# `docextract` Specification
## Overview
`docextract` is a tiny Rust library and CLI for extracting documentation written using a dedicated `////` comment syntax from Rust source files.
The project intentionally **does not parse Rust syntax**. It performs a simple line-based scan of source files. Its purpose is to allow documenting declarative macro DSLs or other generated APIs independently of `rustdoc`.
The project is a **single crate** with two targets:
* **library**: `docextract` — used as a build dependency via `cargo add --build docextract`
* **binary**: `cargo-docextract` — installed via `cargo install docextract`, invoked as `cargo docextract`
The implementation should be lightweight. Zero dependencies without features; the `regex` feature (enabled by default) adds `regex-lite` as an optional dependency.
## Goals
* Extremely small implementation.
* Zero dependency on:
* syn
* quote
* proc-macro2
* clap
* anyhow
* thiserror
* regex (note: `regex-lite` is an allowed optional dependency behind the `regex` feature)
* any argument-parser crate (clap, structopt, pico-args, etc.)
The optional `regex-lite` dependency is the sole exception. When building with `--no-default-features`, the crate has zero dependencies.
* Fast startup.
* Suitable for use from `build.rs`.
* Produce deterministic output.
* Work entirely on stable Rust.
## Non-goals
* Parsing Rust syntax.
* Expanding macros.
* Understanding token trees.
* Markdown rendering.
* rustdoc integration.
* AST generation.
The tool is strictly a documentation extractor.
---
# Documentation Syntax
A line is a documentation line **only** when it begins with optional ASCII whitespace followed by `////` and then a space (0x20), tab (0x09), or end-of-line (i.e., `\n` 0x0A or `\r` 0x0D as part of CRLF). Lines that do not meet this rule are **silently ignored** — no warning is emitted.
Valid documentation lines:
```text
//// Hello
////\tHello
////
//// Hello
\t//// Hello
```
Silently ignored (no whitespace after the four slashes):
```text
////Hello
////Text
```
Silently ignored (five or more slashes):
```text
///// extra slash
////// two extra
```
After confirming the line matches the documentation-line pattern, extraction removes:
* leading ASCII whitespace (space 0x20, tab 0x09)
* the four slashes
* one separator character — the space or tab immediately following the slashes
Everything after the separator is preserved verbatim. If the line was `////` (immediate newline after slashes), `//// ` (space then newline), or `////<TAB>` (tab then newline), the result is an empty string — a verbatim blank line.
Examples:
Input
```text
//// Hello
```
Output
```text
Hello
```
Input
```text
//// ### Heading
```
Output
```text
### Heading
```
(The space between `////` and `###` is the separator — it is stripped by the extractor, leaving `### Heading`. A line written as `////### Heading` with no whitespace after the slashes is **silently ignored**, not extracted.)
Input
```text
//// Two Spaces
```
Output
```text
Two Spaces
```
The leading ASCII whitespace is stripped. The four slashes are stripped. One separator space is stripped. The remaining ` Two Spaces` (space before "Two") is preserved verbatim.
Input (tab indentation)
```text
\t//// text
```
Output
```text
text
```
Input (tab separator after slashes)
```text
////\ttext
```
Output
```text
text
```
Input (verbatim blank line from `////` alone)
```text
////
```
Output
```text
```
(`//// ` — four slashes, space, newline — produces the same empty output. Both are equivalent.)
No Markdown interpretation is performed. Whitespace inside extracted text is treated **opaquely** and passed to output verbatim — only the leading whitespace and separator are parsed; everything else is content.
Only ASCII whitespace is recognized for parsing: space (0x20) and tab (0x09). Any other whitespace characters appearing inside extracted text pass through to output unchanged.
---
# Code-Lifting Syntax
Non-documentation source lines containing `////` preceded by space or tab are **code-capturing lines**. The text left of `////` is *captured code*; the text to the right is a *documentation template* with `{{...}}` placeholders. Placeholders substitute captured code into the template:
- `{{}}` — full substitution with leading/trailing whitespace stripped
- `{{N}}` — remove first N UTF-8 characters, substitute remainder
- `{{XregexX}}` — extract first regex match (requires `regex` feature)
- `{{XregexXsubstX}}` — `replace_all` with `$N` backreferences (requires `regex` feature)
Lines ending with `////` establish a **capture zone** at the indentation of the captured code. Within this zone, a `////` line at deeper indentation is treated as a code-capturing suffix (not a doc prefix) — the leading whitespace becomes empty captured code, and the template resolves normally. The zone ends when a non-code-capturing line at ≤ the zone's indentation is encountered. This allows indented code blocks to flow through without breaking the documentation structure. The captured code has trailing ASCII whitespace stripped before any placeholder resolution. Code-capturing entries are regular persistent documentation entries: they carry the current tag, participate in filtering, and are concatenated with other entries.
See [specs/code-lifting-spec.md](specs/code-lifting-spec.md) for the full specification.
---
# Documentation Blocks
Consecutive documentation lines form one documentation block.
Example:
```rust
//// First
//// Second
let x = 5;
//// Third
```
Produces
```text
First
Second
Third
```
No blank line is automatically inserted between documentation blocks. To separate blocks visually, write an explicit `////` (verbatim blank doc line) between them.
Blank lines inside extracted blocks are produced by explicitly writing empty `////` lines (or `//// ` lines — they are equivalent):
```rust
//// First
////
//// Second
```
Produces
```text
First
Second
```
Non-documentation source lines never appear in the output.
---
# Input
The CLI accepts zero or more input file paths.
A single hyphen `-` as an input filename reads from **stdin**. For example:
```text
cargo docextract output.md -
```
reads from stdin and writes to `output.md`. When both output and the sole input are `-` (`cargo-docextract - -`), the CLI reads stdin and writes stdout. A single `-` as the only argument (`cargo-docextract -`) means the same thing: since output-to-stdout with no inputs would otherwise produce empty output, the lone `-` implies **both** stdout output and stdin input — equivalent to `cargo-docextract - -`.
Each input file is processed independently. The output from all inputs is concatenated in the order given — there is **no separator** added between files. Multi-file output is simply the concatenation of each file's extracted documentation blocks.
An empty input list produces an empty output file. This is well-defined and useful for generated-file pipelines.
Future recursive directory walking is optional but not required for version 1.
---
# Output
Default output is stdout.
The **output path is mandatory** and comes first in the argument list (see CLI section below). To write to stdout, use `-` as the output path.
No leading or trailing blank lines are added to the output. If the source file(s) begin or end with a `////` line, those produce verbatim blank lines in the output at the corresponding position. The extractor does not add or strip blank lines at the edges of the output.
---
# CLI
Executable name:
```text
cargo-docextract
```
Invocation:
```text
cargo-docextract <output> [inputs...]
```
* `<output>` — **mandatory**, must be the first argument. A file path, or `-` for stdout.
* `[inputs...]` — zero or more file paths. A file path, or `-` for stdin.
There are **no options, no flags** (`-o`, `--`, etc. are not recognized). The interface is positional only.
**Help**: Invoking with zero arguments or `--help` prints a short help text and exits with status 0. No other flag or long-flag syntax is supported.
**Stdin**: `-` as an input filename reads from stdin. `cargo-docextract - -` reads stdin and writes stdout. A single `-` alone (`cargo-docextract -`) implies both stdout output and stdin input — equivalent to `cargo-docextract - -`, since output-to-stdout with no inputs makes no sense.
**Empty input list**: `cargo-docextract output.md` (no input files) creates an empty output file. This is well-defined behavior.
**File not found**: If an input file does not exist, print a warning to stderr, **continue** to the next file (do not abort). The exit code reflects the final outcome — success if all existing files were processed, error if none could be opened.
Other I/O errors during input processing (permission denied, read errors) cause immediate abort with exit code 2. Failure to open the output file for writing also produces exit code 2.
---
# Exit Codes
| Code | Meaning |
|------|---------|
| 0 | Success |
| 2 | I/O error |
Exit code 1 is not used: with zero arguments the CLI prints help and exits 0, and with any arguments the first argument *is* the output — a "missing output argument" condition cannot arise. See [specs/cli-spec.md](specs/cli-spec.md) §6.
---
# Library API
The library exposes a single public type: the `Docextract` builder. There are
no standalone functions.
```rust
pub struct Docextract { /* ... */ }
impl Docextract {
/// Create a new builder, writing extracted documentation to `output`.
/// `"-"` means stdout. Auto-detects build-script context (OUT_DIR set)
/// and enables rerun-if-changed automatically.
pub fn new(output: impl AsRef<Path>) -> io::Result<Self>;
/// Extract documentation from a file at `path`. Processes immediately
/// (streaming). File not found → warning to stderr, returns Ok.
/// Other I/O errors → Err. Emits cargo:rerun-if-changed if enabled.
pub fn input(&mut self, path: impl AsRef<Path>) -> io::Result<&mut Self>;
/// Extract documentation from stdin. Processes immediately. No
/// cargo:rerun-if-changed emitted (stdin has no file to watch).
pub fn stdin(&mut self) -> io::Result<&mut Self>;
/// Finalize extraction: flush the output writer. Returns &mut self so the
/// call is chainable. Calling done() multiple times is safe (idempotent).
/// After done(), further .input()/.stdin() calls error with BrokenPipe.
/// Zero prior .input() calls → valid empty output file.
pub fn done(&mut self) -> io::Result<&mut Self>;
/// Control whether cargo:rerun-if-changed is emitted from .input().
/// Default is auto-detected (true in build scripts, false otherwise).
pub fn rerun_if_changed(&mut self, value: bool) -> &mut Self;
}
```
If `.done()` is not called, `Drop` flushes the writer and prints any flush
errors to stderr.
For full details, including error handling semantics, `.done()` vs `Drop`
interaction, and internal algorithms, see [specs/lib-spec.md](specs/lib-spec.md).
No global state. No configuration objects.
---
# Build Script Integration
The intended use is:
```rust
use std::env;
fn main() -> std::io::Result<()> {
let out_dir = env::var("OUT_DIR").unwrap();
docextract::Docextract::new(format!("{out_dir}/macro-docs.md"))?
.input("src/macros.rs")? // auto-detects build script → rerun-if-changed
.done()?;
Ok(())
}
```
The builder automatically detects the build script context (`OUT_DIR` is set)
and emits `cargo:rerun-if-changed=src/macros.rs` via `println!` — no manual
`println!` needed. For multi-file extraction:
```rust
let mut doc = docextract::Docextract::new(format!("{out_dir}/macro-docs.md"))?;
doc.input("src/macros.rs")?;
doc.input("src/other.rs")?;
doc.done()?;
```
Each `.input()` call emits its own `cargo:rerun-if-changed` directive.
---
# Implementation Strategy
Use `BufRead::read_line()` with `BufReader` and `BufWriter`, both configured with **32 KiB buffers**.
Rust sources are always UTF-8. If invalid UTF-8 is encountered, `read_line()` returns an `Err(io::ErrorKind::InvalidData)` which is propagated — this is a hard error, not a recoverable condition.
For every line:
1. Skip leading ASCII whitespace (space 0x20, tab 0x09).
2. Check for the prefix `////`.
3. **Validate**: the character immediately after `////` must be space (0x20), tab (0x09), or end-of-line (i.e., `\n` 0x0A or `\r` 0x0D as part of CRLF). If it is anything else (including another `/`), the line is silently ignored — no output, no warning.
4. Remove the four slashes.
5. Remove one separator character (the space or tab after the slashes, if present; nothing to remove if the line ended immediately after `////`).
6. Emit remaining text.
7. **Preserve line endings** from the input. Whatever line ending the source uses (`\n` or `\r\n`), emit that same line ending in the output. Do not normalize.
Maintain a boolean indicating whether the previous emitted line belonged to the current documentation block.
When transitioning from one block to another, no blank line is emitted — blocks are concatenated directly. Block state tracking is still needed to distinguish lines in the same block from lines in separate blocks (separated by non-documentation lines), but no separator is emitted.
No regular expressions. No parsing beyond simple prefix matching.
---
# Error Handling
Return ordinary `std::io::Error`.
Do not introduce custom error types unless genuinely necessary.
CLI errors should be human-readable and concise.
---
# Dependency Policy
**Zero dependencies by default.** The `regex` feature (enabled by default)
adds `regex-lite` as an optional dependency. `regex-lite` itself has zero
transitive dependencies (maintained by BurntSushi under the `rust-lang/regex`
monorepo, dual-licensed MIT/Apache-2.0, MSRV 1.65.0). When compiled with
`--no-default-features` (disabling the `regex` feature), the crate has
**zero dependencies** — a hand-rolled pattern matcher is used instead.
If a dependency beyond `regex-lite` is genuinely needed, it must be confirmed
by the user before being added. Be very conservative.
The following are explicitly forbidden regardless:
* syn
* quote
* proc-macro2
* clap
* structopt
* pico-args
* regex
* anyhow
* thiserror
TODO: When the tag-filter feature is implemented, update `specs/lib-spec.md` and
`specs/cli-spec.md` to reflect the design changes documented in
`specs/tag-filter-spec.md` — particularly the cache-first parsing model,
block-separator removal, and the optional `regex-lite` dependency.
---
# Testing
Include unit tests covering:
* simple extraction
* indented extraction (spaces and tabs)
* optional separator removal (space and tab after `////`)
* consecutive documentation lines forming blocks
* multiple documentation blocks concatenated without auto-separation
* empty documentation lines (`////` and `//// ` — both produce verbatim blank line)
* `////Text` — no whitespace after slashes → silently ignored
* `/////` — five (or more) slashes → silently ignored
* files with no documentation → empty output
* empty file → empty output
* documentation at EOF (including `////` as last line)
* Windows (`CRLF`) and Unix (`LF`) line endings — both preserved in output
* tabs as indentation
* trailing space after `////` (e.g., `//// ` → empty line, the space after slashes is the separator)
* long documentation line
* whitespace-only documentation line (`//// ` — three spaces: the first is the separator, the remaining two are preserved verbatim)
* invalid UTF-8 / binary input → Err(io::ErrorKind::InvalidData)
* `-` as stdin input
* input file not found → warning to stderr, continue
* zero input arguments → empty output file
* `--help` → prints help, exits 0
* zero arguments → prints help, exits 0
* "missing output argument" is unreachable — the first argument is always the output; zero arguments prints help, exits 0
---
# Future Extensions (Not Part of Version 1)
Potential future additions include:
* recursive directory traversal
* `--extension`
* `--check`
* `--fail-on-empty`
* `--heading-from-file`
* `--strip-indent`
* configurable documentation prefixes
* JSON output
* Markdown table of contents generation
These are explicitly out of scope for the initial implementation.
---
# Packaging
**Cargo.toml metadata**: `license`, `repository`, and `description` fields will be added before publishing. A `README.md` will be written separately.
(End of file)