Truncate Lsp Diagnostics

An OpenCode plugin that trims the LSP diagnostics block that write, edit, and apply_patch append after every edit.

Why this exists

After each edit, OpenCode's edit tools append the current LSP errors to the tool output. In a project with many outstanding errors this block is large, and most of it is residual errors the model already knows about and did not introduce.

OpenCode's own TUI, and most frontends, already collapse or fold long tool output themselves. This plugin is for the ones that do not — for example Zed, which renders tool output raw and leaves it expanded by default.

What it does

1. Session baseline — diagnostics already reported in a previous round are filtered out, so the model only sees newly introduced errors.
2. Cap + spill — if the remaining new diagnostics still exceed a configurable count, only the first few are shown inline and the full list is written to a temp file the model is told to read.

Limits

• The token savings are small. If all you want is to cut tool-output tokens, this plugin is probably not worth installing — tools like magic-context already truncate tool calls and do it more thoroughly.
• Hiding known diagnostics should give the model less noise, but this has not been benchmarked and the expected benefit is modest.

Install

Add the plugin to your opencode.json (or opencode.jsonc):

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["opencode-truncate-lsp-diagnostics"]
}

Then restart OpenCode.

Configuration

Pass options as a tuple:

{
  "plugin": [
    [
      "opencode-truncate-lsp-diagnostics",
      { "cap": 3 }
    ]
  ]
}

Option	Type	Default	Description
cap	number	3	Max new diagnostics to show inline. 0 spills every new diagnostic to a file.
tools	string[]	["write", "edit", "apply_patch"]	Tool ids whose output is inspected.
tmpDir	string	OS temp dir	Directory where spill files are written.

Example

First edit to a file with 5 pre-existing errors, with cap: 2:

Wrote file successfully.

LSP errors detected in this file, please fix:
<diagnostics file="/path/to/file.ts">
ERROR [171:32] 'Promise' only refers to a type...
ERROR [197:20] 'Promise' only refers to a type...
</diagnostics>
(3 additional new LSP diagnostics written to /tmp/opencode-lsp-diagnostics-<session>-<ts>.txt; read it for the full list)

Second edit, where the model fixed two errors but introduced one new one:

Edit applied successfully.

LSP errors detected in this file, please fix:
<diagnostics file="/path/to/file.ts">
ERROR [600:10] Cannot find name 'newSymbol'.
</diagnostics>
(3 previously-seen LSP diagnostics omitted)

How it works

The plugin registers a tool.execute.after hook and rewrites output.output in place. A module-level baseline counts each diagnostic identity (severity + message) per file; the [line:col] location is ignored, so a diagnostic that only shifted lines is still recognized as seen. A message whose count increases from one round to the next has that many new occurrences.

Note: the tool.execute.after hook returns Promise<void>, so the plugin mutates output.output rather than returning a new string. This relies on OpenCode passing the same output object through to the model — current behavior (verified against OpenCode 1.18.21), but not part of the documented contract.

Development

bun install
bun run build    # compile to dist/
bun test         # run unit tests

License

MIT