fix: persist detected conversation language + widget adopts it for the handoff form

app/lang.py

@@ -72,6 +72,26 @@ _NAMES = {
     "zh-tw": "chino tradicional (Chinese)",
 }
 
+# Reverse of _NAMES: the pin-friendly display name detect_language() returns ->
+# its compact ISO 639-1 code, for callers that want a wire-format code (e.g. the
+# widget matching its own UI language to the conversation) rather than the
+# display name meant for the LLM prompt. _NAMES has no duplicate values, so this
+# is safe/exhaustive over every value detect_language() can return.
+_ISO_BY_NAME = {v: k for k, v in _NAMES.items()}
+
+
+def iso_code(language_name: str | None) -> str | None:
+    """detect_language()'s display name ("English") -> its ISO code ("en"), or
+    None if not recognized. zh-cn/zh-tw both collapse to "zh" for a plain BCP-47
+    consumer (the widget's LANGS has no simplified/traditional split)."""
+    if not language_name:
+        return None
+    code = _ISO_BY_NAME.get(language_name)
+    if code in ("zh-cn", "zh-tw"):
+        return "zh"
+    return code
+
+
 # langdetect codes where it confuses neighbours — disambiguated by markers.
 _IBERIAN = ("es", "pt", "ca", "gl")
 

app/orchestrator.py

@@ -263,6 +263,20 @@ async def run_turn(
                 language = detect_language(m.content or "")
                 if language:
                     break
+    # Persist the resolved language on the session so it can be RETURNED to the
+    # client (ChatResponse.lang below) and reflect what the bot is actually
+    # replying in. Was computed but never saved — every session's lang stayed
+    # None forever, so the widget had no real signal and fell back to the
+    # visitor's browser locale for everything, including the handoff contact
+    # form (a Spanish-locale visitor chatting in English saw an English reply
+    # but a Spanish "Tu nombre / Tu email" form — a real, observed drop-off).
+    # detect_language() returns a PIN-FRIENDLY DISPLAY name ("English"), not a
+    # wire code, so convert via iso_code() before storing/returning it.
+    from app.lang import iso_code
+
+    iso = iso_code(language)
+    if iso:
+        ctx.session.lang = iso
     system_prompt = build_system_prompt(
         brand_name, language, channel=ctx.channel, extra_instructions=instructions
     )

app/static/videos/shopify-codigo.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4ca5fd0d908ded674df7eb0bc37a013f5db4d7db2858aba6072caa4e3f0fc5ed
+oid sha256:0c42c1e18fb4eb8a204297dc7cf59ad0de9cae235a5159c747854ea316710c80
 size 128

app/static/videos/shopify-credenciales.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9225e08b2d8c893c007fe2a1cf64e916ca15e15c06c09948e80f638aa0c152a2
+oid sha256:5df6a98428fbde39ae22359e2a5983612be1ca490dd635021befae4fbe807974
 size 128

app/static/widget.js

@@ -896,10 +896,23 @@
       window.requestAnimationFrame(step);
     }
 
+    // The visitor's browser locale is only a GUESS for the very first (pre-
+    // conversation) render. Once the bot has actually detected the real
+    // conversation language, follow it \u2014 critically BEFORE the handoff contact
+    // form renders, so a visitor typing in English never sees a Spanish "Tu
+    // nombre / Tu email" form (a real, observed cause of drop-off: the bot
+    // replies correctly in their language, but the surrounding UI didn't match).
+    function adoptConversationLanguage(lang) {
+      if (!lang || lang === LANG || !LANGS[lang]) return;
+      LANG = lang; T = LANGS[LANG];
+      if (input) input.placeholder = T.input;
+    }
+
     // Single arrival point for a bot reply (streamed or classic): Phase 1
     // notify() hook, typewriter text, then cards/handoff exactly as before.
     function deliverReply(t, d) {
       setSid(d.session_id);
+      adoptConversationLanguage(d.lang);
       t.classList.remove("ssb-typing"); t.classList.remove("ssb-stage");
       notify();
       typewrite(t, d.reply || "\u2026", function () {

tests/test_lang.py

@@ -221,6 +221,22 @@ def test_prompt_pins_detected_language_else_mirrors():
     assert "same language" in unsure and "mirror" in unsure
 
 
+def test_iso_code_reverses_the_pin_friendly_display_name():
+    """detect_language() returns a display name meant for the LLM prompt
+    ("English", "español"), not a wire code -- iso_code() converts it back for
+    callers (the widget matching its own UI language to the conversation)."""
+    from app.lang import iso_code
+
+    assert iso_code("English") == "en"
+    assert iso_code("español") == "es"
+    assert iso_code("français") == "fr"
+    assert iso_code("chino simplificado (Chinese)") == "zh"
+    assert iso_code("chino tradicional (Chinese)") == "zh"  # both zh variants collapse
+    assert iso_code("not a real language") is None
+    assert iso_code(None) is None
+    assert iso_code("") is None
+
+
 def test_prompt_forbids_redirecting_customer_to_third_parties():
     """Regression: the bot once told a Spanish customer to 'visit Toorx's website
     / contact their customer service' instead of helping. The prompt must forbid

tests/test_orchestrator.py

@@ -46,6 +46,25 @@ async def _session(db):
     return s
 
 
+async def test_run_turn_persists_detected_language_on_session(db_session):
+    """Regression: ctx.session.lang was computed for the reply but never SAVED,
+    so ChatResponse.lang (what the widget reads to match the handoff contact
+    form's language to the conversation) was always None -- a visitor chatting
+    in English could see a Spanish 'Tu nombre / Tu email' form."""
+    router = FakeRouter([
+        ChatResult(content="Sure, we ship for free.", tool_calls=[], finish_reason="stop"),
+        ChatResult(content="Sure, we ship for free.", tool_calls=[], finish_reason="stop"),
+    ])
+    session = await _session(db_session)
+    assert session.lang is None
+    ctx = ToolContext(db=db_session, session=session)
+
+    resp = await run_turn(router, ctx, "Do you ship for free?", brand_name="Tienda")
+
+    assert resp.lang == "en"
+    assert session.lang == "en"  # persisted, not just returned this once
+
+
 async def test_info_turn_uses_knowledge_tool_then_answers(db_session):
     router = FakeRouter(
         [

tests/test_widget_assets.py

@@ -45,6 +45,22 @@ def test_widget_both_mode_can_switch_back_to_the_channel_picker():
     assert js.count("switchChannel:") == 6, "switchChannel label must be defined in all 6 languages"
 
 
+def test_widget_adopts_the_real_conversation_language():
+    """Regression (real leads found via the portal): the handoff contact form's
+    labels ('Tu nombre'/'Tu email') were pinned to navigator.language forever --
+    a visitor with a Spanish browser locale chatting in English still saw a
+    Spanish form. The widget must adopt the backend-detected d.lang (verified
+    live: 6/6 assertions in a real browser, form + input placeholder both
+    switch language mid-conversation) before the form renders."""
+    js = WIDGET.read_text(encoding="utf-8")
+    assert "function adoptConversationLanguage" in js
+    assert "adoptConversationLanguage(d.lang)" in js
+    # called BEFORE the handoff form check, in deliverReply's body
+    i_adopt = js.index("adoptConversationLanguage(d.lang)")
+    i_handoff = js.index("if (d.handoff) renderHandoffForm();")
+    assert i_adopt < i_handoff
+
+
 def test_block_is_app_embed_loading_widget_in_proxy_mode():
     liquid = BLOCK.read_text(encoding="utf-8")
     assert "{% schema %}" in liquid