Good call @SomeoneToIgnore, I'll hoist downstream_client.clone() before the loop in both on_buffer_reloaded and invalidate_diagnostic_summaries_for_updated_entries.

(Same pattern exists in stop_local_language_server at line 10745, but I'll at least fix it in my new code.)

You're right, on_buffer_edited already calls pull_workspace_diagnostics (line 7877), so open buffers are covered.

I was trying to handle files not open in any editor that still have stale diagnostic_summaries (e.g. after yarn install modifies files on disk). But thinking about it more, this is the server's job: Zed already sends workspace/didChangeWatchedFiles via update_local_worktree_language_servers, and the server can then push new diagnostics or send workspace/diagnostic/refresh (handled at line 1149).

I'll drop Updated and AddedOrUpdated and keep only Removed, deleted files have no server-driven recovery path.

The scenario: user has a file open with diagnostic errors, does "Revert File" (or the file changes externally). If the resulting content is identical to what's in the buffer, BufferEvent::Reloaded fires unconditionally (buffer.rs:1668), but Edited does not: did_edit (buffer.rs:2845) early-returns when apply_diff produces no text changes.

Since on_buffer_event ignores Reloaded (_ => {} at line 4429), on_buffer_edited never runs, pull_workspace_diagnostics is never triggered, and stale summaries persist in the file tree.

Instead of clearing summaries directly, I could handle Reloaded in on_buffer_event and just call pull_workspace_diagnostics for the relevant servers, same as on_buffer_edited does. That keeps the server in charge. Would that work?

Thank you, all ideas sounds very reasonable — let's start small and ensure language servers do their part of the job.

One small note that a reload scenario is described pretty well, which means we can try and create a test for this use case specifically.

Applied the requested changes :

1. Hoisted downstream_client.clone() before the loop in invalidate_diagnostic_summaries_for_updated_entries, and moved the if let Some(...) check outside the inner server_id loop since it doesn't change between iterations.

2. Removed Updated/AddedOrUpdated from invalidate_diagnostic_summaries_for_updated_entries, only Removed triggers summary clearing now.

3. Rewrote on_buffer_reloaded to call pull_workspace_diagnostics for each relevant language server instead of clearing summaries directly.

4. Rewrote the buffer reload test with a fake LSP that has workspace diagnostics support, verifying the pull is triggered after reload.

One thing to flag: apply_workspace_diagnostic_report skips open buffers (line 12250 "diagnostics from a document pull should win over diagnostics from a workspace pull"). So for the specific scenario where a buffer is open and reloaded without content changes, pull_workspace_diagnostics is triggered but its response won't update that file's summaries. In practice the server should handle this through didChangeWatchedFiles or document-level pulls from the editor, but wanted to call it out.

Thanks for the feedback.

Here's what changed:

• on_buffer_reloaded now calls pull_diagnostics_for_buffer instead of pull_workspace_diagnostics. The task is stored in a HashMap<BufferId, Task> on LspStore (not detached), and cleaned up in the observe_release callback when the buffer refcount drops to zero.
• Renamed to invalidate_diagnostic_summaries_for_removed_entries.
• Replaced the match with .iter().filter(|(_, _, change)| *change == PathChange::Removed).
• Merged the two iterations on summaries_by_server_id into one loop.
• .log_err() → .ok().
• Test rewritten around DocumentDiagnosticRequest since we're doing a document-level pull now.