Checkpoint: CLI Assistant

What this is

No new concepts. This is a checkpoint: a small, real CLI assistant built entirely out of ideas from Lessons 13 through 21, combined into one thing, still without create_agent (that's Lesson 23, the start of Advanced). If you can read lesson.py and understand why every piece is there, you've mastered the Intermediate tier.

What it does

A command-line chat loop with two tools (a calculator and a word counter), manual conversation memory, graceful handling of failing tool calls, automatic retries on transient API errors, and an on-demand structured summary of the conversation so far.

Where each piece came from

@tool
def calculator(...): ...
@tool
def word_counter(...): ...
TOOLS_BY_NAME = {"calculator": calculator, "word_counter": word_counter}

Lessons 13-15: two distinct tools, looked up by name once the model picks one.

model_with_tools = model.bind_tools([calculator, word_counter])

Lesson 14/15: both tools bound to the model at once.

try:
    result = chosen_tool.invoke(call["args"])
except Exception as error:
    result = f"Error: {error}"

Lesson 16: a failing tool call (say, asking the calculator to divide by zero) gets reported back to the model as a ToolMessage, instead of crashing the whole CLI session.

history: list = []
...
history.append(HumanMessage(user_input))
...
history.append(ai_message)

Lesson 17: history is the conversation's actual memory, resent in full to the model on every turn, exactly as introduced back then.

class ConversationSummary(BaseModel):
    topics_discussed: list[str] = ...
    overall_tone: str = ...

structured_model = model.with_structured_output(ConversationSummary)

Lesson 18. Typing "summary" triggers a separate call asking for a real, typed ConversationSummary object, not free text to parse.

def invoke_with_retry(runnable, payload, max_attempts: int = 3):
    ...

Lesson 21's manual retry-with-backoff pattern, wrapped around every real API call in this script (both the tool-using calls and the summary call).

Why tool_calls_made is computed in Python, not asked of the model

tool_calls_made = sum(len(m.tool_calls) for m in history if isinstance(m, AIMessage))

Notice this count is computed directly from history, in plain Python, rather than being one of the fields we ask ConversationSummary to fill in. The model has no reliable way to count exactly how many tool calls happened purely by reading a text transcript, and Lesson 18 already established the difference between things a model can reasonably judge (like the overall tone of a conversation) versus things that need to be exactly correct (a count). Counting is something code should do; judging tone and topic is something worth asking a model for.