Skill v1.0.0
currentLLM-judged scan90/100version: "1.0.0" name: pi-lsp description: Execute deterministic read-only LSP and structural operations directly through the governed Pi gateway, without a model.
Pi LSP
Call lsp_request directly. Supply absolute allowed cwd, one existing file inside it, method, and optional resourceProfile, queueTimeoutSeconds, timeoutSeconds, requestId. No model, provider, thinking or Kether model envelope is needed. Do not delegate deterministic operations to a lower agent.
Methods: diagnostics, hover, definition, references, symbols, completions, code_actions, overview, search. Positioned methods need positive 1-based line and character, or an exact symbol query. Search requires a structural pattern in query plus language. Queries are not natural-language instructions. code_actions returns suggestions only; no edits are applied.
Require ok=true, exact requestedTool/toolsUsed and empty unexpectedTools. Read raw result; diagnostics may contain errors even when the call succeeded. Check backend: tree-sitter is reduced syntax/structure evidence, not full LSP type checking. Diagnostics from LSP must have diagnosticsPublished=true; missing publication cannot prove a clean file.
The tool uses an isolated read-only single-file snapshot, no provider credentials, no external network, fixed resource profiles, independent queue/execution budgets and awaited cgroup cleanup. Cross-file results are limited by that snapshot. Language servers start per request. Python, Java, JS/TS, C/C++, C# configured servers remain available subject to installed runtime and budgets. New tool fields may require refreshing the ChatGPT connector and starting a new conversation.
Result status: success / no-match / degraded / unavailable / failed. unavailable and failed have ok=false. degraded is usable only as explicitly reduced evidence. No-match does not establish an unavailable service.