Cap #rss get summaries to one mesh packet (v0.20.1)

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.
2026-09-25 19:22:13 +00:00
parent a8ce04d3f8
commit e2d712a26c
5 changed files with 88 additions and 9 deletions
@@ -1,5 +1,12 @@
 # Changelog

+## 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`.