });
     if (((_a = this.snapshot) == null ? void 0 : _a.status.type) === "resolved") {
-      return [
-        await this.version.getSystemMessage({
-          globalTools: parameters.tools,
-          objects: parameters.objects,
-          instructions: parameters.instructions,
-          transcript: parameters.transcript,
-          exits: parameters.exits,
-          components: parameters.components
-        }),
+      return withParts([
+        systemMessage,
         this.version.getSnapshotResolvedMessage({
           snapshot: this.snapshot
         })
-      ];
+      ]);
     }
     if (((_b = this.snapshot) == null ? void 0 : _b.status.type) === "rejected") {
-      return [
-        await this.version.getSystemMessage({
-          globalTools: parameters.tools,
-          objects: parameters.objects,
-          instructions: parameters.instructions,
-          transcript: parameters.transcript,
-          exits: parameters.exits,
-          components: parameters.components
-        }),
+      return withParts([
+        systemMessage,
         this.version.getSnapshotRejectedMessage({
           snapshot: this.snapshot
         })
-      ];
+      ]);
     }
     if (!lastIteration) {
-      return [
-        await this.version.getSystemMessage({
-          globalTools: parameters.tools,
-          objects: parameters.objects,
-          instructions: parameters.instructions,
-          transcript: parameters.transcript,
-          exits: parameters.exits,
-          components: parameters.components
-        }),
-        await this.version.getInitialUserMessage({
-          globalTools: parameters.tools,
-          objects: parameters.objects,
-          instructions: parameters.instructions,
-          transcript: parameters.transcript,
-          exits: parameters.exits,
-          components: parameters.components
-        })
-      ];
+      return withParts([systemMessage, await this.version.getInitialUserMessage(promptProps)]);
     }
-    const lastIterationMessages = [
-      await this.version.getSystemMessage({
-        globalTools: parameters.tools,
-        objects: parameters.objects,
-        instructions: parameters.instructions,
-        transcript: parameters.transcript,
-        exits: parameters.exits,
-        components: parameters.components
-      }),
-      ...lastIteration.messages.filter((x) => x.role !== "system")
-    ];
+    const lastIterationMessages = [systemMessage, ...lastIteration.messages.filter((x) => x.role !== "system")];
     if ((lastIteration == null ? void 0 : lastIteration.status.type) === "thinking_requested") {
-      return [
+      return withParts([
         ...lastIterationMessages,
         {
           role: "assistant",
@@ -1230,10 +1292,10 @@ var Context = class {
           reason: lastIteration.status.thinking_requested.reason,
           variables: lastIteration.status.thinking_requested.variables
         })
-      ];
+      ]);
     }
     if ((lastIteration == null ? void 0 : lastIteration.status.type) === "exit_error") {
-      return [
+      return withParts([
         ...lastIterationMessages,
         {
           role: "assistant",
@@ -1243,10 +1305,10 @@ var Context = class {
           code: lastIteration.code ?? "// No code generated",
           message: `Invalid return statement (action: ${lastIteration.status.exit_error.exit}): ${lastIteration.status.exit_error.message}`
         })
-      ];
+      ]);
     }
     if ((lastIteration == null ? void 0 : lastIteration.status.type) === "invalid_code_error") {
-      return [
+      return withParts([
         ...lastIterationMessages,
         {
           role: "assistant",
@@ -1256,10 +1318,10 @@ var Context = class {
           code: lastIteration.code ?? "// No code generated",
           message: lastIteration.status.invalid_code_error.message
         })
-      ];
+      ]);
     }
     if ((lastIteration == null ? void 0 : lastIteration.status.type) === "execution_error") {
-      return [
+      return withParts([
         ...lastIterationMessages,
         {
           role: "assistant",
@@ -1269,7 +1331,7 @@ var Context = class {
           message: lastIteration.status.execution_error.message,
           stacktrace: lastIteration.status.execution_error.stack
         })
-      ];
+      ]);
     }
     throw new Error(
       `Unexpected iteration status: ${lastIteration == null ? void 0 : lastIteration.status.type}. This is likely a bug, please report it.`
@@ -1312,18 +1374,6 @@ var Context = class {
       "boolean",
       "array"
     ];
-    const MessageTool = this.chat && components.length ? new Tool({
-      name: "Message",
-      description: "Send a message to the user",
-      aliases: Array.from(
-        /* @__PURE__ */ new Set(["message", ...components.flatMap((x) => [x.definition.name, ...x.definition.aliases ?? []])])
-      ),
-      handler: async (message) => {
-        var _a2, _b2;
-        return await ((_b2 = (_a2 = this.chat) == null ? void 0 : _a2.handler) == null ? void 0 : _b2.call(_a2, message));
-      }
-    }) : null;
-    const allTools = MessageTool ? [MessageTool, ...tools] : tools;
     for (const tool of tools) {
       for (let name of [...tool.aliases, tool.name]) {
         name = name.toLowerCase();
@@ -1380,7 +1430,7 @@ var Context = class {
     }
     return {
       transcript,
-      tools: allTools,
+      tools,
       objects,
       exits,
       instructions,
@@ -1405,9 +1455,17 @@ var Context = class {
     this.iterations = [];
     this.metadata = props.metadata ?? {};
     this.snapshot = props.snapshot;
+    this.maxTokens = props.maxTokens;
+    this.maxTimeToFirstToken = props.maxTimeToFirstToken;
     if (this.loop < 1 || this.loop > 100) {
       throw new Error("Invalid loop. Expected a number between 1 and 100.");
     }
+    if (this.maxTokens !== void 0 && (!Number.isFinite(this.maxTokens) || this.maxTokens < 1)) {
+      throw new Error("Invalid maxTokens. Expected a positive number.");
+    }
+    if (this.maxTimeToFirstToken !== void 0 && (!Number.isFinite(this.maxTimeToFirstToken) || this.maxTimeToFirstToken < 1)) {
+      throw new Error("Invalid maxTimeToFirstToken. Expected a positive number of milliseconds.");
+    }
   }
   toJSON() {
     var _a;