New project — plainspeak, an esoteric language with English sentence grammar

plainspeak

A new project, built in a single day (2026-08-10, 40 commits start to finish): a deterministic, non-ML esoteric programming language whose surface syntax is a constrained subset of English. Say "Hello, world!". and Set total to 0. are programs, not pseudocode — sentences are the unit of syntax, not symbols.

the pipeline, and the "no LLM" rule

plainspeak source (.eng) goes through a lexer, sentence splitter, recursive-descent/Pratt parser, semantic analysis, and a C code generator, then hands the emitted C99 off to whatever system C compiler is available (cc/gcc/clang) for the native binary. The compiler frontend itself is C++20; that's an implementation detail, not part of the language's public surface.

AGENTS.md is explicit about the one invariant the whole design hangs off: there is no LLM, no statistical model, and no fuzzy/edit-distance matching anywhere in the toolchain, including as a dev tool or fuzzer oracle. "Sounds like English" is a property of a fixed grammar, not of runtime behavior — every valid program is accepted by that grammar, and every rejection is a precise parse error naming the offending word and its source location. Synonym support (set / let / make all resolving to one canonical Set token) is a flat alias-table lookup, never a heuristic.

what's actually in the grammar

By the end of the day the language covers: Say/Set/Add/Subtract/Read/ReadFloat, # and comment-statement comments, Repeat/If/Else/While blocks, Call/Procedure/Return, boolean literals, and a full expression grammar with and/or/not, six comparison operators, plus/minus/times/divided by/mod, to the power of, parentheses, unary minus, and a math-function set (square root of, absolute value of, sine/cosine/tangent/sqrt/log/abs/floor/ceil). Two types exist beyond string: number (a C long) and decimal (a C double), with mixed number/decimal arithmetic promoting to decimal and plus also doing string concatenation against any type.

Every one of those is backed by a golden test (.eng source compiled, run, and diffed against a .expected stdout file) — docs/grammar.md states plainly that a parser accepting something the grammar doc doesn't describe is a bug, and there's a rule that grammar changes and docs/grammar.md changes land in the same commit.

errors have a catalogue, not just messages

Diagnostics carry stable E00xx codes registered in docs/errors.md rather than one-off inline strings — E0001 undeclared variable, E0002/E0003 type mismatches in plus/Add, E0007/E0008 undefined procedure and wrong argument count, and so on. The style guide asks for messages a first-time reader would understand from the grammar docs alone, e.g. pointing at the exact word plainspeak didn't recognize and listing what it expected instead. Negative golden tests under tests/golden/errors/ pin eight of these end to end: undeclared vars, type mismatches in comparisons and not/plus, missing Else colons, undefined procedures, and wrong arg counts.

building it out: floats and math

The last five commits of the day form one coherent feature arc: adding Double/float support from the ground up. A Float token and LParen/RParen land in the lexer; FloatLit, MathCallExpr, PowExpr, ReadFloatStmt, and unary-minus land in the AST; the runtime gets a PS_DOUBLE tag plus ps_sin, ps_cos, ps_tan, ps_sqrt, ps_log, ps_abs, ps_floor, ps_ceil, and ps_pow; sema learns mixed int/decimal promotion and type-checks math calls and power expressions; codegen emits calls into all of it; and a batch of golden/unit tests locks the whole thing down. It closes with examples/calculator.eng, a 15-operation menu-driven scientific calculator (add through ceiling, with divide-by-zero and negative-sqrt/log guards) that exercises effectively the whole language at once — floats, math functions, ReadFloat, procedures, loops, and conditionals together.

status

Working v0 scaffold: ~1,900 lines across the compiler and runtime, CMake as the sole build system (with a fallback raw g++ invocation documented in the README for anyone without CMake), Catch2 for unit tests, and a shell-script golden-test runner. No arrays, lists, or user-defined types yet, and string concatenation results are heap-allocated and never freed — both called out explicitly in docs/grammar.md as known v0 gaps rather than oversights.