llmz 0.8.0 → 0.9.1

package/dist/example-format.d.ts

@@ -0,0 +1,8 @@
+/** Documentation boundaries only; never part of the response protocol. */
+export declare const quoteExample: (output: string) => string;
+/** A single block excerpt; the omitted response is not part of this example. */
+export declare const quotePartialExample: (output: string) => string;
+export declare const exampleBoundaryInstructions = "Triple quotes (\"\"\") mark the beginning and end of examples only. Do NOT emit these delimiters in your response.";
+/** A complete response example, unlike individual block/syntax examples. */
+export declare const quoteResponseExample: (output: string) => string;
+export declare const responseEnvelopeInstructions = "Every response starts with \u25A0start on its own line and ends with \u25A0end on its own line. Between them write only the documented protocol blocks. Nothing goes before \u25A0start or after \u25A0end. The envelope is protocol, not a code fence.";

package/dist/{execute-COIHOIKR.cjs → execute-0bvqakr7.cjs}

@@ -1,13 +1,13 @@
-const require_compiler = require("./compiler-C-OI9sim.cjs");
-const require_common = require("./common-DdArNQhl.cjs");
-const require_dual_modes = require("./dual-modes-Cb7ThvoV.cjs");
-const require_custom_client = require("./custom-client-HkM1Sdzv.cjs");
-const require_utils = require("./utils-3ESwP3yg.cjs");
-const require_jsx = require("./jsx-D5Y2cPo-.cjs");
+const require_compiler = require("./compiler-BJJUmNS9.cjs");
+const require_dual_modes = require("./dual-modes-BNHlj2qM.cjs");
+const require_common = require("./common-DJkSpeqw.cjs");
+const require_custom_client = require("./custom-client-B_o56CB4.cjs");
+const require_utils = require("./utils-DGQ4sRke.cjs");
+const require_jsx = require("./jsx-BJ6GWXI2.cjs");
 const require_stack_traces = require("./stack-traces-ntw_qhoy.cjs");
 const require_errors = require("./errors-C962uJdS.cjs");
-const require_truncator = require("./truncator-CLkouTCO.cjs");
-const require_vm = require("./vm-D6s1zF6M.cjs");
+const require_truncator = require("./truncator-DmCu9M4m.cjs");
+const require_vm = require("./vm-X0uTj5Aa.cjs");
 let _bpinternal_zui = require("@bpinternal/zui");
 let ulid = require("ulid");
 let _botpress_client = require("@botpress/client");
@@ -126,47 +126,21 @@ const STREAM_INACTIVITY_TIMEOUT = 18e4;
 /** A syntactically valid prefix must not execute when generation did not finish successfully. */
 const assertSuccessfulGeneration = (metadata) => {
 	if (metadata.provider === "unknown") throw new require_errors.CognitiveError("LLM generation failed: received error metadata with unknown provider");
-	if (metadata.stopReason === "max_tokens" || metadata.stopReason === "content_filter") throw new require_errors.CognitiveError(`LLM generation did not complete: stopReason=${metadata.stopReason}`);
+	if (metadata.stopReason === "max_tokens" || metadata.stopReason === "content_filter" || metadata.stopReason === "other") throw new require_errors.CognitiveError(`LLM generation did not complete: stopReason=${metadata.stopReason}`);
 };
 const getModelOutputLimit = (inputLength) => clamp(RESPONSE_LENGTH_BUFFER.PERCENTAGE * inputLength, RESPONSE_LENGTH_BUFFER.MIN_TOKENS, RESPONSE_LENGTH_BUFFER.MAX_TOKENS);
-/**
-* Models sometimes wrap their whole response in a code fence. The fence has to
-* be removed before it reaches the incremental parser to avoid a spurious
-* unexpected-text diagnostic. Works on arbitrary chunk boundaries
-* by holding content back until the first newline.
-*/
-var LeadingFenceFilter = class {
-	_buffer = "";
-	_done = false;
-	push(text) {
-		if (this._done) return text;
-		this._buffer += text;
-		const newline = this._buffer.indexOf("\n");
-		if (newline === -1) return "";
-		this._done = true;
-		const firstLine = this._buffer.slice(0, newline);
-		const rest = this._buffer.slice(newline + 1);
-		this._buffer = "";
-		return firstLine.trim().startsWith("```") ? rest : `${firstLine}\n${rest}`;
-	}
-	flush() {
-		if (this._done) return "";
-		this._done = true;
-		const buffered = this._buffer;
-		this._buffer = "";
-		return buffered.trim().startsWith("```") ? "" : buffered;
-	}
-};
-const generateCode = async ({ iteration, ctx, cognitive, controller, metadata, onSend, onSendDelta, onRunStart, onRunComplete }) => {
+const generateCode = async ({ iteration, ctx, cognitive, controller, metadata, onSend, onSendDelta, onRunStart }) => {
 	const startedAt = Date.now();
 	const traces = iteration.traces;
-	const modelRef = Array.isArray(iteration.model) ? iteration.model[0] : iteration.model;
-	const model = await cognitive.getModelDetails(modelRef).catch((thrown) => {
+	controller.signal.throwIfAborted();
+	const modelRefs = Array.isArray(iteration.model) ? iteration.model : [iteration.model];
+	const models = await Promise.all(modelRefs.map((modelRef) => cognitive.getModelDetails(modelRef).catch((thrown) => {
 		throw new require_errors.CognitiveError(`Failed to fetch model details for model "${modelRef}": ${require_utils.getErrorMessage(thrown)}`);
-	});
-	let modelLimit = Math.max(model.input.maxTokens, 8e3);
-	if (ctx.maxTokens) modelLimit = Math.min(ctx.maxTokens, modelLimit);
-	const responseLengthBuffer = getModelOutputLimit(modelLimit);
+	})));
+	controller.signal.throwIfAborted();
+	const model = models[0];
+	const modelLimit = Math.min(...models.map((model) => model.input.maxTokens), ctx.maxTokens ?? Infinity);
+	const responseLengthBuffer = Math.min(getModelOutputLimit(modelLimit), Math.floor(modelLimit / 2));
 	if (iteration.tokens) iteration.tokens.limit = modelLimit;
 	let messages;
 	try {
@@ -220,7 +194,7 @@ const generateCode = async ({ iteration, ctx, cognitive, controller, metadata, o
 	const liveContent = /* @__PURE__ */ new Map();
 	let codeGenerationTraced = false;
 	let runCompleted = false;
-	let awaitsRunResult = false;
+	let completions = [];
 	const dispatchSends = async (events) => {
 		for (const event of events) if (event.type === "item-start") {
 			liveItems.set(event.item.id, event.item);
@@ -234,7 +208,7 @@ const generateCode = async ({ iteration, ctx, cognitive, controller, metadata, o
 			}
 		} else if (event.type === "body-delta" && onSendDelta) {
 			const item = liveItems.get(event.itemId);
-			if ((item === null || item === void 0 ? void 0 : item.kind) !== "send" || awaitsRunResult) continue;
+			if ((item === null || item === void 0 ? void 0 : item.kind) !== "send" || runCompleted) continue;
 			const content = (liveContent.get(item.id) ?? "") + event.delta;
 			liveContent.set(item.id, content);
 			const delta = {
@@ -247,22 +221,19 @@ const generateCode = async ({ iteration, ctx, cognitive, controller, metadata, o
 			};
 			await preview(delta);
 		} else if (event.type === "item-complete") {
-			if (event.item.kind === "send" && onSend && !awaitsRunResult) await onSend({
-				name: event.item.name,
-				props: event.item.props,
-				body: event.item.body
-			}, messageMetadata(event.item.id));
-			else if (event.item.kind === "run" && event.item.status === "complete" && !runCompleted) {
-				runCompleted = true;
-				const code = (event.item.body ?? "").trim();
-				awaitsRunResult = require_compiler.hasTopLevelReturn(code);
-				onRunComplete === null || onRunComplete === void 0 || onRunComplete(code);
-			}
+			if (event.item.kind === "send" && onSend && !runCompleted) {
+				const send = {
+					name: event.item.name,
+					props: event.item.props,
+					body: event.item.body
+				};
+				const metadata = messageMetadata(event.item.id);
+				completions.push(() => onSend(send, metadata));
+			} else if (event.item.kind === "run" && event.item.status === "complete" && !runCompleted) runCompleted = true;
 		}
 	};
 	if (typeof cognitive.generateTextStream === "function") {
-		let parser = new require_common.StreamingMessageParser();
-		let fence = new LeadingFenceFilter();
+		let parser = new require_common.ResponseParser();
 		const streamController = createJoinedAbortController([controller.signal]);
 		const requestedAt = Date.now();
 		const stream = cognitive.generateTextStream({
@@ -290,6 +261,7 @@ const generateCode = async ({ iteration, ctx, cognitive, controller, metadata, o
 		};
 		raw = "";
 		let streamCompleted = false;
+		let accepted = false;
 		try {
 			while (true) {
 				var _chunk$value, _chunk$value2, _chunk$value3, _chunk$value4;
@@ -314,13 +286,13 @@ const generateCode = async ({ iteration, ctx, cognitive, controller, metadata, o
 						...chunk.value.restart
 					});
 					raw = "";
-					parser = new require_common.StreamingMessageParser();
-					fence = new LeadingFenceFilter();
+					completions = [];
+					accepted = false;
+					parser = new require_common.ResponseParser();
 					liveItems.clear();
 					liveContent.clear();
 					codeGenerationTraced = false;
 					runCompleted = false;
-					awaitsRunResult = false;
 					responseMetadata = void 0;
 					attempt = chunk.value.restart.attempt;
 					await preview({
@@ -338,29 +310,63 @@ const generateCode = async ({ iteration, ctx, cognitive, controller, metadata, o
 				timeToLastToken = Date.now() - requestedAt;
 				timeToFirstToken ??= timeToLastToken;
 				raw += delta;
-				await dispatchSends(parser.push(fence.push(delta)));
+				await dispatchSends(parser.push(delta));
 			}
 			if (!responseMetadata) throw new require_errors.CognitiveError("LLM streaming completed without metadata");
 			assertSuccessfulGeneration(responseMetadata);
-			const remaining = fence.flush();
-			if (remaining) await dispatchSends(parser.push(remaining));
-			await dispatchSends(parser.finish());
+			await dispatchSends(parser.finish(responseMetadata.stopReason));
+			if (parser.valid) {
+				for (const complete of completions) await complete();
+				accepted = true;
+			}
 			assistantResponse = require_common.toParsedAssistantResponse(parser.items, raw, parser.diagnostics);
+		} catch (error) {
+			parser.finish();
+			const usage = (responseMetadata === null || responseMetadata === void 0 ? void 0 : responseMetadata.usage) ?? {
+				inputTokens: 0,
+				outputTokens: 0,
+				inputCost: 0,
+				outputCost: 0
+			};
+			iteration.llm = {
+				started_at: startedAt,
+				ended_at: Date.now(),
+				status: "error",
+				cached: (responseMetadata === null || responseMetadata === void 0 ? void 0 : responseMetadata.cached) ?? false,
+				tokens: usage.inputTokens + usage.outputTokens,
+				spend: (responseMetadata === null || responseMetadata === void 0 ? void 0 : responseMetadata.cost) ?? usage.inputCost + usage.outputCost,
+				output: raw,
+				diagnostics: parser.diagnostics,
+				model: (responseMetadata === null || responseMetadata === void 0 ? void 0 : responseMetadata.model) ?? model.id,
+				time_to_first_token: timeToFirstToken,
+				time_to_last_token: timeToLastToken,
+				usage
+			};
+			throw error;
 		} finally {
 			streamController.abort("LLM stream closed");
 			if (!streamCompleted) stream.return(void 0).catch((err) => {});
+			if (!accepted && liveContent.size) await preview({
+				restart: true,
+				iterationId: iteration.id,
+				attempt: attempt + 1,
+				fromModel: model.id,
+				toModel: model.id,
+				reason: "invalid or incomplete response envelope"
+			});
 		}
-		if (midStreamFallback) controller.signal.throwIfAborted();
+		controller.signal.throwIfAborted();
 	} else {
 		const response = await cognitive.generateText(input, { signal: controller.signal }).catch((thrown) => {
 			throw new require_errors.CognitiveError(`LLM generation failed: ${require_utils.getErrorMessage(thrown)}`);
 		});
+		controller.signal.throwIfAborted();
 		if (response.error) throw new require_errors.CognitiveError(`LLM generation failed: ${response.error}`);
 		if (!response.output) throw new require_errors.CognitiveError("LLM did not return any text output");
 		responseMetadata = response.metadata;
 		assertSuccessfulGeneration(responseMetadata);
 		raw = response.output;
-		assistantResponse = ctx.version.parseAssistantResponse(raw);
+		assistantResponse = ctx.version.parseAssistantResponse(raw, responseMetadata.stopReason);
 		for (const [index, send] of assistantResponse.sends.entries()) await (onSend === null || onSend === void 0 ? void 0 : onSend(send, messageMetadata(`send-${index}`)));
 	}
 	iteration.code = assistantResponse.code;
@@ -565,8 +571,10 @@ const finalizeIteration = async ({ iteration, status, controller, onIterationEnd
 //#region src/runtime/tool-wrapper.ts
 const SLOW_TOOL_WARNING = (0, ms.default)("15s");
 function wrapTool({ chat, tool, traces, object, iteration, beforeHook, afterHook, controller }) {
-	const getToolInput = (input) => tool.zInput.safeParse(input).data ?? input;
+	const traceSchema = tool.input ? require_utils.fromJSONSchemaCompat(tool.input) : void 0;
+	const getToolInput = (input) => (traceSchema === null || traceSchema === void 0 ? void 0 : traceSchema.safeParse(input).data) ?? input;
 	return async function(input) {
+		controller.signal.throwIfAborted();
 		const toolCallId = `tcall_${(0, ulid.ulid)()}`;
 		const originalInput = input;
 		let effectiveInput = input;
@@ -642,6 +650,7 @@ function wrapTool({ chat, tool, traces, object, iteration, beforeHook, afterHook
 				object,
 				toolCallId
 			}));
+			controller.signal.throwIfAborted();
 			if (typeof (beforeRes === null || beforeRes === void 0 ? void 0 : beforeRes.input) !== "undefined") effectiveInput = beforeRes.input;
 			output = await tool.execute(effectiveInput, {
 				callId: toolCallId,
@@ -658,9 +667,16 @@ function wrapTool({ chat, tool, traces, object, iteration, beforeHook, afterHook
 			}));
 			if (typeof (afterRes === null || afterRes === void 0 ? void 0 : afterRes.output) !== "undefined") output = afterRes.output;
 		} catch (err) {
-			if (!await handleSignals(err)) {
+			try {
+				if (!await handleSignals(err)) {
+					success = false;
+					error = err;
+					output = void 0;
+				}
+			} catch (hookError) {
 				success = false;
-				error = err;
+				error = hookError;
+				output = void 0;
 			}
 		} finally {
 			clearTimeout(alertSlowTool);
@@ -960,8 +976,6 @@ const executeIteration = async ({ iteration, ctx, cognitive, controller, onExit,
 			};
 		});
 	};
-	const canExecuteEarly = typeof onBeforeExecution !== "function" && !ctx.midStreamFallback;
-	let earlyExecution;
 	await generateCode({
 		iteration,
 		ctx,
@@ -969,15 +983,6 @@ const executeIteration = async ({ iteration, ctx, cognitive, controller, onExit,
 		controller,
 		metadata,
 		onRunStart: () => require_vm.warmupVM(),
-		onRunComplete: canExecuteEarly ? (code) => {
-			if (!code.length || controller.signal.aborted || earlyExecution) return;
-			iteration.code = code;
-			earlyExecution = {
-				code,
-				started_at: Date.now(),
-				promise: runCode(code)
-			};
-		} : void 0,
 		onSend: async (send, messageMetadata) => {
 			if (!ctx.chat) return;
 			const sendStartedAt = Date.now();
@@ -1002,7 +1007,6 @@ const executeIteration = async ({ iteration, ctx, cognitive, controller, onExit,
 		},
 		onSendDelta: ((_ctx$chat = ctx.chat) === null || _ctx$chat === void 0 ? void 0 : _ctx$chat.onMessageDelta) ? (delta) => ctx.chat.onMessageDelta(delta) : void 0
 	});
-	if (earlyExecution) iteration.code = earlyExecution.code;
 	if (typeof onBeforeExecution === "function") try {
 		const hookRes = await onBeforeExecution(iteration, controller);
 		if (typeof (hookRes === null || hookRes === void 0 ? void 0 : hookRes.code) === "string" && hookRes.code.trim().length > 0) iteration.code = hookRes.code.trim();
@@ -1040,7 +1044,7 @@ const executeIteration = async ({ iteration, ctx, cognitive, controller, onExit,
 		return;
 	}
 	if (!iteration.code) {
-		var _iteration$sends;
+		var _iteration$llm;
 		if (iteration.next) {
 			await applyNextExit({
 				iteration,
@@ -1049,26 +1053,14 @@ const executeIteration = async ({ iteration, ctx, cognitive, controller, onExit,
 			});
 			return;
 		}
-		if (ctx.chat && ((_iteration$sends = iteration.sends) === null || _iteration$sends === void 0 ? void 0 : _iteration$sends.length)) {
-			iteration.next = {
-				name: "listen",
-				props: {}
-			};
-			await applyNextExit({
-				iteration,
-				controller,
-				onExit
-			});
-			return;
-		}
 		iteration.end({
 			type: "invalid_code_error",
-			invalid_code_error: { message: "The response did not include a ■run block or a ■next exit. Reply using ■ blocks and end your response with ■run or ■next=<exit>." }
+			invalid_code_error: { message: ((_iteration$llm = iteration.llm) === null || _iteration$llm === void 0 || (_iteration$llm = _iteration$llm.diagnostics) === null || _iteration$llm === void 0 || (_iteration$llm = _iteration$llm.find((diagnostic) => diagnostic.code === "invalid-envelope")) === null || _iteration$llm === void 0 ? void 0 : _iteration$llm.message) ?? "The response did not include a ■run block or a ■next exit. Reply with ■start, the required ■ blocks, and ■end." }
 		});
 		return;
 	}
-	const startedAt = (earlyExecution === null || earlyExecution === void 0 ? void 0 : earlyExecution.started_at) ?? Date.now();
-	const result = earlyExecution ? await earlyExecution.promise : await runCode(iteration.code ?? "");
+	const startedAt = Date.now();
+	const result = await runCode(iteration.code);
 	await interpretVMResult({
 		iteration,
 		result,