tla-kit: the full TLA+ toolchain in one reproducible container, plus a 14-tool MCP server
Scottie-2222 · reddit · 2026-11-19
tla-kit packages the TLA+ tool family (TLC, SANY, TLAPS, cvc5, ethos) into a single digest-pinned container image, solving the "proofs only pass on my machine" reproducibility problem.
Motivation: the author's project has 34 TLAPS proof modules (13.6k obligations) that CI couldn't rerun, because upstream rebuilt the rolling 1.6.0-pre tag under the same name, invalidating the pinned tarball.
Reproducibility design:
- Base image pinned by digest; apt from a dated Ubuntu snapshot.
- TLAPS tarball and cvc5/ethos builds hosted in the repo's releases, each sha256-checked.
- Every build of one commit yields the same image digest.
- Releases carry signed build-provenance and SBOM attestations, verifiable via gh attestation verify.
- PINS.json lists everything in the image; CI fails on mismatch, and tla-kit-selftest checks every digest in 30 seconds.
Contents: tla2tools.jar v1.7.4, tlapm 1.6.0-pre (with Z3, Zenon, Isabelle2025, LS4), cvc5 1.4.0 + ethos 0.2.4 (an SMT obligation counts as proved only when ethos accepts cvc5's proof), Temurin JRE 25.
MCP server: tla-kit mcp runs over stdio using only Python's standard library, exposing 14 tools including parse (SANY errors with locations), symbols, evaluate (TLC constant expressions), tlcrun (structured JSON counterexamples instead of trace text), simulate, coverage (flags actions that never fire), stategraph, proofcheck, and prove, designed to pair with a writing-tla-specs-and-proofs skill.
More from coding & agent
- Engineer uses AI agent to reverse-engineer Dell 6K monitor firmware and fix blinking-light bug — davidcrawshaw · 2026-11-19
- Antigravity Robotics plugin lets anyone build MuJoCo sim tasks with prompts — DynamicWebPaige · 2026-11-19
- Is your MCP server off-spec? Sending a nonexistent method should return JSON-RPC -32601 — Paoli99 · 2026-11-19
- 5 steps to keep your own voice when writing with AI — EXM7777 · 2026-11-19
- Resolve launches SRE-1, a 27B RL-trained model for root-causing production incidents — tianyin_xu · 2026-11-19
- Soleio ships two Grok agents: Matcher chases intros, Polo auto-drafts inbox replies — soleio · 2026-11-19