Built-In Modules¶

Kedi ships a small set of ordinary .kedi modules. They use the same explicit exports, profiles, Python interop, tool metadata, and resolution rules as project modules.

Resolution and Shadowing¶

Import a bundled module by name:

> import: filesystem:
  readonlyfs
  read_text_file

A sibling filesystem.kedi takes precedence over the bundled module. Use this deliberately; an accidental same-name file changes the imported API.

errors¶

The errors module exports ModuleNotInstalledError:

> import: errors

```
raise ModuleNotInstalledError("httpx", required_by="HTTP reports")
```

The exception subclasses ModuleNotFoundError and formats singular or plural missing package names. Use it when a feature has a clear optional Python dependency.

require¶

require exports a Python-callable helper:

> import: require

```
require(["httpx", "pydantic"], required_by="Remote reports")
```

It checks import availability with Python's module discovery and returns True when all names are present. Missing modules raise ModuleNotInstalledError. It does not install packages or validate their versions.

filesystem¶

See filesystem.

sandbox¶

The optional sandbox module exports execute_code and a sandbox profile:

> import: sandbox

[result: Any] = `execute_code("sum(values)", {"values": [1, 2, 3]})`

It requires the Python package pydantic_monty. Importing the module checks that dependency immediately. execute_code(code, inputs={}, fail_fast=False) executes with Monty and returns the native final result; the backtick expression preserves that value. Inputs default to an empty mapping. Execution failures are returned as text by default so an agent can inspect them; pass fail_fast=True to raise instead.

This sandbox is for intentionally constrained generated code. It is not the execution mechanism for ordinary Kedi Python blocks, which use the configured Kedi executor.

helpers¶

helpers exports llm_approval(request: ApprovalRequest) -> ApprovalDecision, an experimental model-backed approval handler. It requires tool reasons to be enabled on the calling tool surface; the model's reason is untrusted context, not proof of user authorization. It does not run for ordinary read-only calls.

> import: helpers:
  llm_approval

> settings:
  tool_reason: enabled
> approval: `llm_approval`

This configures the handler; it does not invoke a tool by itself. Select a compatible model and explicit tool scope before using it. Keep deterministic allowlists or human review for decisions that require stronger guarantees. See Tool Reasons.

this and Example Modules¶

this is a bundled demonstration/easter-egg module whose import executes its encoded output. It is not an application API.

wordle exports its game profile and game procedures; it requires a graphical environment and optional packages. See Example Modules for its actual export list. These are demonstrations, not general-purpose stdlib contracts.