feat: inject today's date into system prompt and add extended thinking support

Fixes age validation errors caused by the LLM not knowing the current date.

Changes:
- prompts.py: inject date.today() at the top of both system prompts so the
  LLM can accurately calculate a child's age from their date of birth
- llm.py: add optional thinking_budget parameter to complete(); when set,
  passes thinking={"type": "enabled", "budget_tokens": N} to litellm and
  raises max_tokens to thinking_budget + 4096 (Anthropic models only)
- config.py: add thinking_budget field, read from THINKING_BUDGET env var
- .env.example: document the THINKING_BUDGET option
- core.py: pass thinking_budget through to llm.complete()
- main.py: pass thinking_budget when constructing EmailAgent
- chat_app.py: switch from stream_complete to asyncio.to_thread(complete)
  so extended thinking works and so only the reply field is shown to
  the parent (not the raw JSON wrapper)

To enable extended thinking set THINKING_BUDGET=8000 in .env.

https://claude.ai/code/session_01SUWzMzFvSfWiHXA2p6rPg9

.env.example

@@ -15,6 +15,12 @@
 #   gemini/gemini-2.0-flash
 AI_MODEL=anthropic/claude-opus-4-6
 
+# Extended thinking — Anthropic models only (leave unset to disable).
+# Enables a reasoning phase before the model's reply, which improves
+# accuracy on age calculations, logic-heavy questions, and edge cases.
+# Recommended value: 8000 (tokens). Must be less than max_tokens.
+# THINKING_BUDGET=8000
+
 # ---------------------------------------------------------------
 # API Keys — set the one matching your chosen model's provider
 # ---------------------------------------------------------------

chat_app.py

@@ -123,17 +123,15 @@ async def on_message(message: cl.Message) -> None:
     # --- Build system prompt ---
     system = build_system_prompt(_kb, state)
 
-    # --- Call LLM natively async to keep the event loop free ---
+    # --- Call LLM natively async (supports extended thinking; no event-loop blocking) ---
-    # litellm.acompletion() is a true coroutine — it does not block the event
-    # loop and does not require asyncio.to_thread (which can drop Chainlit's
-    # contextvars, causing the session to reset mid-conversation).
     try:
-        full_content = await llm.acomplete(_config.ai_model, system, state.messages)
+        full_content = await llm.acomplete(
+            _config.ai_model, system, state.messages, _config.thinking_budget
+        )
     except Exception:
         logger.exception("LLM call failed for session %s", state.conversation_id)
         error_text = fallback_message(state.language)
         await cl.Message(content=error_text).send()
-        # Don't update state — let parent retry
         return
 
     # --- Parse and apply LLM response ---
@@ -151,12 +149,12 @@ async def on_message(message: cl.Message) -> None:
     state.language = language
     state.updated_at = now
 
+    # Send the reply text to the parent (only the human-readable reply, not the JSON wrapper)
+    await cl.Message(content=reply_text).send()
+
     # Append assistant reply to history
     state.messages.append(ChatMessage(role="assistant", content=reply_text))
 
-    # Send the reply text to the user (parsed from the LLM's JSON response)
-    await cl.Message(content=reply_text).send()
-
     # --- Handle registration completion ---
     if is_complete and not state.completed:
         state.completed = True

main.py

@@ -63,7 +63,13 @@ def build_components(config: Config):
         cc_emails=[e.strip() for e in config.admin_email_cc.split(",") if e.strip()],
     )
 
-    agent = EmailAgent(model=config.ai_model, kb=kb, store=store, notifier=notifier)
+    agent = EmailAgent(
+        model=config.ai_model,
+        kb=kb,
+        store=store,
+        notifier=notifier,
+        thinking_budget=config.thinking_budget,
+    )
 
     channel = EmailChannel(
         imap_host=config.imap_host,

src/agent/core.py

@@ -31,8 +31,10 @@ class EmailAgent:
         kb: KnowledgeBase,
         store: ConversationStore,
         notifier: AdminNotifier,
+        thinking_budget: int | None = None,
     ) -> None:
         self._model = model
+        self._thinking_budget = thinking_budget
         self._kb = kb
         self._store = store
         self._notifier = notifier
@@ -98,7 +100,7 @@ class EmailAgent:
         system = build_system_prompt(self._kb, state)
 
         try:
-            content = llm.complete(self._model, system, state.messages)
+            content = llm.complete(self._model, system, state.messages, self._thinking_budget)
             parsed = self._parse_llm_response(content)
         except Exception:
             logger.exception("LLM call failed for %s", state.conversation_id)
@@ -140,7 +142,7 @@ class EmailAgent:
         system = build_system_prompt(self._kb, state)
 
         try:
-            content = llm.complete(self._model, system, state.messages)
+            content = llm.complete(self._model, system, state.messages, self._thinking_budget)
             parsed = self._parse_llm_response(content)
         except Exception:
             logger.exception("LLM call failed (post-completion) for %s", state.conversation_id)

src/agent/prompts.py

@@ -1,6 +1,7 @@
 """Build the system prompt sent to the LLM on every turn."""
 
 import json
+from datetime import date
 
 from ..knowledge_base.loader import KnowledgeBase
 from ..models.conversation import ConversationState
@@ -154,8 +155,12 @@ def _build_registration_prompt(kb: KnowledgeBase, state: ConversationState) -> s
     reg_json = json.dumps(state.registration.to_dict(), ensure_ascii=False, indent=2)
     step_hint = STEP_DESCRIPTIONS.get(state.flow_step, "Continue the conversation.")
 
+    today = date.today().isoformat()
+
     return f"""You are the registration assistant for Spielgruppe Pumuckl, run by Familienverein Fällanden in Fällanden, Switzerland. You help parents register their children for the playgroup and answer questions about it.
 
+**Today's date is {today}.** Use this exact date when calculating a child's age from their date of birth.
+
 {_PERSONALITY}
 
 ## Registration Flow (8 steps)
@@ -200,8 +205,12 @@ def _build_post_completion_prompt(kb: KnowledgeBase, state: ConversationState) -
     reg_json = json.dumps(state.registration.to_dict(), ensure_ascii=False, indent=2)
     child_name = state.registration.child.full_name or "their child"
 
+    today = date.today().isoformat()
+
     return f"""You are the registration assistant for Spielgruppe Pumuckl, run by Familienverein Fällanden in Fällanden, Switzerland.
 
+**Today's date is {today}.**
+
 {_PERSONALITY}
 
 ## Context: Registration Already Complete

src/config.py

@@ -50,6 +50,11 @@ class Config:
     # Polling interval in seconds
     poll_interval: int = 60
 
+    # Extended thinking — Anthropic models only.
+    # When set, enables the thinking phase before the LLM replies.
+    # Recommended value: 8000 (tokens). Set to None/unset to disable.
+    thinking_budget: int | None = None
+
     @classmethod
     def from_env(cls) -> "Config":
         return cls(
@@ -74,4 +79,9 @@ class Config:
                 )
             ),
             poll_interval=int(os.getenv("POLL_INTERVAL", "60")),
+            thinking_budget=(
+                int(os.getenv("THINKING_BUDGET"))
+                if os.getenv("THINKING_BUDGET")
+                else None
+            ),
         )

src/llm.py

@@ -5,7 +5,12 @@ from collections.abc import Generator
 import litellm
 
 
-async def acomplete(model: str, system: str, messages: list) -> str:
+async def acomplete(
+    model: str,
+    system: str,
+    messages: list,
+    thinking_budget: int | None = None,
+) -> str:
     """Call any LLM asynchronously and return the response text.
 
     This is the async equivalent of ``complete()`` — use this from async
@@ -16,17 +21,30 @@ async def acomplete(model: str, system: str, messages: list) -> str:
         model: litellm model string, e.g. "anthropic/claude-opus-4-6".
         system: System prompt text.
         messages: List of objects with .role and .content attributes.
+        thinking_budget: When set, enables extended thinking (Anthropic models
+            only). See ``complete()`` for details.
 
     Returns:
         The model's reply as a plain string.
     """
     api_messages = [{"role": "system", "content": system}]
     api_messages += [{"role": m.role, "content": m.content} for m in messages]
-    response = await litellm.acompletion(model=model, messages=api_messages, max_tokens=2048)
+
+    kwargs: dict = {"model": model, "messages": api_messages, "max_tokens": 2048}
+    if thinking_budget is not None:
+        kwargs["thinking"] = {"type": "enabled", "budget_tokens": thinking_budget}
+        kwargs["max_tokens"] = thinking_budget + 4096
+
+    response = await litellm.acompletion(**kwargs)
     return response.choices[0].message.content
 
 
-def complete(model: str, system: str, messages: list) -> str:
+def complete(
+    model: str,
+    system: str,
+    messages: list,
+    thinking_budget: int | None = None,
+) -> str:
     """Call any LLM and return the response text.
 
     Args:
@@ -35,13 +53,24 @@ def complete(model: str, system: str, messages: list) -> str:
                environment variable (ANTHROPIC_API_KEY, OPENAI_API_KEY, …).
         system: System prompt text.
         messages: List of objects with .role and .content attributes.
+        thinking_budget: When set, enables extended thinking (Anthropic models
+            only). The value is the token budget for the thinking phase; the
+            final ``max_tokens`` is set to ``thinking_budget + 4096`` so the
+            model has enough room to both think and reply.
 
     Returns:
         The model's reply as a plain string.
     """
     api_messages = [{"role": "system", "content": system}]
     api_messages += [{"role": m.role, "content": m.content} for m in messages]
-    response = litellm.completion(model=model, messages=api_messages, max_tokens=2048)
+
+    kwargs: dict = {"model": model, "messages": api_messages, "max_tokens": 2048}
+    if thinking_budget is not None:
+        kwargs["thinking"] = {"type": "enabled", "budget_tokens": thinking_budget}
+        # max_tokens must exceed budget_tokens or the API returns an error.
+        kwargs["max_tokens"] = thinking_budget + 4096
+
+    response = litellm.completion(**kwargs)
     return response.choices[0].message.content