Skillscript

Upgrade impact: none (additive). One thing to know: the new tier-2 lint can surface a pre-existing wrong-arg-name in an existing skill (it was always wrong; it just failed at dispatch before). Tier-2 is a warning — it never blocks compile or execution.

Authoring a skill against an MCP connector tool used to be guess-and-run: the runtime called the connector's tools/list but kept only tool names, discarding the argument schema. Now it retains the schema and puts it to work — without bloating agent context (schemas are surfaced selectively, never dumped).

• Connector-aware input lint. lint_skill validates $ <connector>.<tool> arg=... argument names against the tool's inputSchema: an unknown arg ($ ddg.search querry=...) is unknown-connector-arg and a missing required arg is missing-required-connector-arg, both tier-2. The schema is warmed on demand via a read-only tools/list (protocol introspection — not a tool dispatch, so it crosses no effect boundary) with a short timeout; a connector whose schema isn't reachable degrades to today's arg-agnostic behavior (no false positives). Runs entirely server-side — the schema never enters agent context.
• Selective schema fetch. runtime_capabilities({ tool: "<connector>.<tool>" }) returns that one tool's full descriptor (name, description, input_schema) on demand. The default response stays the compact menu (tool names) — pull the manual only for the tool you're about to call.
• skill_preflight surfaces tool contracts. For the connector tools a skill actually calls, preflight now reports each tool's input schema and — once any approved run has dispatched it — its last-observed output shape (keys + types), so an author knows what -> R binds before running.
• Observed-output-shape capture. A $ connector.tool dispatch records the shape (keys/types, not values) of the value the skill binds, keyed by connector.tool, in an operator-local cache under ~/.skillscript/. Once any approved skill runs a tool, later authors see its return shape. Privacy-first (no values); independent of trace mode.
• Respects allowed_tools. Every discovery surface honors the per-connector allowed_tools gate — a tool the operator gated off is never surfaced or validated against.
• Adopter note: custom McpConnector classes may implement the optional describeTools() method to light up the same discovery + lint; bundled HTTP/remote connectors provide it. A connector without it simply contributes no schema (graceful).