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.
Replies >~190 chars fragment into fire-and-forget packets recipients
often never receive; old prompt produced ~326-334 char replies.
Prompt now asks for at most 3 short bullet lines, each under 45 chars,
bullets only (verified live: exactly 3 compliant bullets, ~110 chars).
fitOnePacket() backstop cuts any over-length reply at complete-bullet
boundaries so delivery is one packet even if the model ignores the prompt.
maxTokens 200 -> 2500 on the summary chat(): empirically verified that
this server's reasoning tokens count against max_tokens (cap 150 ->
completion 150 of which 105 were reasoning; cap 500 fully consumed by
reasoning, zero visible content) and thinking is uncontrollable
server-side (/no_think, chat_template_kwargs.enable_thinking,
thinking_budget all ignored), observed 32..1776 tokens for this task.
2500 covers every observed run + ~80 content tokens, worst case ~42 s
under the 60 s LLM timeout.
#rss subcommands on the command channel: help / add <url> <name> /
remove <name> / tags / sources / get <name|#tag>. Sources are stored in
a new rss_sources table in the existing MESH_DB_PATH file (second
node:sqlite connection, case-insensitive names); feeds parsed with
fast-xml-parser (RSS 2.0 + Atom); AI tags and headline summaries via
injected LlmService.chat() with bounded maxTokens (50/200).
CommandService gains prefix registration so one handler owns the whole
#rss family while #ping stays an exact match; wired in CommandModule's
factory — still no constructor-level cycle, no forwardRef. One dead feed
does not sink a tag summary.
LlmService wraps the local OpenAI-compatible server (default
http://192.168.1.70:1111/v1, no auth) with native fetch only — no new
dependencies. listModels() -> data[].id; chat(messages, opts?) ->
choices[0].message.content (non-2xx throws with status + body snippet);
summarize(text) = chat() with a concise-summary system prompt.
Env: LLM_API_BASE / LLM_MODEL (qwen3.8-27b) / LLM_TIMEOUT_MS (60000,
AbortSignal.timeout). Plain feature module wired into app.module.ts;
docs synced (AGENTS.md, README env list).
Event-driven CommandService dispatches exact case-sensitive #command
matches on COMMAND_CHANNEL (default 2) through a single-registration-point
registry; first command #ping -> #pong!. Replies ride COMMAND_CHANNEL:
directed DM back to the sender when present, else broadcast. Failing
handler/send only logs a warn.
DI: CommandModule's factory injects MeshService and attaches itself via
MeshService.attachCommands() — a constructor-level cycle leaves one side
null in NestJS 12's injector (verified), so no forwardRef is used.
Tests 106 -> 121; docs synced (AGENTS.md, MESHTASTIC.md, README).
kamilkrzywda