Tools and Environments¶

Kedi ships one CLI, an LSP server, editor integrations, package commands, and low-level browser/playground bridges.

Command-Line Interface¶

kedi program.kedi
kedi -c '= hello'
kedi parse program.kedi
kedi program.kedi --test
kedi program.kedi --eval

The CLI loads .env and renders source-aware errors. Failed test cases exit nonzero; a completed eval reports its score without enforcing a quality gate.

Language Server¶

kedi-lsp provides diagnostics, semantic tokens, completion, hover, definition/references, rename, formatting, symbols, signature help, and inlay hints. It also understands Kedi query docstrings in Python.

Editor Extensions¶

VS Code and Zed extensions recognize .kedi, package highlighting/snippets, start kedi-lsp, and forward embedded Python regions to Python tooling.

Browser Execution¶

PlaygroundExecutor moves embedded Python execution through a synchronous worker bridge. PyodideExecutor is its browser-oriented alias/subclass. WebGPUAdapter can route model requests to a browser-hosted model bridge.

Package Tooling¶

kedi install installs a local package.kedi. kedi add supports explicit credential-free GitHub sources. Named public registry installation is not yet available; the local mock registry is for development only.

Validation Workflows¶

Use parse-only checks first, deterministic tests second, evals for measured quality, and optimization only after metrics are trustworthy.

Terminal-Bench¶

The optional Harbor bridge runs Kedi as a custom agent against the official Terminal-Bench 2.1 task containers and graders. It provides an immutable run manifest, bounded terminal and process tools, non-interactive benchmark approval, durable evidence, and Harbor-native resume. See Terminal-Bench 2.1.

The frozen Pydantic AI and LangChain 89-task engineering runs, including cost, latency, token, cache, request, and tool-call distributions, are reported in Terminal-Bench 2.1 Results.

In This Section¶

Command Line

• Run and Parse
• Test, Eval and Optimize
• Package Commands
• CLI Reference
• Environment Variables

Editors

• Language Server
• VS Code
• Zed

Interactive Environments

• Terminal REPL
• Local Notebook
• Notebook Runtimes
• Notebook Files and Recovery
• Browser and Playground

The REPL is a terminal session. Notebook is a separate local browser package with cells and persistent executors. Playground is an integration/demo surface, not a dependency of Notebook or a replacement for its file format.