Skip to content

Write the busy-ack with one short model call instead of templates #1667

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Birfy merged 1 commit into from
Oct 29, 2026
Merged

Write the busy-ack with one short model call instead of templates #1667

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions changelog.d/1665-busy-ack-model.changed.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1 @@
A Bot that is busy now acknowledges a new message with one short, natural sentence written by a quick model call, in the message's language, instead of a templated line that quoted a truncated piece of the message; when the model does not answer within 3 seconds only the pending dots show ([#1665](https://github.com/ThinkFlowLab/FlyBot/issues/1665)).
241 changes: 106 additions & 135 deletions core/src/host/busy-ack.js
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,128 +1,50 @@
// Busy-ack (#1584): a human message that lands while a Bot turn is already running
// gets a short spoken acknowledgment on the live payload immediately. Steer admits
// to next-step, but a blocking job wait may never yield a step (#1509), so without
// this the chat stays silent until the wait ends.
// gets a short spoken acknowledgment on the live payload. Steer admits to next-step,
// but a blocking job wait may never yield a step (#1509), so without this the chat
// stays silent until the wait ends.
//
// The line is composed here, deterministically, from the message itself and what the
// turn is doing right now (its running tool), not a fixed sentence (#1629). A model
// call would wait behind the very turn that is busy, cost a call per message and need
// a fallback anyway; a template is instant, free and cannot fail. Wording rotates and
// never repeats the previous ack of the same chat; the language follows the message.
// The line is written by one short model call (#1665), not a template: the templated
// lines of #1629 quoted a clipped, truncated piece of the message back inside a fixed
// phrase and read as canned. The ack is published at once with an empty text, which the
// client draws as the avatar with blinking dots (#1656/#1657); the model's sentence fills
// it in when it comes. The call is separate from the busy turn, uses the reasoning-off
// setting when the model has one, a tiny output cap, and a hard ACK_TIMEOUT_MS; on a
// timeout, an error or an empty answer no text is ever sent and the dots clear with the
// turn as before. `config.flybot.modelFor['busy-ack']` names a cheaper model for it;
// otherwise the chat's own model is used.
import { BlockAssembler } from '@deepseek-ai/dsh-llm'
import { splitRef } from './text.js'

export const COALESCE_MS = 3_000
export const TOPIC_MAX = { zh: 16, en: 36 }

const CJK = /[\u3400-\u9fff\uf900-\ufaff]/g

/** 'zh' when the message is mostly Chinese, 'en' when mostly Latin, else the UI language. */
export function ackLanguage(text, uiLanguage) {
const value = String(text ?? '')
const cjk = (value.match(CJK) ?? []).length
const latin = (value.match(/[A-Za-z]/g) ?? []).length
if (cjk > 0 && cjk * 2 >= latin) return 'zh'
if (latin > 0) return 'en'
return /chinese|简体|\bzh\b/i.test(String(uiLanguage ?? '')) ? 'zh' : 'en'
export const ACK_TIMEOUT_MS = 3_000
export const ACK_MAX_TOKENS = 80
export const ACK_MAX_CHARS = 120

export const ACK_SYSTEM = [
'You write the one-line acknowledgment a busy assistant sends when a new message arrives while it is still working.',
'Reply in the same language as the new message.',
'Write one natural, short sentence that shows you understood what the new message is about and says briefly what you are doing right now.',
'Do not quote or copy the message, do not use quotation marks, and do not use stock phrases such as "got it" or "noted"; vary your wording.',
'Output only that sentence.',
].join(' ')

/** The user prompt for the ack call: the new message and what the turn is doing. */
export function ackPrompt({ message, tool, working }) {
const doing = typeof tool === 'string' && tool.trim() !== '' && tool !== 'ask_user' ? `running the tool \`${tool}\`` : 'working on the previous request'
const task = String(working ?? '').trim().slice(0, 600)
return [
`Currently: ${doing}.`,
...(task === '' ? [] : [`The request being worked on: ${task}`]),
`New message: ${String(message ?? '').trim().slice(0, 1_500)}`,
].join('\n')
}

/** A short topic out of the message: first line, trimmed, clipped; '' when too thin to echo. */
export function topicOf(text, lang) {
/** The model's answer as one clean line, or '' when unusable. */
export function cleanAck(text) {
const line = String(text ?? '').split(/\r?\n/).map(part => part.trim()).find(Boolean) ?? ''
const clean = line.replace(/\s+/g, ' ').replace(/^[@#]\S+\s*/, '').replace(/[。！？!?.,，、~～…]+$/u, '').trim()
if (clean === '' || /^(hi|hello|hey|ok|okay|嗯+|好的?|在吗|在不在|\?+|？+)$/i.test(clean)) return ''
const max = TOPIC_MAX[lang] ?? TOPIC_MAX.en
const chars = [...clean]
if (chars.length <= max) return clean
let cut = chars.slice(0, max - 1).join('')
if (lang === 'en' && cut.lastIndexOf(' ') > max / 2) cut = cut.slice(0, cut.lastIndexOf(' '))
return `${cut.trimEnd()}…`
}

const TOOL_WORDS = [
[/^(bash|shell|exec|run|command|terminal)$/i, { zh: '跑命令', en: 'running a command' }],
[/^(wait|job|poll|await)$/i, { zh: '等一个任务跑完', en: 'waiting on a job' }],
[/^(search|grep|glob|find)$/i, { zh: '查资料', en: 'searching' }],
[/^(fetch|browse|browser|web|http|url)$/i, { zh: '看网页', en: 'reading a page' }],
[/^(write|edit|patch|apply)$/i, { zh: '改文件', en: 'editing files' }],
[/^(read|cat|open|view|ls)$/i, { zh: '翻文件', en: 'reading files' }],
[/^(message|delegate|subagent|agent)$/i, { zh: '和别的 Bot 对接', en: 'checking with another Bot' }],
[/^(task|todo|plan)$/i, { zh: '整理任务', en: 'sorting out tasks' }],
]

/** What the running tool is, in a few words; undefined with no tool. */
export function doingOf(tool, lang) {
if (typeof tool !== 'string' || tool.trim() === '' || tool === 'ask_user') return undefined
// Match whole name parts (`web_fetch` → web, fetch) so `frobnicate` is not "cat".
const parts = tool.toLowerCase().split(/[^a-z0-9]+/).filter(Boolean)
const hit = TOOL_WORDS.find(([pattern]) => parts.some(part => pattern.test(part)) )
if (hit) return hit[1][lang]
return lang === 'zh' ? `用 ${tool}` : `using ${tool}`
}

const TEMPLATES = {
zh: {
topicDoing: [
'看到你说「{t}」了，我正在{d}，弄完马上接上。',
'「{t}」收到，手上在{d}，这步完了就来。',
'记下「{t}」了，{d}这边一结束就处理。',
],
topic: [
'看到「{t}」了，当前这步一完就回你。',
'「{t}」记下了，马上接着处理。',
'好，「{t}」这个我排上了，稍等一下。',
],
doing: [
'看到了，我正在{d}，弄完就回你。',
'收到新消息，手上在{d}，稍等片刻。',
'嗯，{d}这边一结束就回来。',
],
bare: [
'看到了，这步做完就回你。',
'新消息收到，稍等我一下。',
'好的，马上接上你这条。',
],
},
en: {
topicDoing: [
'Saw "{t}" — I’m {d} right now, back on it in a moment.',
'Noted "{t}" — still {d}, then it’s next.',
'Got "{t}" — I’ll get to it once I’m done {d}.',
],
topic: [
'Saw "{t}" — I’ll pick it up right after this step.',
'Noted "{t}", coming to it shortly.',
'"{t}" is next on my list — one moment.',
],
doing: [
'Seen — I’m {d}, back to you shortly.',
'New message noted; still {d}, one moment.',
'On it as soon as I’m done {d}.',
],
bare: [
'Seen — back to you right after this step.',
'Noted, one moment.',
'Picking this up as soon as the current step ends.',
],
},
}

/**
* One short ack line for `message` while the turn is doing `tool`. `seed` picks the
* variant; `previous` is never returned again (the next variant is taken instead).
*/
export function composeBusyAck({ message, tool, uiLanguage, seed = 0, previous } = {}) {
const lang = ackLanguage(message, uiLanguage)
const topic = topicOf(message, lang)
const doing = doingOf(tool, lang)
const kind = topic && doing ? 'topicDoing' : topic ? 'topic' : doing ? 'doing' : 'bare'
const pool = TEMPLATES[lang][kind]
// One pass only: the topic is the person's own text, so a `{t}`/`{d}` they typed
// must not be rescanned as template markup after it is inserted (#1634).
const fill = index => pool[index % pool.length].replace(/\{(t|d)\}/g, (_, key) => (key === 't' ? topic : (doing ?? '')))
const start = Math.abs(Math.trunc(seed)) % pool.length
for (let step = 0; step < pool.length; step += 1) {
const text = fill(start + step)
if (text !== previous) return text
}
return fill(start)
const clean = line.replace(/^["'“”‘’「」『』]+|["'“”‘’「」『』]+$/gu, '').trim()
if (clean === '' || [...clean].length > ACK_MAX_CHARS) return ''
return clean
}

export function shouldNoteBusyAck({ sourceKind, turnRunning, ownChat }) {
Expand All @@ -139,7 +61,7 @@ export function isHumanRoomLine(source) {
&& source.remote === undefined
}

// The relay prefix delivery.js puts on the line; the ack quotes the person's text, not it.
// The relay prefix delivery.js puts on the line; the ack reads the person's text, not it.
const relayHeader = /^\[[^\]\n]*\]\r?\n/

const textOf = (content) => {
Expand All @@ -149,11 +71,9 @@ const textOf = (content) => {
}

export function install(rt) {
const { ctx } = rt
/** @type {Map<string, { text: string, at: number }>} */
const { ctx, config } = rt
/** @type {Map<string, { text: string, at: number, id: number }>} */
const acks = new Map()
/** Last ack text per chat, kept across clears so consecutive acks differ. */
const lastText = new Map()
let counter = 0

const turnRunning = sessionId => rt.live?.get(sessionId)?.turn !== undefined
Expand All @@ -162,24 +82,73 @@ export function install(rt) {
/** A person's line into this member's running group Session (#1646), once only. */
const roomLine = (sessionId, source) => groupSession(sessionId) && isHumanRoomLine(source)

const note = (sessionId, message) => {
// The route for the ack: a configured economy model first, else the chat's own.
const routeOf = (session) => {
const named = config?.flybot?.modelFor?.['busy-ack']
const own = typeof named === 'string' && named.trim() !== '' ? splitRef(named.trim()) : undefined
if (own?.provider && own?.model) return own
const header = session?.requestHeader?.()?.config
return header?.provider && header?.model ? { provider: header.provider, model: header.model } : undefined
}

// The latest earlier user request in the Session: what the busy turn is about.
const workingOf = (session, message) => {
try {
const messages = session?.deriveMessages?.() ?? []
for (let index = messages.length - 1; index >= 0; index -= 1) {
if (messages[index]?.role !== 'user') continue
const text = textOf(messages[index].content).replace(relayHeader, '').trim()
if (text !== '' && text !== String(message ?? '').trim()) return text
}
} catch { /* no history is fine */ }
return ''
}

// One short call; resolves to the line or '' — never throws, never outlives the timeout.
async function write(session, sessionId, message) {
const route = routeOf(session)
if (route === undefined || typeof ctx.llm?.stream !== 'function') return ''
const signal = AbortSignal.timeout(ACK_TIMEOUT_MS)
const work = (async () => {
const info = await ctx.llm.resolveModelInfo?.(route.provider, route.model)?.catch?.(() => undefined)
const off = info?.reasoning?.efforts?.some(effort => effort.id === 'off')
const messages = [{ role: 'user', content: [{ type: 'text', text: ackPrompt({ message, tool: rt.live?.get(sessionId)?.tool, working: workingOf(session, message) }) }] }]
rt.tagUsage?.(messages, { sessionId, kind: 'other' })
const assembler = new BlockAssembler()
for await (const chunk of ctx.llm.stream({
provider: route.provider,
model: route.model,
...(off ? { reasoningEffort: 'off' } : {}),
system: ACK_SYSTEM,
messages,
maxTokens: ACK_MAX_TOKENS,
purpose: 'other',
signal,
})) assembler.push(chunk)
const finish = assembler.finish
if (finish?.kind === 'error' || finish?.kind === 'aborted' || finish?.kind === 'max-tokens') return ''
return cleanAck(assembler.blocks().filter(block => block.type === 'text').map(block => block.text).join('\n'))
})().catch(() => '')
let timer
const late = new Promise(resolve => { timer = setTimeout(() => resolve(''), ACK_TIMEOUT_MS) })
try { return await Promise.race([work, late]) } finally { clearTimeout(timer) }
}

const note = (sessionId, message, session) => {
const now = Date.now()
const existing = acks.get(sessionId)
if (existing && now - existing.at < COALESCE_MS) {
existing.at = now
return existing
}
counter += 1
const text = composeBusyAck({
message,
tool: rt.live?.get(sessionId)?.tool,
uiLanguage: rt.uiLanguage?.(),
seed: counter + now,
previous: lastText.get(sessionId),
})
lastText.set(sessionId, text)
const entry = { text, at: now }
// Published at once with no text: the client draws the pending dots (#1657).
const entry = { text: '', at: now, id: counter }
acks.set(sessionId, entry)
entry.ready = write(session, sessionId, message).then((text) => {
if (text !== '' && acks.get(sessionId) === entry) entry.text = text
return entry.text
})
return entry
}

Expand All @@ -194,7 +163,7 @@ export function install(rt) {
const own = shouldNoteBusyAck({ sourceKind: source?.kind, turnRunning: turnRunning(id), ownChat: ownChat(id) })
if (!own && !(room && turnRunning(id))) return
const text = textOf(event.data?.content ?? event.data?.message?.content)
note(id, room ? text.replace(relayHeader, '') : text)
note(id, room ? text.replace(relayHeader, '') : text, session)
return
}
if (event.type === 'assistant/message') {
Expand Down Expand Up @@ -222,6 +191,8 @@ export function install(rt) {
return [[botId, ack === undefined ? { text: '', at: 0 } : { text: ack.text, at: ack.at }]]
})),
noteBusyAck: note,
// The pending model answer for a Session's ack ('' when none); for tests and diagnostics.
busyAckReady: sessionId => acks.get(sessionId)?.ready ?? Promise.resolve(''),
clearBusyAck: clear,
})
}
Loading
Loading