rsaeb

rsaeb is a Rust 2024 no_std + alloc, byte-oriented interpreter for A=B ordered rewrite programs.

A=B: https://store.steampowered.com/app/1720850/AB/

Unofficial Project Notice

This project is an unofficial, independently developed interpreter library. It is not affiliated with, endorsed by, or maintained by Artless Games or the original A=B author.

A=B's compact lhs=rhs ordered rewrite system is an unusually elegant programming-puzzle idea. This crate exists because that design is worth studying, testing, and reimplementing. If this interpreter interests you, please support the original game.

Documentation Map

• This README is the package entry point. It explains the interpreter shape, the accepted A=B surface, byte-domain boundaries, and release checks.
• The generated rustdoc is the exact API reference and carries the complete doctested public examples.
• The GitHub Wiki is a short navigation layer for use cases and embedding boundaries.

The crate root intentionally does not re-export duplicate type paths. Public types live under their domain modules, such as source, input, program, limits, execution, inspect, trace, and error.

Quick Start

Parse source into an immutable Program, validate runtime input, admit it into a run seed with explicit execution limits, then run:

use rsaeb::input::{RunSeed, RuntimeInput, RuntimeInputSource};
use rsaeb::limits::{
    DEFAULT_MAX_INPUT_LEN, DEFAULT_MAX_RETURN_LEN, DEFAULT_MAX_STATE_LEN,
    DEFAULT_MAX_STEPS, DEFAULT_PARSE_LIMITS, ExecutionLimits, RuntimeInputLimits,
};
use rsaeb::program::{Program, RunOutcome};
use rsaeb::source::ProgramSource;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let program = Program::parse(ProgramSource::from_text("a=b"), DEFAULT_PARSE_LIMITS)?;
    let input_limits = RuntimeInputLimits::new(DEFAULT_MAX_INPUT_LEN);
    let execution_limits = ExecutionLimits::new(
        DEFAULT_MAX_STEPS,
        DEFAULT_MAX_STATE_LEN,
        DEFAULT_MAX_RETURN_LEN,
    );
    let input = RuntimeInput::validate(RuntimeInputSource::from_bytes(b"a"), input_limits)?;
    let seed = RunSeed::admit(input, execution_limits)?;
    let result = program.run(seed)?;

    if !matches!(
        result.outcome(),
        RunOutcome::Stable(output) if output.as_slice() == b"b"
    ) {
        return Err("unexpected stable output".into());
    }

    Ok(())
}

ProgramSource::from_text and ProgramSource::from_bytes only label source input; Program::parse performs source validation. RuntimeInputSource and RuntimeInput::validate do the same for runtime input bytes. Reuse parsed programs freely: a Program is immutable, and (once) consumption is local to each execution.

Execution Shape

The normal host flow is:

1. Load source bytes or text outside the interpreter.
2. Construct ProgramSource.
3. Parse with Program::parse.
4. Label host input bytes with RuntimeInputSource::from_bytes.
5. Validate with RuntimeInput::validate.
6. Admit with RunSeed::admit and ExecutionLimits.
7. Execute through run-to-completion, stepwise execution, tracing, or rule-attempt stepping.

The crate intentionally contains no filesystem, process, argument parsing, environment access, stdout/stderr, or lossy display boundary. Hosts perform I/O outside the interpreter and pass already-loaded bytes into typed boundaries.

Program::run is the borrowed run-to-completion API. Program::start_run is the borrowed stepwise API for hosts that keep a reusable parsed program. Program::into_run is the explicit ownership-transfer stepwise API for cases where the execution session must own the parsed program. Rule-attempt execution is separate: Program::start_rule_attempt_run and Program::into_rule_attempt_run observe executable rule-line attempts, including misses, without changing normal committed-step semantics.

The exact typestate names, transition variants, owned recovery methods, tracing events, and error variants are documented in rustdoc.

A=B Language Reference

A program source is a byte sequence containing one rewrite rule per non-empty code line:

lhs=rhs

Each line is parsed in this order:

1. # starts a comment. Everything from # to the end of the line is ignored.
2. Non-ASCII bytes are rejected in the remaining code part.
3. ASCII whitespace in the code part is removed completely.
4. Remaining non-whitespace code bytes must be printable ASCII.
5. Empty compact code is ignored.
6. Non-empty compact code must contain exactly one =.
7. The left side and right side are parsed as compact rule syntax.

Examples:

a=b# this is parsed as a=b
#a=b  this whole line is a comment
a b = b b  # this is parsed as ab=bb

Comments may contain arbitrary non-ASCII or non-UTF-8 bytes when source is provided with ProgramSource::from_bytes. Executable code outside comments must be ASCII. ASCII control bytes are invalid in executable code except for ASCII whitespace that is removed during compaction.

Parse error columns are one-based byte positions in the original source line before whitespace compaction. Diagnostics point at the user's source text, not at the internal compacted representation.

Reserved Characters

The following characters are reserved in program code:

= # ( )

Their meanings are fixed:

• = separates the left side from the right side.
• # starts a comment.
• ( and ) are only allowed as part of supported modifier/action tokens.

A second = in compact code is a parse error:

a=b=c

A second = inside a comment is ignored:

a=b#=c

Reserved syntax where payload data is expected is always a parse error:

a=b(
a=b)
a=b()
a=()
a=b(start)
a=(once)b
a(once)=b

Because whitespace is removed from program code, spaces cannot be represented as rule data. Because =, #, (, and ) are reserved, program payloads also refuse them as rule data.

Runtime input is different. Input bytes are runtime data, not program code. Input must be ASCII, but it may contain whitespace, ASCII control bytes, and reserved characters. Ordinary rewrite actions cannot match, create, or delete those bytes directly.

program: a=b
input:   a=()#c
output:  b=()#c

Rules cannot match across preserved runtime-only bytes:

program: ab=bb
input:   a bc
output:  a bc

(return) stops execution and replaces the final output with its return payload, so runtime-only input bytes are not preserved after a matching return rule:

program: a=(return)x
input:   a=()#c
output:  x

Left-Side Modifiers

The left side may start with one repeat modifier and one anchor modifier:

• (once): the rule may be used at most once per runtime execution.
• (start): the rule only matches at the start of the current state.
• (end): the rule only matches at the end of the current state.

Supported modifier order is (once) first, then an optional anchor. Duplicated or unsupported left-side modifier order is a parse error.

Examples:

a=b
(once)a=b
(start)a=b
(end)a=b
(once)(start)a=b

Because code whitespace is ignored, this is also valid and equivalent to (once)(start)a=b:

( once ) ( start ) a = b

Right-Side Actions

The right side selects the action for a matching rule:

• text: replace the matched left side with text.
• (start)text: remove the match and insert text at the start of the state.
• (end)text: remove the match and append text to the end of the state.
• (return)text: stop execution immediately and output text, discarding the current runtime state.

The action payload is still program data, so it cannot contain whitespace, reserved characters, non-ASCII bytes, or ASCII control bytes. (return) can therefore output only program-representable bytes, even if the discarded runtime state contained spaces or reserved characters from the original input.

Examples:

a=b
x=(start)y
x=(end)y
x=(return)y

Empty Sides

The left side and right side may be empty.

An empty right side deletes the matched left side:

a=

An empty left side matches an empty byte sequence. For unanchored rules and (start) rules, it matches at the start of the current state:

(once)=x

With input ab, this inserts x at the start and produces xab.

For (end) rules, an empty left side matches at the end of the current state:

(once)(end)=x

With input ab, this inserts x at the end and produces abx.

An unanchored empty-left rule without (once), (return), or some later rule that makes execution stop can rewrite forever until the step limit is reached. That is legal syntax; execution remains governed by ExecutionLimits.

Ordered Execution

Execution is ordered and single-step.

On each step, the runtime scans rules from top to bottom and applies the first rule that matches the current state. For an unanchored non-empty left side, the leftmost match in the current state is used. After one applied step, scanning restarts from the first rule.

Example:

program:
aa=x
a=y

input:
aaaa

output:
xx

The first rule is preferred over the second rule, and each application rewrites the leftmost matching aa.

Byte-Domain Boundary

Program source and runtime input are deliberately different byte domains:

• Program code is compact printable ASCII syntax.
• ASCII whitespace in program code is ignored before parsing.
• # starts a comment for the rest of the source line.
• Comments may contain non-ASCII or non-UTF-8 bytes.
• Executable code outside comments must be ASCII.
• Program payloads cannot contain whitespace, =, #, (, ), non-ASCII bytes, or ASCII control bytes.
• Runtime input is ASCII data and may contain spaces, ASCII control bytes, and reserved syntax bytes.
• Normal rewrites preserve runtime-only bytes that program code cannot construct or match.
• (return) stops execution and replaces the whole output with its return payload.

Internally, parser and runtime phases stay separate instead of passing raw byte buffers through every stage:

raw line bytes
  -> RawSourceLine
  -> CodeLine                # comment removed, executable code ASCII validated
  -> CompactCodeLine         # whitespace removed, SourceColumn retained
  -> NonEmptyCompactCodeLine # empty compact lines cannot enter rule parsing
  -> RuleSyntaxLine          # exactly one '=' has been proven
  -> LeftSyntax / RightSyntax
  -> ProgramByte             # bytes that program code may construct and match

runtime input bytes
  -> AsciiByte         # runtime input domain validation
  -> RuntimeByte       # private ProgramConstructible(ProgramByte) or Opaque(NonProgramAsciiByte)
  -> execution session # consumes RuntimeInput and owns mutable execution state

Program payloads are stored as ProgramByte, not raw u8. Runtime state is stored as RuntimeByte: payload-compatible input and rule output become editable program bytes, while whitespace, control bytes, and reserved syntax bytes from input become opaque ASCII bytes. Ordinary rules match only editable bytes. Opaque input bytes are preserved by surrounding rewrites but cannot be directly matched, created, or deleted by program payloads.

Public observation crosses explicit materialization boundaries. Runtime state views materialize to snapshots only when requested, stable run results own final state bytes, (return) outputs use a separate return-output domain, parsed payload inspection materializes explicitly, and snapshot tracing has its own byte limit. During execution, the active state and rewrite scratch buffer remain separate typed buffers until a successful continuation step commits.

(once) rules carry private slots assigned during parsing. Each execution allocates only those slot states, and only a committed application can consume its slot.

no_std + alloc Boundary

The library crate is #![no_std] and uses alloc only at owned-buffer boundaries such as parsed rules, runtime input validation, per-run (once) state, run results, canonical rule source, explicit view materialization, and trace snapshots. It requires an allocator, but not std.

Allocation is explicit and fallible. Parser/runtime paths reserve explicitly and report AllocationError instead of relying on accidental Vec growth. Runtime expansion is budgeted through ExecutionLimits; the runtime checks size limits before allocating oversized states or return outputs. Step budget is reserved before rewrite or return-output materialization, so an exhausted step limit cannot allocate a candidate state or return buffer. Trace snapshot materialization is budgeted separately through TraceSnapshotByteLimit.

Owned public values that contain byte buffers intentionally do not implement Clone; copying bytes is an explicit materialization step, not a hidden infallible API. Parser payload validation is reported before payload storage allocation, so invalid source bytes are not hidden behind allocation failures.

A downstream std application can use the library normally. A downstream no_std application must provide an allocator before calling APIs that allocate.

Error Model

The library error model is intentionally split. Parse errors, runtime input errors, run-admission errors, runtime execution errors, allocation errors, and trace materialization errors have separate structured types under rsaeb::error.

Allocation failures preserve the allocation boundary as AllocationContext. Reservation failures also report a typed RequestedCapacity, so hosts can distinguish failures while validating input, materializing state views, building canonical rule source, producing final output, or retaining trace snapshots without parsing display strings.

Configured byte budgets and step budgets are reported through concrete errors such as ParseLimitError, RuntimeStateLimitError, ReturnOutputLimitError, StepLimitError, and RuleAttemptLimitError. Trace snapshot byte limits are reported through TraceSnapshotError, because snapshot materialization is outside runtime execution.

Filesystem failures are not part of the library error model. External I/O must be handled before bytes enter ProgramSource::from_bytes, ProgramSource::from_text, or RuntimeInputSource::from_bytes.

Development Checks

Run the public documentation and package checks before publishing changes:

rustup target add thumbv7em-none-eabihf
cargo fmt --check
cargo check --lib --no-default-features
cargo check --lib --all-features --target thumbv7em-none-eabihf
cargo clippy --all-targets --all-features -- -D warnings
cargo test --all-targets --all-features
cargo test --doc --all-features
latest_rlib="$(find target/debug/deps -maxdepth 1 -name 'librsaeb-*.rlib' -printf '%T@ %p\n' | sort -nr | awk 'NR == 1 { print $2 }')"
rustdoc --edition=2024 --test README.md -L dependency=target/debug/deps --extern "rsaeb=${latest_rlib}"
RUSTDOCFLAGS="-D warnings" cargo doc --all-features --no-deps
cargo package --list
cargo package