Using the Salesforce LSP

The salesforce-development plugin hosts a local MCP server named salesforce-lsp that lazily spawns Salesforce Language Server children and exposes their semantic capabilities as MCP tools. This skill is the contract other skills follow when they call those tools, and the answer to "how do I use the Salesforce LSP?" — what each tool does, how to read its result, what every error code means, and how to fall back when the host is absent.

This is a documentation/reference skill. It does not author or deploy code; it tells you (and other skills) how to drive the LSP tools correctly.

This plugin build vendors Apex + SOQL only. The salesforce-lsp host ships the Apex language server (@salesforce/apex-ls) and the SOQL language server. The LWC language server is intentionally not bundled in this plugin — the lwc.* tools are registered by the host but will always return an unavailable-class envelope here. Treat any lwc.* call as unavailable and use the fallback (read the component source / deploy-compile).

When to Use This Skill

• A user asks how to use the Salesforce LSP, or which LSP/MCP tools are available.
• A user asks what a specific tool does (apex.diagnostics, validate_soql, complete_soql, etc.) or how to read its output.
• An LSP tool returned an error envelope (lsp_disabled, spawn_timeout, circuit_open, no_apex_workspace, no_org_connected) and you need to know what it means and how to recover.
• You're authoring or reviewing another skill that calls an LSP tool and need the canonical call/fallback pattern.
• The LSP seems broken and you need to debug it (lsp.health, ${CLAUDE_PLUGIN_ROOT}/bin/lsp-doctor, the kill switch).

The Tools

All tools are served by the MCP server salesforce-lsp (invoke names follow the plugin-prefixed pattern: mcp__plugin_salesforce-development_salesforce-lsp__<tool_name>, where dots in tool names become underscores — e.g. apex.diagnostics becomes mcp__plugin_salesforce-development_salesforce-lsp__apex_diagnostics). Spawning is lazy: a tool that needs a language server brings the child up on first call (a one-time cold start of a few seconds for Apex), then reuses it. The pure static-analysis tools never spawn anything.

Apex (spawns the Apex LSP)

SOQL

Just deployed a field/object and it won't resolve? When a SOQL or Apex reference to a freshly-deployed field fails (e.g. No such column 'Foo__c', or complete_soql doesn't offer it) right after a deploy, the cached org describe is stale — call refresh_org_schema to invalidate it, then re-run the check before assuming a code error or renaming anything. This is the lever for post-deploy schema lag; reach for it before treating the failure as a bug in your query/class. (It only clears the local cache; it can't speed up server-side propagation, so if the org itself hasn't finished publishing the field, re-running after a moment is the fallback.)

LWC — not available in this build

The lwc.* tools (lwc.diagnostics, lwc.hover, lwc.definition, lwc.completion, lwc.workspace_symbols) are registered by the host but the LWC language server is not vendored in this plugin. Every lwc.* call returns an unavailable envelope. For LWC work, fall back to reading the component source/templates directly or deploy-compiling and reading the CLI errors.

Diagnostics / health

Coordinates everywhere are one-based (line/column). validate_soql diagnostic coordinates are relative to the query string, not a file — when a query came from extract_soql_from_apex, translate back using that query's range.

The Call / Fallback Contract

Every skill that calls an LSP tool follows the same three rules:

1. Prefer the LSP tool over guessing. If a tool exists for the step (validate a query, compile-check a class, complete against the org schema), call it before falling back to hand-analysis.
2. Treat an error envelope as "unavailable," never as "passed." A tool may return { error: <code> } instead of a result (see Error Codes). On any such code, record <tool>=unavailable: <code> in your report and continue down the fallback path — never report the input as valid/clean just because the check didn't run.
3. Degrade, don't fail. The LSP is an accelerator, not a hard dependency. If the host isn't available at all, the salesforce-lsp tools simply won't exist — fall back to the CLI/manual path for that step and say so.

When the LSP host is absent

If the plugin/host isn't available, the MCP tools above are not registered, so a call to (e.g.) validate_soql will not resolve. Detect this the same way you detect an error envelope — the tool is unavailable — and use the documented fallback:

Error Codes

A tool returns { error: <code> } (or, for complete_soql, a hint) instead of a result. Each maps to a recovery:

The first three are transient/config states shared by every spawning tool; the rest are expected, healthy states for a particular workspace/org/file context — not signs of a broken install. (no_lwc_bundles / unsupported_lwc_file are LWC-only codes; in this build the LWC server is not vendored, so lwc.* returns unavailable regardless.)

Debugging the LSP

Three layers, cheapest first:

1. lsp.health (MCP tool). The fastest check — never spawns a child. Reports the kill-switch mode, resolved workspace, per-server status, circuit-breaker state, last cold-spawn timing, and the vendored apex-ls version. Ask Claude to "run lsp.health," or call the lsp.health tool directly.

2. bin/lsp-doctor (CLI). A deeper, install-level diagnostic for support and onboarding. Verifies the committed bundles exist, the vendored apex-ls artifacts resolve, the org-schema cache is parseable, and each LSP child can actually spawn. Exits 0 when healthy, 1 with structured per-check output when something is wrong.

   "${CLAUDE_PLUGIN_ROOT}"/bin/lsp-doctor            # human-readable report
   "${CLAUDE_PLUGIN_ROOT}"/bin/lsp-doctor --json     # machine-readable
   "${CLAUDE_PLUGIN_ROOT}"/bin/lsp-doctor --no-spawn # skip child-spawn probes (CI/restricted)
   

   (In this build lsp-doctor will report the LWC server as missing — that is expected; only Apex + SOQL are vendored here.)

3. SFDX_LSP_DEBUG=1. Emits single-line JSON telemetry (spawn timing, cache hits, circuit events, per-tool latency) to stderr. Off by default (zero overhead).

The kill switch — SFDX_LSP

The fast way to control or disable the LSP. Set the environment variable:

To rule the LSP in or out of a problem, set SFDX_LSP=disabled and re-run: if the issue persists it isn't the LSP, and skills will have fallen back to their non-LSP paths automatically. An unknown value defaults to all (with a warning) so a typo never silently disables the feature.

Pre-deploy diagnostics gate

A PreToolUse hook (bin/lsp-precheck) runs Apex diagnostics on the .cls/.trigger files a sf project deploy start/validate is about to push, and emits a decision. It is fail-open: any error (including a missing/slow LSP) allows the deploy. Mode is controlled by SFDX_LSP_DEPLOY_GATE (off | warn | block, default warn) — warn surfaces diagnostics without blocking; block denies a deploy that has Apex compile errors.

Verification

• Asked "how do I use the Salesforce LSP tools?", this skill is the match and lists the tools, their inputs/outputs, and the fallback contract.
• Appears in the /skills listing as platform-lsp-integrate.
• Given an error code (e.g. no_apex_workspace), it explains the meaning and the correct recovery without treating the unrun check as a pass.

Cross-Skill Integration