Literate Programming Skill

CRITICAL: This skill MUST be activated BEFORE making any changes to .nw files!

You are an expert in literate programming using the noweb system.

Reference Files

This skill includes detailed references in references/:

When to Use This Skill

Correct Workflow

1. User asks to modify a .nw file
2. YOU ACTIVATE THIS SKILL IMMEDIATELY
3. You plan the changes with literate programming principles
4. You make the changes following the principles
5. You regenerate code with make/notangle

Anti-pattern (NEVER do this)

1. User asks to modify a .nw file
2. You directly edit the .nw file ← WRONG
3. Later review finds literate quality problems
4. You have to redo everything

Remember

• .nw files are NOT regular source code files
• They combine documentation and code for human readers
• Literate quality is AS IMPORTANT as code correctness
• Bad literate quality = failed task, even if code works

Planning Changes

When making changes to a .nw file:

1. Read the existing file to understand structure and narrative
2. Plan with literate programming in mind:

- What is the "why" behind this change? - Why does this approach work, not just why was it chosen? - How does this fit into the existing narrative? - Does the document need an early whole-picture overview before details? - Which overview, intro, roadmap figure, or relevant README.md summarizes this area, and does this change make it stale? - What new chunks are needed? What are their meaningful names? - Where in the pedagogical order should this be explained?

1. Design documentation BEFORE writing code:

- Write prose explaining the problem and solution - Use subsections to structure complex explanations - For substantial .nw files, give maintainers the whole before the parts. If relationships are hard to hold in prose, add a structural figure. See references/overview-patterns.md.

1. Decompose code into well-named chunks:

- Each chunk = one coherent concept - Names describe purpose, not syntax (like pseudocode)

1. Write the code chunks
2. Regenerate and test

Key principle: If you find yourself writing code comments to explain logic, that explanation belongs in the documentation chunks instead.

Reviewing Literate Programs

When reviewing, evaluate:

1. Narrative flow: Coherent story? Pedagogical order?
2. Variation theory: Contrasts used? "Whole, parts, whole" structure?

Maintainer-facing .nw files should orient the reader to the system, then drill into parts, then reconnect changes to the whole.

1. Chunk quality: Meaningful names? Focused on single concepts? Bucket

chunks named with nouns (<<option completions>>), not verb phrases (<<wire option completions>>)?

1. Explanation quality: Explains "why" not just "what"? The

explanation should also say why the chosen approach works. Red flags: prose that begins "We [verb] the [noun]" matching a function name; prose that describes parameter types visible in the signature; prose that restates conditionals without explaining why they matter; prose that says an approach is better without explaining the mechanism, invariant, or constraint that makes it work.

1. Test organization: Tests after implementation, not before?
2. Proper noweb syntax: [[code]] notation in prose? Identifiers in

chunk titles escaped with [[...]]? Valid chunk references?

1. Overview quality: For non-trivial files, does the overview identify

the main parts, how they relate, and where the document will go next? Is the level appropriate for a maintainer who knows the language but not this codebase?

1. Visual clarity: When the structure is distributed, non-linear, or

hard to track in prose, does the document add a roadmap or diagram?

1. Overview maintenance: When structure, flow, chunk organization,

entry points, or likely edit locations changed, were the overview, local intro prose, roadmap figures, and any relevant README.md kept in sync?

1. Co-location and consistency: Are wiring chunks (attribute

assignments, register/install calls, decorator-style mutation) placed next to the entities they mutate, not the helpers they call? When the same task is solved in multiple places in the file, do they share a pattern — or does prose explain why a particular site diverges?

Core Philosophy

Literate programming (Knuth) has two goals:

1. Explain to human beings what we want a computer to do
2. Present concepts in order best for human understanding (psychological order, not compiler order)

Variation Theory

Apply variation-theory skill when structuring explanations:

• Contrast: Show what something IS vs what it is NOT
• Separation: Start with whole (module outline), then parts (chunks)
• Generalization: Show pattern across different contexts
• Fusion: Integrate parts back into coherent whole

When the literate document is student-facing educational LaTeX — the students read the woven output — keep pedagogical meta-commentary such as variation/invariance labels and sequencing rationale out of the visible narrative. Put that reasoning in \ltnote{...} via the didactic-notes skill.

\ltnote{...} is also available in an ordinary (maintainer-facing) literate program, but scope it tightly: it is for reasoning about the exposition, not the code. A note on why the narrative is ordered as it is — why function A is explained before function B, why a concept is introduced in this chapter rather than later, why the whole precedes the parts — is commentary about the writing and belongs in an \ltnote. A design decision about the code itself (an architecture choice, why a data structure was picked, the rationale and any citations behind it) is content: it belongs in the visible body prose. The test: if removing the note would lose information about the program, it is body text; if it would only lose information about how the explanation was authored, it is an \ltnote. See the didactic-notes skill for details.

When the tangled artifact is itself teaching material — a .nw that tangles to a student-facing tutorial (e.g. a pytorial .md) and weaves to a maintainer/educator chapter — apply the maintainer-facing rule with the tutorial in the role of the program. Each layer teaches the layer below it, so each has its own object of learning: the tangled tutorial teaches the student the subject; the woven prose teaches the teacher (the educator/maintainer) the tutorial's design — how the tutorial teaches; an \ltnote teaches the future author how the prose targets the teacher's learning. Pedagogical rationale about the tutorial (why this step order, what varies, the critical aspects) is therefore the woven chapter's own object of learning and belongs in the visible body prose, not in \ltnote — even though it is variation-theory commentary. \ltnote sits one level above the document's own object of learning: reserve it for how the woven exposition itself is authored. The test carries over: removing it loses information about the tutorial → body text; it loses only information about how the chapter's prose was authored → \ltnote.

If the .nw document generates slides or handouts with live questions, Mentipy is a suitable tool for embedding those prompts in the LaTeX output.

CRITICAL: Show concrete examples FIRST, then state general principles. Readers cannot discern a pattern without first experiencing variation.

Noweb File Format

Documentation Chunks

• Begin with @ followed by space or newline
• Contain explanatory text (LaTeX, Markdown, etc.)
• Copied verbatim by noweave

Code Chunks

• Begin with <<chunk name>>= on a line by itself (column 1)
• End when another chunk begins or at end of file
• Reference other chunks using <<chunk name>>
• Multiple chunks with same name are concatenated

Syntax Rules

• Quote code in documentation using [[code]] (escapes LaTeX special chars).

Never manually escape characters (e.g. \) inside [[...]] — noweb handles all escaping automatically. Writing [[\]] double-escapes.

• Mention a chunk by name in prose as [[<<chunk name>>]] — the outer

[[...]] typesets the reference as code. A bare <<chunk name>> in prose is not the convention; conversely, when reviewing, do not flag [[<<...>>]] as redundant.

• [[...]] works inside \item[...] labels (and other moving arguments),

but separate the inner ]] from the outer ] with at least one character — typically a space. The failure mode is three brackets in a row (]]]), where noweb's ]] terminator and LaTeX's ] argument terminator collide and produce a Runaway argument? ! Paragraph ended before \@item was complete. error.

GOOD — ]] is followed by a space or by other text: ``latex \item[The [[--restart]] flag] explains the idempotence rule. \item[ [[shell-basics $]] ] anchors on the nested shell prompt. ``

BAD — [[...]] butts directly against the closing ]: ``latex \item[The required pattern [[shell-basics $]]] % Runaway argument ``

The same rule applies to any other LaTeX macro argument that is delimited with brackets (\caption[...], \section[...] short forms, etc.). When in doubt, add a trailing space inside the outer brackets.

• Escape: @<< for literal <<, @@ in column 1 for literal @

Don't Write @ Between Adjacent Code Chunks

@ starts a documentation chunk; it is not a code-chunk terminator. A code chunk already ends the moment any chunk begins — including another named code chunk. So when one code chunk is immediately followed by another with no prose between them, no @ is needed: the second chunk closes the first on its own.

Worse, a stray @ (followed by a blank line) opens an empty documentation chunk, which noweave renders as unwanted vertical space in the woven PDF.

Rule of thumb: write @ only when documentation prose follows the code. This is the complement of Writing Guideline 14 (prose-follows ⇒ @; code-follows ⇒ no @).

GOOD — the second chunk ends the first; no @, no empty doc chunk:

<<constants>>=
CONST = 3.14
<<functions>>=
def area_of_circle(radius):
    return CONST * radius ** 2
@

BAD — redundant @ opens an empty doc chunk → unwanted space in the PDF:

<<constants>>=
CONST = 3.14
@

<<functions>>=
def area_of_circle(radius):
    return CONST * radius ** 2
@

Embedding Heredocs — Keep the Delimiter Pair in the Wrapper Chunk

When a tangled script embeds a heredoc (a shell <<EOF ... EOF, an inline python3 - <<EOF, a cat > file <<EOF, etc.), put both the opening <<DELIM and the closing DELIM in the same wrapper chunk, with a single chunk reference for the body in between. Define the body in its own chunk that contains only the payload — never the delimiter.

This matters because noweb chunk boundaries and the heredoc are two independent layers: noweb blindly concatenates every definition of the body chunk, and the shell only sees the tangled text. If the closing delimiter is written inside the body chunk (especially when the body is built from several concatenated definitions), the heredoc terminates after the first piece and everything after it tangles as code instead of data — a syntax error far from its cause.

GOOD — delimiter pair together in the wrapper; body chunk holds only the payload (and may be split across several definitions safely):

<<write the config>>=
cat > "$file" <<'_EOF'
<<config body>>
_EOF
@

<<config body>>=
first line of payload
@

<<config body>>=
more payload, concatenated by noweb into the same heredoc
@

BAD — the closing delimiter lives in the body chunk, so a second definition (or any following chunk) escapes the heredoc:

<<write the config>>=
cat > "$file" <<'_EOF'
<<config body>>
@

<<config body>>=
first line of payload
_EOF
@

<<config body>>=
this never reaches the heredoc — it tangles as shell
@

Quote the opening delimiter (<<'EOF') when the payload must not undergo shell expansion — for example embedded Python or JSON containing $. A distinctive delimiter such as EOF or PYEOF keeps it easy to see the pair at a glance.

Writing Guidelines

1. Start with the human story - problem, approach, design decisions
2. Give the whole before the parts - substantial maintainer-facing

documents should start with an overview of the moving pieces, their relationships, and the route through the file

1. Introduce concepts in pedagogical order - not compiler order
2. Use meaningful chunk names - 2-5 word summary of purpose (like pseudocode).

Bucket chunks are named with nouns; phase chunks may be imperative. A bucket chunk holds the same kind of content across many definitions and is named as a category: <<constants>>, <<functions>>, <<imports>>, <<test functions>>, <<repo methods>>, <<option completions>>. Phase sub-chunks (Guideline 8) are different — they describe a step in an algorithm and read naturally as imperatives (<<pick random starting point>>, <<collect sentences until target length>>).

BAD — verb-phrase bucket name: ``noweb <<wire option completions>>= LAYOUTOPTION.autocompletion = choicecompleter(LAYOUTCHOICES) @ ``

GOOD — noun bucket name: ``noweb <<option completions>>= LAYOUTOPTION.autocompletion = choicecompleter(LAYOUTCHOICES) @ ``

1. Escape all identifiers in chunk names — any identifier (variable,

function, parameter, attribute) that appears in a chunk title must be wrapped in [[...]]. This tells noweave to render it as code (monospace) and prevents LaTeX errors from underscores.

BAD — bare identifier, underscore breaks LaTeX: ``noweb <<restrict to top-level when not showtree>>= if not showtree: ... @ ``

GOOD — identifier escaped with [[...]]: ``noweb <<restrict to top-level when not [[showtree]]>>= if not showtree: ... @ ``

Other examples: <<add graders to [[graders]] list>>, <<initialise [[default_username]]>>.

IMPORTANT: Do not escape underscores or other special characters inside [[...]]. The brackets already tell noweb to escape everything. [[init.py]] is correct; [[\\init\\.py]] is wrong and will double-escape.

1. Decompose by concept, not syntax
2. Explain the "why" - don't just describe what the code does.

Prose that merely restates the code in English teaches nothing. Good prose explains why a design choice was made and why this approach works: what alternative was rejected, what property makes the chosen approach effective, what would break without it, or what constraint drives the implementation.

Self-test: If your prose could be mechanically generated from the function signature, it's "what" not "why." Ask yourself: What design decision does this paragraph justify? What alternative did we reject and why? Why does the chosen approach work here? If the paragraph doesn't answer those questions, rewrite it.

BAD — prose restates code in English: ```noweb \subsection{Counting $n$-grams}

We count overlapping $n$-grams. If $n$ is larger than the input, the result is empty.

<<functions>>= def ngram_counts(text, *, n): ... @ ```

GOOD — prose explains why this design choice: ```noweb \subsection{Counting $n$-grams}

We use overlapping $n$-grams because they capture all positional contexts---in \enquote{THE}, overlapping bigrams yield TH and HE, whereas non-overlapping would only yield TH. This matches the standard definition used in cryptanalysis.

<<functions>>= def ngram_counts(text, *, n): ... @ ```

Red flags that prose is "what" not "why": - Begins "We [verb] the [noun]" where the verb matches a function name - Describes parameter types or return values already in the signature - Restates conditional logic ("If X, we do Y") without explaining why X matters

1. **Keep chunks focused — one function per <<functions>>= chunk with

prose before it.** Each function (or small group of tightly related functions) gets its own <<functions>>= chunk preceded by explanatory prose. Never put multiple unrelated functions in a single chunk.

BAD — four functions crammed into one chunk with minimal prose: ```noweb \subsection{Helper Functions}

We provide several utility functions.

<<functions>>= def normalize_text(text): ...

def letters_only(text): ...

def key_shifts(key): ...

def indexofcoincidence(text): ... @ ```

GOOD — each function with its own subsection and prose: ```noweb \subsection{Text Normalization}

Before analysis, we strip non-alphabetic characters and convert to lowercase so that frequency counts are meaningful.

<<functions>>= def normalize_text(text): ... @

\subsection{Index of Coincidence}

The index of coincidence measures how likely two randomly chosen letters from a text are identical ...

<<functions>>= def indexofcoincidence(text): ... @ ```

1. Decompose long functions into named sub-chunks — If a function has

more than ~25 lines and contains two or more distinct algorithmic phases, decompose it into named sub-chunks. Each sub-chunk name should read like a step in an algorithm description. The prose before each sub-chunk explains why that phase works the way it does and what property of the data or algorithm makes the approach succeed. This is the classic Knuth technique.

BAD — 80-line function with one line of prose: ```noweb We generate plaintext by concatenating sentences.

<<functions>>= def generateplaintext(size, *, sources, seed=None): """...""" if size <= 0: raise ValueError(...) paragraphs = extractparagraphs(sources, ...) ... # 75 more lines return normalize(prefix, options) @ ```

GOOD — function body decomposed into named sub-chunks with prose: ```noweb <<functions>>= def generate_plaintext(size, *, sources, seed=None): """...""" <<prepare filtered paragraphs>> <<pick random starting point>> <<collect sentences until target length>> <<select closest sentence boundary>> @

We extract paragraphs from the corpus, removing headings and ToC entries. Paragraphs lacking sentence-ending punctuation are discarded---they are typically list items or table rows.

<<prepare filtered paragraphs>>= if size <= 0: raise ValueError("size must be positive") ... @

To avoid always starting at the beginning of the corpus, we rotate to a random paragraph.

<<pick random starting point>>= rng = random.Random(seed) ... @ ```

1. Decompose classes by method — Introduce the class shell in one

chunk with a placeholder like <<repo methods>>, then define each method in its own \section or \subsection with prose + method chunk + tests. This is the class-level analogue of guideline 7 (one function per chunk) and guideline 8 (decompose long functions). The class shell gives the reader the whole picture; the method sections fill in each part with full explanations and verification.

BAD — entire class in one chunk: ```noweb \section{The Repository}

<<classes>>= class Repo: def init(self, path): ...

def save(self, data): ...

def load(self, key): ... @

\section{Tests} <<test functions>>= def testsave(): ... def testload(): ... @ ```

GOOD — class shell + one section per method, tests distributed: ```noweb \section{The Repository}

The class needs two operations: saving and loading. We introduce the class shell here and fill in each method in its own section.

<<classes>>= class Repo: def init(self, path): self.path = pathlib.Path(path)

<<repo methods>> @

\subsection{Saving data}

We serialise to JSON because students can inspect the files manually, unlike a binary format.

<<repo methods>>= def save(self, data): """Persist `data` to disk as JSON.""" ... @

Let's verify that saving round-trips correctly:

<<test functions>>= def testsavecreatesfile(tmppath): repo = Repo(tmppath) repo.save({"key": "value"}) assert (tmppath / "data.json").exists() @

\subsection{Loading data}

We return [[None]] for missing keys rather than raising, because a missing key is a normal condition during first run.

<<repo methods>>= def load(self, key): """Load data for `key, or None` if absent.""" ... @

<<test functions>>= def testloadmissingreturnsnone(tmppath): repo = Repo(tmppath) assert repo.load("absent") is None @ ```

The chunk name <<repo methods>> is a bucket chunk scoped to the class: noweb concatenates all definitions, so each <<repo methods>>= adds another method to the class body. Choose a name that reflects the class (e.g. <<iolog methods>>, <<stream capture methods>>).

Apply the same pattern to test classes. If <<test functions>>= introduces class Test...:, put only the class shell there and delegate the body to a class-specific bucket chunk such as <<feature a test methods>>. Do not concatenate indented test methods directly into <<test functions>>=; nested scopes are clearer and less error-prone when they use their own bucket chunk.

1. **Use bucket chunks — distribute <<constants>>= near their relevant

code** - Define each constant in the section where it is conceptually relevant. Never group all constants into a single \subsection{Constants}.

IMPORTANT: When a constant exists only to support one helper or one small cluster of related helpers, place its <<constants>>= bucket immediately adjacent to the bucket that contains those helpers. Do not hide configuration keys in one distant global block if the reader only needs them to understand one local section.

Preferred pattern — prose, then helper bucket, then local constants bucket for that helper family: ```noweb We read the debug flag from configuration rather than the process environment so detached hooks and interactive commands see the same value.

<<helper functions>>= def trackdebugenabled(): return config.get(TRACKDEBUGCONFIG) <<constants>>= TRACKDEBUGCONFIG = "track.debug" @ ```

This keeps the constant close enough to the narrative that readers meet it when they need it, without forcing them to search a giant global constants block.

BAD — all constants dumped in one subsection: ```noweb \subsection{Constants}

<<constants>>= DATADIR = ... % used in loading section GUTENBERGSTART = ... % used in extraction section SENTENCERE = ... % used in sentence splitting section KEEPPUNCT = ... % used in normalization section @ ```

GOOD — each constant near the code that uses it: ```noweb \subsection{Loading Texts}

<<constants>>= DATADIR = Path(file).parent / "data" <<functions>>= def loadtext(path): ... @

\subsection{Extracting Body Text}

<<constants>>= GUTENBERGSTART = " START OF" GUTENBERGEND = " END OF" <<functions>>= def extract_body(text): ... @ ```

1. **Co-locate wiring chunks with the entities they mutate, not the helpers

they call.** A wiring chunk (attribute assignment, register()/install() call, decorator-style mutation) has two natural neighbours: the helper functions it calls and the objects it mutates. Place it next to the objects it mutates so a reader who lands on the option/class/registry sees what is attached to it; the helper has its own section and does not benefit from a trailing assignment block. Chunk source order is independent of tangle order, so co-locating the source costs nothing at runtime.

BAD — wiring sits next to the helper it calls, far from the option it mutates: ```noweb <<constants>>= LAYOUTOPTION = typer.Option("slide", "--layout", help=LAYOUTHELP) @

...300 lines later...

<<functions>>= def choicecompleter(values): ... <<option completions>>= LAYOUTOPTION.autocompletion = choicecompleter(LAYOUTCHOICES) @ ```

GOOD — wiring sits next to the object it mutates; the helper moves to live near the same neighbourhood: ``noweb <<constants>>= LAYOUTOPTION = typer.Option("slide", "--layout", help=LAYOUTHELP) <<functions>>= def choicecompleter(values): ... <<option completions>>= LAYOUTOPTION.autocompletion = choicecompleter(LAYOUTCHOICES) @ ``

1. Define constants for magic numbers - never hardcode values
2. Co-locate dependencies with features - feature's imports in feature's section
3. Never leave prose inside an open code chunk — When inserting local

<<constants>>= buckets or explanatory paragraphs, first close the current code chunk with @. Documentation between two function sections must be outside code mode; otherwise noweb tangles the prose into Python and produces syntax errors.

BAD — prose accidentally tangled as Python: ```noweb <<helper functions>>= def getdefaultdaily_limit(): ...

This helper uses the daily-limit config key.

<<constants>>= DEFAULTDAILYLIMITCONFIG = "track.dailylimit" @ ```

GOOD — close the code chunk before prose, then reopen a bucket: ```noweb <<helper functions>>= def getdefaultdaily_limit(): ... @

This helper uses the daily-limit config key.

<<constants>>= DEFAULTDAILYLIMITCONFIG = "track.dailylimit" @ ```

Conversely, when a code chunk is immediately followed by another code chunk with no prose between them, omit the @ — see "Don't Write @ Between Adjacent Code Chunks" above.

1. Prefer public functions - Default to making functions public with

docstrings. Only use -prefixed private functions for true internal helpers tightly coupled to a single caller. Public utilities (e.g., normalizetext, lettersonly) are reusable across modules and discoverable via help(). Duplicated private helpers across modules (e.g., to_ascii in both vigenere.nw and plaintexts.nw) are a sign the function should be public in a shared module.

1. Keep lines under 80 characters - both prose and code
2. Survey siblings before introducing a new pattern. Before reaching

for a new mechanism in a .nw file, grep the same file (and sibling .nw files) for places that already solve the same task. If three command-local options already attach completion via autocompletion=... at construction, the fourth option should match — or the prose must explain why the divergence is necessary (forward-reference constraint, chunk-tangle order, dependency on a later-defined symbol, etc.). The cost of this survey is one grep; the cost of skipping it is an asymmetric file where readers cannot tell whether the new mechanism is load-bearing or accidental.

LaTeX Documentation Quality

Apply latex-writing skill. Most common anti-patterns in .nw files:

Lists with bold labels: Use \begin{description} with \item[Label], NOT \begin{itemize} with \item \textbf{Label}:

Code with manual escaping: Use [[code]], NOT \texttt{...\_...}

Manual quotes: Use \enquote{...}, NOT "..." or ` ...'' `

Manual cross-references: Use \cref{...}, NOT Section~\ref{...}

TikZ without preamble support: If a .nw file introduces a TikZ diagram, add \usepackage{tikz} to the project's preamble.tex.

Non-ASCII bytes in code chunks weave into the .tex: noweave copies code chunks verbatim, so any non-ASCII character inside a chunk — a box-drawing diagram in a test fixture, a Unicode normalization table, fancy quotes — lands in the woven LaTeX and must be typeset by the engine. How that behaves is engine-dependent: under pdfLaTeX an unmapped byte is a fatal ! LaTeX Error: Unicode character … (U+XXXX) not set up (at the woven line); under XeLaTeX/LuaLaTeX it is at worst a non-fatal Missing character warning. Check the build engine first — a literate project's build rules are often tangled from a .nw (e.g. tex.mk), so a make/submodule update can flip latexmk -pdf to -xelatex and change which fix is correct.

Fix it in the project's preamble.tex, not by deleting or escaping the byte in the .nw — the tangled program usually needs the real character at runtime (the test that asserts an ASCII-art figure is filtered, the table that maps real typographic punctuation). Before touching any non-ASCII byte in a code chunk, confirm whether the generated code depends on it. The preamble fix differs per engine (\DeclareUnicodeCharacter under pdfLaTeX — defined via \usepackage[utf8]{inputenc}; \newunicodechar under XeLaTeX/LuaLaTeX; guard with iftex). See the latex-writing skill's references/unicode-and-fonts.md for the full engine matrix, the T1-fontenc requirement for pdfLaTeX monospace fonts like Bera Mono, and the pdffonts/pdftotext diagnostics.

Progressive Disclosure Pattern

When introducing high-level structure, use abstract placeholder chunks that defer specifics:

def cli_show(user_regex,
             <<options for filtering>>):
  <<implementation>>
@

[... later, explain each option ...]

\paragraph{The --all option}
<<options for filtering>>=
all: Annotated[bool, all_opt] = False,
@

Benefits: readable high-level structure, pedagogical ordering, maintainability.

The same technique applies to function bodies: long functions can use <<phase name>> sub-chunks to present algorithmic steps in pedagogical order with prose between them (see Writing Guideline 8, "Decompose long functions").

Chunk Concatenation Patterns

Use multiple definitions when building up a parameter list pedagogically:

\subsection{Adding the diff flag}
<<args for diff>>=
diff=args.diff,
@

[... later ...]

\subsection{Fine-tuning thresholds}
<<args for diff>>=
threshold=args.threshold
@

Use separate chunks when contexts differ (different scopes):

<<args from command line>>=  # Has args object
diff=args.diff,
@

<<params for recursion>>=    # No args, only parameters
diff=diff,
@

Chunk Dependency Hazards

When a chunk references a variable defined in another chunk, there is an implicit dependency between them. Unlike function calls, noweb does not enforce that prerequisite chunks are included — the compiler only sees the tangled output. This makes it easy to reuse a chunk in a new code path while accidentally omitting the chunk that defines a variable it needs.

Rule: When reusing a chunk in a new code path, verify that all variables it references are defined by preceding chunks in that path.

BAD — chunk B depends on a variable set in chunk A, but a new path omits A:

<<path one>>=
<<chunk A>>
<<chunk B>>
@

<<path two>>=
<<chunk B>>      % ← UnboundLocalError: variable from A missing
@

GOOD — either include the prerequisite or document the dependency:

<<path two>>=
<<chunk A>>
<<chunk B>>
@

Tip: If a chunk both defines variables AND performs side effects that are not always wanted, consider splitting it so the variable-defining part can be included independently.

Design rule: Prefer extracting shared logic into a function rather than a reusable chunk when the logic is used across multiple code paths. Functions make dependencies explicit through parameters — a missing argument is a compile-time error, while a missing prerequisite chunk is only caught at runtime. Reserve chunks for pedagogical decomposition (presenting code in narrative order); use functions for operational decomposition (sharing logic between code paths).

Test Organization

CRITICAL: Tests MUST appear AFTER implementation, distributed throughout the file near the code they verify. NEVER create a \section{Tests} or \section{Unit Tests} that groups all tests at the end of the file.

See references/testing-patterns.md for detailed patterns.

Key rules:

• Each implementation section is followed by its <<test functions>>= chunk
• Use single <<test functions>> chunk name — noweb concatenates them
• If <<test functions>>= introduces a test class, treat it as the class

shell and accumulate methods in a class-specific bucket chunk such as <<feature a test methods>>

• Use from module import * in the test file header
• Frame tests pedagogically: "Let's verify this works..."
• Test files tangle with bare notangle (no black pass), so the .nw

indentation lands verbatim in tests/test_*.py. Indent new test chunks 2 spaces per level; when extending an existing chunk, match its indentation instead of mixing styles (module chunks differ: black reformats them, so their .nw indentation is cosmetic). See references/testing-patterns.md.

BAD — all tests collected at the end:

\section{Encryption}
<<functions>>=
def encrypt(text, key): ...
@

\section{Decryption}
<<functions>>=
def decrypt(text, key): ...
@

\section{Tests}          % ← NEVER do this

<<test functions>>=
def test_encrypt(): ...
def test_decrypt(): ...
@

GOOD — each test immediately after its implementation:

\section{Encryption}
<<functions>>=
def encrypt(text, key): ...
@

Let's verify that encryption produces the expected ciphertext:

<<test functions>>=
def test_encrypt(): ...
@

\section{Decryption}
<<functions>>=
def decrypt(text, key): ...
@

We can verify that decryption inverts encryption:

<<test functions>>=
def test_decrypt(): ...
@

If a test section needs a test class, introduce the class in <<test functions>>= and collect its methods in a class-specific bucket chunk rather than concatenating indented methods directly into <<test functions>>=. See references/testing-patterns.md for a worked example and anti-pattern.

Multi-Directory Projects

For large projects (5+ .nw files), see references/multi-directory-projects.md.

Key structure:

project/
├── Makefile       # Root orchestrator (compile → test → docs)
├── pyproject.toml # Poetry packaging configuration
├── src/           # .nw files → .py + .tex
├── doc/           # Document wrapper (.nw), preamble.tex
├── tests/         # Extracted test files (unit/ subdir)
└── makefiles/     # Shared build rules (noweb.mk, subdir.mk)

Initializing a New Project

See references/project-initialization.md for full details. Quick checklist:

1. Create pyproject.toml with [tool.poetry] packages/include/exclude
2. Create src/.gitignore (.py, .tex) and tests/.gitignore (*.py)
3. Create src/packagename/Makefile with explicit init.py rule
4. Create src/packagename/packagename.nw with <<[[init.py]]>> and

<<test [[packagename.py]]>> chunks

1. Create tests/Makefile with auto-discovery (uses %20 encoding, cpif,

unit/ subdirectory)

1. Create doc/packagename.nw wrapper, doc/Makefile, doc/preamble.tex
2. Create root Makefile orchestrating compile → test → docs

Adding a New Module to an Existing Project

CRITICAL: a new .nw module must be registered in several places, and a missed step fails silently — the module builds and tests fine but never appears in the PDF documentation. When creating src/package/newmodule.nw, register it in ALL of:

1. src/package/Makefile — MODULES+= newmodule.py
2. doc/Makefile — main.pdf: ../src/package/newmodule.tex
3. The master document (doc/main.tex / contents.tex / doc wrapper

.nw) — \input{../src/package/newmodule.tex}

1. Verify tests were auto-discovered (make -C tests)

Then build the docs and verify the module is actually in the PDF (pdftotext doc/main.pdf - | grep -c newmodule — a count of 0 means a step was missed). Apply the same check when reviewing a change that adds a .nw file. Also ensure the doc Makefile declares the .nw behind each cross-directory .tex prerequisite — a generic ../%:: recursion rule never re-weaves an existing stale .tex. See "Registering a New Module" in references/multi-directory-projects.md for details and the stale-weave failure mode.

LaTeX-Safe Chunk Names

Use [[...]] notation for Python chunks with underscores:

<<[[module_name.py]]>>=
def my_function():
    pass
@

Extract with: notangle -R"[[modulename.py]]" file.nw > modulename.py

Filename Chunk Names Drive Language Inference

Naming root chunks exactly like their output filenames is load-bearing, not just a convention: the autolang and tominted filters infer each chunk's language from filename-like names — Python for <<[[fib.py]]>>, Make for <<[[Makefile]]>> — and chunks whose names are not filenames (<<functions>>, <<test functions>>) inherit the language of the chunks that use them. This is what lets one .nw file mix languages (a program plus its Makefile) while each chunk is highlighted by its own lexer and the identifier index stays free of foreign identifiers. See references/noweb-commands.md for the inference rules and .ext=lexer / name=lexer mapping overrides.

Best Practices Summary

1. Write documentation first - then add code
2. Keep lines under 80 characters
3. Check for unused chunks - run noroots to find typos
4. Keep tangled code in .gitignore - .nw is source of truth
5. NEVER commit generated files - .py and .tex from .nw are build artifacts
6. Test your tangles - ensure extracted code runs
7. Require PEP-257 docstrings on all public functions - Prose in .nw

is for maintainers reading the literate source; docstrings are for users of the compiled .py who never see the .nw file. Both are needed. Private functions (prefixed _) may omit docstrings. Never use \cref or other LaTeX commands inside docstrings — and never use noweb [[...]] code-quoting there either. [[...]] is only interpreted by noweave in documentation chunks; inside a docstring (a string in a code chunk) it is copied verbatim, so it leaks into the generated .py and help() output. Quote code in docstrings with markdown single backticks (` x ) — not RST double-backticks, and never [[...]]`. (Only deviate when a project's existing docstrings consistently use another style.)

Chunk references (<<...>>) inside docstrings are fine: the custom lexer tominted uses by default keeps them hyperlinked in the woven PDF even though they sit inside a Python string literal.

BAD — noweb [[...]] in a docstring leaks into help(): ``noweb <<functions>>= def setattr(self, name, value): """Keep [[labels]] a sorted, duplicate-free list.""" ... @ ``

GOOD — quote code with the project's docstring convention: ``noweb <<functions>>= def setattr(self, name, value): """Keep labels a sorted, duplicate-free list.""" ... @ ``

BAD — function with prose but no docstring: ```noweb We convert text to lowercase ASCII for uniform comparison.

<<functions>>= def normalize_text(text): return text.lower().encode("ascii", "ignore").decode() @ ```

GOOD — prose for maintainers AND docstring for users: ```noweb We convert text to lowercase ASCII for uniform comparison.

<<functions>>= def normalize_text(text): """Return lowercase ASCII version of `text`.

Non-ASCII characters are silently dropped. """ return text.lower().encode("ascii", "ignore").decode() @ ```

1. Keep the maps in sync - When a change affects structure, flow, entry

points, chunk grouping, or likely edit locations, update the overview, local intro prose, roadmap diagrams, and any relevant README.md in the same change

1. Include table of contents - add \tableofcontents in documentation

Git Workflow

See references/git-workflow.md for details — including the mandatory post-merge-conflict checks (noroots + re-tangle + grep a known symbol): a resolution that glues prose onto a <<name>>= line makes notangle silently emit an empty file.

Core rules:

• Only commit .nw files to git
• Add generated files to .gitignore immediately
• Regenerate code with make after checkout/pull
• Never commit generated .py or .tex files

Noweb Commands Quick Reference

See references/noweb-commands.md for details, including the one-time custom-lexer setup the standard weave below depends on.

# Tangling
notangle -R"[[module.py]]" file.nw > module.py
notangle -t8 -R"[[Makefile]]" file.nw > Makefile  # -t8 keeps tabs
noroots file.nw                              # List root chunks

# Weaving (standard recipe: highlighted with minted, clean index)
# For inclusion; the master preamble loads minted via \usepackage[minted]{noweb}
# tominted finds its bundled custom lexer itself (installed next to
# the filter); no -lexer argument or local noweb_lexer.py copy needed.
noweave -n -delay -autolang -autodefs python3 -autodefs sh \
    -autodefs make -index -filter tominted file.nw > file.tex
# compile with -shell-escape (pdflatex -shell-escape or latexmk option)
# Standalone (noweave writes the preamble): add -minted so it loads minted
noweave -autolang -autodefs python3 -autodefs sh -autodefs make \
    -index -minted file.nw > file.tex

# Classic fallback (no minted/patched noweb; identifier uses inside
# code stay hyperlinked, at the price of no syntax highlighting)
noweave -n -delay -x -t2 file.nw > file.tex  # For inclusion
noweave -latex -x file.nw > file.tex         # Standalone

Pipeline rules baked into the standard recipe: -autolang always runs first no matter where it appears on the command line; tominted must be the last filter, after -index. tominted is LaTeX-only — never combine it with -html.

Stacking several -autodefs works because -autolang's @language annotations gate each filter to its own chunks. Available filters: python3, sh/bash, make, haskell, rust, java, c (ANSI C chunks only), cpp (C++ chunks: classes, templates, namespaces, qualified methods), plus the classic icon/sml/tex/yacc/etc. — run noweave -showautodefs for the installed list. Filter budget: the dbosk fork's pipeline is unbounded — it accumulates filters in a single variable and joins them with |, so any number of -autodefs filters can stack alongside -autolang, -index (which inserts two filters, finduses and noidx) and tominted in one weave. A six-language document is one pass, not two. (Stock and pre-fork noweave enumerated only seven filter slots, a portability artifact of the array-less /bin/sh pipeline builder; that ceiling — three -autodefs for a highlighted indexed weave, four for a classic one — applies only if you run against an older noweave, in which case split the languages across a second weave; see the multilang example in noweb's examples/.)

[[code]] in headings + hyperref (the \hyper@@link / \Hy@reserved@a error)

A [[identifier]] quote inside a \chapter/\section/\subsection heading is a hazard when -index is active. -index turns the quote into a hyperlinked identifier use (\nwlinkedidentq → \nwhyperreference → \hyperlink). In a heading the title is expanded into moving arguments — the .toc, the running header, and the PDF bookmark — and hyperref breaks there. Under newer TeX Live (≥2025) the symptom is a fatal error, typically:

! Undefined control sequence.
\hyper@@link ->\let \Hy@reserved@a ...
l.379 ...}}{NW4Ar8dO-3qo3Hb-8}\nwendquote} method}

(preceded by a cascade of Token not allowed in a PDF string ... removing '\Hy@reserved@a' warnings), or a stack overflow from noweb's self-redefining \nwhyperreference (its terminating \global\let only fires when the macro is executed, never when it is merely expanded into a moving argument). This is a toolchain regression: the same document built cleanly under older hyperref/TeX Live, so a make/submodule update that bumps TeX Live can surface it with no source change. Only headings whose quoted text is an indexed identifier (a name defined in a code chunk, e.g. a dunder like str) become links and fail; plain words like [[command]] weave to a non-linked \Tt and are harmless.

Preferred fix — keep the literate [[...]]-in-headings convention. noweb.sty from the dbosk fork (integration branch, Aug 2026 onwards) fixes this itself: \nwhyperreference is declared robust and the bookmark rules below are registered at \begin{document}, so with that noweb.sty installed no preamble change is needed. For a stock noweb.sty, the shipped references/preamble.tex applies the same two cooperating fixes project-wide (after hyperref is loaded), guarded so they take effect only when the loaded noweb.sty lacks them:

\ifcsname nwhyperreference \endcsname\else   % already robust (fork)? skip
  % (1) robust \nwhyperreference: written literally into moving arguments
  %     instead of expanding (and misfiring its self-\let) there
  \DeclareRobustCommand\nwhyperreference[2]{\hyperlink{noweb.#1}{#2}\relax}
  % (2) strip noweb's code-quote macros from PDF bookmark strings
  \pdfstringdefDisableCommands{%
    \def\nwlinkedidentq#1#2{#1}%
    \def\nwlinkedidentc#1#2{#1}%
    \let\Tt\relax
    \let\nwendquote\relax
  }
\fi

Two details matter when hand-copying this block. The guard: \DeclareRobustCommand stores the real macro under the name with a trailing space, which stock noweb.sty never defines, so the test tells a fixed noweb.sty from a stock one. The trailing \relax: beamerarticle redefines \hyperlink to accept an overlay specification after its arguments, and in a code chunk a linked identifier is followed by the rest of the line — close<enter> weaves to \hyperlink{noweb.X}{close}<enter>, and without the \relax beamer takes <enter> as an overlay and drops both the identifier and the <enter> (dbosk/noweb#11). An unguarded override without \relax reintroduces that bug on top of a fixed noweb.sty.

The narrower alternative — rewriting the heading to use \texttt{...} instead of [[...]] — also works but loses the cross-reference link and must be repeated per heading, so prefer the preamble fix.

While debugging such a build, note latexmk can mask the real LaTeX error behind a later toolchain failure (e.g. biber: malformed bcf from a stale aux). Run one bare pdflatex -halt-on-error pass to see the first genuine error before chasing bibliography/pythontex noise.

When Literate Programming Is Valuable

• Complex algorithms requiring detailed explanation
• Educational code where understanding is paramount
• Code maintained by others
• Programs where design decisions need documentation
• Projects combining multiple languages/tools