@pithy-sh/core 0.10.0 → 0.12.0

package/src/privacy/report.ts

@@ -0,0 +1,64 @@
+// SPDX-FileCopyrightText: 2026 Pithy
+// SPDX-License-Identifier: MIT
+
+import type { PrivacyField } from "./classify";
+
+/**
+ * **The inventory's shape, apart from the walk that fills it** (#132).
+ *
+ * Its own module so a browser may import it. The manifest's response schema names {@link PRIVACY_SOURCES}, and
+ * every capability's scope declarations reach that schema; the walk in `./inventory` takes a `Capability`, which
+ * reaches Hono, Kysely and the Workers types. A type-only edge is still an edge to the compiler a browser build
+ * uses, so the shape lives here and the walk imports it, never the other way.
+ */
+
+/** The storage the inventory walks, in the order it reports them. */
+export const PRIVACY_SOURCES = ["d1", "kv"] as const;
+
+/** One storage kind the inventory walks. */
+export type PrivacySource = (typeof PRIVACY_SOURCES)[number];
+
+/** One schema holding at least one marked field. */
+export interface PrivacyInventorySchema {
+  /** The capability that declared it. */
+  capability: string;
+  /** Where it is stored. */
+  storage: PrivacySource;
+  /** The D1 database name or the KV namespace name it lives in. */
+  container: string;
+  /** The table name or the KV store name. */
+  name: string;
+  /** Every marked field in it, with its class. */
+  fields: PrivacyField[];
+  /**
+   * Every field in it marked `encrypted()`, classed or not (#390): encryption may be marked alone, so this is where an
+   * unclassed one is listed. `unrotatable` where the storage is not D1.
+   */
+  encrypted: EncryptedFieldRecord[];
+}
+
+/** One encrypted field of a schema. */
+export interface EncryptedFieldRecord {
+  /** Where it sits, on the same paths as {@link PrivacyField.path}. */
+  path: string;
+  /** Set where the storage is not D1, so it can never be rotated. */
+  unrotatable?: true;
+}
+
+/** Everything one Worker's storage schemas say about privacy. */
+export interface PrivacyInventory {
+  /** The storage kinds walked. Anything not listed was not. */
+  sources: PrivacySource[];
+  /** How many schemas were walked, marked or not. */
+  schemasWalked: number;
+  /** Every schema that holds a marked field, in composition order. */
+  schemas: PrivacyInventorySchema[];
+  /** Marked fields per class. A class with none is absent rather than zero. */
+  counts: Record<string, number>;
+  /** Every class with at least one marked field: the built-ins in their fixed order, then the rest by name. */
+  observed: string[];
+  /** How many encrypted fields the walked schemas hold, classed or not (#390). */
+  encryptedFields: number;
+  /** How many of them rest outside D1 and so can never be rotated. */
+  unrotatable: number;
+}

package/src/privacy/retirement.ts

@@ -0,0 +1,102 @@
+// SPDX-FileCopyrightText: 2026 Pithy
+// SPDX-License-Identifier: MIT
+
+import { z } from "zod";
+
+/**
+ * **The retirement guard's verdict, apart from the counts that feed it** (#390).
+ *
+ * Its own module, and a leaf, so a browser may import it: the manifest's posture block names
+ * {@link RetirementVerdict}, and the management client validates that block in a browser. The counts are read by
+ * `./fieldRotation`, which reaches D1 and the Workers types; it imports this, never the other way.
+ */
+
+/** Why the guard is not clean: what holds a version that is due to leave. */
+export const RetirementReason = z
+  .enum(["unrotatable", "incomplete-coverage", "stale"])
+  .describe(
+    "`unrotatable` — an encrypted field rests outside D1, so nothing can prove it re-sealed. `incomplete-coverage` — a schema source the guard derives from was not read. `stale` — a row still holds an envelope or an index under a version due to leave.",
+  );
+export type RetirementReason = z.output<typeof RetirementReason>;
+
+/** The guard's verdict, as the manifest's posture block reports it. */
+export const RetirementVerdict = z
+  .object({
+    blocked: z.boolean().describe("Whether a version due to leave the window is being held live by one of `reasons`."),
+    reasons: z
+      .array(RetirementReason)
+      .describe("Every reason the guard is not clean now, empty when it is. Read directly: never inferred."),
+    liveVersions: z
+      .number()
+      .int()
+      .nonnegative()
+      .describe("How many field-key versions are live. Two is the normal state: current and previous."),
+  })
+  .describe("Whether the field key's oldest live version can be retired, and if not, why.");
+export type RetirementVerdict = z.output<typeof RetirementVerdict>;
+
+/** What the guard reads. */
+export interface RetirementInput {
+  /** Every live version, newest first. */
+  versions: readonly number[];
+  /** The version every write seals under. Absent where the key set could not be read: then nothing is due. */
+  current?: number | undefined;
+  /** Rows still under any version due to leave, summed over every stamped envelope and index column. */
+  stale: number;
+  /** Whether every schema source the stamp derives from was read. */
+  coverageComplete: boolean;
+  /** How many encrypted fields rest outside D1. */
+  unrotatable: number;
+}
+
+/**
+ * The version `current` replaced: the highest live version below it. A version above current is staged and never
+ * promoted, so it replaced nothing and is never the previous one.
+ */
+export function previousVersion(versions: readonly number[], current: number): number | undefined {
+  const below = versions.filter((version) => version < current);
+  return below.length === 0 ? undefined : Math.max(...below);
+}
+
+/**
+ * The versions staged over `current` and never promoted, oldest first. A rotation whose confirm failed leaves one in
+ * the entry; the next rotation promotes it rather than minting past it, so the window stays anchored on current.
+ */
+export function stagedVersions(versions: readonly number[], current: number): number[] {
+  return versions.filter((version) => version > current).sort((left, right) => left - right);
+}
+
+/**
+ * The versions due to leave the window: every live version older than the one `current` replaced. Current and
+ * previous are the normal state, and the previous version is still sealing rows lazily, so only an older one is
+ * due. Anchored on the pointer, never on the two highest numbers: a version staged over current and never promoted
+ * would otherwise take the previous slot, and make the version that was current, or current itself, due.
+ */
+export function versionsDue(versions: readonly number[], current: number): number[] {
+  const previous = previousVersion(versions, current);
+  if (previous === undefined) return [];
+  return versions.filter((version) => version < previous).sort((left, right) => right - left);
+}
+
+/**
+ * **The retirement guard.** Clean only when all three hold: the stale count is zero across every stamped envelope
+ * and index column, coverage is complete, and no unrotatable field was observed. `blocked` is whether a version is
+ * due and the guard is not clean. Retiring a version is the one irreversible act in the feature — every ciphertext
+ * still under it becomes unopenable — so nothing short of clean retires one, and nothing implies it.
+ */
+export function retirementVerdict(input: RetirementInput): RetirementVerdict {
+  const reasons: RetirementReason[] = [];
+  if (input.unrotatable > 0) reasons.push("unrotatable");
+  if (!input.coverageComplete) reasons.push("incomplete-coverage");
+  if (input.stale > 0) reasons.push("stale");
+  return {
+    blocked: input.current !== undefined && versionsDue(input.versions, input.current).length > 0 && reasons.length > 0,
+    reasons,
+    liveVersions: input.versions.length,
+  };
+}
+
+/** Whether a verdict lets the rotator retire: nothing at all holds a version. */
+export function guardIsClean(verdict: RetirementVerdict): boolean {
+  return verdict.reasons.length === 0;
+}

package/src/privacy/sealed.ts

@@ -0,0 +1,323 @@
+// SPDX-FileCopyrightText: 2026 Pithy
+// SPDX-License-Identifier: MIT
+
+import type { z } from "zod";
+import { decryptValue, type EncryptedEnvelope, type EncryptionConfig, encryptValue } from "../crypto/envelope";
+import { FIELD_KEYS_BINDING, type FieldKeys, liveVersions, resolveFieldKeys } from "../crypto/fieldKeys";
+import { isSqliteJson, type SqliteJsonSchema, sqliteJson } from "../data/codecs";
+import { fromZodError, ValidationError } from "../error/pithyError";
+import { blindIndexValue } from "./blindIndex";
+import {
+  type EncryptedFieldLayout,
+  type EncryptedIndexOf,
+  type EncryptedInner,
+  encryptedLayout,
+  encryptedSpecOf,
+  hasEncryptedFields,
+} from "./encrypted";
+
+/**
+ * **`sealed()`: the round trip for a table with an encrypted field** (#390).
+ *
+ * The codecs stay synchronous (`./encrypted`), so AES-GCM happens one step above them, in a seam the repository builds
+ * per request from `env` — which is how the key reaches the encode without a module-level schema ever holding one.
+ *
+ * ```ts
+ * const patients = sealed(Patient, { keys: env.PITHY_FIELD_KEYS });
+ *
+ * const record = await patients.seal(value);   // seal each marked field, then Patient.encode
+ * await db.insertInto("patient").values(record).execute();
+ *
+ * const value = await patients.open(row);      // Patient.parse, then open each marked field
+ *
+ * const hashes = await patients.match("email", input);   // one per live version
+ * await db.selectFrom("patient").where("emailIndex", "in", hashes).selectAll().execute();
+ * ```
+ *
+ * ## What each seal binds
+ *
+ * The AAD is `pithy/enc/1:<name>:<field path>:<pk>` ({@link fieldContext}): the stable name `encryptedTable()`
+ * declared, the field's path, and the row's primary key. A ciphertext moved to another field, row or table does not
+ * open. The declared name is not the SQL table name, so renaming a table in a migration leaves every row openable. A
+ * column's SQL name is always its key in snake case, so renaming an encrypted column changes its field path. The key
+ * is app-assigned and never updated, which is the immutability guarantee. Changing the name, a path (a column rename
+ * included) or a key is a re-seal migration through `sealed()`: read with the old schema, write with the new.
+ *
+ * ## What it refuses, at build
+ *
+ * A schema with an encrypted field and no `encryptedTable()` declaration, and every problem `encryptedLayout` names:
+ * an encrypted field no path of object keys reaches, an indexed field with no index column, an index column naming no
+ * indexed field. Each is refused when `sealed()` is called, before any key is read, naming the schema.
+ *
+ * Every write seals under the key set's current version. Every read opens under whichever version sealed the value,
+ * for as long as that version is live.
+ */
+
+/** The versioned prefix of every field's AAD. Changing what is bound changes this, so the break is deliberate. */
+export const FIELD_CONTEXT_PREFIX = "pithy/enc/1";
+
+/** The AAD one field binds: `pithy/enc/1:<name>:<field path>:<pk>`. */
+export function fieldContext(name: string, path: string, pk: string): string {
+  return `${FIELD_CONTEXT_PREFIX}:${name}:${path}:${pk}`;
+}
+
+/** Depth countdown for {@link Opened}. */
+type Prev = [never, 0, 1, 2, 3, 4, 5, 6, 7, 8];
+
+/**
+ * The value `open` returns and `seal` takes: the schema's output, with every encrypted field the output of the schema
+ * it seals — the plaintext, carrying that schema's privacy brand — instead of an envelope. Read through the presence
+ * wrappers, objects, and the kit's JSON columns, to eight levels.
+ */
+export type Opened<S, D extends number = 8> = [D] extends [never]
+  ? z.output<S>
+  : [EncryptedInner<S>] extends [never]
+    ? S extends z.ZodOptional<infer I>
+      ? Opened<I, Prev[D]> | undefined
+      : S extends z.ZodNullable<infer I>
+        ? Opened<I, Prev[D]> | null
+        : S extends z.ZodDefault<infer I>
+          ? Exclude<Opened<I, Prev[D]>, undefined>
+          : S extends z.ZodReadonly<infer I>
+            ? Readonly<Opened<I, Prev[D]>>
+            : [SqliteJsonSchema<S>] extends [never]
+              ? S extends z.ZodObject<infer Shape>
+                ? { [K in keyof z.output<S>]: K extends keyof Shape ? Opened<Shape[K], Prev[D]> : z.output<S>[K] }
+                : z.output<S>
+              : Opened<SqliteJsonSchema<S>, Prev[D]>
+    : z.output<EncryptedInner<S>>;
+
+/** A field through its presence wrappers. */
+type Unwrapped<S> =
+  S extends z.ZodOptional<infer I>
+    ? Unwrapped<I>
+    : S extends z.ZodNullable<infer I>
+      ? Unwrapped<I>
+      : S extends z.ZodDefault<infer I>
+        ? Unwrapped<I>
+        : S;
+
+/** The top-level fields of a table schema that declare a blind index: what `match` takes. */
+export type IndexedField<S> =
+  S extends z.ZodObject<infer Shape>
+    ? { [K in keyof Shape & string]: [EncryptedIndexOf<Unwrapped<Shape[K]>>] extends [never] ? never : K }[keyof Shape &
+        string]
+    : never;
+
+/** What {@link sealed} takes. */
+export interface SealedOptions {
+  /** The field key: `env.PITHY_FIELD_KEYS`, or a resolved key set. Read on the first seal, open or match. */
+  keys: FieldKeys | undefined;
+}
+
+/** The repository seam over one table with encrypted fields. */
+export interface Sealed<S extends z.ZodType> {
+  /** The table schema. */
+  readonly schema: S;
+  /** The schema of the opened value: every encrypted field replaced by the schema it seals. `redact` takes it. */
+  readonly opened: z.ZodType<Opened<S>>;
+  /** Seal every encrypted field under the current version, write each blind index, then `Schema.encode`. */
+  seal(value: Opened<S>): Promise<z.input<S>>;
+  /** `Schema.parse`, then open every encrypted field and validate it against the schema it seals. */
+  open(row: z.input<S>): Promise<Opened<S>>;
+  /** The blind index of `value` for `field` under every live version, newest first: what an `IN (…)` looks for. */
+  match(field: IndexedField<S>, value: string): Promise<string[]>;
+}
+
+/** How a refusal names a schema: its description, or what it holds. */
+function schemaLabel(schema: z.ZodType, paths: readonly string[]): string {
+  const description = schema.description;
+  if (description) return `The schema "${description}"`;
+  return paths.length > 0 ? `The schema holding ${paths.join(", ")}` : "This schema";
+}
+
+/** The value at a path, or `undefined` where any step is absent. */
+function readPath(value: unknown, segments: readonly string[]): unknown {
+  let at: unknown = value;
+  for (const segment of segments) {
+    if (typeof at !== "object" || at === null) return undefined;
+    at = (at as Record<string, unknown>)[segment];
+  }
+  return at;
+}
+
+/**
+ * A copy of `value` with `next` at `segments`, copying each object on the way so the caller's value is never
+ * written to. A step that is absent stays absent: there is nothing there to seal.
+ */
+function writePath(value: unknown, segments: readonly string[], next: unknown): unknown {
+  const [head, ...rest] = segments;
+  if (head === undefined) return next;
+  if (typeof value !== "object" || value === null) return value;
+  const copy = { ...(value as Record<string, unknown>) };
+  copy[head] = rest.length === 0 ? next : writePath(copy[head], rest, next);
+  return copy;
+}
+
+/** The plaintext a value seals to: its schema's encoded form, as JSON. Refused without echoing the value. */
+function plaintextOf(field: EncryptedFieldLayout, value: unknown): string {
+  const encoded = field.spec.inner.safeEncode(value);
+  if (!encoded.success) {
+    throw fromZodError(encoded.error, {
+      message: `The encrypted field "${field.path}" does not hold a value its schema accepts.`,
+    });
+  }
+  const json = JSON.stringify(encoded.data);
+  if (typeof json !== "string") {
+    throw new ValidationError({ message: `The encrypted field "${field.path}" holds a value JSON cannot carry.` });
+  }
+  return json;
+}
+
+/** The opened schema: each encrypted field replaced by the schema it seals, along the paths that reach one. */
+function openedSchemaOf(schema: z.ZodType): z.ZodType {
+  if (encryptedSpecOf(schema)) {
+    // Keep the presence the field was given outside the marker.
+    const def = schema._zod.def as unknown as { type: string; innerType?: z.ZodType };
+    if (def.innerType && (def.type === "optional" || def.type === "nullable" || def.type === "readonly")) {
+      return schema.clone({ ...def, innerType: openedSchemaOf(def.innerType) } as never);
+    }
+    return (encryptedSpecOf(schema) as { inner: z.ZodType }).inner;
+  }
+  if (!hasEncryptedFields(schema)) return schema;
+  const def = schema._zod.def as unknown as { type: string; innerType?: z.ZodType; out?: z.ZodType };
+  if (def.innerType && ["optional", "nullable", "default", "prefault", "readonly"].includes(def.type)) {
+    return schema.clone({ ...def, innerType: openedSchemaOf(def.innerType) } as never);
+  }
+  if (isSqliteJson(schema) && def.out) {
+    const meta = schema.meta();
+    const column = sqliteJson(openedSchemaOf(def.out));
+    return meta ? column.meta(meta) : column;
+  }
+  if (def.type === "object") {
+    const object = schema as z.ZodObject;
+    const changed: Record<string, z.ZodType> = {};
+    for (const [key, field] of Object.entries(object.shape)) {
+      if (hasEncryptedFields(field)) changed[key] = openedSchemaOf(field);
+    }
+    return object.extend(changed);
+  }
+  return schema;
+}
+
+/**
+ * Build the seam for one table, from the request's `env`. Refuses, before any key is read, a schema with an encrypted
+ * field and no `encryptedTable()` declaration, and every layout problem. See the module note.
+ */
+export function sealed<S extends z.ZodType>(schema: S, options: SealedOptions): Sealed<S> {
+  const layout = encryptedLayout(schema);
+  const paths = layout.fields.map((field) => field.path);
+  if (layout.declaration === null) {
+    if (layout.fields.length > 0 || hasEncryptedFields(schema)) {
+      throw new ValidationError({
+        message: `${schemaLabel(schema, paths)} holds an encrypted field and declares no table.`,
+        action: "Wrap it in encryptedTable(schema, { name, key }): the name and key are what every sealed field binds.",
+      });
+    }
+    throw new ValidationError({
+      message: `${schemaLabel(schema, paths)} holds no encrypted field.`,
+      action: "Read and write it with Schema.parse and Schema.encode; sealed() is for a table with an encrypted field.",
+    });
+  }
+  if (layout.problems.length > 0) {
+    throw new ValidationError({
+      message: `${schemaLabel(schema, paths)} cannot be sealed as declared. ${layout.problems.join(" ")}`,
+      action: "Fix the schema's encrypted fields and their blindIndex columns.",
+    });
+  }
+  const { name, key } = layout.declaration;
+  let resolved: Promise<EncryptionConfig> | undefined;
+  const config = (): Promise<EncryptionConfig> => {
+    resolved ??= resolveFieldKeys(options.keys);
+    return resolved;
+  };
+  let opened: z.ZodType | undefined;
+
+  const primaryKey = (value: unknown, doing: string): string => {
+    const pk = readPath(value, [key]);
+    if (typeof pk !== "string" || pk.length === 0) {
+      throw new ValidationError({
+        message: `The ${name} row has no ${key} to bind, so it cannot be ${doing}.`,
+        action: `Assign ${key} in the app before the insert: every encrypted field binds it.`,
+      });
+    }
+    return pk;
+  };
+
+  return {
+    schema,
+    get opened() {
+      opened ??= openedSchemaOf(schema);
+      return opened as z.ZodType<Opened<S>>;
+    },
+
+    async seal(value) {
+      const pk = primaryKey(value, "sealed");
+      const keys = await config();
+      const current = Number(keys.currentVersion);
+      let out: unknown = value;
+      for (const field of layout.fields) {
+        const plain = readPath(value, field.segments);
+        if (plain === undefined || plain === null) {
+          if (field.indexField) out = writePath(out, [field.indexField], plain === null ? null : undefined);
+          continue;
+        }
+        const envelope = await encryptValue(keys, fieldContext(name, field.path, pk), plaintextOf(field, plain), {
+          keys: FIELD_KEYS_BINDING,
+        });
+        out = writePath(out, field.segments, envelope);
+        if (field.indexField && field.spec.index) {
+          if (typeof plain !== "string") {
+            throw new ValidationError({
+              message: `The encrypted field "${field.path}" has a blind index, so it holds a string.`,
+            });
+          }
+          out = writePath(out, [field.indexField], await blindIndexValue(keys, current, field.spec.index, plain));
+        }
+      }
+      const encoded = schema.safeEncode(out as z.output<S>);
+      if (!encoded.success) throw fromZodError(encoded.error);
+      return encoded.data;
+    },
+
+    async open(row) {
+      const parsed = schema.safeParse(row);
+      if (!parsed.success) throw fromZodError(parsed.error);
+      const pk = primaryKey(parsed.data, "opened");
+      const keys = await config();
+      let out: unknown = parsed.data;
+      for (const field of layout.fields) {
+        const envelope = readPath(parsed.data, field.segments) as EncryptedEnvelope | null | undefined;
+        if (envelope === undefined || envelope === null) continue;
+        const plaintext = await decryptValue(keys, fieldContext(name, field.path, pk), envelope, {
+          keys: FIELD_KEYS_BINDING,
+        });
+        let json: unknown;
+        try {
+          json = JSON.parse(plaintext);
+        } catch {
+          throw new ValidationError({
+            message: `The encrypted field "${field.path}" opened to something that is not JSON.`,
+          });
+        }
+        const value = field.spec.inner.safeParse(json);
+        if (!value.success) throw fromZodError(value.error);
+        out = writePath(out, field.segments, value.data);
+      }
+      return out as Opened<S>;
+    },
+
+    async match(field, value) {
+      const located = layout.fields.find((one) => one.path === field && one.indexField !== null);
+      if (!located?.spec.index) {
+        throw new ValidationError({
+          message: `"${String(field)}" is not a field of the ${name} table with a blind index.`,
+          action: `Declare it encrypted(x, { index: "exact" | "email" }) with a blindIndex("${String(field)}") column.`,
+        });
+      }
+      const keys = await config();
+      const normalizer = located.spec.index;
+      return Promise.all(liveVersions(keys).map((version) => blindIndexValue(keys, version, normalizer, value)));
+    },
+  };
+}