/**
 * The `instructions` string returned in graft's MCP `initialize` response.
 *
 * This is the one piece of graft prose that survives **tool deferral**. When a
 * host has more tools than its schema budget allows, it sends tool *names* only
 * and withholds the JSONSchemas until a `ToolSearch`-style lookup fetches them —
 * measured in a real session: 111 tools deferred, of which graft's six arrived as
 * six bare strings with no descriptions at all. The MCP spec's `instructions`
 * field is delivered on a separate track (Claude Code records it as its own
 * `mcp_instructions_delta` context layer), so it lands whole even then. Other
 * servers already rely on this — `claude-in-chrome` uses it for exactly the
 * batch-your-ToolSearch instruction below; graft used to send nothing.
 *
 * Two jobs, in this order:
 *   1. Defuse the deferral tax. One lookup loads all six tools for the whole
 *      session, so the cost is a single round trip, not two calls per use. An
 *      agent that doesn't know this can only assume the worse reading.
 *   2. Say what each tool is FOR as a decision rule, not a feature summary —
 *      because in a host with no hooks and no skill listing (a plain chat client
 *      with an MCP config), this string plus the tool descriptions are the entire
 *      steering budget graft gets.
 *
 * Budget: keep this under ~1,000 characters. Observed sibling servers sit at
 * 660–984, and nothing proves a longer one survives un-truncated.
 */
/** The `select:` argument that loads every graft tool in one lookup. */
export declare function toolSearchQuery(prefix?: string): string;
export declare function mcpInstructions(): string;
//# sourceMappingURL=instructions.d.ts.map