Edit-Time Tactics in Idris

Abstract

Metaprogramming allows users to write programs that write programs. In dependently-typed languages such as Idris, recent work on elaborator reflection [17] paved the way for new applications of metaprogramming by showing that a tacticbased proof language can be substituted with a monadic interface that exposes the internal elaborator. The goal of our work is to use elaborator reflection to write editor interaction actions in Idris. Currently in Idris modes of editors such as Emacs, users can perform actions like type-checking holes, case-splitting, and lemma extraction. Implementations of all of these Idris editor actions are hard-coded in the compiler, and they are written in Haskell. Our work will allow us to rewrite them in Idris as metaprograms and to move them into an Idris library, instead of having them embedded into the compiler. 4.1. New constructors for the new Elab primitives, and function definitions based on them. 4.2. The relationship between reflection, reification, quotation, unquotation, elaboration and delaboration. 4.3. The function to reflect Haskell terms of the type SExp to the internal representations of Idris terms of the Idris type SExp. 4.4. The function to reify the internal representations of Idris terms of the Idris type SExp to Haskell terms of the type SExp. 4.5. The runElabAction function in the compiler, how Elab action terms are run under the hood. viii 4.6. Adding cases for Prim__FromEditor and Prim__ToEditor in runTacTm in the compiler. 4.7. The function elabEditAt that runs editor actions in the compiler. 4.8. The function collectTypes to dissect a type signature to its components. 5.1. Implementation of the edit-time tactic for "add clause". 5.2. Example function to run addClause on. 5.3. Result of running addClause on the example function. 5.4. Definitions of Context and Sequent for Hezarfen. 5.5. The "Either implies" case in Hezarfen 5.6. Rudimentary implementation of reduce in Hezarfen. 6.1. Definition of hole_command in Lean. ix CHAPTER 1 Especially as dependent types gradually sneak into mainstream languages, like they already have to Haskell [26] and Scala [3], we predict that TDD will only get more popular. 2 Propositions and types are distinct in these systems.

An Integrative Human-Centered Architecture for Interactive Programming Assistants

“…One key direction is to directly incorporate AI techniques, in particular deep reinforcement learning, which also is oriented around an action space and a reward/scoring structure (human acceptance, tests passing, type errors resolution). We are also working to incorporate more complex multistage refactorings via interactive monadic edit actions, extending work on edit-time tactics in proof assistants [37].…”