Async Remote File Explorer

Problem

The file explorer blocks the main editor thread on remote filesystem I/O. When the remote host becomes unreachable (e.g. DHCP lease change, network drop), the editor freezes permanently because these blocking calls have no timeout.

Root cause

poll_file_tree_changes() and file_explorer_toggle_expand() (plus 5 other call sites) use runtime.block_on(...) to perform remote filesystem operations synchronously on the main thread. The underlying AgentChannel::request() has no timeout — if the remote never responds, the main thread blocks forever.

Affected code

file_operations.rs — poll_file_tree_changes()

• Calls self.filesystem.metadata() synchronously for every expanded dir
• Calls runtime.block_on(tree.refresh_node()) for changed dirs

file_explorer.rs — 6 block_on call sites:

Line	Operation	Context
275	toggle_node	User expands/collapses a dir
406	refresh_node	User manually refreshes
442	refresh_node	After creating a file
498	refresh_node	After creating a directory
586	refresh_node	After deleting (trash)
707	refresh_node	After renaming

Solution

Three changes, layered for defense in depth.

1. Channel-level request timeout (30s)

Add tokio::time::timeout(Duration::from_secs(30), ...) around the response wait in AgentChannel::request(). Returns ChannelError::Timeout (variant already exists) on expiry. request_blocking() inherits this automatically.

This protects all remote operations as a safety net — file open, save, metadata, list_dir, etc. — not just file explorer calls.

2. Async poll_file_tree_changes (two-phase)

Replace the synchronous polling loop with a background task:

Phase 1 — main thread (non-blocking):

• Collect expanded dirs + their stored mtimes
• Spawn a tokio task
• Set file_tree_poll_in_progress = true, return immediately

Phase 2 — tokio task:

• Call metadata_async() for each dir (with the 30s channel timeout)
• Wrap the whole batch in a 5s timeout so we don't wait 30s x N dirs
• Compare mtimes, send changed dirs back via AsyncMessage::FileTreePollResult(Vec<(NodeId, PathBuf, SystemTime)>)

Phase 3 — main thread (on receiving AsyncMessage):

• Update dir_mod_times
• Spawn async refresh tasks for changed dirs (see below)
• Clear file_tree_poll_in_progress

3. Async file explorer operations

Replace all 6 block_on calls in file_explorer.rs with spawned tasks.

Prerequisites:

• Derive Clone on FileTree (all fields are owned data + Arc<FsManager>)

Pattern for toggle_node (expand/collapse):

1. Set node to NodeState::Loading immediately (already exists)
2. Clone the FileTreeView and spawn a tokio task
3. Task calls toggle_node / refresh_node on the cloned tree
4. Task sends result back via AsyncMessage with the updated tree
5. Main thread swaps in the new tree, does post-processing (gitignore load, decoration rebuild, selection update, status message)

Pattern for post-mutation refreshes (create/delete/rename):

• The filesystem mutation itself (create, rename, delete) remains synchronous — these are user-initiated one-shot operations, protected by the 30s channel timeout
• Only the refresh_node call after the mutation becomes async
• If the refresh fails, the mutation already succeeded; the tree is just stale (user can manually refresh)

UX behavior

Normal operation (remote responsive)

No visible change. Expand/collapse may feel slightly more responsive since the main thread isn't blocked during directory listing.

Remote becomes unresponsive

Editor never freezes. Input, rendering, and all non-remote features continue working normally.

Feature	Behavior
File tree auto-poll	Stops updating silently. Tree shows last known state. Debug log on timeout.
Expand/collapse dir	Shows "Loading..." state. Times out after 30s, shows error in status bar, node reverts.
Create/delete/rename	Mutation times out after 30s, error in status bar.
Open/save file	Times out after 30s, error in status bar.

No new UI elements

A stale file tree is an acceptable degraded state. The user discovers the remote is down when they try an active operation (open, save, expand) and see the timeout error.

Out of scope

• Reconnection logic (separate feature)
• Making file open/save async (much larger refactor, 30s timeout is sufficient)

Implementation order

1. Channel timeout — smallest change, biggest safety improvement
2. poll_file_tree_changes async — fixes the deadlock that prompted this work
3. file_explorer_toggle_expand + the 5 refresh sites — completes the picture