fix(lang): pin the detected customer language explicitly (Spanish-heavy context was overriding plain mirror); None->mirror fallback

app/prompts.py

@@ -101,23 +101,31 @@ def build_system_prompt(
     channel: str = "web",
     extra_instructions: str = "",
 ) -> str:
-    # LANGUAGE = mirror the customer, the way ChatGPT/Claude do it. We do NOT pin a
-    # heuristically-detected language and command the model to use it: a wrong guess
-    # (e.g. all-caps Spanish misread as Hungarian) would FORCE the wrong language.
-    # Modern LLMs are natively multilingual and reply in the language they're
-    # addressed in — so we just instruct them to do exactly that and trust them.
-    # (A separate, reliable script-based guard repairs only NON-Latin drift on weak
-    # providers; see app.lang.needs_language_repair.)
-    language_rule = (
-        "Read the customer's most recent message and reply in EXACTLY the same "
-        "language they wrote it in — mirror it (English->English, Spanish->Spanish, "
-        "German->German, Arabic->Arabic, and so on). Detect the language yourself "
-        "from what they actually wrote; do not assume. Never switch languages unless "
-        "the customer does. Even if the store's catalog, product sheets or tool "
-        "results are in another language, translate that information into the "
-        "customer's language. This rule is absolute."
-    )
-    lang_short = "the SAME language the customer wrote in"
+    # LANGUAGE: reply in the customer's language. The store's catalog, tool results,
+    # brand and even these instructions are heavily Spanish, which DRAGS the model
+    # into Spanish even for an English customer (observed live) — so a plain "mirror"
+    # hint is too weak. We pin the customer's language EXPLICITLY when we detected it
+    # (detection is upstream and now reliable: English->English, Spanish->Spanish,
+    # and the old all-caps-Spanish->Hungarian misfire is fixed at the detector). When
+    # detection is unsure (None) we fall back to "mirror" and trust the model. The
+    # non-Latin script guard (app.lang.needs_language_repair) is the last safety net.
+    if language:
+        language_rule = (
+            f"The customer is writing in {language}. Write your ENTIRE reply in "
+            f"{language}. The store's catalog, product sheets, tool results and even "
+            f"these instructions may be in a DIFFERENT language — IGNORE that for your "
+            f"choice of language and TRANSLATE everything into {language}. Never reply "
+            f"in any other language and never mix languages. This rule is absolute."
+        )
+        lang_short = language
+    else:
+        language_rule = (
+            "Reply in EXACTLY the same language the customer wrote their last message "
+            "in — detect it yourself and mirror it. The store's data and these "
+            "instructions may be in another language; translate them into the "
+            "customer's language. Never switch languages unless the customer does."
+        )
+        lang_short = "the customer's language"
     channel_handoff = _WA_HANDOFF if channel == "whatsapp" else _WEB_HANDOFF
     # The merchant's own rules go HIGH in the prompt (right after the language
     # rule, before GROUNDING) and are framed as TOP priority that overrides the

tests/test_lang.py

@@ -208,16 +208,17 @@ def test_prompt_forbids_emojis():
     assert "emoji" in build_system_prompt("X").lower()
 
 
-def test_prompt_mirrors_language_and_never_force_pins():
-    # Root-cause fix: the prompt must tell the model to MIRROR the customer's
-    # language (how capable LLMs natively work), NOT command "write everything in
-    # <detected language>" — a wrong detection (Hungarian) must never force a wrong
-    # reply language. Even when a (wrong) language is passed, it is NOT pinned.
+def test_prompt_pins_detected_language_else_mirrors():
+    # The store context is heavily Spanish and drags the model into Spanish even for
+    # an English customer (observed live), so a detected language is pinned EXPLICITLY;
+    # when detection is unsure (None) we mirror. The Hungarian misfire is fixed at the
+    # DETECTOR (see test_spanish_detected), not by refusing to pin.
     from app.prompts import build_system_prompt
 
-    low = build_system_prompt("Tienda", "húngaro (Hungarian)").lower()
-    assert "same language" in low and "mirror" in low
-    assert "húngaro" not in low  # the (possibly wrong) detected name is never injected/forced
+    pinned = build_system_prompt("Tienda", "English").lower()
+    assert "english" in pinned and "entire reply" in pinned
+    unsure = build_system_prompt("Tienda", None).lower()
+    assert "same language" in unsure and "mirror" in unsure
 
 
 def test_prompt_forbids_redirecting_customer_to_third_parties():

tests/test_prompts.py

@@ -9,11 +9,13 @@ def test_system_prompt_includes_rules_and_brand():
     # language is the prominent RULE #1
     assert "RULE #1" in p
     assert "language" in p.lower()
-    # LANGUAGE = mirror the customer (like ChatGPT). A (possibly wrong) detected
-    # language is NEVER force-pinned into the prompt — that was the Hungarian bug.
+    # A DETECTED language is pinned explicitly (the Spanish-heavy context otherwise
+    # drags the model into Spanish even for an English customer). The Hungarian bug
+    # is fixed at the DETECTOR, not by refusing to pin.
     pinned = build_system_prompt("X", language="portugués de Portugal (pt-PT)")
-    assert "pt-PT" not in pinned
-    assert "same language" in pinned.lower() and "mirror" in pinned.lower()
+    assert "pt-PT" in pinned
+    # when detection is unsure (None) we fall back to mirroring
+    assert "mirror" in build_system_prompt("X", language=None).lower()
     # anti-hallucination grounding
     assert "invent" in p.lower()
     # order verification (email + order number / postal code)