traceback-coach 0.3.0

A visual error-literacy coach for Jupyter: explains tracebacks with diagrams and worked examples so students learn to read errors themselves

traceback-coach

A visual error-literacy coach for Jupyter notebooks. When a cell raises, it turns the traceback into a teaching moment: a plain-language translation, a causal diagram of why it broke, a worked example, and one guiding question — so students learn to read errors themselves and not repeat them.

It never shows the fix. The goal is independence: read the error, understand the cause, fix it yourself.

"An expert is a person who has made all the mistakes that can be made in a very narrow field." — Niels Bohr

Works fully offline — no account, no network required. An LLM is optional and only personalizes the guiding question.

Features

Feature	What it does
%%coach cell magic	Explain the error in a single cell
%coach_watch on	Auto-explain every failing cell the moment it breaks
Anatomy card	Plain translation + causal diagram + worked example + one question
Causal diagram	Mermaid flowchart of the call chain — recursion drawn as a loop
Never the fix	Only a guiding question; you fix it yourself
Progressive fade	Full card → brief → minimal nudge as you repeat an error type
Quiz / active recall	Guess the error type before the analysis is revealed
Bilingual	The whole card in English or Traditional Chinese (zh-HK)
Collapsible traceback	Fold the giant red Python traceback into a click-to-expand line
Error stats	See which error types you hit most this session
Lessons on demand	Open a lesson for any error family — no error needed
Offline diagrams	Mermaid is vendored — renders with no network

Install

pip install traceback-coach

Then in a notebook:

%load_ext traceback_coach

Quick start

%%coach
print(total)        # `total` was never defined

Instead of a wall of red, the Coach shows an anatomy card — something like:

🔴 What happened — the name total is used before it's defined; Python has no value bound to it yet (NameError: name 'total' is not defined). 📊 Why it breaks — a diagram tracing the cell down to the failing line. 🏷️ NameError · 🔎 a NameError means a name is used before it exists — check for a typo, or whether you assigned it earlier. ❓ Question: Where should total receive its first value, before this line?

…and never the corrected code. Or watch every cell automatically:

%coach_watch on
xs = [1, 2, 3]
xs[5]               # IndexError is explained the moment it happens

How it works

┌───────────────┐   ┌───────────────┐   ┌──────────────────────────┐   ┌───────────────┐
│ Cell raises   │ → │ Parse the     │ → │ Build the card:          │ → │ Render inline │
│ (type, value, │   │ traceback     │   │  · plain translation     │   │ in the        │
│  frames)      │   │ (drop noise   │   │  · causal diagram        │   │ notebook —    │
│               │   │  frames)      │   │  · worked example        │   │ never a fix   │
│               │   │               │   │  · ONE guiding question  │   │               │
└───────────────┘   └───────────────┘   └──────────────────────────┘   └───────────────┘

The card is built to teach, not to solve. On every error the engine:

1. Translates the traceback into one plain sentence.
2. Draws where it broke — recursion becomes a visible loop, not a 1000-line wall.
3. Shows the same error on minimal code, plus how to avoid it next time.
4. Asks exactly one guiding question — never the corrected code.

Commands

Command	What it does
%load_ext traceback_coach	load the extension
%%coach	run a cell; if it raises, show the anatomy card
%coach_watch on / off	auto-explain every failing cell
%coach_explain	re-explain the most recent error
%coach_lesson <Type>	open a lesson for an error family (no error needed), e.g. %coach_lesson IndexError
%coach_level full|brief|min|auto	detail level; auto fades as you repeat an error (full → brief → minimal)
%coach_stats	a session map of which error types you hit most
%coach_quiz on / off	active recall — guess the error type before revealing the analysis
%coach_lang en|zh	explanation language: English or Traditional Chinese (zh-HK)
%coach_llm / on / off	LLM status / toggle personalized vs template questions
%coach_compact on / off	fold the red Python traceback into a click-to-expand line
%coach_help, %coach_off	help / stop watching

The anatomy card

On any error the Coach shows:

• 🔴 What happened — the traceback in one plain sentence
• 📊 Why it breaks — a Mermaid flowchart of the call chain with the code at each step, down to the line that broke. Recursion is drawn as a loop — a self-loop labelled calls itself ×N for direct recursion, or a cycle that loops back ×N for mutual recursion. Falls back to a pure HTML/CSS diagram if Mermaid can't render.
• 📍 Where — the line + offending token
• 🏷️ Error family + 🔎 how to read this kind of error yourself
• 📖 Worked example (collapsible) — the same error on minimal code + how to avoid it next time
• ❓ One guiding question — never the fix

Covers 12 beginner error families: NameError, TypeError, ValueError, IndexError, KeyError, AttributeError, IndentationError, SyntaxError, ZeroDivisionError, ModuleNotFoundError/ImportError, RecursionError, UnboundLocalError.

Feature highlights

• Progressive fade (%coach_level auto, default) — the 1st time you hit an error type you get the full card; the 2nd a brief one; the 3rd+ a minimal nudge. Force a level with %coach_level full|brief|min.
• Error stats (%coach_stats) — see your most common mistakes this session.
• Quiz / active recall (%coach_quiz on) — predict the error type before revealing the analysis; recalling first aids retention.
• Bilingual (%coach_lang zh) — the whole card (translation, family, “read it yourself”, worked example, question, UI labels) in Traditional Chinese for Hong Kong learners. English is the default and unchanged.
• Collapsible traceback (%coach_compact on) — fold the giant red Python traceback into one click-to-expand line; the Coach card stays untouched.
• Personalized questions — with an LLM configured, the guiding question references the student's actual variables (and is asked in the chosen language); without one, a built-in template is used. Either way, never a fix.

Optional: personalized questions via an LLM

Set any OpenAI-compatible endpoint. With nothing set, everything still works on templates.

Variable	Purpose
TRACEBACK_COACH_LLM_API_KEY / OPENAI_API_KEY / DEEPSEEK_API_KEY	API key (any one)
OPENAI_BASE_URL / TRACEBACK_COACH_LLM_BASE_URL	endpoint (any OpenAI-compatible gateway)
OPENAI_MODEL / TRACEBACK_COACH_LLM_MODEL	model id (default deepseek-chat)

Check it live with %coach_llm. The model is told to never output a fix.

Notebooks

The package ships demo + lesson notebooks — ready to run, start with 00_start_here.ipynb:

• 00_start_here.ipynb — run it yourself (start here, hands-on): the whole package end-to-end — collapsible traceback (fold the red wall), recursion-as-a-loop, progressive fade, stats, quiz, bilingual EN/中文, LLM, lessons. Restart Kernel & Run All Cells.
• 01_showcase.ipynb — the full feature tour: anatomy card, causal diagram, recursion-as-a-loop, progressive fade, stats, quiz, bilingual EN/中文, LLM-personalized questions, collapsible tracebacks, watch mode, lessons.
• 02_basics_tour.ipynb — basic tour (nested calls, recursion, watch mode)
• 03_features_tour.ipynb — fade before/after, stats, quiz, language, LLM toggle
• lesson_01_names_and_types, lesson_02_lists_and_dicts, lesson_03_functions_and_recursion, lesson_04_values_and_math — short beginner Python lessons, each with a broken cell to read and fix.

Developer guide

traceback_coach/
├── __init__.py     # version, load/unload_ipython_extension
├── knowledge.py    # ErrorFamily + 12 English FAMILIES + lookup()      [no IPython]
├── i18n.py         # zh-HK FAMILIES_ZH + bilingual LABELS              [no IPython]
├── _core.py        # parse_traceback, diagrams, question, card, render [no IPython]
├── magics.py       # CoachMagics, post_run_cell hook, display glue     [IPython]
└── static/mermaid.min.js   # vendored, offline

• knowledge.py, i18n.py, _core.py import no IPython — they're unit-testable headless. magics.py is the only IPython-facing module.
• Add an error family: add an ErrorFamily(...) to FAMILIES in knowledge.py (and a zh entry in i18n.py). Templates may use {token}, {message}, {error_type}, {line_no}.
• Add a language: add a FAMILIES_<lang> + a LABELS["<lang>"] map and a lookup(error_type, lang) branch; thread the lang code through %coach_lang.
• Run tests: pip install -e ".[dev]" && pytest -v (112 tests).
• Build & verify in a clean env: python -m build, then install the wheel in a fresh venv and import (two-venv pattern).

License

MIT