agentlanguages.dev
agentlanguages/catalogue/lom
verification camp

Lom.

An LLM-repair-native dynamic language: stable diagnostic codes, a confidence-graded fix engine with a written auto-apply safety axiom, and a repair-loop benchmark are the language's reason to exist.

authorwytyKen (with LLM co-authorship throughout)
implementationRust
targettree-walking interpreter and WASM (dual backends, byte-identical output); self-hosted subset compiler bootstraps to a byte-identical quine
licenceApache-2.0
first seenAugust 2026
maturityworking compiler
sitegithub.com
repolom-lang/lom
RepairLoopBenchRepairLoopBench
markdownlom.md

The thesis.

Most agent languages optimise how agents write code. Lom optimises what happens when the agent gets it wrong: the repair loop. The language ships with stable diagnostic codes across nine namespaces (plus an eight-family code space for its self-hosted subset compiler), a fix engine whose suggestions are graded high/medium/low, and a written safety axiom: auto-apply is restricted to high-confidence non-hint actions, and cognitive notices are never auto-rewritten, because rewriting them is a correctness regression, not a fix. That boundary is locked by an end-to-end fix corpus — thirteen pairs, five asserting fixed == bad, including the two pairs that pin the cognitive warnings (a deterministic name-collision notice and a closure-capture notice).

What it looks like.

A small dynamically typed language with a frozen surface (20 keywords, 43 builtins), run by a tree-walking interpreter and compiled to WASM, with both backends required to agree byte-for-byte. A self-hosted subset compiler written in the language itself (12,671 lines) regenerates itself as a byte-identical 272,925-byte quine — the repair-loop tooling it exercises is bootstrapped by that compiler, in-language.

Distinctive moves.

The repair loop is not a toolchain afterthought here — it is the reason the language exists.

Diagnostics are the deliverable: –check –json emits machine-readable diagnostics, fix –plan grades suggestions, fix –apply obeys the axiom, and codes –json exports the full registry as an agent bootstrap file. The repo also maintains ACRI, a small open specification for agent-compiler repair interfaces (object model, CLI protocol, benchmark methodology), with Lom as its reference implementation — and RepairLoopBench, whose repair tasks hand models verbatim diagnostics captured from the real compiler, machine-checked against it in CI.

lom fix broken.lom --plan
共 1 个诊断（1 个有可应用修复，0 个仅提示）

--- 诊断 #1 [NAM003] (error) ---
  位置: broken.lom:3:13
  消息: 未定义变量 'y'
  修复 (1):
    [1] 将 'y' 替换为 'x' [medium]
      action: replace at 3:13..3:14 text="x"
  retry: true

A real session, captured verbatim from the toolchain. The suggestion for the undefined-variable diagnostic is medium confidence, so fix --apply will skip it — the safety axiom grades what may be rewritten automatically versus advised. A wrong-variable rewrite that happens to compile is exactly the class of fix the axiom exists to refuse.

Maturity.

A working toolchain: dual backends, formatter, LSP, local package manifests, CLI fix tooling, and a browser playground. CI runs the full self-hosting regression on Linux and the build/test matrix on three platforms. The dependency and claim surface is machine-checked — a zero-third-party-crate CI gate, 368 self-compilation cases, a 14-case three-layer bootstrap check, 621 Rust tests — and the codebase contains no unsafe code.

Agent tooling.

Beyond the JSON surfaces in the frontmatter: the repair tasks in RepairLoopBench embed diagnostics captured verbatim from the live tool rather than hand-written, with a CI gate re-checking the embedded code multiset against live output — the same contamination discipline the benchmark card states plainly, including what the suite is not (not repo-scale, not comparable to SWE-bench-family leaderboards). The reported headline — full-set pass@1 of 98.3% / 98.5% across two frontier models, ten samples each — carries the qualification that raw model outputs are not archived in-repo; per-task pass matrices and methodology are.

design DNA
  • ZeroverificationZero frames its diagnostics as a repair contract and ships typed repair plans; Lom pushes the same idea into the language proper — suggestions are graded executable actions with an auto-apply safety axiom — and adds a published, contamination-safe benchmark for the loop.
  • VeraverificationVera attaches a fix hint and a spec reference to every diagnostic; Lom's suggestions are executable actions with confidence grades, and RepairLoopBench measures the loop end-to-end rather than asserting it.
  • ValeaunclassifiedThe closest blueprint on paper — structured JSON diagnostics plus an agent fix loop, written in Rust. Lom differs in shipping the loop with a safety axiom, a self-hosted byte-identical quine, and a reproducible benchmark.
  • X07syntacticX07 applies quickfixes deterministically at the toolchain level; Lom grades what may be auto-applied versus advised, and locks that boundary in a corpus of end-to-end fix pairs.