Source practice

This section turns the preceding knowledge into observable, verifiable tasks. Every exercise states its objective, required source reading, expected observations, completion criteria, verification method, and review questions.

Prerequisites

• Complete at least the “One Turn” learning path.
• Set up an isolated configuration by following Your first request with DeepSeek.
• Know how to set a Python function breakpoint in your IDE.
• Use a separate workspace and test Sessions for every experiment; never use important chat history.

Practice levels

Difficulty	Exercise	Modify nanobot source?
1	Trace a Tool-free Turn	No
1	Trace a Tool Call and Session	No
2	Reproduce streaming, cancellation, and retry	No
2	Write a Workspace Skill	No source change; writes to the lab workspace
3	Build a read-only Tool plugin	No core change; installs an isolated plugin
4	Capstone extension design review	Design first; implementation is optional

The learning loop

flowchart TD
    A["Write down a prediction"] --> B["Run or debug"]
    B --> C["Record real objects and state"]
    C --> D["Return to the source to explain differences"]
    D --> E["Reproduce with a test"]
    E --> F["Write a short retrospective"]
    F --> A

Prediction comes first. Single-stepping through thousands of lines without one often leaves only the impression that “I saw the code,” not a reusable mental model.

Shared experiment-note template

Create a note section like this for each exercise:

## My prediction
 
## Actual call chain
 
## Key object snapshots
 
## Differences from my prediction
 
## Source evidence
 
## Test evidence
 
## Still uncertain

“Still uncertain” is not a failure. It separates facts established by the code from questions that require another experiment.

Safety rules

• Do not screenshot or copy API keys from a debugger’s variable view.
• Do not use an important real workspace for write or cancellation experiments.
• Do not test Tools that can cause external side effects such as payment, email, deletion, or publishing.
• Put modification exercises in an isolated plugin or example directory.
• Change only one variable at a time.
• A smoke test against a real model does not replace deterministic tests.

Next: Trace a Tool-free Turn