@mandujs/core 0.47.0 → 0.49.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mandujs/core",
-  "version": "0.47.0",
+  "version": "0.49.0",
   "description": "Mandu Framework Core - Spec, Generator, Guard, Runtime",
   "type": "module",
   "main": "./src/index.ts",

package/src/deploy/index.ts

@@ -52,6 +52,20 @@ export {
   type InferenceResult,
 } from "./inference/heuristic";
 
+export {
+  inferDeployIntentWithBrain,
+  type BrainInferenceOptions,
+} from "./inference/brain";
+
+export {
+  extractExplicitIntents,
+  mergeExplicitIntents,
+  type ExplicitIntentEntry,
+  type ExplicitIntentError,
+  type ExtractExplicitIntentsOptions,
+  type ExtractExplicitIntentsResult,
+} from "./inference/filling-extract";
+
 export {
   planDeploy,
   planHasChanges,

package/src/deploy/inference/brain.ts

@@ -0,0 +1,268 @@
+/**
+ * Brain-based deploy intent inferer.
+ *
+ * Issue #250 — Phase 1 M4.
+ *
+ * The brain inferer is **not** a replacement for the heuristic — it
+ * runs AFTER the heuristic and decides whether to confirm or refine
+ * the first-pass result. Three reasons for the wrap-not-replace shape:
+ *
+ *   1. **Cost cap.** The heuristic answers ~80% of routes correctly
+ *      and pays nothing. Brain only weighs in on the rest.
+ *   2. **Determinism floor.** When the brain is offline, fails to
+ *      authenticate, returns malformed JSON, or its output fails Zod
+ *      validation, the heuristic result stands. The pipeline never
+ *      blocks on the brain.
+ *   3. **Auditability.** Every intent the cache stores carries a
+ *      `rationale` string. Wrapping lets the brain rationale read
+ *      "agreed with heuristic: ..." or "overridden because ..." —
+ *      reviewers can see the reasoning in the diff.
+ *
+ * Failure modes (all fall back to heuristic):
+ *
+ *   - LLM call throws (network, auth, rate limit) → heuristic.
+ *   - LLM returns empty / non-JSON content → heuristic.
+ *   - LLM returns JSON that doesn't satisfy `DeployIntent` Zod → heuristic.
+ *   - LLM returns valid intent but `runtime: "static"` on a route
+ *     the manifest knows can't prerender → heuristic (caught by
+ *     `isStaticIntentValidFor`).
+ *
+ * The fallback is silent on `failOnError: false` (default — agents
+ * should still get a useful plan even with brain hiccups). Tests can
+ * pass `failOnError: true` to surface every brain miss.
+ */
+
+import { z } from "zod";
+import {
+  DeployIntent,
+  isStaticIntentValidFor,
+  type DeployIntent as DeployIntentType,
+} from "../intent";
+import type { LLMAdapter } from "../../brain/adapters/base";
+import type { DeployInferenceContext } from "./context";
+import { inferDeployIntentHeuristic, type InferenceResult } from "./heuristic";
+
+// ─── Public surface ───────────────────────────────────────────────────
+
+export interface BrainInferenceOptions {
+  /**
+   * The LLM adapter to call. Get one from
+   * `resolveBrainAdapter({ adapter: "auto" })`.
+   */
+  adapter: LLMAdapter;
+  /**
+   * Hard-fail on any brain error instead of falling back to the
+   * heuristic. Tests use this; production leaves it false.
+   */
+  failOnError?: boolean;
+  /**
+   * Timeout in ms for the LLM call. Defaults to 15s — deploy:plan is
+   * usually run interactively, agents prefer quick failure over long
+   * waits when the network is bad.
+   */
+  timeoutMs?: number;
+  /**
+   * Maximum source bytes sent to the brain. Anything beyond this is
+   * truncated with a `[…truncated]` marker. Default 4096 — enough for
+   * a route handler, well under any cloud token cap.
+   */
+  maxSourceBytes?: number;
+}
+
+/**
+ * Build the brain-based inferer. The returned function has the same
+ * signature `planDeploy({ infer })` expects, so it plugs in without
+ * any changes to the plan engine.
+ */
+export function inferDeployIntentWithBrain(
+  options: BrainInferenceOptions,
+): (ctx: DeployInferenceContext) => Promise<InferenceResult> {
+  return async (ctx) => {
+    const heuristic = inferDeployIntentHeuristic(ctx);
+    try {
+      const refined = await callBrain(ctx, heuristic, options);
+      if (!refined) return heuristic;
+
+      // Re-validate the brain's intent against the route shape — the
+      // brain is good at semantics but may miss the "static requires
+      // generateStaticParams" rule.
+      const validation = isStaticIntentValidFor(refined.intent, {
+        isDynamic: ctx.isDynamic,
+        hasGenerateStaticParams: ctx.hasGenerateStaticParams,
+        kind: ctx.kind,
+      });
+      if (!validation.ok) {
+        return {
+          intent: heuristic.intent,
+          rationale: `${heuristic.rationale} (brain proposed ${refined.intent.runtime} but it conflicts with route shape: ${validation.reason})`,
+        };
+      }
+
+      return refined;
+    } catch (err) {
+      if (options.failOnError) throw err;
+      return {
+        intent: heuristic.intent,
+        rationale: `${heuristic.rationale} (brain unavailable: ${err instanceof Error ? err.message : String(err)})`,
+      };
+    }
+  };
+}
+
+// ─── Internals ────────────────────────────────────────────────────────
+
+/** Schema for the brain's JSON response — slightly looser than DeployIntent
+ *  so we can validate field-by-field with helpful errors. */
+const BrainInferenceResponse = z
+  .object({
+    runtime: z.enum(["static", "edge", "node", "bun"]),
+    cache: z.union([
+      z.literal("no-store"),
+      z.literal("public"),
+      z.object({
+        maxAge: z.number().int().nonnegative().optional(),
+        sMaxAge: z.number().int().nonnegative().optional(),
+        swr: z.number().int().nonnegative().optional(),
+      }),
+    ]).optional(),
+    regions: z.array(z.string().min(1)).optional(),
+    timeout: z.number().int().positive().optional(),
+    visibility: z.enum(["public", "private"]).optional(),
+    rationale: z.string().min(1),
+    /** Brain's confidence — used only for telemetry; not persisted. */
+    confidence: z.number().min(0).max(1).optional(),
+  })
+  .strict();
+
+async function callBrain(
+  ctx: DeployInferenceContext,
+  heuristic: InferenceResult,
+  options: BrainInferenceOptions,
+): Promise<InferenceResult | null> {
+  const timeout = options.timeoutMs ?? 15_000;
+  // maxBytes reserved for future source-passing; for now the prompt
+  // works off context metadata (imports, deps) without raw source.
+  void options.maxSourceBytes;
+
+  const prompt = buildPrompt(ctx, heuristic);
+
+  // Race the LLM call against a timeout.
+  const controller = new AbortController();
+  const tHandle = setTimeout(() => controller.abort(), timeout);
+  let raw: string;
+  try {
+    raw = await options.adapter.generate(prompt, {
+      // CompletionOptions doesn't formally include `signal` today; we
+      // pass it via the loose extras object so adapters that honour it
+      // get the cancel signal, while older ones gracefully ignore.
+      ...({ signal: controller.signal } as object),
+      temperature: 0.0,
+      maxTokens: 512,
+    });
+  } finally {
+    clearTimeout(tHandle);
+  }
+
+  const parsed = parseBrainResponse(raw);
+  if (!parsed) return null;
+
+  // Merge against heuristic defaults — brain doesn't have to repeat
+  // every field, only the ones it overrides.
+  const merged: DeployIntentType = DeployIntent.parse({
+    runtime: parsed.runtime,
+    cache: parsed.cache ?? heuristic.intent.cache,
+    regions: parsed.regions ?? heuristic.intent.regions,
+    timeout: parsed.timeout ?? heuristic.intent.timeout,
+    visibility: parsed.visibility ?? heuristic.intent.visibility,
+  });
+
+  const rationale =
+    parsed.runtime === heuristic.intent.runtime &&
+    sameCache(parsed.cache ?? heuristic.intent.cache, heuristic.intent.cache)
+      ? `agreed with heuristic: ${parsed.rationale}`
+      : `brain refined: ${parsed.rationale}`;
+
+  return { intent: merged, rationale };
+}
+
+function buildPrompt(
+  ctx: DeployInferenceContext,
+  heuristic: InferenceResult,
+): string {
+  const truncatedImports = ctx.imports.slice(0, 30).join(", ");
+  const dependencyClasses = [...ctx.dependencyClasses].join(", ");
+
+  return [
+    "You are an expert deploy engineer for the Mandu framework.",
+    "",
+    "TASK: Decide the deploy runtime + cache directive for one route.",
+    "Output strict JSON ONLY (no markdown fences, no commentary).",
+    "",
+    "SCHEMA:",
+    `{
+  "runtime": "static" | "edge" | "node" | "bun",
+  "cache": "no-store" | "public" | { "maxAge"?: number, "sMaxAge"?: number, "swr"?: number },
+  "regions": string[],
+  "timeout": number_in_ms,
+  "visibility": "public" | "private",
+  "rationale": "1-sentence explanation",
+  "confidence": 0..1
+}`,
+    "",
+    "RULES:",
+    "- runtime=static is only valid for prerenderable pages (no dynamic segments OR has generateStaticParams).",
+    "- API routes (kind=api) MUST use edge / node / bun, never static.",
+    "- Prefer edge for low-latency stateless work; node/bun for DB / native deps / long-latency calls.",
+    "- bun is required when the route imports `bun:*` modules.",
+    "- Default cache for API: no-store. Default for static pages: { sMaxAge: 31536000, swr: 86400 }.",
+    "- Override the heuristic only when there is CLEAR evidence in the source. Otherwise echo the heuristic.",
+    "",
+    `ROUTE: ${ctx.routeId} (${ctx.pattern})`,
+    `KIND: ${ctx.kind}`,
+    `DYNAMIC: ${ctx.isDynamic}`,
+    `HAS_GENERATE_STATIC_PARAMS: ${ctx.hasGenerateStaticParams}`,
+    `IMPORTS: ${truncatedImports || "(none)"}`,
+    `DEPENDENCY_CLASSES: ${dependencyClasses || "(none)"}`,
+    "",
+    "HEURISTIC FIRST-PASS:",
+    `- runtime: ${heuristic.intent.runtime}`,
+    `- cache: ${formatCacheForPrompt(heuristic.intent.cache)}`,
+    `- visibility: ${heuristic.intent.visibility}`,
+    `- rationale: ${heuristic.rationale}`,
+    "",
+    "RESPOND WITH JSON ONLY:",
+  ].join("\n");
+}
+
+function parseBrainResponse(raw: string): z.infer<typeof BrainInferenceResponse> | null {
+  if (!raw || raw.trim() === "") return null;
+  // The brain might wrap JSON in fences despite the rules — strip the
+  // most common wrappers before parsing.
+  const stripped = stripCodeFences(raw).trim();
+  let json: unknown;
+  try {
+    json = JSON.parse(stripped);
+  } catch {
+    return null;
+  }
+  const result = BrainInferenceResponse.safeParse(json);
+  if (!result.success) return null;
+  return result.data;
+}
+
+function stripCodeFences(s: string): string {
+  // ```json ... ``` or ``` ... ``` — capture the inner block.
+  const fenced = /^```(?:json)?\s*\n([\s\S]*?)\n```\s*$/m.exec(s.trim());
+  if (fenced) return fenced[1]!;
+  return s;
+}
+
+function formatCacheForPrompt(cache: DeployIntentType["cache"]): string {
+  if (cache === "no-store" || cache === "public") return cache;
+  return JSON.stringify(cache);
+}
+
+function sameCache(a: DeployIntentType["cache"], b: DeployIntentType["cache"]): boolean {
+  if (typeof a === "string" || typeof b === "string") return a === b;
+  return JSON.stringify(a) === JSON.stringify(b);
+}

package/src/deploy/inference/filling-extract.ts

@@ -0,0 +1,245 @@
+/**
+ * Build-time extractor for `.deploy()` intents on `Mandu.filling()`
+ * route modules.
+ *
+ * Issue #250 M5. Bridges the runtime DSL (filling.deploy) and the
+ * build-time cache (`.mandu/deploy.intent.json`):
+ *
+ *   1. For every route in the manifest, dynamically `import()` the
+ *      module file.
+ *   2. If the default export looks like a `ManduFilling` instance and
+ *      its `getDeployIntent()` returns a non-null value, capture it
+ *      as a Zod-validated `DeployIntent`.
+ *   3. The CLI / MCP merges captured entries into the previous cache
+ *      with `source: "explicit"` BEFORE calling `planDeploy()` —
+ *      explicit entries are protected from inference (M1) so the
+ *      user's `.deploy()` always wins.
+ *
+ * Failure modes (non-fatal — extractor returns the offending route in
+ * `errors[]`, planner falls back to inference for it):
+ *
+ *   - Module file cannot be imported (syntax error, missing dep).
+ *   - Default export is not a filling instance.
+ *   - `.deploy({...})` payload fails Zod validation (this throws at
+ *     module load, so the import error path catches it).
+ *
+ * The extractor uses dynamic import rather than AST parsing because
+ * the manifest is built by importing routes anyway (for static
+ * params, slot metadata, etc.) — adding another loader would
+ * duplicate work. Side effects in route modules ARE a concern, but
+ * Mandu's convention is to put side effects in `loader()` / `get()`
+ * callbacks (which the extractor never invokes).
+ */
+
+import path from "node:path";
+import { pathToFileURL } from "node:url";
+import { promises as fs } from "node:fs";
+import { DeployIntent, type DeployIntent as DeployIntentType } from "../intent";
+import type { RoutesManifest, RouteSpec } from "../../spec/schema";
+
+export interface ExplicitIntentEntry {
+  routeId: string;
+  pattern: string;
+  /** Fully-validated, defaults-applied intent. */
+  intent: DeployIntentType;
+}
+
+export interface ExplicitIntentError {
+  routeId: string;
+  pattern: string;
+  module: string;
+  reason: string;
+}
+
+export interface ExtractExplicitIntentsResult {
+  entries: ExplicitIntentEntry[];
+  errors: ExplicitIntentError[];
+}
+
+export interface ExtractExplicitIntentsOptions {
+  /**
+   * Override the dynamic import for tests. Receives the absolute file
+   * URL the extractor would have imported and returns whatever module
+   * shape the test wants to inject.
+   */
+  importer?: (fileUrl: string) => Promise<unknown>;
+  /**
+   * Skip routes whose modules don't exist on disk. Default `true` —
+   * a missing route is a manifest staleness issue, not an extractor
+   * concern.
+   */
+  skipMissing?: boolean;
+}
+
+/**
+ * Walk the manifest, import each route, and capture explicit deploy
+ * intents from `Mandu.filling().deploy({...})` declarations.
+ */
+export async function extractExplicitIntents(
+  rootDir: string,
+  manifest: RoutesManifest,
+  options: ExtractExplicitIntentsOptions = {},
+): Promise<ExtractExplicitIntentsResult> {
+  const importer = options.importer ?? defaultImporter;
+  const skipMissing = options.skipMissing ?? true;
+  const entries: ExplicitIntentEntry[] = [];
+  const errors: ExplicitIntentError[] = [];
+
+  for (const route of manifest.routes) {
+    const modulePath = path.resolve(rootDir, route.module);
+    if (skipMissing) {
+      const exists = await pathExists(modulePath);
+      if (!exists) continue;
+    }
+
+    let mod: unknown;
+    try {
+      const url = pathToFileURL(modulePath).href;
+      mod = await importer(url);
+    } catch (err) {
+      errors.push({
+        routeId: route.id,
+        pattern: route.pattern,
+        module: route.module,
+        reason: `import failed: ${err instanceof Error ? err.message : String(err)}`,
+      });
+      continue;
+    }
+
+    const filling = pickDefaultFilling(mod);
+    if (!filling) continue;
+
+    let raw: unknown;
+    try {
+      raw = filling.getDeployIntent();
+    } catch (err) {
+      errors.push({
+        routeId: route.id,
+        pattern: route.pattern,
+        module: route.module,
+        reason: `getDeployIntent threw: ${err instanceof Error ? err.message : String(err)}`,
+      });
+      continue;
+    }
+
+    if (raw === undefined || raw === null) continue;
+
+    // Validate + apply schema defaults so downstream code (cache,
+    // adapters) never sees half-typed shapes.
+    const parsed = DeployIntent.safeParse(raw);
+    if (!parsed.success) {
+      errors.push({
+        routeId: route.id,
+        pattern: route.pattern,
+        module: route.module,
+        reason: `.deploy() validation failed: ${parsed.error.errors.map((e) => e.message).join("; ")}`,
+      });
+      continue;
+    }
+
+    entries.push({
+      routeId: route.id,
+      pattern: route.pattern,
+      intent: parsed.data,
+    });
+  }
+
+  return { entries, errors };
+}
+
+// ─── Internals ────────────────────────────────────────────────────────
+
+/**
+ * Loose duck-type check for ManduFilling. We can't `instanceof`
+ * because the imported module's `ManduFilling` constructor may not be
+ * the one this package was bundled with (workspace symlinks, package
+ * upgrades, etc.). Instead, we look for the public `getDeployIntent`
+ * method that ships in the same release as `.deploy()` — versions
+ * without the M5 method silently no-op (no entry captured).
+ */
+type FillingLike = { getDeployIntent: () => unknown };
+
+function isFillingLike(value: unknown): value is FillingLike {
+  return (
+    value != null &&
+    typeof value === "object" &&
+    "getDeployIntent" in value &&
+    typeof (value as { getDeployIntent: unknown }).getDeployIntent === "function"
+  );
+}
+
+function pickDefaultFilling(mod: unknown): FillingLike | null {
+  if (!mod || typeof mod !== "object") return null;
+  const candidate = (mod as { default?: unknown }).default;
+  return isFillingLike(candidate) ? candidate : null;
+}
+
+async function defaultImporter(fileUrl: string): Promise<unknown> {
+  return await import(fileUrl);
+}
+
+async function pathExists(p: string): Promise<boolean> {
+  try {
+    await fs.access(p);
+    return true;
+  } catch {
+    return false;
+  }
+}
+
+// ─── Cache merge helper ───────────────────────────────────────────────
+
+import type { DeployIntentCache, DeployIntentCacheEntry } from "../cache";
+import { hashSource } from "./context";
+
+/**
+ * Merge extractor entries into a cache, marking each one as
+ * `source: "explicit"`. Existing entries with the same route id are
+ * overwritten — the user's `.deploy()` always wins.
+ *
+ * `sourceHash` is set from the file's current bytes so the planner
+ * can detect drift between the cached intent and the live source.
+ * The cache's `generatedAt` and `brainModel` are left untouched so
+ * the merge step doesn't pretend to be a re-plan.
+ */
+export async function mergeExplicitIntents(
+  cache: DeployIntentCache,
+  entries: ReadonlyArray<ExplicitIntentEntry>,
+  rootDir: string,
+  manifest: RoutesManifest,
+): Promise<DeployIntentCache> {
+  if (entries.length === 0) return cache;
+  const next: DeployIntentCache = {
+    ...cache,
+    intents: { ...cache.intents },
+  };
+  const routeById = new Map<string, RouteSpec>(
+    manifest.routes.map((r) => [r.id, r]),
+  );
+
+  for (const entry of entries) {
+    const route = routeById.get(entry.routeId);
+    if (!route) continue;
+    const modulePath = path.resolve(rootDir, route.module);
+    let source = "";
+    try {
+      source = await fs.readFile(modulePath, "utf8");
+    } catch {
+      // Route file vanished between extraction and merge — keep
+      // going with an empty hash; the planner will catch the
+      // missing-source case on its own.
+    }
+    const cacheEntry: DeployIntentCacheEntry = {
+      intent: entry.intent,
+      source: "explicit",
+      rationale:
+        cache.intents[entry.routeId]?.source === "explicit"
+          ? cache.intents[entry.routeId]?.rationale ?? "explicit .deploy() override"
+          : "explicit .deploy() override",
+      sourceHash: hashSource(source),
+    };
+    next.intents[entry.routeId] = cacheEntry;
+  }
+
+  return next;
+}

package/src/filling/filling.ts

@@ -31,6 +31,10 @@ import {
   type ExecuteOptions,
 } from "../runtime/lifecycle";
 import type { SlotMetadata, SlotConstraints } from "../guard/semantic-slots";
+import {
+  DeployIntentInput,
+  type DeployIntentInput as DeployIntentInputType,
+} from "../deploy/intent";
 
 /** Handler function type */
 export type Handler = (ctx: ManduContext) => Response | Promise<Response>;
@@ -112,6 +116,19 @@ interface FillingConfig<TLoaderData = unknown> {
   middleware: MiddlewareEntry[];
   /** Semantic slot metadata */
   semantic: SlotMetadata;
+  /**
+   * Issue #250 M5 — explicit deploy intent override. When present,
+   * `mandu deploy:plan` records this entry as `source: "explicit"`,
+   * which the planner never overwrites via inference. The shape is
+   * the partial-input form so users can declare just the fields they
+   * care about (typically `runtime` + maybe `regions`).
+   *
+   * Use this to escape-hatch the heuristic / brain when you know
+   * better than they do — e.g. an API route that imports a DB driver
+   * but is actually called from a build-only script and you want it
+   * to ship as `static`.
+   */
+  deploy?: DeployIntentInputType;
 }
 
 export class ManduFilling<TLoaderData = unknown> {
@@ -201,6 +218,49 @@ export class ManduFilling<TLoaderData = unknown> {
     return { ...this.config.semantic };
   }
 
+  /**
+   * Deploy intent override — pin the runtime / cache / regions for this
+   * route so `mandu deploy:plan` never overrides it via inference.
+   *
+   * Issue #250 M5. The shape mirrors the cache schema's partial-input
+   * form: declare only the fields you care about; the rest fall to
+   * the heuristic / brain. The `deploy:plan` command records the
+   * entry as `source: "explicit"` and protects it from re-inference.
+   *
+   * @example Pin an API route to bun + Seoul region:
+   * ```typescript
+   * Mandu.filling()
+   *   .deploy({ runtime: "bun", regions: ["icn1"] })
+   *   .post(async (ctx) => { ... });
+   * ```
+   *
+   * @example Force a dynamic page to render statically (only valid when
+   * `generateStaticParams` covers every parameter set):
+   * ```typescript
+   * Mandu.filling()
+   *   .deploy({ runtime: "static" })
+   *   .loader(async () => ({ ... }));
+   * ```
+   *
+   * The intent is validated immediately so a typo (`runtime: "lambdda"`)
+   * fails at module load instead of silently shipping the wrong shape.
+   */
+  deploy(intent: DeployIntentInputType): this {
+    this.config.deploy = DeployIntentInput.parse(intent);
+    return this;
+  }
+
+  /**
+   * Read the explicit deploy intent for this route, if any.
+   *
+   * Returns `undefined` when the user did not call `.deploy()`. The
+   * build-time extractor consults this to decide whether to mark the
+   * cache entry as `source: "explicit"` or pass through to inference.
+   */
+  getDeployIntent(): DeployIntentInputType | undefined {
+    return this.config.deploy;
+  }
+
   /**
    * SSR 데이터 로더 등록
    *