LemmaScript (Tech Preview)

A verification toolchain for TypeScript. Write ordinary TypeScript with //@ specification annotations. The toolchain generates verifiable code from your TypeScript — either in Dafny or Lean 4 (with Velvet/Loom).

See SPEC.md, DESIGN.md, and GETTING_STARTED.md.

This is a Tech Preview: the core idea is there, but support, semantics, and ergonomics are still evolving.

See our blog post.

Examples and Case Studies

Each example and case study is verified in Lean 4 and/or Dafny from the same annotated TypeScript source.

See the internal examples.

See the external case studies:

• collab-todo-lemmascript — collaborative task management web app (React + Supabase) with a verified domain model. Single domain.ts imported directly by the UI, hooks, and edge functions — no adapter layer. 123 Dafny lemmas (120 in a separate domain.proofs.dfy): 16-conjunct invariant preserved across 25 single-project + 3 cross-project actions, NoOp completeness/soundness, initialization. Dafny only.
• colorwheel-lemmascript — verified color palette generator with mood + harmony constraints. 31 Lean proofs + 18 behavioral properties, 115 Dafny lemmas (invariant preservation, commutativity, NoOp completeness).
• clear-split-lemmascript — greenfield verified expense splitting web app. Conservation theorem, invariant preservation, delta laws — all proven in both Lean (no sorry) and Dafny (56 lemmas).
• github-star-checker-lemmascript — small verified CLI that tracks GitHub star counts across repos and reports per-run deltas. Verifies: per-row diff correctness and totalDiff == sumDiffs(rows) (via an inductive SumDiffs_append lemma); three sign-classified extractors (gainers / losers / unchanged) with soundness, completeness, ordered completeness (gainers appear in the notification in the same order they were listed on the command line), and count/sum equalities against prefix-indexed upTo helpers; conservation theorem decompose(r) — the three splits partition every row exactly once, and sumDiffs(increases) + sumDiffs(decreases) == totalDiff. 33 Dafny VCs, 0 errors; proof additions include a head/tail bridge (sumDiffs ↔ sumDiffsUpTo) and two partition-on-n inductions. Dafny only.
• equality-game-lemmascript — greenfield verified arithmetic equality card game (React + Tailwind). Sound + complete decision procedure for "can these two card lists be combined into equal expressions": canEqualize(L, R) ⟺ ∃ eL, eR. eval(eL) == eval(eR) ∧ multiset(leaves(eL)) == multiset(L) ∧ same for R. Algorithm is subset-DP over a bitmask m ∈ [1, 2^n − 1); the proof composes a PopCount upper/lower bound chain (with stdlib LemmaDivDenominator / LemmaFundamentalDivModConverse), a splitLeft/splitRight ↔ imperative-loop connection, a WitnessCombine lemma threading existential Expr witnesses through the cross-product loops, and a ChooseMask combinatorial constructor that, given any sub-multiset of cards, produces the realizing mask. Capped by CompletenessFromMaskCoverage. 753 verification conditions, 0 errors, 0 assumes, 0 axioms under --isolate-assertions --verification-time-limit 180. Dafny only.
• talktimer-lemmascript — verified talk timer React app, ported from a Dafny-only talktimer-lemmafit twin. 17-variant Action state machine + verified History (undo/redo/preview/commitFrom) all in one domain.ts — the original Dafny's Domain refines Kernel abstract-module pattern inlined since LS has no abstract modules. 108 VCs in domain.dfy (invariant preservation) + 123 in domain.proofs.dfy (behavioral lemmas + Kernel round-trip). Dafny only.
• node-casbin-lemmascript — brownfield verification of node-casbin. 5 functions verified, 217 existing tests pass. End-to-end correctness and order independence for all 4 effect modes in both Lean and Dafny (39 lemmas).
• hono-lemmascript — brownfield verification of hono's security middleware. Four CVEs covered: IP restriction bypass (CVE-2026-39409) and cookie name bypass (CVE-2026-39410) — 51 Dafny lemmas, cookie verification in-place; plus serveStatic's URL-encoded directory traversal (CVE-2024-32869) + repeated-slash bypass (CVE-2026-39407), proved as a composition — decode(rawPath) before check(decoded), so a buggy implementation that reordered the steps would fail the proof. First use of //@ assume + //@ havoc-on-assign. Dafny only.
• charmchat — brownfield verification of an AI agent orchestration backend. isEmptyResult (string emptiness predicate, 8 postconditions, <1s) and topologicalSort (Kahn's algorithm — memory safety, output bounds, completeness via acyclicity ranking witness, termination). Full completeness proof: 23 helper lemmas, 14 opaque ghost predicates, 115 loop invariants; 736 VCs verified under --isolate-assertions --verification-time-limit 600. Key technique: snapshot-based inner invariants (ghost var originalRemDeps := remDeps) replace the mid-iteration SEEN/UNSEEN split so preservation is frame reasoning against a ghost-constant rather than set-subtraction against mutating state. Dafny only.
• xyflow-lemmascript — brownfield verification of xyflow's core edge and geometry utilities. 9 functions verified: addEdge (dedup — never loses edges, adds at most one), reconnectEdge (semantic: under a unique-id precondition, the result is in-place — |result| ≤ |edges|, no insertion — and when a matching edge existed with non-empty new endpoints, the output contains an edge with those endpoints. Uses //@ assume to characterize destructuring, find, and the constructed edge), connectionExists, getEdgeCenter (midpoint correctness), clamp (bounds), rectToBox/boxToRect (field arithmetic), getBoundsOfBoxes (enclosure), getOverlappingArea (non-negative), areSetsEqual (subset + same size). 14 Dafny proof obligations. Dafny only.
• rallly-lemmascript — brownfield verification of rallly's meeting-poll Next.js app. 2 functions: validateRedirectUrl (in-place — open-redirect predicate; non-undefined outputs start with / but not //) and scorePoll (extracted ranking core — length preservation, score bounds, top-choice characterization, score-formula equality, within-poll monotonicity, tiebreaker injectivity). The injectivity proof surfaced a real spec-level constraint on the existing (yes + ifNeedBe) * 1000 + yes encoding: it overflows when an option has ≥ 1000 yes votes. 10 Dafny VCs, 0 errors. Drove four toolchain additions: s.startsWith(), T | null nullability, \result narrowing under ==>, Math.max(...arr) spread. Dafny only.
• opencode-lemmascript — brownfield verification of opencode's permission system and unified-diff patch parser. Highlights: (1) Patch.parsePatch carries conservation loop invariants over local ghost state — a parser bug here would silently corrupt user files when an AI applies a patch, and (2) the permission-engine work mechanically closes opencode bug #26514 (subagents bypassing Plan Mode's file-edit restrictions). 9 functions verified in-place, 0 errors. Dafny only.
• pi-lemmascript — brownfield, in-place verification of the context-compaction cut-point selector in pi (the earendil-works agent harness). When the context window fills, pi discards history before a chosen cut; a provider API rejects a retained prefix containing an orphaned toolResult (a tool result whose tool call was cut away). Both selector functions proven: the cut never lets the kept suffix start with — nor split a tool-use/tool-result run into — an orphaned tool result, even across the backward metadata snap. The no-orphan result forced the session tree's tool-pairing ordering into an explicit requires. 4 VCs, 0 errors. Drove five toolchain additions, headlined by an opaque fall-through type: a union LemmaScript can't discriminate (here an array-element union of unreachable imports) becomes a single opaque type — the field stays present so distinct values stay distinct, and with no constructor or tag predicate it can only be passed through, never unsoundly observed. Dafny only.
• balanced-match-lemmascript — brownfield, in-place verification of balanced-match, the ~70-line balanced-bracket finder pulled in by npm, webpack, and most of the JS tooling stack (1B+ downloads/month). The stack-based range core is verified by refinement: a pure recursive spec range_spec mirrors the loop one branch per recursive case, so the single equivalence range == range_spec transfers every property automatically — including an unconditional Dyck body-balance theorem for the interior of every returned pair. 2233 VCs, 0 errors under --isolate-assertions (registered on the dafny-slow track). Dafny only.
• guardians-lemmascript — greenfield verification of the core safety argument behind Guardians (Erik Meijer, "Guardians of the Agents", CACM Jan 2026), a generate-verify-execute checker for AI-agent workflows. Instead of verifying an app, it proves the agent guardrail itself sound — that a static taint/automaton check over the real recursive workflow AST can never admit an unsafe plan. Highlights: taint over nested conditionals as a sound branch-union over-approximation; per-source provenance with a join (a multi-input tool is tainted if any input was); unbounded loops discharged by a one-step pre-fixpoint (sat = t0 ‖ bodyTaint(t0)) that bounds taint over any iteration count without iterating to a fixpoint; and a unified capstone (verifyWfSound) — one clean verdict rules out, on every path, both a tainted-data-to-sink leak and a security-automaton error. 54 Dafny obligations, 0 errors. The verified cores are reached from a Guardians-style Workflow/Policy through a thin unverified adapter, differentially tested against the real Python Guardians (used as the oracle, not a porting target). Dafny only.
• quorum-lemmascript — greenfield verified when2meet/Doodle-style group scheduler (React + Cloudflare Durable Objects), with one domain.ts running unchanged in the browser, the in-app query, and the server. The standout is that the proof licenses the architecture: countFree is a homomorphism from participant-list concatenation to integer addition (so the heatmap is order-independent) plus same-participant last-writer-wins convergence — which is exactly what makes the lock-free, no-login, optimistic multi-device backend safe, with the Durable Object and the browser applying the same verified applyOp (server-authoritatively, client-optimistically) with no rollback or operational transform. Also: heatmap is exactly the per-slot count and isBest exactly its argmax; monotonicity; invariant-preserving mutations + op-log replay; a sparse export codec round-trip; an in-app whoIsFree(e, s) whose length provably equals the cell's count; and a separate grid.ts proving the (day, time) → slot map in-range + injective — which makes specific-dates-vs-days-of-the-week pure shell labeling at zero proof cost; and full element-level permutation invariance (heatmapPermInvariant — the heatmap depends only on the multiset of participant rows), which drove the perm(...) spec predicate into LemmaScript itself. 100 Dafny VCs (90 + 10), 0 errors. The aggregate is proven; the React UI, WebSocket/DO I/O, and timezone labeling are the stated trust boundary. Dafny only.
• quota-lemmascript — greenfield verified booking app: providers publish a page of limited-capacity featured slots, signed-in users grab them (React + Cloudflare Durable Objects + D1). A deliberate inverse of quorum-lemmascript: where Quorum counts up to a threshold over data partitioned per participant (no conflicts ⇒ optimistic, lock-free, no rollback), Quota's bookings contend for shared inventory, so the load-bearing fact flips from a count to a bound — for every slot j, confirmedCount(bookings, j) <= slots[j].capacity — and so does the concurrency story: the same domain.ts runs in the browser and the Durable Object, but server-authoritatively (it never oversells under contention; no optimistic client apply). Proven: no overbooking (invariant preservation across tryBook / cancel), accept-iff-room, an idempotent three-way tryBook (a retry reads as success, not rejection — only "confirmed" mutates), cancellation frees seats, replay determinism, and full order-invariance of availability under contention — confirmedCountPerm / hasRoomPermInvariant show availability depends only on the multiset of the booking log (any reordering, not just a pairwise swap), via the same perm(...) predicate Quorum drove into LemmaScript. NDJSON export is built on the verified confirmedOnly. The counting kernel is Quorum's countFree re-pointed at bookings — total and precondition-free so it composes. 80 Dafny VCs, 0 errors. Trust boundary (stated plainly): auth, the React UI, WebSocket/DO/D1 I/O, email, slot date/time labeling, and abuse/rate-limiting. Dafny only.
• henri-lemmascript — a runnable TypeScript coding-agent CLI (a port of henri; multi-provider, Anthropic + AWS Bedrock) whose security- and protocol-critical core is verified and imported directly by the live agent — decide() gates every real tool call and the conversation invariant is asserted on every turn (it streams end-to-end against Bedrock). Three modules, 48 Dafny VCs, 0 errors. (1) Permission gate: soundness (decide == Allow ⟺ isAllowed), path-traversal containment — auto-allow-in-cwd can never resolve outside cwd, with ./.. normalization proved in-core so the shell is trusted only to resolve().split('/') — grant monotonicity, and rejectPrompts is deny-only. (2) Conversation protocol: tool-call/result pairing plus the pi-lemmascript-style no-orphaned-tool_result property, proved as an invariant preserved by the loop — wellFormed(msgs + [assistant(calls), tool(makeResults(calls))]) — not checked after the fact. (3) Hook/config merge: removal, tool-name uniqueness — a fix (henri concatenated hook tool lists with no dedup, so two hooks could shadow a name), order-independence, and additivity composed cross-module with the gate's monotonicity (merging only grows the allow-sets, which by P3 never revokes an Allow). The merge is verified in place via //@ declare-type Tool { name: string }, shadowing the real Tool's function-valued execute so the actual mergeTools(Tool[]) is the proof target rather than a parallel model. Dafny only.

Setup

Prerequisites: Node.js >= 18. For the Lean backend: elan. For the Dafny backend: Dafny >= 4.x.

Install from npm:

npm install lemmascript

Or from source:

git clone https://github.com/midspiral/LemmaScript.git
cd LemmaScript && npm install && npm run build

Lean backend additionally requires the Loom and Velvet forks:

git clone https://github.com/namin/loom.git -b lemma ../loom
git clone https://github.com/namin/velvet.git -b lemma ../velvet

Usage

Dafny backend

npx lsc gen --backend=dafny src/myModule.ts
npx lsc check --backend=dafny src/myModule.ts
npx lsc regen --backend=dafny src/myModule.ts

The Dafny backend generates two files per TS source: foo.dfy.gen (always regeneratable) and foo.dfy (source of truth, with LLM/user proof additions). The diff between them must be additions-only.

Lean backend

npx lsc gen --backend=lean src/myModule.ts
lake build

Annotations

//@ requires arr.length > 0
//@ ensures \result >= -1 && \result < arr.length
//@ invariant 0 <= i && i <= arr.length
//@ decreases arr.length - i
//@ type i nat

For the full surface, see SPEC.md.

File Structure

Dafny backend

File	Generated?	Purpose
.ts	—	TypeScript source with //@ annotations
.dfy.gen	Yes	Generated Dafny (merge base, always regeneratable)
.dfy	Yes (initial)	Annotated Dafny (gen + proof additions)

Lean backend

File	Generated?	Purpose
.ts	—	TypeScript source with //@ annotations
.types.lean	Yes	Lean types, namespace Pure defs
.spec.lean	No	Ghost definitions, helper lemmas
.def.lean	Yes	Velvet method definitions
.proof.lean	No	prove_correct with proof tactics