specula-org/Specula

Specula: An agentic tool for finding deep bugs in system code using TLA+

What it solves

Specula finds deep, complex bugs in concurrent and distributed system code that are typically difficult to detect through traditional testing. It automates the process of formal verification by bridging the gap between high-level system specifications and actual code implementation.

How it works

Specula uses AI coding agents to analyze a target system's source code and automatically write TLA+ specifications. These specifications include formal models of the system's implementation and invariants that define what "correctness" looks like for that system. The tool then model-checks these specifications to find potential violations. Once a violation is found, Specula works to reproduce those violations at the actual code level to confirm the bug.

Who it’s for

Software engineers and researchers working on concurrent or distributed systems who need to ensure high levels of correctness and find deep architectural or logic bugs.

Highlights

  • Automated Formal Specs: Uses LLM-based coding agents to generate TLA+ models from source code.
  • Code-Level Reproduction: Not only finds theoretical violations in a model but reproduces them in the actual implementation.
  • CI Integration: Supports continuous verification to check source changes incrementally.
  • Agent Agnostic: Compatible with various coding agents including Claude Code, Codex, and Copilot CLI via a system of skills and MCP tools.

Related

  • Project
    kaochenlong/spectra-app

    A Spec-Driven Development tool consisting of a desktop app, CLI, and AI agent skills to visually track and manage specifications, change proposals, and tasks.

  • Project
    reticlehq/reticle

    A verification tool for AI coding agents that drives real running applications to confirm if a feature actually works, providing verdicts based on internal app state and network traffic.

  • Project
    github/spec-kit

    Spec Kit is an open-source toolkit that provides structured, repeatable processes for AI coding agents to handle feature development, bug fixing, and project assessment.

  • Project
    tanweai/pua

    pua is a multi‑platform skill that injects corporate‑style “performance‑improvement‑plan” rhetoric and a 7‑point debugging checklist into AI coding assistants (Claude Code, OpenAI Codex, Cursor, Kiro, etc.). It auto‑triggers when the model shows lazy patterns (retries, blaming the user, passive waiting) or can be invoked manually with `/pua`. The skill forces the model to exhaust solutions, verify results, and act proactively, offering language‑specific packs (Chinese, English, Japanese). Benchmarks on a handful of bugs show higher fix and verification counts, though speed gains are modest. Installation is via the respective tool’s plugin/skill system (e.g., `npx skills add tanweai/pua --skill pua-en` for Claude Code).

  • Project
    NanmiCoder/cc-haha

    cc‑haha is a cross‑platform desktop workbench (Electron + React) that lets you chat with Claude, ChatGPT, Grok and many other LLMs, manage multiple coding sessions, review diffs, preview web pages, generate images, run multi‑agent workflows, and even let agents control other desktop apps (Computer Use) – all from a single app.