Tacit.

The thesis.

Tacit’s diagnosis is that text-based source forces models to maintain stylistic conventions, name choices, and whitespace that the compiler immediately discards. The project’s response is to make the AST the canonical artefact and treat text as a derived view. The repository’s README states the position directly: “The AST is the source of truth. Tacit does not treat a human-oriented surface syntax as the authoritative program representation.”

If human readability is not the primary constraint, a language can optimise for three things at once.

Two views render the same tree: a dense authoring view sized for model token budgets, and an inspection view layered for debugging and human review. Both round-trip losslessly through a JSON sidecar (.tacd). Canonical text is byte-exact — there is exactly one serialisation per AST, which eliminates formatter debates and makes hashing meaningful. Definitions are identified by the BLAKE3 hash of their canonical text; display names live in the sidecar and carry no semantic weight. Variable references use DeBruijn indices in canonical form. The cross-camp move is the verification-adjacent effect lattice: Tacit-Lite tracks IO / Alloc / Mut / Div in signatures, and unit boundaries must declare type and effect rows explicitly.

What it looks like.

unit Math {
  import inc : Int -> Int
    from blake3:0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef;

  private double : Int -> Int =
    lambda x. @add x x;

  export public add_two : Int -> Int =
    lambda x. inc (inc x);
}

Imports name exact blake3:<64-hex> definition hashes, not paths or version ranges. Visibility (public / package / private) is part of the artefact. The display names x, inc, double are sidecar metadata; in canonical form references are DeBruijn indices.

Distinctive moves.

• AST as authoritative source. The on-disk .tac file is the byte-exact canonical projection of the tree; comments and free-form formatting are not in the language. Names, field order, and type/effect hints live in a .tacd JSON sidecar.
• Content-addressed definitions. Imports resolve to BLAKE3 hashes, not names. Renaming a definition leaves its hash unchanged; changing its signature, body, or referenced hashes produces a new identity. Package manifests pin a hash-indexed local cache.
• Typed Hole nodes instead of parse failures. Malformed code reduces to a Hole node with a structured diagnostic and a type slot, so downstream tools can keep operating on partial programs (ADR 0040, landed in Phase 2).
• Explicit effect rows on boundaries. Tacit-Lite’s fixed lattice is IO / Alloc / Mut / Div; effect signatures are mandatory at unit exports and inferred locally elsewhere. Tacit-Full (refinement types, capability-based security, handlers) is reserved as a roadmap, not shipped.
• Toolchain pin as a first-class artefact. tacit init writes a tacit-toolchain.toml that pins toolchain, primer, and bundled stdlib hashes; every package-aware command refuses to run on a mismatched pin and surfaces a toolchain-pin-* diagnostic.

Maturity.

A Rust workspace (five crates: tacit-canonical, tacit-views, tacit-typecheck, tacit-codegen, tacit-cli) at v0.7.7, released 19 May 2026. Apache-2.0 / MIT dual-licensed; 237 commits, 3 stars, 2 forks at time of cataloguing. The decision log runs to ~90 ADRs; Phase 6 was frozen by ADR 0089 on 2026-05-17, closing modules/units, package manifests with hash-pinned lockfiles, package tests with stable tacit-test-v1 JSON output, fixed-width integers with wrapping/checked/saturating arithmetic, typed mutable-memory handles, source-level stdlib packages (tacit.core, .bytes, .array, .text, .collections, .io), and a constrained host-interface ABI with generated C headers and Rust bindings. A Rust embedding demo links a Tacit kernel as a static library. Phase 7 is the next planned phase; debugger, diff/blame, IDE, public registry, and arbitrary FFI are explicitly out of scope until a later ADR. LLVM 19 is pinned via inkwell; published release artefacts target Linux x86_64 with a glibc 2.35 floor.

Agent tooling.

AGENTS.md (1.7 KB) carries the Codex-facing sealed-corpus guardrail and a pointer to CLAUDE.md. CLAUDE.md (~20 KB) functions as a full development guide rather than a SKILL.md — it enumerates frozen artefacts, ground rules, the file-extension contract (.tac / .tacd / .taca), and the per-phase delivered surface. The toolchain ships its own primer: tacit primer prints the byte-pinned Tacit-Lite primer, and tacit primer --search / --list-sections / --section <id> support selective disclosure designed to fit a model’s context window without flooding it. Diagnostics, package tests, and tacit version all emit stable JSON.