Large language models (LLMs) have reshaped the landscape of program synthesis. However, contemporary LLM-based code completion systems often hallucinate broken code because they lack appropriate code context, particularly when working with definitions that are neither in the training data nor near the cursor. This paper demonstrates that tighter integration with the type and binding structure of the programming language in use, as exposed by its language server, can help address this contextualization problem in a token-efficient manner. In short, we contend that AIs need IDEs, too! In particular, we integrate LLM code generation into the Hazel live program sketching environment. The Hazel Language Server is able to identify the type and typing context of the hole that the programmer is filling, with Hazel’s total syntax and type error correction ensuring that a meaningful program sketch is available whenever the developer requests a completion. This allows the system to prompt the LLM with codebase-wide contextual information that is not lexically local to the cursor, nor necessarily in the same file, but that is likely to be semantically local to the developer’s goal. Completions synthesized by the LLM are then iteratively refined via further dialog with the language server, which provides error localization and error messages. To evaluate these techniques, we introduce MVUBench, a dataset of model-view-update (MVU) web applications with accompanying unit tests that have been written from scratch to avoid data contamination, and that can easily be ported to new languages because they do not have large external library dependencies. These applications serve as challenge problems due to their extensive reliance on application-specific data structures. Through an ablation study, we examine the impact of contextualization with type definitions, function headers, and errors messages, individually and in combination. We find that contextualization with type definitions is particularly impactful. After introducing our ideas in the context of Hazel, a low-resource language, we duplicate our techniques and port MVUBench to TypeScript in order to validate the applicability of these methods to higher-resource mainstream languages. Finally, we outline ChatLSP, a conservative extension to the Language Server Protocol (LSP) that language servers can implement to expose capabilities that AI code completion systems of various designs can use to incorporate static context when generating prompts for an LLM.

Type systems typically only define the conditions under which an expression is well-typed, leaving ill-typed expressions formally meaningless. This approach is insufficient as the basis for language servers driving modern programming environments, which are expected to recover from simultaneously localized errors and continue to provide a variety of downstream semantic services. This paper addresses this problem, contributing the first comprehensive formal account of total type error localization and recovery: the marked lambda calculus. In particular, we define a gradual type system for expressions with marked errors, which operate as non-empty holes, together with a total procedure for marking arbitrary unmarked expressions. We mechanize the metatheory of the marked lambda calculus in Agda and implement it, scaled up, as the new basis for Hazel, a full-scale live functional programming environment with, uniquely, no meaningless editor states. The marked lambda calculus is bidirectionally typed, so localization decisions are systematically predictable based on a local flow of typing information. Constraint-based type inference can bring more distant information to bear in discovering inconsistencies but this notoriously complicates error localization. We approach this problem by deploying constraint solving as a type-hole-filling layer atop this gradual bidirectionally typed core. Errors arising from inconsistent unification constraints are localized exclusively to type and expression holes, i.e., the system identifies unfillable holes using a system of traced provenances, rather than localized in an ad hoc manner to particular expressions. The user can then interactively shift these errors to particular downstream expressions by selecting from suggested partially consistent type hole fillings, which returns control back to the bidirectional system. We implement this type hole inference system in Hazel.

Several modern programming systems, including GHC Haskell, Agda, Idris, and Hazel, support typed holes . Assigning static and, to varying degree, dynamic meaning to programs with holes allows program editors and other tools to offer meaningful feedback and assistance throughout editing, i.e. in a live manner. Prior work, however, has considered only holes appearing in expressions and types. This paper considers, from type theoretic and logical first principles, the problem of typed pattern holes. We confront two main difficulties, (1) statically reasoning about exhaustiveness and irredundancy when patterns are not fully known, and (2) live evaluation of expressions containing both pattern and expression holes. In both cases, this requires reasoning conservatively about all possible hole fillings. We develop a typed lambda calculus, Peanut, where reasoning about exhaustiveness and redundancy is mapped to the problem of deriving first order entailments. We equip Peanut with an operational semantics in the style of Hazelnut Live that allows us to evaluate around holes in both expressions and patterns. We mechanize the metatheory of Peanut in Agda and formalize a procedure capable of deciding the necessary entailments. Finally, we scale up and implement these mechanisms within Hazel, a programming environment for a dialect of Elm that automatically inserts holes during editing to provide static and dynamic feedback to the programmer in a maximally live manner, i.e. for every possible editor state. Hazel is the first maximally live environment for a general-purpose functional language.

We present a system called Smyth for program sketching in a typed functional language whereby the concrete evaluation of ordinary assertions gives rise to input-output examples, which are then used to guide the search to complete the holes. The key innovation, called live bidirectional evaluation, propagates examples “backward” through partially evaluated sketches. Live bidirectional evaluation enables Smyth to (a) synthesize recursive functions without trace-complete sets of examples and (b) specify and solve interdependent synthesis goals. Eliminating the trace-completeness requirement resolves a significant limitation faced by prior synthesis techniques when given partial specifications in the form of input-output examples. To assess the practical implications of our techniques, we ran several experiments on benchmarks used to evaluate Myth, a state-of-the-art example-based synthesis tool. First, given expert examples (and no partial implementations), we find that Smyth requires on average 66% of the number of expert examples required by Myth. Second, we find that Smyth is robust to randomly-generated examples, synthesizing many tasks with relatively few more random examples than those provided by an expert. Third, we create a suite of small sketching tasks by systematically employing a simple sketching strategy to the Myth benchmarks; we find that user-provided sketches in Smyth often further reduce the total specification burden (i.e. the combination of partial implementations and examples). Lastly, we find that Leon and Synquid, two state-of-the-art logic-based synthesis tools, fail to complete several tasks on which Smyth succeeds.