diff --git a/plugins/narrator/README.md b/plugins/narrator/README.md
index 3d6afb8..bfb83e0 100644
--- a/plugins/narrator/README.md
+++ b/plugins/narrator/README.md
@@ -15,7 +15,7 @@ With a language set (see Settings) the switch shows it after its words: `⏵ nar
 
 Where the footer is short the switch gives up `on` or `off`, then the language, then `narrator` (or `narrating`), and stays a `⏵`, `⏸︎`, `⏹︎` or `⏭︎` in its color, so the row never breaks in two. Beside [Sub Meter](../sub-meter), the meter shrinks first.
 
-When a turn ends, Haiku (at low effort, so it is quick) reads your request, the tools Claude used and its final message, and writes one or two short spoken sentences in first person ("I added the narrator and the tests pass"), in the language you wrote in, or the one you set (see Settings). Haiku also names that language, and the mod picks the best voice installed for it; where there is none, the English one. A short answer already says it, so it is read out as it is, with no Haiku call: a path by its file's name (`hooks/register.tsx`: "register") and no commit hashes, in the language its words or your request tell, else the one the last narration was in (on Windows and Linux, Haiku narrates it while neither tells; macOS tells the language from the text). With a language set, a short answer is read as it is only when its words or your request tell it is in that language; any other, Haiku says in it.
+When a turn ends, Haiku (at low effort, so it is quick) reads your request, the tools Claude used and its final message, and writes one or two short spoken sentences in first person ("I added the narrator and the tests pass"), in the language you wrote in, or the one you set (see Settings). Haiku also names that language, and the mod picks the best voice installed for it; where there is none, the English one. When the request's words tell its language, Haiku is told to speak it, and a narration in another language is not said: the answer itself is read instead. A short answer already says it, so it is read out as it is, with no Haiku call: a path by its file's name (`hooks/register.tsx`: "register") and no commit hashes, in the language its words or your request tell, else the one the last narration was in (on Windows and Linux, Haiku narrates it while neither tells; macOS tells the language from the text). With a language set, a short answer is read as it is only when its words or your request tell it is in that language; any other, Haiku says in it.
 
 | System | Voice |
 | --- | --- |
diff --git a/plugins/narrator/hooks/register.test.tsx b/plugins/narrator/hooks/register.test.tsx
index ec133aa..3c09acc 100644
--- a/plugins/narrator/hooks/register.test.tsx
+++ b/plugins/narrator/hooks/register.test.tsx
@@ -442,6 +442,29 @@ test('the language setting shows beside the switch, tells the narrator and reach
   expect((await ui.find({ key: 'narrator' }))?.text).toBe('⏹︎ narrating · ES')
 })
 
+/** A Spanish answer too long to be read as it is. */
+const LARGO = `He añadido el **narrador** en \`plugins/narrator\`. ${'Habla después de cada turno, en tu idioma, con la mejor voz que tiene el sistema. '.repeat(3)}`
+
+test('a request whose words tell its language: the narrator is told to speak it', async ($, on) => {
+  const w = world(on, { store: { enabled: true } })
+  await startSession($)
+  const speech = w.nextSpeech()
+  await turn($, { prompt: 'añade el narrador y que hable en el idioma de la petición', answer: LARGO })
+  await speech
+  expect(w.systems[0]).toContain('Always speak Spanish (es)')
+  expect(w.spoken[0]!.env.NARRATOR_TEXT).toBe('He añadido el narrador y los tests pasan.')
+})
+
+test('a narration in another language than the request is not said: the answer is read instead', async ($, on) => {
+  const w = world(on, { reply: 'es\nI added the narrator and the tests pass, so it is ready for you.', store: { enabled: true } })
+  await startSession($)
+  const speech = w.nextSpeech()
+  await turn($, { prompt: 'añade el narrador y que hable en el idioma de la petición', answer: LARGO })
+  await speech
+  expect(w.spoken[0]!.env.NARRATOR_TEXT).toStartWith('He añadido el narrador en narrator.')
+  expect(w.spoken[0]!.env.NARRATOR_LANG).toBe('es')
+})
+
 test('a region is dropped, and a value that is no known language means automatic', () => {
   expect(languageOf('es-MX')).toBe('es')
   expect(languageOf(' PT ')).toBe('pt')
diff --git a/plugins/narrator/hooks/register.tsx b/plugins/narrator/hooks/register.tsx
index 71e5232..7b2a253 100644
--- a/plugins/narrator/hooks/register.tsx
+++ b/plugins/narrator/hooks/register.tsx
@@ -412,15 +412,24 @@ async function narrate ($: EngineInterface, answer: string) {
       await speak($, narration, direct, directLanguage)
       return
     }
+    // The language to speak: the setting, else the request's when its words tell it. The model is told
+    // it in so many words: left to tell it from a request in one language and tool calls in another, it
+    // sometimes narrates in the wrong one.
+    const asked = settings.language || guessLanguage(request)
     const reply = await $.model.complete(
       // A short summary: no deep thinking, so it is said right after the turn.
-      { model: settings.model, system: NARRATOR(settings.language), prompt: brief(answer), maxTokens: 150, effort: 'low', timeoutMs: 30_000 },
+      { model: settings.model, system: NARRATOR(asked), prompt: brief(answer), maxTokens: 150, effort: 'low', timeoutMs: 30_000 },
       { signal: narration.abort.signal }
     ).catch(() => undefined)
     if (!narration.isActive) return
     const told = reply?.isAnswered ? parseNarration(reply.text) : undefined
-    // No narration written: the answer itself, as far as it reads aloud.
-    const { text, language } = told?.text ? told : { text: cut(direct, 250), language: directLanguage }
+    // A narration in another language than the request's is not said, whatever code the model wrote:
+    // its words tell. The setting is left to the model, as the answer may be in neither.
+    const toldLanguage = told ? guessLanguage(told.text) || told.language : ''
+    const isOtherLanguage = !settings.language && asked !== '' && toldLanguage !== '' && toldLanguage !== asked
+    if (isOtherLanguage) $.ui.log(`narrator: the model narrated in ${toldLanguage}, not ${asked}: reading the answer instead`, { to: 'debug' })
+    // No narration written, or not in the request's language: the answer itself, as far as it reads aloud.
+    const { text, language } = told?.text && !isOtherLanguage ? told : { text: cut(direct, 250), language: directLanguage }
     if (language) spokenLanguage = language
     if (text) await speak($, narration, text, language)
   } catch (error) {