Skip to content

feat(api): route ai bot replies through an llm router - #22

Merged
Datzu712 merged 19 commits into
refactor/ai-module-vercel-sdkfrom
feat/ai-bot-router
Nov 17, 2026
Merged

Datzu712 merged 19 commits into
refactor/ai-module-vercel-sdkfrom
feat/ai-bot-router

Conversation

@Datzu712

@Datzu712 Datzu712 commented Nov 17, 2026 •

Copy link
Copy Markdown
Owner

Stacked on #21. Bots no longer need to be named: an LLM router reads the last 20 messages and decides who, if anyone, replies. Named bots always answer, addressed bots reply in parallel, one other bot may react to their replies (teasing allowed), spontaneous replies are capped (1 per bot/channel/10 min), and replies are short and paced like typing. AI_ROUTER=off falls back to name detection.

Verified: pnpm check, unit tests, full api suite 112/112 (--runInBand), manual tests with claude-haiku-5-5.

Datzu712 and others added 14 commits November 16, 2026 17:12
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Add SpontaneousCaps, which records the last time a bot replied on its own
in a channel and blocks another unprompted reply for 10 minutes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
BotConversationService replaces MessageService.handleAIResponse. Human
messages are debounced per channel (2.5 s of silence), then the router
picks the speakers, merged with the bots named in the new messages by
planSpeakers. Speakers reply in turns, each seeing the previous replies;
a new human message skips the remaining turns and starts a new cycle.
Spontaneous speakers after the first are re-confirmed with the router.
After the turns, at most one bot that has not spoken may react.
Spontaneous replies and reactions are capped per channel and bot.

With AI_ROUTER=off or a router error, only the named bots reply, as
before. stopBotTyping now takes the channel, so a reply ending in one
channel does not clear the bot's typing indicator in another.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A cycle queued behind a running one took the pending messages as soon
as it started, so a message that arrived meanwhile skipped its 2.5 s of
silence, and the next timer then ran a cycle with no new messages. A
cycle now returns without consuming anything while a newer timer is
waiting or when nothing is pending.

A human message during the router re-ask for a spontaneous speaker did
not stop that turn; the re-ask now reports the speaker as not picked
once the cycle was interrupted.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An interrupt now skips only the remaining turns of bots nobody named, plus
the reaction; a named bot still answers. planSpeakers drops unnamed speakers
from the tail instead of slicing, so a named bot that sits after router
picks is never cut. A running cycle stops at its next turn and skips the
reaction after onModuleDestroy.

Also removes the unused detectBot, moving its name-variant cases under
detectNamedBots, and adds tests for routerEnabled false, the spontaneous
re-ask, and the router prompt without exclusions.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Style: the three personas are shorter, and each one now ends with a shared
CHAT_STYLE block that asks for one or two short sentences, a casual tone,
no actions between asterisks, no commentary on the conversation, varied
openings, no question at the end of every message and no words put in
other people's mouths. Replies are capped at 150 output tokens.

Router: questions about other bots go to whoever was talking with the
person or to the bot not mentioned, "era para X" / "iba para X" makes X
answer what it was asked, a greeting to the group gets one bot, and
laughter, thanks and goodbyes normally get none.

Pacing: the debounce drops to 1.5 s. The typing indicator stays on until
the reply is posted, held after generation for typingDelayMs (27 chars/s,
at most 5 s, minus the generation time). A bot pauses 1.2 s to read
before answering another bot's reply, including the reaction, and the
interrupt and destroy checks run again after every pause.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The named-bots interrupt test now lets pedro post and checks that the
reaction never routes. The destroy test checks that nothing is posted
after the typing delay. New tests cover a human message during the read
pause: an unnamed second speaker is skipped, a named one still answers,
and a reaction is skipped.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Live tests felt slow: a reply took 6.5-8 s, and a second bot asked the
same question came 4 s after the first.

Pacing: the debounce drops from 1.5 s to 0.6 s, typing speed goes from
27 to 40 chars/s and the typing cap from 5 s to 3 s. The read pause stays
at 1.2 s and now only applies to sequential turns.

Parallel rule: the leading run of addressed speakers (named or picked as
addressed by the router) starts at once, each with its own typing
indicator and history, and each posts when its own reply is typed. No
read pause inside the run. Interrupt and destroy rules apply per
speaker. Spontaneous speakers keep their sequential turns after the run
(read pause, recheck, re-ask, cap) and see its replies; the reaction
still comes last.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Pacing constants, history limits and helpers move out of the service into
bot-conversation.util.ts, and both files live in message/bot-conversation/.
Also drops the comment and leading blank line of CHAT_STYLE.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@Datzu712
Datzu712 added this pull request to stack #23 November 17, 2026 04:20
Datzu712 and others added 5 commits November 16, 2026 22:54
- Presence tracks every channel a bot is typing in, so a bot typing in two
  channels keeps its indicator until both replies are posted.
- The spontaneous cap is only used up by a reply that was actually posted.
- A reply cut off by the token limit is trimmed to its last full sentence.
- compose.yaml passes AI_ROUTER through; the router history size lives in
  one constant; AGENTS.md states the real speaker limit (2 in total, more
  only when more bots are named) and the README notes that the router sends
  channel history to the provider.
- BotConversationService passes one BotCycle object around instead of many
  parameters, and helpers used once are inlined.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The bot picked to react is now told whose replies it reacts to, so it teases
or disagrees with them instead of answering the human again. CHAT_STYLE
allows friendly banter between bots, and the router is invited to pick a
reaction when a bot has something witty to say, still none otherwise.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@Datzu712
Datzu712 merged commit c53b4ac into main Nov 17, 2026
7 of 8 checks passed
@Datzu712
Datzu712 deleted the feat/ai-bot-router branch November 17, 2026 23:12
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant