technicalpickles/actually-lsp-doctor

/actually-lsp-doctor

Parse the user's args from the invocation:

• fix as the first arg means skip the diagnostic report and jump straight to action.
• rust | typescript | ruby as an arg narrows focus to that ecosystem.
• No args: full diagnostic report followed by interactive fix.

Step 1: Read project state

Read .claude/actually-lsp.json in the current project root. If missing, run detection: source lib/detect.sh and lib/ecosystems.sh from the plugin root (the env var CLAUDE_PLUGIN_ROOT points there), call detect_ecosystems "$PWD", and for each detected ecosystem compute the current state per the rules in CONTEXT.md.

Step 2: Diagnose (unless fix arg present)

Output a per-ecosystem report. For each ecosystem:

• State (from CONTEXT.md's six states)
• For non-ready ecosystems: what's needed to reach ready
• For dismissed ecosystems: note them but don't propose action

Keep tone terse. No celebration messaging.

Step 3: Act

For each ecosystem in no-lsp-plugin, server-not-runnable, or error:

no-lsp-plugin: try claude plugin install <recommended_plugin>@claude-plugins-official via Bash. The user gets a permission prompt. If denied, output the slash command form (/plugin install <recommended_plugin>@claude-plugins-official) and ask the user to run it themselves.

server-not-runnable: run env fixes per ecosystem via Bash. All env fixes auto-run; the user has implicit project consent.

• Rust: cargo build
• TypeScript: npm install
• Ruby: bundle install, plus gem install ruby-lsp if command -v ruby-lsp is empty

error: surface the cached last_error from the state file. Ask the user how to proceed.

Step 4: Re-detect

Re-run detection (same as Step 1's "if missing" path) and compute the new state per ecosystem. Hold the result in memory; don't write the state file yet. Step 5 may downgrade ready ecosystems before persistence.

Step 5: Probe ready ecosystems via LSP

The goal is to ground a ready verdict in an actual LSP response, not just env state. Env can say "server should run" while the server fails to answer queries (not indexed yet, wrong workspace root, crashed).

For each ecosystem whose computed state is ready:

1. Ensure the LSP tool is loaded. If it isn't available in this session, call ToolSearch with query select:LSP. If ToolSearch returns no match, skip the probe for every ecosystem and note "LSP tool unavailable in session" in the report. Do not downgrade any state. This is an environment issue, not an ecosystem failure.

2. Find a sample source file under $PROJECT_DIR:

   • rust: find "$PROJECT_DIR" -maxdepth 4 -type f -name '*.rs' -print -quit
   • typescript: find "$PROJECT_DIR" -maxdepth 4 -type f \( -name '*.ts' -o -name '*.tsx' \) -print -quit
   • ruby: find "$PROJECT_DIR" -maxdepth 4 -type f -name '*.rb' -print -quit

   If no file is found, skip the probe for this ecosystem and note "no sample file" in the report. Do not downgrade.

3. Call LSP documentSymbol against the sample file at line: 1, character: 1.

4. On error or non-array response: downgrade the ecosystem to error and set last_error to the stringified LSP response (or a short summary if the response is large). Env said the server should be runnable, so a probe failure is a real LSP failure worth surfacing on the next SessionStart.

5. On success: count the symbols in the response array. State stays ready. The count is for the report only; nothing extra goes into the state file.

Step 6: Write state and final report

Write the (possibly probe-updated) per-ecosystem state to .claude/actually-lsp.json using write_state from lib/state.sh. Output one status line per ecosystem:

• ready with successful probe: <ecosystem>: ready (LSP: N symbols)
• ready with skipped probe: <ecosystem>: ready (LSP: skipped, <reason>)
• error from probe failure: <ecosystem>: error (LSP probe failed: <summary>)
• Other states: <ecosystem>: <state> plus the recovery hint from Step 2

Keep tone terse. No celebration messaging.