pydantic-ai-filesystem-sandbox 0.1.0

Filesystem sandbox toolset for PydanticAI agents with LLM-friendly errors

pydantic-ai-filesystem-sandbox

Filesystem sandbox toolset for PydanticAI agents with LLM-friendly errors.

Why This Package?

When building LLM agents that interact with the filesystem, you need:

1. Sandboxing - Restrict which directories the agent can access
2. Read/Write Control - Fine-grained permissions per path
3. LLM-Friendly Errors - Error messages that help the LLM correct its behavior
4. Approval Integration - Works with human-in-the-loop approval flows

This package provides a FileSandboxImpl toolset that implements all of these as a PydanticAI AbstractToolset.

Installation

pip install pydantic-ai-filesystem-sandbox

Quick Start

from pydantic_ai import Agent
from pydantic_ai_filesystem_sandbox import (
    FileSandboxImpl,
    FileSandboxConfig,
    PathConfig,
)

# Configure sandbox paths
config = FileSandboxConfig(paths={
    "input": PathConfig(root="./data/input", mode="ro"),   # Read-only
    "output": PathConfig(root="./data/output", mode="rw"), # Read-write
})

# Create the sandbox toolset
sandbox = FileSandboxImpl(config)

# Use with PydanticAI agent
agent = Agent("openai:gpt-4", toolsets=[sandbox])

Configuration

PathConfig Options

Option	Type	Default	Description
root	str	required	Root directory path
mode	"ro" | "rw"	"ro"	Access mode
suffixes	list[str] | None	None	Allowed file extensions (None = all)
max_file_bytes	int | None	None	Maximum file size limit
write_approval	bool	True	Require approval for writes
read_approval	bool	False	Require approval for reads

Example Configuration

config = FileSandboxConfig(paths={
    # Read-only input - any file type
    "input": PathConfig(
        root="./data/input",
        mode="ro",
    ),
    # Read-write output - only markdown and text
    "output": PathConfig(
        root="./data/output",
        mode="rw",
        suffixes=[".md", ".txt"],
        max_file_bytes=1_000_000,  # 1MB limit
    ),
    # Config files - read-only, requires approval
    "config": PathConfig(
        root="./config",
        mode="ro",
        read_approval=True,
    ),
})

Available Tools

The sandbox provides three tools to the agent:

read_file

Read a text file from the sandbox.

Path format: 'sandbox_name/relative/path'
Parameters:
  - path: str (required)
  - max_chars: int (default: 20,000)
  - offset: int (default: 0)

write_file

Write a text file to the sandbox (requires mode="rw").

Path format: 'sandbox_name/relative/path'
Parameters:
  - path: str (required)
  - content: str (required)

list_files

List files matching a glob pattern.

Parameters:
  - path: str (default: "." for all sandboxes)
  - pattern: str (default: "**/*")

LLM-Friendly Errors

All errors include guidance on what IS allowed:

# PathNotInSandboxError
"Cannot access 'secret/file.txt': path is outside sandbox.
Readable paths: input, output"

# PathNotWritableError
"Cannot write to 'input/file.txt': path is read-only.
Writable paths: output"

# SuffixNotAllowedError
"Cannot access 'output/data.json': suffix '.json' not allowed.
Allowed suffixes: .md, .txt"

# FileTooLargeError
"Cannot read 'output/huge.txt': file too large (5,000,000 bytes).
Maximum allowed: 1,000,000 bytes"

Approval Integration

Works with pydantic-ai-blocking-approval for human-in-the-loop:

from pydantic_ai_filesystem_sandbox import FileSandboxImpl, FileSandboxConfig, PathConfig
from pydantic_ai_blocking_approval import ApprovalToolset, ApprovalController

# Create sandbox
config = FileSandboxConfig(paths={
    "output": PathConfig(root="./output", mode="rw", write_approval=True),
})
sandbox = FileSandboxImpl(config)

# Wrap with approval
controller = ApprovalController(mode="interactive", approval_callback=my_prompt_fn)
approved_sandbox = ApprovalToolset(
    inner=sandbox,
    prompt_fn=controller.approval_callback,
    memory=controller.memory,
    require_approval=["write_file", "read_file"],
)

agent = Agent(..., toolsets=[approved_sandbox])

The sandbox implements needs_approval() and present_for_approval() for fine-grained approval control.

ReadResult

The read_file tool returns a ReadResult object:

class ReadResult(BaseModel):
    content: str        # The file content read
    truncated: bool     # True if more content exists
    total_chars: int    # Total file size in characters
    offset: int         # Starting position used
    chars_read: int     # Characters actually returned

This allows agents to handle large files by reading in chunks:

# First read
result = sandbox.read("input/large.txt", max_chars=10000)
if result.truncated:
    # Continue reading
    result2 = sandbox.read("input/large.txt", max_chars=10000, offset=10000)

API Reference

Configuration

• FileSandboxConfig - Top-level configuration with named paths
• PathConfig - Configuration for a single sandbox path

Toolset

• FileSandboxImpl - PydanticAI AbstractToolset implementation

Errors

• FileSandboxError - Base class for all sandbox errors
• PathNotInSandboxError - Path outside sandbox boundaries
• PathNotWritableError - Write to read-only path
• SuffixNotAllowedError - File extension not allowed
• FileTooLargeError - File exceeds size limit

Types

• ReadResult - Result of read operations with metadata

License

MIT