Merge pull request '#rss get: 3-5 separate messages with short headers (v0.21.0)' (#43) from feat/rss-multi-message into main

The single-message summary was too small (owner request). The summarizer
prompt now asks for exactly 3-5 lines, one item per line, in the format
'<short header>: <one or two short sentences>' with a <=20-char header,
no intro/outro/numbering. maxTokens stays 2500 (reasoning-token headroom).

The reply is parsed into items: drop empty lines and lines without a ':'
header separator, cap at 5; fewer than 3 valid lines sends what we got
(no failure/retry), zero valid lines sends nothing. Each item goes out as
its own mesh message via respond(id, line) - all threaded to the trigger
packet id on its channel. Per-item backstop fitOneLine() hard-cuts any
line >185 chars at a word boundary near 180 so every item fits one packet;
the old whole-reply fitOnePacket() cut is gone.

CommandService sends array replies line by line through the same routing
as single replies (respond per line, or legacy sendText fallback per line
when there is no packet id). LLM failure path unchanged: single
'summary failed' reply.

AGENTS.md

@@ -27,5 +27,5 @@
 
 - MESHTASTIC.md — rules for Meshtastic communication (owner authority, language, response protocol); read before acting on mesh messages.
 - Hourly cron weather report job (`src/weather`): fetches Open-Meteo current conditions and broadcasts a one-line report to the mesh on the pinned send channel; env `WEATHER_CRON`/`WEATHER_LAT`/`WEATHER_LON`/`WEATHER_CITY`.
-- Channel-2 auto-responder (`src/commands`): event-driven — `CommandService.onRx()` is called by `MeshService` for every decoded rx text (right after the notify gate) and dispatches exact, case-sensitive `#command` matches on `COMMAND_CHANNEL` (env, default 2) through a single-registration-point registry (handlers registered in the constructor via `register(name, fn)`; handler gets `{ text, from?, id? }`, returns reply string or void = no reply). First command: `#ping` → replies `#pong!`. Replies go through `MeshService.respond(id, reply)`: auto-routed (rx → DM back to the original sender unless the original was a broadcast; tx → its destination), riding the original record's channel with `replyId = id` so the recipient app renders it as a threaded quote — `COMMAND_CHANNEL` is only the trigger/listen gate. Only when the trigger has no packet id does it fall back to legacy sendText routing on `COMMAND_CHANNEL`. A failing handler/respond only logs a warn. DI: provided by `CommandModule`'s factory (injects `MeshService`, attaches itself via `attachCommands()`) — no constructor-level cycle; `forwardRef` was tried and left one side null in NestJS 12's injector. Prefix dispatch (`registerPrefix(name, fn)`) lets one handler own a whole family: the `#rss` subcommands live in `src/rss` (`RssService.handle`, registered as prefix `'rss'` from the CommandModule factory — exact matches still win over prefixes). Subcommands: `#rss help`; `#rss add <url> <name>` (validates URL, fetches with 15 s timeout, requires ≥1 titled item, rejects duplicate names case-insensitively; LLM generates 2-4 lowercase tags from up to 8 latest titles — on LLM failure the source is added with empty tags and the reply says so); `#rss remove <name>`; `#rss tags` (all distinct tags); `#rss sources`; `#rss get <name|#tag>` (LLM headline summary, maxTokens 200; `#tag` collects items across ALL sources carrying that tag — one dead feed does not sink the summary). Storage: `rss_sources(name TEXT PRIMARY KEY, url, tags, created_at)` table in the EXISTING MESH_DB_PATH file via a second node:sqlite connection (names matched case-insensitively); feed parsing via `fast-xml-parser` (RSS 2.0 + Atom; malformed/no-items → "not a valid feed"); LLM via injected `LlmService.chat()` with explicit maxTokens (50 tags / 200 summary) to keep mesh replies short.
+- Channel-2 auto-responder (`src/commands`): event-driven — `CommandService.onRx()` is called by `MeshService` for every decoded rx text (right after the notify gate) and dispatches exact, case-sensitive `#command` matches on `COMMAND_CHANNEL` (env, default 2) through a single-registration-point registry (handlers registered in the constructor via `register(name, fn)`; handler gets `{ text, from?, id? }`, returns a reply string — or an array of lines, sent as one message per line through the same routing — or void = no reply). First command: `#ping` → replies `#pong!`. Replies go through `MeshService.respond(id, reply)`: auto-routed (rx → DM back to the original sender unless the original was a broadcast; tx → its destination), riding the original record's channel with `replyId = id` so the recipient app renders it as a threaded quote — `COMMAND_CHANNEL` is only the trigger/listen gate. Only when the trigger has no packet id does it fall back to legacy sendText routing on `COMMAND_CHANNEL`. A failing handler/respond only logs a warn. DI: provided by `CommandModule`'s factory (injects `MeshService`, attaches itself via `attachCommands()`) — no constructor-level cycle; `forwardRef` was tried and left one side null in NestJS 12's injector. Prefix dispatch (`registerPrefix(name, fn)`) lets one handler own a whole family: the `#rss` subcommands live in `src/rss` (`RssService.handle`, registered as prefix `'rss'` from the CommandModule factory — exact matches still win over prefixes). Subcommands: `#rss help`; `#rss add <url> <name>` (validates URL, fetches with 15 s timeout, requires ≥1 titled item, rejects duplicate names case-insensitively; LLM generates 2-4 lowercase tags from up to 8 latest titles — on LLM failure the source is added with empty tags and the reply says so); `#rss remove <name>`; `#rss tags` (all distinct tags); `#rss sources`; `#rss get <name|#tag>` (LLM headline summary as 3-5 SEPARATE mesh messages: prompt asks for exactly 3-5 lines, one item per line, `<short header>: <one or two short sentences>` with ≤20-char header, no intro/outro/numbering; reply parsed into items — drop empty/separator-less lines, cap at 5, fewer than 3 → send what we got (no retry), zero → no message; each item sent via its own respond() threaded to the trigger packet id, per-line backstop `fitOneLine()` hard-cuts >185-char lines at a word boundary near 180 so every item fits one packet; maxTokens 2500 is headroom because the LLM server's reasoning tokens count against it — see rss.service.ts comment; `#tag` collects items across ALL sources carrying that tag — one dead feed does not sink the summary). Storage: `rss_sources(name TEXT PRIMARY KEY, url, tags, created_at)` table in the EXISTING MESH_DB_PATH file via a second node:sqlite connection (names matched case-insensitively); feed parsing via `fast-xml-parser` (RSS 2.0 + Atom; malformed/no-items → "not a valid feed"); LLM via injected `LlmService.chat()` with explicit maxTokens (50 tags / 2500 summary headroom).
 - LLM client (`src/llm`): `LlmService` wraps the local OpenAI-compatible server (default `http://192.168.1.70:1111/v1`, no auth) with native fetch — plain feature module, no Mesh/Schedule dependency. Methods: `listModels()` → `data[].id`; `chat(messages, { maxTokens?, temperature? })` → `choices[0].message.content` (non-2xx throws with status + body snippet); `summarize(text)` = chat() with system prompt "Summarize the following text concisely." Env: `LLM_API_BASE`, `LLM_MODEL` (default `qwen3.8-27b`), `LLM_TIMEOUT_MS` (default 60000, via `AbortSignal.timeout`).

CHANGELOG.md

@@ -1,5 +1,22 @@
 # Changelog
 
+## v0.21.0 — 2026-09-25 19:59
+
+- `#rss get <name|#tag>` now replies with **3–5 SEPARATE mesh messages** instead of one small summary (owner request: the single-message summary was too small). Each item is its own packet, so no fragmentation — and each rides the trigger's channel threaded as a quote of the command packet.
+- New summarizer prompt asks for exactly 3-5 lines, one item per line, format `<short header>: <one or two short sentences>` with a ≤20-char header (e.g. `Culture: Museum opens new wing`), no intro/outro/numbering. maxTokens stays 2500 (reasoning-token headroom — see v0.20.1).
+- The LLM reply is parsed into items: split on newlines, trim, drop empty lines and lines without a `:` header separator, cap at 5. Fewer than 3 valid lines → send what we got (no failure, no retry — the model is non-deterministic); zero valid lines → no message sent.
+- Per-item backstop replaces the old whole-reply cut: any single line >185 chars is hard-cut at a word boundary near 180 (`fitOneLine`), so every item fits one packet even if the model ignores the prompt. `fitOnePacket()` (whole-reply ≤190) is gone — it no longer applies to joined text.
+- CommandService now sends array replies line by line: each element goes through the same routing as a single reply (`respond(id, line)` threaded to the trigger packet id; legacy sendText fallback per line when there is no packet id). A failing send stops the remaining lines (warn log), like any failed command.
+- LLM failure path unchanged: single `summary failed` reply via respond(). `#rss get #tag` gets the same multi-message treatment automatically (shared summary path — covered by tests). All other subcommands unchanged.
+- Tests: 158 → 165 — prompt shape + maxTokens asserted on the new call, 5-item cap, invalid-line filtering, fewer-than-3 fallback, per-line over-length cut through `handle()`, direct `parseSummaryItems`/`fitOneLine` specs, and CommandService array-reply routing (threaded respond ×N to one packet id; legacy sendText fallback).
+
+## v0.20.1 — 2026-09-25 19:20
+
+- `#rss get` summaries now fit ONE mesh packet (≤190 chars). Mesh fragments longer replies into fire-and-forget packets that recipients frequently never receive, and the old prompt ("a few short one-line highlights") produced ~326-334 char replies.
+- The summary prompt now asks for "at most 3 short bullet lines, each under 45 characters … bullets only - no intro, no other text" (verified against the live LLM: exactly 3 compliant bullets, ~110 chars). A `fitOnePacket()` backstop additionally cuts any over-length reply at complete-bullet boundaries so delivery is one packet even if the model ignores the prompt.
+- The summary `chat()` call now uses an explicit `maxTokens = 2500` (was 200). Empirically verified on this server: qwen3.8-27b emits `reasoning_content` and those tokens COUNT against max_tokens (cap 150 → completion_tokens 150, of which 105 were reasoning, finish_reason 'length'; a cap of 500 was fully consumed by reasoning with zero visible content). Thinking length for this task is unbounded and uncontrollable server-side (`/no_think`, `chat_template_kwargs.enable_thinking` and `thinking_budget` are all ignored) — observed 32..1776 tokens across runs. 2500 covers every observed run plus ~80 content tokens, and keeps the worst case (~42 s at ~60 tok/s) under the 60 s LLM timeout. The old cap of 200 guaranteed empty replies whenever reasoning exceeded it.
+- Tests: 154 → 158 — prompt shape + maxTokens asserted on the summary call, over-length reply cut through `handle()`, and direct `fitOnePacket` specs (pass-through at/under 190, bullet-boundary cut, hard-cut fallback for non-bulleted paragraphs).
+
 ## v0.20.0 — 2026-09-25 18:40
 
 - New `#rss` command family on the command channel (ch2): `#rss help`, `#rss add <url> <name>` (validates URL + fetches with a 15 s timeout, requires ≥1 item with a non-empty title, rejects duplicate names case-insensitively; tags are generated by the LLM from up to 8 latest titles — on LLM failure the source is still added with empty tags and the reply says so), `#rss remove <name>`, `#rss tags` (all distinct tags), `#rss sources`, `#rss get <name>` and `#rss get #<tag>` (LLM headline summary, maxTokens 200; tag form collects items across ALL sources carrying the tag — one dead feed does not sink the summary). Bare `#rss` / unknown subcommand → help. All replies ride the existing respond() path like `#ping`.

package.json

@@ -1,6 +1,6 @@
 {
   "name": "mesh-bot",
-  "version": "0.20.0",
+  "version": "0.21.0",
   "description": "",
   "author": "",
   "private": true,

src/commands/command.service.spec.ts

@@ -114,6 +114,30 @@ describe('CommandService.onRx', () => {
     expect(mesh.respond).toHaveBeenCalledWith(5, 'up from 99');
   });
 
+  it('sends each line of an array reply as its own threaded respond() to the same packet id', async () => {
+    const [svc, mesh] = makeService();
+    svc.register('multi', () => ['one: first item', 'two: second item']);
+    await svc.onRx(2, '#multi', 42, 9);
+    expect(mesh.respond).toHaveBeenCalledTimes(2);
+    expect(mesh.respond.mock.calls[0]).toEqual([9, 'one: first item']);
+    expect(mesh.respond.mock.calls[1]).toEqual([9, 'two: second item']);
+    expect(mesh.sendText).not.toHaveBeenCalled();
+  });
+
+  it('sends each line of an array reply via the legacy sendText fallback when there is no packet id', async () => {
+    const [svc, mesh] = makeService();
+    svc.register('multi', () => ['one: first item', 'two: second item']);
+    await svc.onRx(2, '#multi', 42); // no id → legacy path on COMMAND_CHANNEL, DM back to the sender
+    expect(mesh.sendText).toHaveBeenCalledTimes(2);
+    expect(mesh.sendText.mock.calls[0]).toEqual([
+      42,
+      'one: first item',
+      undefined,
+      2,
+    ]);
+    expect(mesh.respond).not.toHaveBeenCalled();
+  });
+
   it('swallows a failing handler (a bad command must not crash the app)', async () => {
     const [svc, mesh] = makeService();
     svc.register('boom', () => {

src/commands/command.service.ts

@@ -14,10 +14,10 @@ export interface CommandContext {
   id?: number;
 }
 
-/** A command handler: returns the reply text to send back, or void/undefined for no reply. */
+/** A command handler: returns the reply text(s) to send back — an array sends one message per line — or void/undefined for no reply. */
 export type CommandHandler = (
   ctx: CommandContext,
-) => string | undefined | Promise<string | undefined>;
+) => string | string[] | undefined | Promise<string | string[] | undefined>;
 
 @Injectable()
 export class CommandService {
@@ -83,17 +83,20 @@ export class CommandService {
     try {
       const reply = await handler({ text: trimmed, from, id });
       if (reply === undefined || reply === null) return; // void = no reply
-      if (id !== undefined && this.mesh) {
-        // respond() auto-routes the reply: rides the original record's channel with replyId = id (threaded quote)
-        await this.mesh.respond(id, reply);
-      } else {
-        // defensive fallback when the trigger has no packet id: legacy routing on COMMAND_CHANNEL
-        await this.mesh?.sendText(
-          from ?? BROADCAST_ADDR,
-          reply,
-          undefined,
-          this.commandChannel,
-        );
+      // an array reply sends one message per line, all threaded to the same trigger packet id
+      for (const line of Array.isArray(reply) ? reply : [reply]) {
+        if (id !== undefined && this.mesh) {
+          // respond() auto-routes each line: rides the original record's channel with replyId = id (threaded quote)
+          await this.mesh.respond(id, line);
+        } else {
+          // defensive fallback when the trigger has no packet id: legacy routing on COMMAND_CHANNEL
+          await this.mesh?.sendText(
+            from ?? BROADCAST_ADDR,
+            line,
+            undefined,
+            this.commandChannel,
+          );
+        }
       }
     } catch (err) {
       this.logger.warn(

src/rss/rss.service.spec.ts

@@ -5,7 +5,7 @@ import { DatabaseSync } from 'node:sqlite';
 import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
 import type { LlmService } from '../llm/llm.service.js';
 import { RssStore } from './rss.store.js';
-import { RssService } from './rss.service.js';
+import { fitOneLine, parseSummaryItems, RssService } from './rss.service.js';
 
 const FEED_A = `<rss version="2.0"><channel>
   <item><title>Alpha headline</title></item>
@@ -149,18 +149,79 @@ describe('RssService.handle', () => {
     expect(await svc.handle(ctx('#rss sources'))).toBe('Aye, Bee');
   });
 
-  it('get <name>: summarizes the latest titles via chat with a bounded maxTokens', async () => {
+  it('get <name>: summarizes via chat with a bounded maxTokens and replies with headered lines', async () => {
     stored().add({ name: 'myfeed', url: 'https://a.example/rss', tags: [] });
     fetchMock.mockResolvedValueOnce(feedResponse(FEED_A));
-    llm.chat.mockResolvedValueOnce('Alpha and beta happened.');
+    llm.chat.mockResolvedValueOnce(
+      'Chips: EU funds foundries.\nPark: Riverfront expansion approved.',
+    );
 
     const reply = await svc.handle(ctx('#rss get myfeed'));
 
-    expect(reply).toBe('Alpha and beta happened.');
+    expect(reply).toEqual([
+      'Chips: EU funds foundries.',
+      'Park: Riverfront expansion approved.',
+    ]);
     const [messages, opts] = llm.chat.mock.calls[0];
-    expect(opts).toEqual({ maxTokens: 200 });
-    expect(messages[0].content).toContain('Alpha headline');
-    expect(messages[0].content).toContain('Beta headline');
+    // headroom for the server's reasoning tokens (they count against max_tokens) — see rss.service.ts
+    expect(opts).toEqual({ maxTokens: 2500 });
+    const prompt = messages[0].content;
+    expect(prompt).toContain('exactly 3-5 lines');
+    expect(prompt).toContain('<short header>: <one or two short sentences>');
+    expect(prompt).toContain('at most 20 characters');
+    expect(prompt).toContain('No intro, no outro, no numbering');
+    expect(prompt).toContain('Alpha headline');
+    expect(prompt).toContain('Beta headline');
+  });
+
+  it('get <name>: caps items at 5', async () => {
+    stored().add({ name: 'myfeed', url: 'https://a.example/rss', tags: [] });
+    fetchMock.mockResolvedValueOnce(feedResponse(FEED_A));
+    const seven = ['a', 'b', 'c', 'd', 'e', 'f', 'g']
+      .map((c) => `${c}: item ${c}`)
+      .join('\n');
+    llm.chat.mockResolvedValueOnce(seven);
+
+    const reply = await svc.handle(ctx('#rss get myfeed'));
+
+    expect(reply).toEqual([
+      'a: item a',
+      'b: item b',
+      'c: item c',
+      'd: item d',
+      'e: item e',
+    ]);
+  });
+
+  it('get <name>: drops empty and separator-less lines, sends what remains when fewer than 3 are valid', async () => {
+    stored().add({ name: 'myfeed', url: 'https://a.example/rss', tags: [] });
+    fetchMock.mockResolvedValueOnce(feedResponse(FEED_A));
+    llm.chat.mockResolvedValueOnce(
+      'Chips: EU funds foundries.\n\nJust a sentence without a header\n   \nPark: expansion approved.',
+    );
+
+    const reply = await svc.handle(ctx('#rss get myfeed'));
+
+    expect(reply).toEqual([
+      'Chips: EU funds foundries.',
+      'Park: expansion approved.',
+    ]);
+  });
+
+  it('get <name>: hard-cuts an over-length line at a word boundary so each item fits one packet', async () => {
+    stored().add({ name: 'myfeed', url: 'https://a.example/rss', tags: [] });
+    fetchMock.mockResolvedValueOnce(feedResponse(FEED_A));
+    const longLine = `Chips: ${'word '.repeat(40)}`; // 207 chars — over the per-line cap
+    llm.chat.mockResolvedValueOnce(`${longLine}\nPark: short line.`);
+
+    const reply = await svc.handle(ctx('#rss get myfeed'));
+
+    expect(Array.isArray(reply)).toBe(true);
+    const [cut, ok] = reply as string[];
+    expect(cut.length).toBeLessThanOrEqual(180);
+    expect(longLine.startsWith(cut)).toBe(true); // prefix — cut at a word boundary, not mid-word
+    expect(cut.endsWith(' ')).toBe(false);
+    expect(ok).toBe('Park: short line.'); // other lines untouched
   });
 
   it('get #tag: collects items across all sources with the tag; one dead feed does not sink it', async () => {
@@ -171,11 +232,11 @@ describe('RssService.handle', () => {
       .mockResolvedValueOnce(feedResponse(FEED_A)) // A answers (Alpha, Beta)
       .mockRejectedValueOnce(new Error('down')); // B is dead
 
-    llm.chat.mockResolvedValueOnce('Only alpha survived.');
+    llm.chat.mockResolvedValueOnce('Chips: only alpha survived.');
     const reply = await svc.handle(ctx('#rss get #tech'));
 
     // B's fetch rejected, yet the summary still went out from A's titles alone
-    expect(reply).toBe('Only alpha survived.');
+    expect(reply).toEqual(['Chips: only alpha survived.']);
     expect(llm.chat).toHaveBeenCalledTimes(1);
     const content = llm.chat.mock.calls[0][0][0].content;
     expect(content).toContain('Alpha headline');
@@ -188,9 +249,15 @@ describe('RssService.handle', () => {
       .mockResolvedValueOnce(feedResponse(FEED_A)) // A → Alpha, Beta
       .mockResolvedValueOnce(feedResponse(FEED_B)); // B → Gamma
 
-    llm.chat.mockResolvedValueOnce('all three.');
-    await svc.handle(ctx('#rss get #tech'));
+    llm.chat.mockResolvedValueOnce(
+      'Alpha: first headline.\nGamma: second headline.',
+    );
+    const reply = await svc.handle(ctx('#rss get #tech'));
 
+    expect(reply).toEqual([
+      'Alpha: first headline.',
+      'Gamma: second headline.',
+    ]);
     const content = llm.chat.mock.calls[0][0][0].content;
     expect(content).toContain('Alpha headline');
     expect(content).toContain('Gamma headline'); // proves B's feed was collected too
@@ -204,3 +271,47 @@ describe('RssService.handle', () => {
     );
   });
 });
+
+describe('parseSummaryItems (per-line items for multi-message summaries)', () => {
+  it('keeps lines with a header separator, drops empty and separator-less ones', () => {
+    expect(
+      parseSummaryItems(
+        'Chips: EU funds foundries.\n\nNo header here\n   \nPark: expansion approved.',
+      ),
+    ).toEqual(['Chips: EU funds foundries.', 'Park: expansion approved.']);
+  });
+
+  it('caps at 5 items', () => {
+    const seven = ['a', 'b', 'c', 'd', 'e', 'f', 'g']
+      .map((c) => `${c}: item`)
+      .join('\n');
+    expect(parseSummaryItems(seven)).toHaveLength(5);
+    expect(parseSummaryItems(seven)[4]).toBe('e: item');
+  });
+
+  it('returns [] when nothing is valid', () => {
+    expect(parseSummaryItems('no headers at all\n')).toEqual([]);
+  });
+});
+
+describe('fitOneLine (one mesh packet per line)', () => {
+  it('passes through lines that already fit, including exactly 185 chars', () => {
+    const short = 'Chips: EU funds foundries.';
+    expect(fitOneLine(short)).toBe(short);
+    const exact = 'x'.repeat(185);
+    expect(fitOneLine(exact)).toBe(exact);
+  });
+
+  it('cuts over-long lines at a word boundary near 180', () => {
+    const long = `Chips: ${'word '.repeat(40)}`; // 207 chars
+    const cut = fitOneLine(long);
+    expect(cut.length).toBeLessThanOrEqual(180);
+    expect(long.startsWith(cut)).toBe(true); // prefix — not mid-word
+    expect(cut.endsWith(' ')).toBe(false);
+  });
+
+  it('hard-cuts at 180 when there is no word boundary in reach', () => {
+    const long = 'w'.repeat(250);
+    expect(fitOneLine(long)).toBe('w'.repeat(180));
+  });
+});

src/rss/rss.service.ts

@@ -10,12 +10,24 @@ import { RssStore } from './rss.store.js';
 const FETCH_TIMEOUT_MS = 15000; // feed fetches — same budget as the opencode notifier's POSTs
 const MAX_TITLES = 8; // "up to 8 latest item titles" per the design
 const TAG_MAX_TOKENS = 50;
-const SUMMARY_MAX_TOKENS = 200;
+// The LLM server (qwen3.8-27b) emits reasoning_content and those tokens COUNT against max_tokens
+// (verified: cap 150 → completion_tokens 150 of which 105 were reasoning, finish_reason 'length').
+// Thinking length for this task is unbounded and uncontrollable server-side (/no_think,
+// chat_template_kwargs.enable_thinking and thinking_budget are all ignored) — observed 32..1776
+// tokens across runs. So the cap must be headroom for reasoning + ~80 content tokens:
+// 2500 covers every observed run (worst case ≈42 s at ~60 tok/s, under LLM_TIMEOUT_MS=60 s).
+const SUMMARY_MAX_TOKENS = 2500;
+// Mesh fragments replies longer than ~190 chars into fire-and-forget packets that recipients
+// often never get — each summary line must fit ONE packet on its own. The prompt asks for short
+// headered lines; this is the hard backstop if the model ignores it (cut at a word boundary).
+const MAX_LINE_CHARS = 185; // per-line one-packet threshold
+const LINE_CUT_AT = 180; // where over-long lines get cut, snapped back to a word boundary
+const MAX_ITEMS = 5; // cap on summary items sent as separate messages
 
 /**
  * The `#rss` command family: stored RSS sources (SQLite, shared MESH_DB_PATH file) with
- * LLM-generated tags and on-demand headline summaries. All replies are short strings —
- * CommandService routes them through MeshService.respond() like any other command reply.
+ * LLM-generated tags and on-demand headline summaries. Replies are short strings — or an array
+ * of lines for multi-message summaries, which CommandService sends one respond() per line.
  */
 @Injectable()
 export class RssService {
@@ -29,8 +41,8 @@ export class RssService {
     this.store = new RssStore(db);
   }
 
-  /** Entry point for the `#rss` prefix — always replies (errors included). */
-  async handle(ctx: CommandContext): Promise<string> {
+  /** Entry point for the `#rss` prefix — always replies (errors included); `get` may reply with several lines. */
+  async handle(ctx: CommandContext): Promise<string | string[]> {
     const words = ctx.text.trim().split(/\s+/);
     switch (words[1]) {
       case undefined: // bare `#rss`
@@ -131,8 +143,8 @@ export class RssService {
     return names.length ? names.join(', ') : 'no sources yet';
   }
 
-  /** `#rss get <name>` or `#rss get #<tag>`. */
-  private async get(target: string): Promise<string> {
+  /** `#rss get <name>` or `#rss get #<tag>` — replies with 3-5 separate one-packet lines. */
+  private async get(target: string): Promise<string | string[]> {
     if (!target) return 'usage: #rss get <name|#tag>';
 
     let urls: string[];
@@ -163,15 +175,18 @@ export class RssService {
     if (!titles.length) return 'no items found';
 
     try {
-      return await this.llm.chat(
+      const reply = await this.llm.chat(
         [
           {
             role: 'user',
-            content: `Summarize these RSS headlines into a few short one-line highlights. Be concise:\n${titles.join('\n')}`,
+            content: `Summarize these RSS headlines into exactly 3-5 lines, one item per line, each in the format "<short header>: <one or two short sentences>" where the header is at most 20 characters (e.g. "Culture: Museum opens new wing"). No intro, no outro, no numbering - items only:\n${titles.join('\n')}`,
           },
         ],
         { maxTokens: SUMMARY_MAX_TOKENS },
       );
+      // each item goes out as its own mesh message — CommandService sends the array line by line;
+      // fewer than 3 valid lines is fine (the model is non-deterministic), we send what we got
+      return parseSummaryItems(reply).map(fitOneLine);
     } catch (err) {
       this.logger.warn(`[rss] summary failed: ${(err as Error).message}`);
       return 'summary failed';
@@ -190,6 +205,23 @@ export class RssService {
   }
 }
 
+/** Parse an LLM summary into per-line items: drop empty lines and lines without a `:` header separator, cap at MAX_ITEMS. */
+export function parseSummaryItems(reply: string): string[] {
+  return reply
+    .split('\n')
+    .map((l) => l.trim())
+    .filter((l) => l && l.includes(':'))
+    .slice(0, MAX_ITEMS);
+}
+
+/** Hard-cut a single summary line at a word boundary near LINE_CUT_AT when it exceeds MAX_LINE_CHARS. */
+export function fitOneLine(line: string): string {
+  if (line.length <= MAX_LINE_CHARS) return line;
+  const slice = line.slice(0, LINE_CUT_AT);
+  const at = slice.lastIndexOf(' ');
+  return (at > 0 ? slice.slice(0, at) : slice).trimEnd();
+}
+
 /** Normalize LLM tag output: lowercase, no #, whitespace → '-', deduped. */
 export function sanitizeTags(raw: string): string[] {
   const seen = new Set<string>();