@valbuild/server 0.122.0 → 0.123.2

package/package.json

@@ -16,7 +16,7 @@
     "./package.json": "./package.json"
   },
   "types": "dist/valbuild-server.cjs.d.ts",
-  "version": "0.122.0",
+  "version": "0.123.2",
   "devDependencies": {
     "@prettier/sync": "^0.6.1",
     "@types/jest": "^30.0.0"
@@ -29,9 +29,9 @@
     "typescript": "^6.0.3",
     "zod": "^4.4.3",
     "zod-validation-error": "^5.0.0",
+    "@valbuild/shared": "0.123.2",
     "@valbuild/core": "0.121.0",
-    "@valbuild/shared": "0.122.0",
-    "@valbuild/ui": "0.122.0"
+    "@valbuild/ui": "0.123.2"
   },
   "engines": {
     "node": "^20.19.0 || >=22"

package/dist/declarations/src/tools/createValTools.d.ts

@@ -1,41 +0,0 @@
-import type { ValModules } from "@valbuild/core";
-import type { ValServerConfig } from "../ValServer.js";
-import type { ValOps } from "../ValOps.js";
-import type { ValToolState } from "./defineTool.js";
-import { type ValToolResult, type ValTools } from "./types.js";
-export type ValToolsOptions = ValServerConfig;
-/**
- * Val's server-side tool registry.
- *
- * This is the piece Val did not have: the Studio's chat tools are defined *and
- * executed in the browser*, against its client stores, so nothing here could be
- * re-exposed. These tools run against {@link ValOps} instead, which is what lets
- * an MCP server — or a stdio transport, or anything else — drive Val content
- * without a browser.
- *
- * Transport-agnostic on purpose: nothing under `tools/` imports an MCP SDK. A
- * host adapts {@link ValToolResult} at its own edge.
- */
-export declare function createValTools(valModules: ValModules, options: ValToolsOptions): ValTools;
-/**
- * The content as the caller should see it, loaded once per call.
- *
- * Pending patches are applied, because an agent looking at a project mid-edit
- * should see what the Studio would show rather than the last published state.
- *
- * Deliberately not cached across calls. In fs mode a save recomputes the base
- * sha within the same process, so a cached view would go stale silently — and
- * the cost of being wrong here is an agent writing a patch against content that
- * has already moved.
- *
- * Exported for `toolsFixture`, not for consumers: `tools/index.ts` re-exports by
- * name, so this stays inside the package. A test that assembled its own state
- * would be asserting against a view no real call ever sees.
- */
-export declare function loadState(ops: ValOps): Promise<{
-    status: "ok";
-    state: ValToolState;
-} | {
-    status: "error";
-    result: ValToolResult;
-}>;

package/dist/declarations/src/tools/defineTool.d.ts

@@ -1,69 +0,0 @@
-import type { Json, ModuleFilePath, PatchId, SerializedSchema } from "@valbuild/core";
-import type { z } from "zod";
-import type { ValOps, OrderedPatches, PatchAnalysis, Schemas, Sources } from "../ValOps.js";
-import type { ValToolContext, ValToolDefinition, ValToolErrorCode, ValToolResult } from "./types.js";
-/**
- * How a tool is written, and what it is handed.
- *
- * Tools are defined with {@link defineTool} so that the handler's `args` are
- * inferred from the tool's own `inputSchema`. Without that the array of tools
- * would have to be typed at its widest and every handler would start by
- * re-narrowing `unknown`, which is exactly where a tool and its schema drift
- * apart unnoticed.
- */
-/** Everything a tool is allowed to reach. Deliberately narrow. */
-export type ValToolDeps = {
-    /**
-     * The data layer for *this call*. In proxy mode it is authenticated as the
-     * caller, so a tool must never reach for an ambient instance instead.
-     */
-    ops: ValOps;
-    ctx: ValToolContext;
-    /**
-     * The content as the caller should see it: base sources with pending patches
-     * applied, plus the schemas and the patch analysis that produced them.
-     *
-     * Loaded once per call and shared, because a tool that re-derived it would
-     * both pay for it again and risk reading a different revision than the one it
-     * validated against.
-     */
-    state: ValToolState;
-};
-export type ValToolState = {
-    schemas: Schemas;
-    serializedSchemas: Record<ModuleFilePath, SerializedSchema>;
-    sources: Sources;
-    patches: OrderedPatches;
-    analysis: PatchAnalysis;
-    /**
-     * Pending patches that would not apply, by the module they belong to.
-     *
-     * A module listed here has `sources` that silently lack those changes, so its
-     * content is not what publishing would produce. Reads still return it — it is
-     * what the Studio would show too, and a project has to stay diagnosable — but
-     * a write to such a module is refused, because it would be based on a state
-     * that does not exist.
-     */
-    unappliedPatches: Record<ModuleFilePath, {
-        patchId: PatchId;
-        skipped: boolean;
-        error: {
-            message: string;
-        };
-    }[]>;
-};
-export type ValToolImpl = ValToolDefinition & {
-    handler: (args: unknown, deps: ValToolDeps) => Promise<ValToolResult>;
-};
-/**
- * Declare a tool, binding its handler to its input schema.
- *
- * The handler receives already-parsed arguments: the registry validates against
- * `inputSchema` before calling, so a handler never sees input its schema would
- * have rejected.
- */
-export declare function defineTool<S extends z.ZodType>(definition: Omit<ValToolDefinition, "inputSchema"> & {
-    inputSchema: S;
-}, handler: (args: z.infer<S>, deps: ValToolDeps) => Promise<ValToolResult>): ValToolImpl;
-export declare function ok(data: Json): ValToolResult;
-export declare function err(code: ValToolErrorCode, message: string): ValToolResult;

package/dist/declarations/src/tools/index.d.ts

@@ -1,3 +0,0 @@
-export { createValTools, type ValToolsOptions } from "./createValTools.js";
-export { VAL_SCOPE_READ, VAL_SCOPE_WRITE, authorIdFromVerifiedSubject, } from "./types.js";
-export type { ValScope, ValToolAuth, ValToolContext, ValToolDefinition, ValToolDefinitionJson, ValToolErrorCode, ValToolResult, ValTools, } from "./types.js";

package/dist/declarations/src/tools/types.d.ts

@@ -1,163 +0,0 @@
-import type { AuthorId } from "../ValOps.js";
-import type { Json } from "@valbuild/core";
-import type { z } from "zod";
-/**
- * The public surface of Val's server-side tool registry.
- *
- * Types only, deliberately: this file is the contract that the MCP hosts, the
- * CLI's stdio transport and the tools themselves are all written against, and
- * keeping it free of implementation means those can be built in any order
- * without one of them owning the shape.
- *
- * The design this implements is `docs/plans/mcp.md`, Part A. Two constraints
- * from it are load-bearing and easy to break by accident:
- *
- * 1. **Nothing here may import an MCP SDK.** That is what lets hosts other
- *    than the template consume these tools, and it is not hypothetical
- *    hygiene — the TypeScript SDK reorganised itself at v2.0.0, and a registry
- *    coupled to it would have moved with it.
- * 2. **The result type is deliberately not MCP's `CallToolResult`.** Each host
- *    adapts {@link ValToolResult} at its own edge, which is also where an
- *    error becomes an in-band `isError` result the model can recover from
- *    rather than a transport failure.
- */
-/** Why a tool call failed, in a form a host can map onto its own errors. */
-export type ValToolErrorCode = 
-/** No tool of that name is registered. */
-"unknown-tool"
-/** The arguments did not satisfy the tool's `inputSchema`. */
- | "invalid-args"
-/** The path, module or key the arguments name does not exist. */
- | "not-found"
-/** The caller is not allowed to do this — wrong credential, or none. */
- | "forbidden"
-/**
- * The write was rejected because applying it would leave content invalid.
- * Nothing was stored. See `docs/plans/mcp.md` Part C on speculative
- * validation.
- */
- | "validation-failed"
-/**
- * Another writer moved the patch chain first. Callers may re-derive the
- * parent ref and retry once; the registry does that itself before
- * surfacing this.
- */
- | "conflict"
-/** The tool exists but is not available in this mode (e.g. fs vs http). */
- | "unsupported"
-/** Anything else, including an upstream failure. */
- | "internal";
-export type ValToolDefinition = {
-    name: string;
-    title?: string;
-    description: string;
-    /** zod v4 — a Standard Schema, which is what the MCP SDKs consume. */
-    inputSchema: z.ZodType;
-    /**
-     * MCP tool annotations. Hints, not enforcement: a host may show them to a
-     * user or use them to decide what to auto-approve, so they have to be
-     * honest about what the tool does.
-     */
-    annotations?: {
-        readOnlyHint?: boolean;
-        destructiveHint?: boolean;
-        idempotentHint?: boolean;
-    };
-};
-/**
- * The same definition with `inputSchema` as JSON Schema, for hosts that want the
- * wire shape rather than a Standard Schema.
- *
- * Typed as whatever zod's own converter produces, so deriving it needs no cast
- * and no second hand-written description of the same input.
- */
-export type ValToolDefinitionJson = Omit<ValToolDefinition, "inputSchema"> & {
-    inputSchema: ReturnType<typeof z.toJSONSchema<z.ZodType>>;
-};
-/**
- * How the caller was established, and there is one acceptable answer: the host
- * **checked a signature**.
- *
- * A union of one, deliberately. It carried a second variant — a personal access
- * token relayed to the backend unchecked, on the reasoning that the app cannot
- * resolve one and the backend can. The reasoning held; the shape did not. A
- * credential the host cannot check is one it also cannot refuse, so accepting
- * one made "a deployed endpoint that authenticates nobody" a supported
- * configuration, and it let a host serve these tools without ever being told
- * where callers should authorize. The discriminant stays so that adding a
- * second *verified* kind stays a one-line change at every call site.
- */
-export type ValToolAuth = {
-    type: "verified-profile";
-    /**
-     * The profile the host **verified** — the `sub` of an access token whose
-     * signature, issuer, audience and expiry were all checked against the
-     * authorization server's published key.
-     *
-     * This field is why an identity field exists at all, and an earlier version
-     * of this file argued none should. That argument was about a specific case
-     * and stated too broadly: an id the host *asserts* on the strength of a
-     * credential it cannot check is an unverified claim dressed as a checked one,
-     * and that is still refused — it is why a relayed token never carried a
-     * profile, and now cannot reach here at all. An id the host *verified*
-     * cryptographically is a different thing, and it is the same standing the
-     * Studio has when it re-signs a session it established itself.
-     */
-    profileId: AuthorId;
-    /**
-     * The token's granted scopes, as the authorization server issued them.
-     *
-     * Enforced here as well as by the backend, deliberately. Two checks on one
-     * grant is not redundancy for its own sake: this one can refuse a write
-     * before it is attempted, so a token that may only read never reaches the
-     * code that builds a patch.
-     */
-    scopes: string[];
-};
-/**
- * Who is calling, established once per request by the host.
- *
- * `null` means local fs mode, where there is no credential to hold and patches
- * are written with no author, exactly as the Studio does locally (D.1). In
- * proxy mode `null` is refused rather than falling back to the app's own API
- * key: that key can do more than any single user, and quietly substituting it
- * would turn a missing credential into full access.
- */
-export type ValToolContext = {
-    auth: ValToolAuth | null;
-    /** Groups a run of related edits, when the host has such a notion. */
-    sessionId: string | null;
-};
-/**
- * Brand a verified subject as an {@link AuthorId}.
- *
- * `AuthorId` is a branded string so that an id cannot be conjured from any
- * string that happens to be lying around — which is exactly the mistake this
- * type is guarding against. That makes one assertion unavoidable at the boundary
- * where a real id enters the system, so it lives here, once, with a name that
- * says what makes it legitimate: the caller has *verified* this subject, not
- * received it.
- *
- * Do not reach for this to satisfy a type. If you are holding a string you did
- * not verify, the honest value is `null`.
- */
-export declare function authorIdFromVerifiedSubject(subject: string): AuthorId;
-/** Read access. Every call needs it, the writes included. */
-export declare const VAL_SCOPE_READ = "val:read";
-/** Write access. Needed *in addition* by any tool not marked `readOnlyHint`. */
-export declare const VAL_SCOPE_WRITE = "val:write";
-export type ValScope = typeof VAL_SCOPE_READ | typeof VAL_SCOPE_WRITE;
-export type ValToolResult = {
-    status: "ok";
-    data: Json;
-} | {
-    status: "error";
-    code: ValToolErrorCode;
-    message: string;
-};
-export type ValTools = {
-    list(): ValToolDefinition[];
-    listJsonSchema(): ValToolDefinitionJson[];
-    call(name: string, args: unknown, ctx: ValToolContext): Promise<ValToolResult>;
-    dispose(): Promise<void>;
-};