Auto-reply: smarter chunking breaks
This commit is contained in:
46
src/auto-reply/chunk.test.ts
Normal file
46
src/auto-reply/chunk.test.ts
Normal file
@@ -0,0 +1,46 @@
|
||||
import { describe, expect, it } from "vitest";
|
||||
|
||||
import { chunkText } from "./chunk.js";
|
||||
|
||||
describe("chunkText", () => {
|
||||
it("keeps multi-line text in one chunk when under limit", () => {
|
||||
const text = "Line one\n\nLine two\n\nLine three";
|
||||
const chunks = chunkText(text, 1600);
|
||||
expect(chunks).toEqual([text]);
|
||||
});
|
||||
|
||||
it("splits only when text exceeds the limit", () => {
|
||||
const part = "a".repeat(20);
|
||||
const text = part.repeat(5); // 100 chars
|
||||
const chunks = chunkText(text, 60);
|
||||
expect(chunks.length).toBe(2);
|
||||
expect(chunks[0].length).toBe(60);
|
||||
expect(chunks[1].length).toBe(40);
|
||||
expect(chunks.join("")).toBe(text);
|
||||
});
|
||||
|
||||
it("prefers breaking at a newline before the limit", () => {
|
||||
const text = `paragraph one line\n\nparagraph two starts here and continues`;
|
||||
const chunks = chunkText(text, 40);
|
||||
expect(chunks).toEqual([
|
||||
"paragraph one line",
|
||||
"paragraph two starts here and continues",
|
||||
]);
|
||||
});
|
||||
|
||||
it("otherwise breaks at the last whitespace under the limit", () => {
|
||||
const text = "This is a message that should break nicely near a word boundary.";
|
||||
const chunks = chunkText(text, 30);
|
||||
expect(chunks[0].length).toBeLessThanOrEqual(30);
|
||||
expect(chunks[1].length).toBeLessThanOrEqual(30);
|
||||
expect(chunks.join(" ").replace(/\s+/g, " ").trim()).toBe(
|
||||
text.replace(/\s+/g, " ").trim(),
|
||||
);
|
||||
});
|
||||
|
||||
it("falls back to a hard break when no whitespace is present", () => {
|
||||
const text = "Supercalifragilisticexpialidocious"; // 34 chars
|
||||
const chunks = chunkText(text, 10);
|
||||
expect(chunks).toEqual(["Supercalif", "ragilistic", "expialidoc", "ious"]);
|
||||
});
|
||||
});
|
||||
48
src/auto-reply/chunk.ts
Normal file
48
src/auto-reply/chunk.ts
Normal file
@@ -0,0 +1,48 @@
|
||||
// Utilities for splitting outbound text into platform-sized chunks without
|
||||
// unintentionally breaking on newlines. Using [\s\S] keeps newlines inside
|
||||
// the chunk so messages are only split when they truly exceed the limit.
|
||||
|
||||
export function chunkText(text: string, limit: number): string[] {
|
||||
if (!text) return [];
|
||||
if (limit <= 0) return [text];
|
||||
if (text.length <= limit) return [text];
|
||||
|
||||
const chunks: string[] = [];
|
||||
let remaining = text;
|
||||
|
||||
while (remaining.length > limit) {
|
||||
const window = remaining.slice(0, limit);
|
||||
|
||||
// 1) Prefer a newline break inside the window.
|
||||
let breakIdx = window.lastIndexOf("\n");
|
||||
|
||||
// 2) Otherwise prefer the last whitespace (word boundary) inside the window.
|
||||
if (breakIdx <= 0) {
|
||||
for (let i = window.length - 1; i >= 0; i--) {
|
||||
if (/\s/.test(window[i])) {
|
||||
breakIdx = i;
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// 3) Fallback: hard break exactly at the limit.
|
||||
if (breakIdx <= 0) breakIdx = limit;
|
||||
|
||||
const rawChunk = remaining.slice(0, breakIdx);
|
||||
const chunk = rawChunk.trimEnd();
|
||||
if (chunk.length > 0) {
|
||||
chunks.push(chunk);
|
||||
}
|
||||
|
||||
// If we broke on whitespace/newline, skip that separator; for hard breaks keep it.
|
||||
const brokeOnSeparator =
|
||||
breakIdx < remaining.length && /\s/.test(remaining[breakIdx]);
|
||||
const nextStart = Math.min(remaining.length, breakIdx + (brokeOnSeparator ? 1 : 0));
|
||||
remaining = remaining.slice(nextStart).trimStart();
|
||||
}
|
||||
|
||||
if (remaining.length) chunks.push(remaining);
|
||||
|
||||
return chunks;
|
||||
}
|
||||
@@ -255,11 +255,11 @@ export async function runCommandReply(
|
||||
}
|
||||
|
||||
const parsed = trimmed ? agent.parseOutput(trimmed) : undefined;
|
||||
const parserProvided = !!parsed;
|
||||
|
||||
// Collect one message per assistant text from parseOutput (tau RPC can emit many).
|
||||
const parsedTexts =
|
||||
parsed?.texts?.map((t) => t.trim()).filter(Boolean) ??
|
||||
(parsed?.text ? [parsed.text.trim()] : []);
|
||||
parsed?.texts?.map((t) => t.trim()).filter(Boolean) ?? [];
|
||||
|
||||
type ReplyItem = { text: string; media?: string[] };
|
||||
const replyItems: ReplyItem[] = [];
|
||||
@@ -274,7 +274,7 @@ export async function runCommandReply(
|
||||
}
|
||||
|
||||
// If parser gave nothing, fall back to raw stdout as a single message.
|
||||
if (replyItems.length === 0 && trimmed) {
|
||||
if (replyItems.length === 0 && trimmed && !parserProvided) {
|
||||
const { text: cleanedText, mediaUrls: mediaFound } =
|
||||
splitMediaFromOutput(trimmed);
|
||||
if (cleanedText || mediaFound?.length) {
|
||||
@@ -401,7 +401,7 @@ export async function runCommandReply(
|
||||
}
|
||||
|
||||
verboseLog(`Command auto-reply meta: ${JSON.stringify(meta)}`);
|
||||
return { payloads, meta };
|
||||
return { payloads, payload: payloads[0], meta };
|
||||
} catch (err) {
|
||||
const elapsed = Date.now() - started;
|
||||
logger.info(
|
||||
|
||||
49
src/auto-reply/reply.chunking.test.ts
Normal file
49
src/auto-reply/reply.chunking.test.ts
Normal file
@@ -0,0 +1,49 @@
|
||||
import { describe, expect, it, vi } from "vitest";
|
||||
|
||||
import type { WarelayConfig } from "../config/config.js";
|
||||
import { autoReplyIfConfigured } from "./reply.js";
|
||||
|
||||
describe("autoReplyIfConfigured chunking", () => {
|
||||
it("sends a single Twilio message for multi-line text under limit", async () => {
|
||||
const body = [
|
||||
"Oh! Hi Peter! 🦞",
|
||||
"",
|
||||
"Sorry, I got a bit trigger-happy with the heartbeat response there. What's up?",
|
||||
"",
|
||||
"Everything working on your end?",
|
||||
].join("\n");
|
||||
|
||||
const config: WarelayConfig = {
|
||||
inbound: {
|
||||
reply: {
|
||||
mode: "text",
|
||||
text: body,
|
||||
},
|
||||
},
|
||||
};
|
||||
|
||||
const create = vi.fn().mockResolvedValue({});
|
||||
const client = { messages: { create } } as unknown as Parameters<
|
||||
typeof autoReplyIfConfigured
|
||||
>[0];
|
||||
|
||||
const message = {
|
||||
body: "ping",
|
||||
from: "+15551234567",
|
||||
to: "+15557654321",
|
||||
sid: "SM123",
|
||||
} as Parameters<typeof autoReplyIfConfigured>[1];
|
||||
|
||||
await autoReplyIfConfigured(client, message, config);
|
||||
|
||||
expect(create).toHaveBeenCalledTimes(1);
|
||||
expect(create).toHaveBeenCalledWith(
|
||||
expect.objectContaining({
|
||||
body,
|
||||
from: message.to,
|
||||
to: message.from,
|
||||
}),
|
||||
);
|
||||
});
|
||||
});
|
||||
|
||||
@@ -1,5 +1,4 @@
|
||||
import crypto from "node:crypto";
|
||||
|
||||
import type { MessageInstance } from "twilio/lib/rest/api/v2010/account/message.js";
|
||||
import { loadConfig, type WarelayConfig } from "../config/config.js";
|
||||
import {
|
||||
@@ -18,6 +17,7 @@ import { defaultRuntime, type RuntimeEnv } from "../runtime.js";
|
||||
import type { TwilioRequester } from "../twilio/types.js";
|
||||
import { sendTypingIndicator } from "../twilio/typing.js";
|
||||
import { runCommandReply } from "./command-reply.js";
|
||||
import { chunkText } from "./chunk.js";
|
||||
import {
|
||||
applyTemplate,
|
||||
type MsgContext,
|
||||
@@ -307,7 +307,7 @@ export async function getReplyFromConfig(
|
||||
mediaUrl: reply.mediaUrl,
|
||||
};
|
||||
cleanupTyping();
|
||||
return [result];
|
||||
return result;
|
||||
}
|
||||
|
||||
if (reply && reply.mode === "command" && reply.command?.length) {
|
||||
@@ -318,7 +318,7 @@ export async function getReplyFromConfig(
|
||||
mode: "command" as const,
|
||||
};
|
||||
try {
|
||||
const { payloads, meta } = await runCommandReply({
|
||||
const runResult = await runCommandReply({
|
||||
reply: commandReply,
|
||||
templatingCtx,
|
||||
sendSystemOnce,
|
||||
@@ -329,6 +329,17 @@ export async function getReplyFromConfig(
|
||||
timeoutSeconds,
|
||||
commandRunner,
|
||||
});
|
||||
const payloadArray =
|
||||
runResult.payloads ?? (runResult.payload ? [runResult.payload] : []);
|
||||
const meta = runResult.meta;
|
||||
const normalizedPayloads =
|
||||
payloadArray.length === 1 ? payloadArray[0] : payloadArray;
|
||||
if (
|
||||
!normalizedPayloads ||
|
||||
(Array.isArray(normalizedPayloads) && normalizedPayloads.length === 0)
|
||||
) {
|
||||
return undefined;
|
||||
}
|
||||
if (sessionCfg && sessionStore && sessionKey) {
|
||||
const returnedSessionId = meta.agentMeta?.sessionId;
|
||||
if (returnedSessionId && returnedSessionId !== sessionId) {
|
||||
@@ -357,7 +368,7 @@ export async function getReplyFromConfig(
|
||||
if (meta.agentMeta && isVerbose()) {
|
||||
logVerbose(`Agent meta: ${JSON.stringify(meta.agentMeta)}`);
|
||||
}
|
||||
return payloads;
|
||||
return normalizedPayloads;
|
||||
} finally {
|
||||
cleanupTyping();
|
||||
}
|
||||
@@ -459,10 +470,8 @@ export async function autoReplyIfConfigured(
|
||||
: [];
|
||||
|
||||
const text = replyPayload.text ?? "";
|
||||
const chunks =
|
||||
text.length > 0
|
||||
? (text.match(new RegExp(`.{1,${TWILIO_TEXT_LIMIT}}`, "g")) ?? [])
|
||||
: [""];
|
||||
const chunks = chunkText(text, TWILIO_TEXT_LIMIT);
|
||||
if (chunks.length === 0) chunks.push("");
|
||||
|
||||
for (let i = 0; i < chunks.length; i++) {
|
||||
const body = chunks[i];
|
||||
|
||||
Reference in New Issue
Block a user