Fix empty-content crash with reasoning models (GLM-5.2 default)

Reasoning models spend tokens thinking before writing the visible answer,
and those tokens count against max_tokens. The seed-keyword step's
max_tokens=400 left GLM-5.2 (the default model) no room to answer: the
API returned content=None with finish_reason=length, and re.search(None)
crashed every default-settings plan generation with 'Unexpected error:
expected string or bytes-like object'.

- app.py: raise the seed-keyword budget 400 -> 2000 (verified sufficient
against the live API with the real prompt).
- llm.py chat(): retry once with +3000 headroom when content is empty at
finish_reason=length, and raise a clear LLMError instead of returning
None if content is still empty - covers every call site centrally.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

app.py

@@ -81,7 +81,10 @@ Product/service: {product_description}
         hf_token=hf_token,
         model=model,
         messages=[{"role": "user", "content": prompt}],
-        max_tokens=400,
+        # The answer itself is ~100 tokens, but reasoning models (the default
+        # GLM-5.2 included) think before answering and that counts against
+        # max_tokens — 400 left no room and made content come back empty.
+        max_tokens=2000,
         temperature=0.3,
     )
     match = re.search(r"\[.*\]", raw, re.DOTALL)

modules/llm.py

@@ -51,13 +51,25 @@ def chat(
         raise LLMError("Please enter your Hugging Face access token.")
 
     client = InferenceClient(api_key=hf_token.strip())
-    try:
-        response = client.chat_completion(
+
+    def _call(budget: int):
+        return client.chat_completion(
             model=model,
             messages=messages,
-            max_tokens=max_tokens,
+            max_tokens=budget,
             temperature=temperature,
         )
+
+    try:
+        response = _call(max_tokens)
+        choice = response.choices[0]
+        # Reasoning models (GLM-5.2, DeepSeek-V3.2, ...) spend tokens thinking
+        # before emitting the visible answer, and those tokens count against
+        # max_tokens — a budget that fits the answer can still come back with
+        # content=None and finish_reason="length". Retry once with headroom.
+        if not choice.message.content and choice.finish_reason == "length":
+            response = _call(max_tokens + 3000)
+            choice = response.choices[0]
     except HfHubHTTPError as exc:
         status = getattr(exc.response, "status_code", None)
         if status == 401:
@@ -77,4 +89,15 @@ def chat(
     except Exception as exc:  # network errors, timeouts, etc.
         raise LLMError(f"LLM request failed: {exc}") from exc
 
-    return response.choices[0].message.content
+    content = choice.message.content
+    if not content:
+        detail = (
+            " (it ran out of tokens while reasoning)"
+            if choice.finish_reason == "length"
+            else ""
+        )
+        raise LLMError(
+            f"The model ({model}) returned an empty response{detail}. "
+            "Try again or pick a different model."
+        )
+    return content