Custom LSP server starts and quiesces but never delivers diagnostics on edit or inspect

[acramsay]

What happened?

I registered rumdl 0.2.64 (Markdown linter with a built-in LSP) under lsp.servers and turned on lsp.diagnostics_on_edit:

{
  "format_on_edit": true,
  "lsp": {
    "diagnostics_on_edit": true,
    "servers": {
      "rumdl": {
        "extensions": ["md"],
        "binary": "rumdl",
        "args": ["server"],
        "root_markers": [".git", ".rumdl.toml"]
      }
    }
  }
}

Observations on 0.55.1, OpenCode 1.18.27, linux x64:

1. aft_inspect reports lsp_start (producer: rumdl) and lsp_quiescence (producer: rumdl), so the binary spawns the server. But it reports zero diagnostics for .md files that rumdl check flags with 14 issues.
2. Every edit of a .md file reports LSP server(s) did not respond in time: rumdl. Bash and yaml built-ins respond normally in the same sessions.
3. The server itself is healthy. A minimal stdio client against rumdl server gets initialize in 7ms and full publishDiagnostics 5ms after didOpen, against the same file in the same repo.
4. Nothing mentioning rumdl appears in aft-plugin.log or the bridge log, so the failure is silent.

My read: the post-edit wait in crates/aft/src/lsp/manager.rs never sends textDocument/didOpen to ServerKind::Custom clients, or drops their publishDiagnostics on the freshness check. Built-ins work end to end, so the custom kind is where the path diverges. Happy to test a fix or gather more data (a wrapper binary logging the protocol would show whether didOpen ever goes out).

Expected: custom servers participate in post-edit diagnostics the same way built-ins do.

Diagnostics

AFT CLI and binary 0.55.1, OpenCode host 1.18.27, plugin registered, version 0.55.1, config at ~/.config/cortexkit/aft.jsonc, storage and LSP cache healthy per doctor. Full doctor output available on request.

Plugin version

0.55.1

AFT binary version

0.55.1

Platform

linux x64

[aft-alfonso]

Thanks - this is a precise report and the minimal-client observation (initialize 7 ms, publishDiagnostics 5 ms after didOpen) narrows it to our side. Tracing the ServerKind::Custom path in the post-edit wait and the inspect producer now; the two candidates you name (didOpen never sent to custom clients, or their publishDiagnostics dropped on the version/freshness check) are the first things I am checking. The silence in the logs is its own defect and gets fixed alongside. I will post the mechanism here once reproduced with rumdl.

[aft-alfonso]

Reproduced with rumdl 0.2.64 on your exact config, and the mechanism is not the one either of us guessed - it is a capability inversion, and it is not specific to custom servers:

• AFT advertises LSP 3.17 pull diagnostics (textDocument/diagnostic) in its client capabilities. rumdl respects that and stops pushing publishDiagnostics (it sends one empty, unversioned compatibility publish after didOpen).
• AFT's post-edit wait only accepted a versioned publish and never sent a pull, so the empty publish was rejected by the freshness gate and the wait timed out - hence did not respond in time on every edit and zero findings on inspect. The built-in servers you compared against still push regardless of the capability, which is why they looked fine. The lack of any log line was a second defect.

Fix on main: after an edit, AFT now pulls diagnostics from every pull-capable server (built-in or custom), tags the result with the document version, keeps the push path as fallback, sends .md as markdown rather than plaintext, and logs didOpen / pull / publishDiagnostics / dropped-because per server. On your file rumdl now reports its findings on both edit and inspect; the fake pull-only server test pins it. Present on 0.55.1 (confirmed against the tag); ships with the next release. Thanks for the minimal-client check - it is what ruled out the server.