feat: W7 功能標籤分類、對白檢索與注入 (#8)

用 $.model.complete（Haiku）判斷回應的功能標籤，從角色索引取同標籤對白，以只增不改的訊息附在使用者訊息之前；失敗一律退回不注入。 Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
2026-10-14 17:44:09 +00:00
parent 1c4e0c5d98
commit d9e50d196b
12 changed files with 385 additions and 4 deletions
@@ -0,0 +1,62 @@
+import { describe, expect, test } from 'claude-code/testing'
+
+import { buildClassifyPrompt, classify, parseFunction, type Complete } from './classify'
+import { FUNCTIONS } from './tags'
+
+describe('parseFunction', () => {
+  test('白名單內的代號', () => {
+    expect(parseFunction('comfort')).toBe('comfort')
+    expect(parseFunction('  Report_Fail\n')).toBe('report_fail')
+    expect(parseFunction('tease。')).toBe('tease')
+  })
+
+  test('白名單外或空回覆視為失敗', () => {
+    expect(parseFunction('flirt')).toBeNull()
+    expect(parseFunction('')).toBeNull()
+    expect(parseFunction('我覺得是安慰')).toBeNull()
+  })
+
+  test('只認開頭第一個詞', () => {
+    expect(parseFunction('comfort or tease')).toBe('comfort')
+    expect(parseFunction('他說 comfort')).toBeNull()
+  })
+})
+
+describe('buildClassifyPrompt', () => {
+  test('列出所有功能代號（other 除外）並包住使用者的話', () => {
+    const p = buildClassifyPrompt('測試失敗了')
+    for (const key of Object.keys(FUNCTIONS).filter(k => k !== 'other')) expect(p).toContain(`${key}：`)
+    expect(p).toContain('<使用者的話>\n測試失敗了\n</使用者的話>')
+  })
+
+  test('過長的使用者輸入截斷', () => {
+    expect(buildClassifyPrompt('字'.repeat(5000)).length).toBeLessThan(1500)
+  })
+})
+
+describe('classify', () => {
+  const ok = (text: string): Complete => async () => ({ isAnswered: true, text })
+
+  test('成功回傳標籤，並用 haiku、低 effort、有逾時', async () => {
+    let seen: Parameters<Complete>[0] | undefined
+    const r = await classify(async req => ((seen = req), { isAnswered: true, text: 'comfort' }), '好累')
+    expect(r).toBe('comfort')
+    expect(seen?.model).toBe('haiku')
+    expect(seen?.effort).toBe('low')
+    expect(seen?.timeoutMs).toBeGreaterThan(0)
+  })
+
+  test('未回答、空回覆、白名單外都回傳 null', async () => {
+    expect(await classify(async () => ({ isAnswered: false }), 'x')).toBeNull()
+    expect(await classify(ok(''), 'x')).toBeNull()
+    expect(await classify(ok('flirt'), 'x')).toBeNull()
+  })
+
+  test('呼叫丟出例外也回傳 null', async () => {
+    expect(
+      await classify(async () => {
+        throw new Error('blocked model')
+      }, 'x'),
+    ).toBeNull()
+  })
+})
@@ -0,0 +1,51 @@
+import { FUNCTIONS, isFunctionTag, type FunctionTag } from './tags'
+
+export const CLASSIFY_TIMEOUT_MS = 3000
+
+// 判斷的是「角色接下來最自然的回應」屬於哪種功能，不是使用者這句話本身的功能。
+export function buildClassifyPrompt(userText: string): string {
+  const choices = Object.entries(FUNCTIONS)
+    .filter(([key]) => key !== 'other')
+    .map(([key, label]) => `${key}：${label}`)
+    .join('\n')
+  return [
+    '使用者正在請助理協助做事。判斷在他這句話之後，一位親切的角色最自然的第一句回應屬於下列哪一種功能。',
+    '只回答功能代號一個詞，不要解釋；都不適合就回答 other。',
+    '',
+    choices,
+    '',
+    '<使用者的話>',
+    userText.slice(0, 600),
+    '</使用者的話>',
+  ].join('\n')
+}
+
+// 只認回覆開頭的第一個詞，且必須在白名單內；其餘一律視為判斷失敗。
+export function parseFunction(reply: string): FunctionTag | null {
+  const word = reply.trim().toLowerCase().match(/^[a-z_]+/)?.[0]
+  return isFunctionTag(word) ? word : null
+}
+
+export type Complete = (request: {
+  model: string
+  prompt: string
+  maxTokens: number
+  effort: 'low'
+  timeoutMs: number
+}) => Promise<{ isAnswered: boolean; text?: string }>
+
+// 任何失敗（逾時、API 錯誤、空回覆、白名單外）都回傳 null，由呼叫端退回不注入。
+export async function classify(complete: Complete, userText: string): Promise<FunctionTag | null> {
+  try {
+    const r = await complete({
+      model: 'haiku',
+      prompt: buildClassifyPrompt(userText),
+      maxTokens: 16,
+      effort: 'low',
+      timeoutMs: CLASSIFY_TIMEOUT_MS,
+    })
+    return r.isAnswered && r.text ? parseFunction(r.text) : null
+  } catch {
+    return null
+  }
+}
@@ -0,0 +1,8 @@
+import type { CharacterCode } from '../characters'
+import type { IndexEntry } from '../retrieval'
+
+// 各角色的對白索引。W5、W6 由建置流水線產生並填入；目前為空，空索引時 hook 不呼叫模型。
+export const INDEXES: Readonly<Record<CharacterCode, readonly IndexEntry[]>> = {
+  'sao-asuna': [],
+  'sao-suguha': [],
+}
@@ -0,0 +1,13 @@
+import type { IndexEntry } from './retrieval'
+
+export const MAX_LINE = 80
+
+// 注入對話的語氣參考。明說只學語氣，不照抄，也不要提到這份參考。
+export function formatExamples(name: string, entries: readonly IndexEntry[]): string | null {
+  if (entries.length === 0) return null
+  const lines = entries.map(e => `- 「${e.text.length > MAX_LINE ? e.text.slice(0, MAX_LINE) + '…' : e.text}」`)
+  return [
+    `以下是「${name}」在類似情境說過的話，只用來掌握她的語氣、用字與節奏。不要照抄，也不要向使用者提到這份參考。`,
+    ...lines,
+  ].join('\n')
+}
@@ -0,0 +1,54 @@
+import { describe, expect, test } from 'claude-code/testing'
+
+import type { Complete } from './classify'
+import { resolveConfig } from './config'
+import { buildInjection } from './inject'
+import type { IndexEntry } from './retrieval'
+
+const entry = (id: string, fn: IndexEntry['function']): IndexEntry => ({ id, function: fn, emotion: 'neutral', mode: null, text: `句${id}` })
+const indexes = {
+  'sao-asuna': [entry('a', 'comfort'), entry('b', 'comfort')],
+  'sao-suguha': [] as IndexEntry[],
+}
+const answer = (text: string): Complete => async () => ({ isAnswered: true, text })
+const deps = (opts: Record<string, string>, complete: Complete, calls = { n: 0 }) => ({
+  config: resolveConfig(opts),
+  indexes,
+  complete: async (r: Parameters<Complete>[0]) => (calls.n++, complete(r)),
+  now: () => 0,
+  turn: 0,
+})
+
+describe('buildInjection', () => {
+  const on = { defaultCharacter: 'sao-asuna', intensity: '2' }
+
+  test('分類成功時回傳同標籤的對白', async () => {
+    const r = await buildInjection(deps(on, answer('comfort')), '測試失敗了')
+    expect(r?.fn).toBe('comfort')
+    expect(r?.text).toContain('句a')
+  })
+
+  test('沒選角色或強度 0 不呼叫模型', async () => {
+    const calls = { n: 0 }
+    expect(await buildInjection(deps({ intensity: '2' }, answer('comfort'), calls), 'x')).toBeNull()
+    expect(await buildInjection(deps({ ...on, intensity: '0' }, answer('comfort'), calls), 'x')).toBeNull()
+    expect(calls.n).toBe(0)
+  })
+
+  test('索引為空不呼叫模型', async () => {
+    const calls = { n: 0 }
+    expect(await buildInjection(deps({ defaultCharacter: 'sao-suguha', intensity: '2' }, answer('comfort'), calls), 'x')).toBeNull()
+    expect(calls.n).toBe(0)
+  })
+
+  test('分類失敗退回不注入，但仍回報耗時', async () => {
+    const r = await buildInjection(deps(on, async () => ({ isAnswered: false })), 'x')
+    expect(r).toEqual({ text: null, fn: null, ms: 0 })
+  })
+
+  test('該標籤沒有對白時不注入', async () => {
+    const r = await buildInjection(deps(on, answer('greet')), 'x')
+    expect(r?.fn).toBe('greet')
+    expect(r?.text).toBeNull()
+  })
+})
@@ -0,0 +1,31 @@
+import { CHARACTERS, type CharacterCode } from './characters'
+import { classify, type Complete } from './classify'
+import type { Config } from './config'
+import { formatExamples } from './examples'
+import { retrieve, type IndexEntry } from './retrieval'
+
+export type InjectDeps = {
+  config: Config
+  indexes: Readonly<Record<CharacterCode, readonly IndexEntry[]>>
+  complete: Complete
+  append: (text: string) => Promise<unknown>
+  now: () => number
+  turn: number
+}
+
+// 回傳 null 表示這一輪不注入（沒選角色、強度 0、索引為空、分類失敗或沒有同標籤的對白）。
+// 索引為空時不呼叫模型，省下一次無用的分類。
+export async function buildInjection(
+  { config, indexes, complete, now, turn }: Omit<InjectDeps, 'append'>,
+  userText: string,
+): Promise<{ text: string | null; fn: string | null; ms: number } | null> {
+  const character = config.character
+  if (character === null || config.intensity === 0) return null
+  const index = indexes[character]
+  if (index.length === 0) return null
+  const started = now()
+  const fn = await classify(complete, userText)
+  const ms = now() - started
+  if (fn === null) return { text: null, fn: null, ms }
+  return { text: formatExamples(CHARACTERS[character].name, retrieve(index, fn, { turn })), fn, ms }
+}
@@ -1,15 +1,33 @@
 import type { Register } from 'claude-code'

 import { resolveConfig } from './config'
+import { INDEXES } from './data/index'
+import { buildInjection } from './inject'
 import { buildRules, SECTION_ID } from './rules'

-// 依強度與稱呼把規則段落加到系統提示最後；人格卡與檢索見 W5～W7。
+// 規則段落（W8）加到系統提示最後；每則使用者訊息再依功能標籤檢索角色的對白當語氣參考（W7）。
 export const register: Register = (on, options) => {
  const config = resolveConfig(options)
+  const showTiming = options.showTiming === true
+  let turn = 0
+
  on('prompt.compose', async ($, e, next) => {
    const result = await next(e)
    const text = buildRules(config)
    if (text === null) return result
    return { sections: [...result.sections, { id: SECTION_ID, text, scope: 'session' }] }
  })
+
+  on('prompt.submit', async ($, e, next) => {
+    if (e.command || e.isMeta) return next(e)
+    const r = await buildInjection(
+      { config, indexes: INDEXES, complete: req => $.model.complete(req), now: Date.now, turn: turn++ },
+      e.text,
+    )
+    if (r === null) return next(e)
+    if (showTiming) $.ui.status(`anime-companion 分類 ${r.ms}ms → ${r.fn ?? '失敗'}`)
+    // 附加在使用者訊息之前，只增不改，不影響對話的快取。
+    if (r.text !== null) await $.session.append({ message: { type: 'user', content: [{ type: 'text', text: r.text }] } })
+    return next(e)
+  })
 }
@@ -0,0 +1,61 @@
+import { describe, expect, test } from 'claude-code/testing'
+
+import { formatExamples } from './examples'
+import { retrieve, type IndexEntry } from './retrieval'
+
+const e = (id: string, fn: IndexEntry['function'], mode: string | null = null): IndexEntry => ({
+  id,
+  function: fn,
+  emotion: 'neutral',
+  mode,
+  text: `句${id}`,
+})
+const INDEX = [e('c', 'comfort'), e('a', 'comfort'), e('b', 'comfort'), e('d', 'tease'), e('x', 'comfort', 'leafa')]
+
+describe('retrieve', () => {
+  test('只取同功能標籤，依 id 排序', () => {
+    expect(retrieve(INDEX, 'comfort', { k: 10 }).map(r => r.id)).toEqual(['a', 'b', 'c', 'x'])
+  })
+
+  test('預設取 3 句', () => {
+    expect(retrieve(INDEX, 'comfort')).toHaveLength(3)
+  })
+
+  test('沒有同標籤的對白回傳空陣列', () => {
+    expect(retrieve(INDEX, 'greet')).toEqual([])
+  })
+
+  test('依模式過濾', () => {
+    expect(retrieve(INDEX, 'comfort', { mode: 'leafa' }).map(r => r.id)).toEqual(['x'])
+  })
+
+  test('同一輪結果固定，不同輪輪流取', () => {
+    const ids = (turn: number) => retrieve(INDEX, 'comfort', { k: 2, turn }).map(r => r.id)
+    expect(ids(0)).toEqual(ids(0))
+    expect(ids(0)).toEqual(['a', 'b'])
+    expect(ids(1)).toEqual(['b', 'c'])
+    expect(ids(3)).toEqual(['x', 'a'])
+  })
+
+  test('k 大於池子時不重複', () => {
+    expect(new Set(retrieve(INDEX, 'tease', { k: 5 }).map(r => r.id)).size).toBe(1)
+  })
+})
+
+describe('formatExamples', () => {
+  test('沒有對白不注入', () => {
+    expect(formatExamples('亞絲娜', [])).toBeNull()
+  })
+
+  test('列出對白並說明只學語氣', () => {
+    const text = formatExamples('亞絲娜', [e('a', 'comfort')])!
+    expect(text).toContain('句a')
+    expect(text).toContain('不要照抄')
+    expect(text).toContain('亞絲娜')
+  })
+
+  test('過長的句子截斷', () => {
+    const long = { ...e('a', 'comfort'), text: 'あ'.repeat(200) }
+    expect(formatExamples('亞絲娜', [long])!.length).toBeLessThan(300)
+  })
+})
@@ -0,0 +1,28 @@
+import type { FunctionTag } from './tags'
+
+// 對白索引的一筆；由建置流水線產生，見 data/。text 不超過 80 字。
+export type IndexEntry = {
+  id: string
+  function: FunctionTag
+  emotion: string
+  mode: string | null
+  text: string
+}
+
+export const DEFAULT_K = 3
+
+// 取同功能標籤的對白 k 句。依 id 排序後從 turn 起輪流取，讓每一輪撈到的句子不同，
+// 同一輪結果固定（便於測試）。沒有同標籤的對白就回傳空陣列，不退而求其次。
+export function retrieve(
+  index: readonly IndexEntry[],
+  fn: FunctionTag,
+  { k = DEFAULT_K, mode = null, turn = 0 }: { k?: number; mode?: string | null; turn?: number } = {},
+): IndexEntry[] {
+  const pool = index
+    .filter(e => e.function === fn && (mode === null || e.mode === mode))
+    .sort((a, b) => (a.id < b.id ? -1 : a.id > b.id ? 1 : 0))
+  if (pool.length === 0) return []
+  const n = Math.min(k, pool.length)
+  const start = ((turn % pool.length) + pool.length) % pool.length
+  return Array.from({ length: n }, (_, i) => pool[(start + i) % pool.length]!)
+}
@@ -0,0 +1,26 @@
+// 溝通功能標籤白名單，與建置流水線的 tools/corpus/tags.json 的 functions 一致。
+export const FUNCTIONS = {
+  greet: '打招呼、道別',
+  refuse: '拒絕、否定請求',
+  comfort: '安慰、安撫',
+  tease: '吐槽、調侃、挖苦',
+  celebrate: '慶祝、開心、稱讚',
+  report_fail: '報告失敗、壞消息、承認錯誤',
+  confirm_request: '請求確認、確認對方意願或事實',
+  shy_cover: '害羞掩飾、嘴硬、岔開話題',
+  explain: '說明、陳述事實或計畫',
+  ask: '發問',
+  agree: '同意、答應',
+  apologize: '道歉',
+  thank: '道謝',
+  encourage: '鼓勵、打氣',
+  worry: '擔心、關心',
+  urge: '催促、命令、警告',
+  other: '以上皆非',
+} as const
+
+export type FunctionTag = keyof typeof FUNCTIONS
+
+export function isFunctionTag(value: unknown): value is FunctionTag {
+  return typeof value === 'string' && Object.hasOwn(FUNCTIONS, value)
+}