agentdir 0.1.6

Python bindings for agentdir — virtual file tree infrastructure for agent-ready file layouts

agentdir

Virtual filesystem for agent-optimized exploration of general-purpose files using CoW reflinks.

agentdir is a Python binding for the agentdir Rust library. It lets you map real directories of documents, media, datasets, generated artifacts, plain text, binaries, or any other OS-visible files into a virtual file tree, move and copy entries without touching the originals, track original-file changes, and fork the tree into isolated snapshots via copy-on-write.

• Version: 0.1.6
• License: MIT
• Python: >= 3.9
• Built with: PyO3 + maturin (native Rust extension, abi3 wheels)

---

Installation

pip install agentdir

No extra dependencies. The package ships pre-built abi3 wheels for Linux, macOS, and Windows.

Installation only installs the binding. After mapping original files, call Workspace.refresh() whenever the original file tree may have changed. The Python binding exposes reconciliation through refresh() and refresh_with_hash_verification(); it does not start the CLI watch loop for you.

---

Quick Start

from agentdir import Workspace

ws = Workspace.init("./workspace")
summary = ws.map("./team-files", "/files")
print(f"Mapped {summary['entries_added']} entries")

content = ws.read_bytes("/files/q1-report.txt")
print(content.decode())

ws.mv("/files/q1-report.txt", "/reports/q1-report.txt")  # original files are untouched

# Reconcile source changes before an agent/session depends on the view.
sync = ws.refresh()
print(sync)

All methods are synchronous. There is no async API. Internally the library runs a Tokio runtime, but the Python surface is fully blocking.

If you also install the CLI, long-running workflows can run agentdir -w ./workspace watch --interval 60 in a separate process. The watcher combines filesystem events with periodic full rescans; Python applications should otherwise call refresh() on their own schedule or at session boundaries.

---

API Reference

Import

from agentdir import Workspace, SnapshotWorkspace

---

Workspace

Static methods

Workspace.init(path: str, strategy: str = "reflink") -> Workspace

Initialize a new workspace at path. The strategy controls how files are materialized:

Value	Behavior
"reflink"	CoW reflink, falls back to byte-copy if unsupported (default)
"symlink"	Symbolic links
"virtual"	Metadata-only, no materialization

Workspace.open(path: str) -> Workspace

Open an existing workspace. Raises FileNotFoundError if the workspace does not exist at path.

---

Instance methods

map(source: str, mount: str) -> dict[str, int]

Map a source directory into the virtual tree at mount. Returns a summary dict:

{
    "entries_added": int,
    "reflinked":     int,
    "copied":        int,
    "symlinked":     int,
    "dirs_created":  int,
    "errors":        int,
}

unmap(mount: str) -> dict[str, int]

Remove the mapping at mount and clean up its entries. Returns:

{"entries_removed": int}

mv(from_path: str, to_path: str) -> None

Move a virtual entry. The original file on disk is not touched.

cp(from_path: str, to_path: str) -> None

Copy a virtual entry. The original file on disk is not touched.

mkdir(path: str) -> None

Create a virtual directory.

rmdir(path: str, recursive: bool) -> None

Remove a virtual directory. Pass recursive=True to remove non-empty directories.

rename(path: str, new_name: str) -> None

Rename the last path component of a virtual entry. new_name is a bare name, not a full path.

exists(path: str) -> bool

Return True if the virtual path exists.

stat(path: str) -> dict[str, object]

Return metadata for a virtual path:

{
    "virtual_path":  str,
    "source_path":   str,
    "size_bytes":    int,
    "mtime_ns":      int,
    "entry_type":    str,   # "File" or "Directory"
    "materialized":  bool,
}

read_bytes(path: str) -> bytes

Read the raw bytes of a file at the given virtual path.

refresh() -> dict[str, int]

Detect changes in source directories and apply them to the virtual tree. Returns:

{
    "added":     int,
    "refreshed": int,
    "removed":   int,
    "errors":    int,
}

refresh_with_hash_verification(verify_hashes: bool = False) -> dict[str, int]

Same as refresh(), with an optional SHA-256 pass. When verify_hashes=True, files whose mtime and size are unchanged are additionally verified by content hash to catch silent modifications. Returns the same shape as refresh().

status() -> dict[str, object]

Return workspace-level metadata:

{
    "total_entries":           int,
    "source_roots":            int,
    "materialized_root":       str,
    "last_updated_epoch_secs": int,
}

export_mapping(reverse: bool = False, relative_to: str | None = None) -> dict[str, str]

Export the source-to-virtual path mapping as a plain dict. Pass reverse=True to get virtual-to-source instead. Pass relative_to to relativize source paths against a base directory.

map_batch(mappings: list[tuple[str, str]]) -> dict[str, object]

Map multiple files in one call. Each tuple is (source_path, mount_point). Note: batch map accepts files only, not directories. Returns:

{
    "entries_added": int,
    "reflinked":     int,
    "copied":        int,
    "symlinked":     int,
    "dirs_created":  int,
}

rglob(pattern: str) -> list[str]

Match virtual paths against a glob pattern. Supports * and ** wildcards (e.g. "/files/*.pdf", "/media/**/*.png"). Returns a list of matching virtual paths.

list_snapshots() -> list[str]

Return the names of all snapshots attached to this workspace.

snapshot(name: str) -> SnapshotWorkspace

Create a named CoW snapshot of the current virtual tree. The snapshot starts as a fork of the workspace and accepts isolated writes.

open_snapshot(name: str) -> SnapshotWorkspace

Open an existing named snapshot.

destroy_snapshot(name: str) -> None

Destroy a named snapshot and remove its files from disk.

---

SnapshotWorkspace

A CoW fork of a Workspace. Writes to a snapshot are isolated and do not affect the base workspace or any original files.

exists(path: str) -> bool

Return True if the virtual path exists in this snapshot.

stat(path: str) -> dict[str, object]

Return metadata for a virtual path. Same shape as Workspace.stat().

read_bytes(path: str) -> bytes

Read the raw bytes of a file in this snapshot.

write(path: str, content: bytes) -> None

Write content to a file in this snapshot. The write is copy-on-write and does not affect the base workspace.

export_mapping(reverse: bool = False, relative_to: str | None = None) -> dict[str, str]

Export the path mapping for this snapshot. Same semantics as Workspace.export_mapping().

destroy() -> None

Destroy this snapshot and remove all its files from disk.

---

Examples

Map a directory and read files

from agentdir import Workspace

ws = Workspace.init("./workspace")
summary = ws.map("./team-files", "/files")
print(f"Mapped {summary['entries_added']} entries")

content = ws.read_bytes("/files/q1-report.txt")
print(content.decode())

ws.mv("/files/q1-report.txt", "/reports/q1-report.txt")  # original files untouched

Snapshots with isolated writes

from agentdir import Workspace

ws = Workspace.init("./workspace")
ws.map("./team-files", "/files")

snap = ws.snapshot("experiment")
snap.write("/files/q1-report.txt", b"snapshot-only draft")

# Base workspace is unaffected:
original = ws.read_bytes("/files/q1-report.txt")
modified = snap.read_bytes("/files/q1-report.txt")

snap.destroy()

---

License

MIT. See LICENSE for details.

For the CLI and Rust library, see the main repository: https://github.com/NomaDamas/agentdir