prose-guard

1import type { Register } from 'claude-code'
2
3const DASH = String.fromCharCode(0x2014)
4const countDashes = (s: unknown) => (typeof s === 'string' ? s.split(DASH).length - 1 : 0)
5const SPACED_DASH = new RegExp(`[ \\t]*${DASH}[ \\t]*`, 'g')
6const TRAILING_DASH = new RegExp(`[ \\t${DASH}]+$`)
7
8const DASH_DENY =
9  'This adds an em dash. Rewrite it with periods or commas. ' +
10  "To match the character in a shell command, write it as an escape such as $'\\xe2\\x80\\x94'."
11
12const REPLY_BLOCK =
13  'Your reply contained an em dash. Send the reply again, rewritten with periods or commas and no em dashes.'
14
15// British spelling -> American spelling, one entry per inflected form.
16const BRITISH = new Map<string, string>()
17const add = (uk: string, us: string) => BRITISH.set(uk, us)
18
19for (const w of ['colour', 'behaviour', 'favour', 'honour', 'labour', 'neighbour', 'flavour', 'humour', 'rumour', 'harbour'])
20  for (const end of ['', 's', 'ed', 'ing', 'er', 'ers', 'able', 'al', 'ful', 'hood', 'ite', 'ites'])
21    add(w + end, w.replace('our', 'or') + end)
22
23for (const w of ['centre', 'theatre', 'metre', 'litre', 'fibre'])
24  for (const [uk, us] of [['re', 'er'], ['res', 'ers'], ['red', 'ered'], ['ring', 'ering']])
25    add(w.slice(0, -2) + uk, w.slice(0, -2) + us)
26
27for (const stem of [
28  'organis', 'recognis', 'realis', 'apologis', 'prioritis', 'summaris', 'optimis', 'initialis', 'serialis',
29  'normalis', 'customis', 'minimis', 'maximis', 'categoris', 'finalis', 'utilis', 'authoris', 'emphasis',
30  'standardis', 'synchronis', 'visualis',
31])
32  for (const end of ['e', 'es', 'ed', 'ing', 'er', 'ers', 'able', 'ation', 'ations'])
33    add(stem + end, stem.slice(0, -1) + 'z' + end)
34
35// "analyses" is left out: it is also the plural of "analysis".
36for (const end of ['e', 'ed', 'ing', 'er', 'ers']) add('analys' + end, 'analyz' + end)
37
38for (const stem of ['cancel', 'travel', 'model', 'label', 'level', 'signal', 'fuel'])
39  for (const end of ['ed', 'ing', 'er', 'ers']) add(stem + 'l' + end, stem + end)
40
41for (const w of ['licence', 'defence', 'offence', 'pretence']) {
42  add(w, w.slice(0, -2) + 'se')
43  add(w + 's', w.slice(0, -2) + 'ses')
44}
45
46for (const [uk, us] of [
47  ['grey', 'gray'], ['greys', 'grays'], ['greyed', 'grayed'], ['greying', 'graying'], ['greyer', 'grayer'], ['greyish', 'grayish'],
48  ['catalogue', 'catalog'], ['catalogues', 'catalogs'], ['catalogued', 'cataloged'], ['cataloguing', 'cataloging'],
49  ['programme', 'program'], ['programmes', 'programs'],
50  ['aluminium', 'aluminum'], ['whilst', 'while'], ['ageing', 'aging'],
51  ['artefact', 'artifact'], ['artefacts', 'artifacts'],
52  ['cheque', 'check'], ['cheques', 'checks'],
53  ['tyre', 'tire'], ['tyres', 'tires'],
54  ['manoeuvre', 'maneuver'], ['manoeuvres', 'maneuvers'], ['manoeuvred', 'maneuvered'], ['manoeuvring', 'maneuvering'],
55  ['sceptic', 'skeptic'], ['sceptics', 'skeptics'], ['sceptical', 'skeptical'], ['scepticism', 'skepticism'],
56] as [string, string][])
57  add(uk, us)
58
59const britishCounts = (s: unknown) => {
60  const counts = new Map<string, number>()
61  if (typeof s !== 'string') return counts
62  for (const word of s.toLowerCase().match(/[a-z]+/g) ?? [])
63    if (BRITISH.has(word)) counts.set(word, (counts.get(word) ?? 0) + 1)
64  return counts
65}
66
67// British words whose count rises from the old texts to the new ones.
68const addedBritish = (pairs: [unknown, unknown][]) => {
69  const delta = new Map<string, number>()
70  for (const [oldText, newText] of pairs) {
71    for (const [w, n] of britishCounts(newText)) delta.set(w, (delta.get(w) ?? 0) + n)
72    for (const [w, n] of britishCounts(oldText)) delta.set(w, (delta.get(w) ?? 0) - n)
73  }
74  return [...delta].filter(([, n]) => n > 0).map(([w]) => w)
75}
76
77const britishDeny = (words: string[]) =>
78  'This adds British spelling. Use American spelling: ' +
79  words.map(w => `${w} -> ${BRITISH.get(w)}`).join(', ') + '.'
80
81const PROSE_FILE = /\.(md|mdx|markdown|txt|rst)$/i
82const COMMIT_OR_PR = /\bgit\s+(?:-[Cc]\s+\S+\s+)*commit\b|\bgh\s+pr\s+(?:create|edit)\b/
83const PR_TOOLS = ['mcp__github__create_pull_request', 'mcp__github__update_pull_request']
84
85export const register: Register = on => {
86  on('tool.call', async ($, e, next) => {
87    // MultiEdit and the GitHub tools may not be in this build's typed tool list.
88    const t = e as unknown as Record<string, any>
89    let dashes = 0
90    let british: string[] = []
91    switch (t.tool) {
92      case 'Write': {
93        const before = await $.fs.read(t.file_path).catch(() => '')
94        dashes = countDashes(t.content) - countDashes(before)
95        if (PROSE_FILE.test(t.file_path)) british = addedBritish([[before, t.content]])
96        break
97      }
98      case 'Edit':
99        dashes = countDashes(t.new_string) - countDashes(t.old_string)
100        if (PROSE_FILE.test(t.file_path)) british = addedBritish([[t.old_string, t.new_string]])
101        break
102      case 'MultiEdit': {
103        const edits: { old_string?: string; new_string?: string }[] = t.edits ?? []
104        dashes = edits.reduce((n, x) => n + countDashes(x.new_string) - countDashes(x.old_string), 0)
105        if (PROSE_FILE.test(t.file_path)) british = addedBritish(edits.map(x => [x.old_string, x.new_string]))
106        break
107      }
108      case 'NotebookEdit':
109        dashes = countDashes(t.new_source)
110        break
111      case 'Bash':
112        dashes = countDashes(t.command)
113        if (COMMIT_OR_PR.test(t.command ?? '')) british = addedBritish([['', t.command]])
114        break
115      default:
116        if (PR_TOOLS.includes(t.tool)) {
117          dashes = countDashes(t.title) + countDashes(t.body)
118          british = addedBritish([['', t.title], ['', t.body]])
119        }
120    }
121    const reasons = []
122    if (dashes > 0) reasons.push(DASH_DENY)
123    if (british.length > 0) reasons.push(britishDeny(british))
124    return reasons.length > 0 ? { deny: reasons.join(' ') } : next(e)
125  })
126
127  // Each em dash in a reply becomes a comma as the reply streams, so none is
128  // drawn or recorded. Thinking gets the same rewrite, since the app shows it
129  // as progress notes; the engine still records the signed original. A
130  // piece's trailing spaces and dashes wait for the next piece of the same
131  // block, since the rest of the dash may arrive there.
132  on('turn.step', async function* ($, e, next) {
133    let held = ''
134    let heldIndex = -1
135    for await (const c of next(e)) {
136      if (c.kind !== 'text' && c.kind !== 'thinking') {
137        yield c
138        continue
139      }
140      const joined = (c.index === heldIndex ? held : '') + c.text
141      held = joined.match(TRAILING_DASH)?.[0] ?? ''
142      heldIndex = c.index
143      const text = joined.slice(0, joined.length - held.length)
144      yield { ...c, text: text.replace(SPACED_DASH, ', ') }
145    }
146  })
147
148  // Stop is the one event that can send a finished reply back to the model.
149  // The stream above leaves no dash for it to find; it stays as a backstop.
150  on('classic.Stop', async ($, e, next) => {
151    const result = await next(e)
152    // One rewrite per reply, so a stubborn dash cannot loop forever.
153    if (e.stop_hook_active || result.block !== undefined) return result
154    if (countDashes(await turnText($, e.last_assistant_message)) === 0) return result
155    return { ...result, block: REPLY_BLOCK }
156  })
157}
158
159// Text of every assistant message since the last real user message, plus the
160// final reply, which the transcript may not hold yet.
161async function turnText($: { session: { messages: () => Promise<unknown> } }, last: string | undefined) {
162  const texts = [last ?? '']
163  const messages = (await $.session.messages()) as { role: string; text: string; toolResults?: unknown[] }[]
164  for (let i = messages.length - 1; i >= 0; i--) {
165    const m = messages[i]!
166    if (m.role === 'user') {
167      if (!m.toolResults?.length) break
168    } else texts.push(m.text)
169  }
170  return texts.join('\n')
171}
172