extension-oql

작성자: caffeinelabs

Make a canister's data queryable by the Caffeine Data Intelligence agent. Use whenever an app stores structured data (Maps/Lists/arrays of records) that should be answerable in natural language — "top customers", "revenue by region", "active projects". Adds a discoverable `schema()` and a JSON `execute()` query endpoint via the `caffeineai-oql` mops package's `Expose` mixin.

npx skills add https://github.com/caffeinelabs/skills --skill extension-oql

ZIP 다운로드 GitHub이 스킬 사용해 보기스폰서

OQL — Object Query Layer

Go over the actor's fields (non-transient) and, for each collection worth querying, consider how its data maps to a table in a database (an entity). You only declare one entity per table — the Expose mixin makes them queryable.

Backend

Each entity carries an authorization level; the default .controllerOnly() is safe (private to users, still readable by the Data Intelligence agent). Model your entities first, then pick a level per entity — see ## Auth.

Setup

Run mops add caffeineai-oql@0.5.3 in the same write batch as your first mo:caffeineai-oql/... import. Auto-derivation requires moc >= 1.11 (the generated-app template already satisfies this).

Build flags

--default-persistent-actors and --implicit-package=core are mandatory — without them the library does not compile. If the app uses OQL.Table and needs more than 4 GiB of Region, add --max-stable-pages 1638400 as well; a dependency's own flags are not applied to the project that depends on it, so it has to be set in the app's own build.

Imports — one per resolver module

.toEntity, the builder chain (.sample / .build / .public_ / …), and record _toRow derivation are resolved from modules imported top-level in the file that declares entities — the resolver does not walk submodules, so importing only mo:caffeineai-oql is not enough. Import exactly the resolver modules your code uses:

mo:caffeineai-oql/Entity — always (the .sample / .build / .edge / .ownedBy / auth-level builder chain, and .payload / .flatten in manual mode).
the collection module for each .toEntity / .toEntityManual receiver — MapEntity, SetEntity, ListEntity, ArrayEntity, or VarArrayEntity.
for each auto-derived (.toEntity) record: RecordValue, plus one <Type>Value per primitive field type present — NatValue, TextValue, PrincipalValue, BoolValue, IntValue, FloatValue, the sized Nat/Int widths, BlobValue. Manual .payload return types need their <Type>Value too; manual-only entities need no RecordValue.

When a collection module is missing the compiler names it — "field toEntity does not exist … Did you mean to import mo:caffeineai-oql/MapEntity?" — add the named import. A missing Entity import gets no such hint: it surfaces as a bare "field payload does not exist in type Builder<…>" (M0072). Treat any field <builderMethod> does not exist error as a missing top-level import from this list, never as a wrong package version.

Declare entities and install

.toEntity(name, typeName, primaryKey) turns a collection of records into a queryable entity; the compiler auto-derives the fields. Each entity sets its own authorization level (see ## Auth); the example below shows one table per level. Expose adds only the OQL query methods (schema / execute) — your existing state, types, and shared methods are untouched.

Always call .sample({...}) on every .toEntity / Entity.manual chain; dummy values are fine. Empty collection + no sample → empty schema (fields: [] / "record { }").

include Expose({ entities = [tasks.toEntity("task", "Task", "id").sample({ id = 0; title = "" }).public_().build()] })

import Map       "mo:core/Map";
import Nat       "mo:core/Nat";
import Principal "mo:core/Principal";
import OQL       "mo:caffeineai-oql";
import Expose    "mo:caffeineai-oql/Expose";
// Resolver modules, imported top-level (see "Imports" above). This app derives
// Map entities over records of Nat / Text / Principal fields:
import MapEntity      "mo:caffeineai-oql/MapEntity";
import Entity         "mo:caffeineai-oql/Entity";
import RecordValue    "mo:caffeineai-oql/RecordValue";
import NatValue       "mo:caffeineai-oql/NatValue";
import TextValue      "mo:caffeineai-oql/TextValue";
import PrincipalValue "mo:caffeineai-oql/PrincipalValue";

actor {
  type Product  = { id : Nat; name : Text; priceUsd : Nat };
  type Vendor   = { id : Nat; name : Text };
  type AuditLog = { id : Nat; action : Text; atNs : Nat };
  type Note     = { id : Nat; user : Principal; body : Text };
  type Document = { id : Nat; owner : Principal; title : Text; ciphertext : Text };
  type User     = { id : Principal; isAdmin : Bool };

  let products  = Map.empty<Nat, Product>();
  let vendors   = Map.empty<Nat, Vendor>();
  let supplies  = Map.empty<Product, Vendor>();
  let auditLogs = Map.empty<Nat, AuditLog>();
  let notes     = Map.empty<Nat, Note>();
  let documents = Map.empty<Nat, Document>();
  // not all collections need to be exposed if there is no need — `users` backs
  // auth only, so it is intentionally never turned into an entity below
  let users     = Map.empty<Principal, User>();

  let anyP = Principal.fromText("aaaaa-aa");   // sample owner; the value is ignored

  // Look up whether a caller is an admin.
  func isAdmin(p : Principal) : Bool =
    switch (users.get(p)) { case (?u) u.isAdmin; case null false };

  // A custom .ownedByWith rule: admins see every document, everyone else only
  // their own. `owner` is the field's Value — a Principal column arrives as #text.
  func canSeeDocument(caller : Principal, owner : OQL.Value) : Bool =
    isAdmin(caller) or owner == #text(caller.toText());

  include Expose({
    entities = [
      // #public_ — anyone, incl. anonymous, reads the whole catalogue
      products.toEntity("product", "Product", "id")
        .sample({ id = 0; name = ""; priceUsd = 0 })
        .public_()
        .build(),
      vendors.toEntity("vendor", "Vendor", "id")
        .sample({ id = 0; name = "" })
        .public_()
        .build(),
      // `supplies : Map<Product, Vendor>` — a map between two non-primitive types.
      // The identity lives in the key/value records, not a field, so iterate
      // .entries() in manual mode, promote each side's id, and .edge both — a
      // query can then traverse "product.name" and "vendor.name".
      OQL.Entity.manual<(Product, Vendor)>("supply", func () = supplies.entries(), "Supply", "key")
        .sample(({ id = 0; name = ""; priceUsd = 0 }, { id = 0; name = "" }))
        .payload("key",     func ((p, v)) = p.id.toText() # ":" # v.id.toText())
        .payload("product", func ((p, _)) = p.id) .edge("product", "product")
        .payload("vendor",  func ((_, v)) = v.id) .edge("vendor",  "vendor")
        .controllerOnly()
        .build(),
      // #controllerOnly (the default, shown explicitly) — only the platform reads
      auditLogs.toEntity("auditLog", "AuditLog", "id")
        .sample({ id = 0; action = ""; atNs = 0 })
        .controllerOnly()
        .build(),
      // #scopedPerUser — each signed-in user reads only their own rows
      notes.toEntity("note", "Note", "id")
        .sample({ id = 0; user = anyP; body = "" })
        .ownedBy("user")
        .scopedPerUser()
        .build(),
      // #controllerOrScoped — controller reads all; scoped reads use canSeeDocument.
      // `.hidden` — opaque column absent from schema + default projection
      documents.toEntity("document", "Document", "id")
        .sample({ id = 0; owner = anyP; title = ""; ciphertext = "" })
        .hidden("ciphertext")
        .ownedByWith("owner", canSeeDocument)
        .controllerOrScoped()
        .build(),
    ];
  });
}

Auth

Authorization is per entity — each builder declares a level, and schema() and execute() both run the check against the live caller. No app-wide config, no tokens. The default when none is set is #controllerOnly.

Builder call	Who reads	Rows returned
`.public_()`	anyone (incl. anonymous)	all
`.controllerOnly()` (default)	controllers only	all
`.scopedPerUser()`	any signed-in caller	only the caller's own
`.controllerOrScoped()`	controllers + signed-in callers	controller: all; user: own

Choosing a level

Pick per entity by who should read its rows — when in doubt, keep the default.

.controllerOnly() (default) — private app data the agent should answer over, but no end user reads directly (orders, metrics, audit logs, config). The agent calls as the controller, so it reads everything while the data stays private to users.
.public_() — world-readable data, including logged-out visitors (public catalogue, published content, leaderboards).
.controllerOrScoped() — per-user data where each user reads only their own rows, but the agent must still answer aggregate questions (profiles, a user's orders). Requires an owner column.
.scopedPerUser() — strictly private per-user data: each user reads only their own, and the agent is scoped too, so it cannot answer over this table (DMs, private journals). Requires an owner column — prefer .controllerOrScoped() unless the agent must be blind to it.

The user may override per entity; if a request implies per-user data but is ambiguous, ask.

Per-user (row-level) scoping

Scoped levels (.scopedPerUser(), .controllerOrScoped()) need a way to know which rows belong to the caller — an owner column or a subject-honouring source. .build() traps if a scoped entity has neither, and also traps if a .public_() entity declares an owner (the check would never run). This is the guardrail against the common data-leak footgun.

When to tag: a Principal field is the signal.

.ownedBy(field) — the field is the owner; visibility is identity equality.
.ownedByWith(field, canSee) — custom visibility (teams, admins, sharing). canSee : (caller : Principal, owner : Value) -> Bool decides per row; field need not be a Principal, and the closure can read actor state.

A scoped caller sees only its owned rows — both as the query target and through a join — so traversal can never leak another owner's rows.

// Per-user notes: each signed-in user reads only their own rows.
notes.toEntity("note", "Note", "id")
  .sample({ id = 0; owner = Principal.fromText("aaaaa-aa") /* any principal */; body = "" })
  .ownedBy("owner")
  .scopedPerUser()
  .build()

// .ownedByWith custom rule: the owner sees their own docs, listed admins see
// everyone's, and the platform controller sees all (#controllerOrScoped).
// `owner` is the field's Value — a Principal column arrives as #text(principal).
docs.toEntity("doc", "Doc", "id")
  .sample({ id = 0; owner = Principal.fromText("aaaaa-aa"); title = "" })
  .ownedByWith("owner", func (caller, owner) =
    admins.get(caller) != null or owner == #text(caller.toText()))
  .controllerOrScoped()
  .build()

Where ownership decides which rows a scoped caller sees, .viewWith(view) decides what shape it sees them in — a per-subject redaction that runs only on rows the ownership check already admitted:

// Everyone sees their own bookings; exact amounts only on their own rows is
// not needed here — but coarsen the contact field for non-owners of a shared
// calendar, say:
bookings.toEntity("booking", "Booking", "id")
  .sample({ id = 0; calendarId = 0; contact = "" })
  .ownedByWith("calendarId", canSeeCalendar)
  .viewWith(func (subject, b) = if (isOwner(subject, b)) b else { b with contact = "" })
  .scopedPerUser()
  .build()

view : (subject : Principal, row : T) -> T reshapes the typed row; the whole query pipeline (filters included) evaluates the VIEWED row, so a predicate can never probe a value the view hides. Views run for scoped subjects only — pairing .viewWith with .public_() traps at .build() (unrestricted reads always see raw rows).

.ownedBy(f) is exactly .ownedByWith(f, OQL.Entity.ownerIsCaller). At most one owner column; it must be a real field, not also .edge / .hidden. For owner-keyed storage (Map<Principal, List<T>>) use OQL.Entity.newScoped(name, scopedIter, typeName, primaryKey) so the scan is O(user rows): scopedIter(?p) returns only p's rows, scopedIter(null) all (schema seeding).

Entity builder

Two modes, picked by the row type T.

Auto-derivation — `.toEntity`

For records whose fields are all primitives with a built-in _toRow (Nat, Int, Float, Text, Bool, the sized Nat/Int widths, Principal):

customers.toEntity(name, typeName, primaryKey)
  .sample(template)              // REQUIRED — empty collection + no sample → empty schema
  .edge(field, targetEntity)     // tag an existing field as a foreign key
  .ownedBy(field)                // (or .ownedByWith(field, canSee)) per-user scoping
  .scopedPerUser()               // auth level: .public_ / .controllerOnly (default) / .scopedPerUser / .controllerOrScoped
  .hidden(field)                 // opaque/sensitive — drop from schema + default projection
  .build()

.toEntity is sugar for OQL.Entity.new<T>(name, func () = coll.values(), …); it exists on Map, Set, List, [T], and [var T]. It iterates values only — if a row's identity (PK or owner) lives in the Map key, it is not a field: promote it via manual mode over .entries(), or OQL.Entity.newScoped when it's the owner.
primaryKey, and any .edge / .ownedBy field, must name a real, non-.hidden column of the row.
.edge(name, target) tags an existing field (it does not add one) as an FK, enabling dotted-path traversal "name.targetField" in queries. FK/PK types must be Text, Nat/Int, or Bool (Float keys are rejected), and the target's primary key must not be .hidden.
The edge target must be an entity registered in THIS canister's Expose. An edge to an absent entity (a typo, or an FK into another canister) silently drops the whole field from schema() — the column still stores and filters, but no schema-driven client can discover it. A cross-canister FK belongs as a plain payload field, not an .edge.
.sample(template) seeds schema discovery. Always call it; without it an empty collection yields an empty schema (fields: []). Only the shape matters, not the values.
.hidden(name) drops a derivable field from schema + default projection; it does not skip _toRow — unsupported field types still need manual mode or a <Type>Value.

Schema fields are listed in lexicographic order (the __record combiner's canonical form); sort client-side if display order matters.

Manual mode — `.toEntityManual` / `OQL.Entity.manual`

For non-record T, computed fields, or records with nested / variant / option / collection fields:

// REQUIRED, top-level in this file — `.payload` / `.flatten` are Entity
// functions reached by receiver notation, not fields of the builder:
import Entity "mo:caffeineai-oql/Entity";

authors.toEntityManual<Author>("author", "Author", "id")
  .sample({ id = 0; name = ""; address = { street = ""; city = "" }; tags = [] })
  .payload("name", func a = a.name)        // one field; extract returns a _toRow value
  .flatten(func a = a.address)             // splice a nested record's fields as columns
  .payload("tagCount", func a = a.tags.size())
  // .edge tags a declared column; .hidden drops one you already added via
  // `.payload` / `.flatten` — omit fields by not adding them
  .build()

field payload does not exist in type Builder<…> (M0072) means the Entity import above is missing — it is never a package-version problem. .payload and .flatten have been Entity functions since 0.1.0; mops add-ing a different caffeineai-oql version will not fix this error. Importing only mo:caffeineai-oql is not enough, and neither is reaching the module through the OQL.Entity.… re-export: receiver notation resolves against top-level imports only, so a file that calls OQL.Entity.manual(...).payload(...) still needs its own import Entity "mo:caffeineai-oql/Entity";. Unlike .toEntity, this failure carries no "Did you mean to import …?" hint.

.payload(name, extract) — name must not contain .. Prefer func r = r.field (let Motoko infer; avoid redundant annotations). For options/variants, return Text/Nat with a sentinel (see below).
.flatten(extract : T -> S) — S must be flat; each of its fields becomes a top-level column. Drop unwanted ones with .hidden. Name collisions get __1, __2 suffixes (nothing is dropped).
OQL.Entity.manual<T>(name, iter, typeName, primaryKey) for arbitrary row sources (custom flatteners, filtered iterators). Always chain .sample(...) with one dummy row of type T. The qualified call resolves through the OQL import, but any .payload / .flatten chained onto its result still needs the top-level Entity import.

OQL.Value is { #null_; #bool; #nat; #int; #float; #text }. Numeric variants compare across each other, so a JSON integer threshold matches a Float value.

Row type `T`	Mode
All-primitive record	`.toEntity`
Record with `?` / variant / nested field	`.toEntity` once you ship `<Type>Value.mo` (below); else manual
Record with a collection field	manual — `.size()` or `Text.join` into a payload
Tuple / primitive / computed	manual

Converting non-primitive fields

To keep a record on the auto-derive path, give each non-primitive field type a _toRow : T -> OQL.Value: one file per type named <TypeName>Value.mo, a single public func _toRow, imported top-level in the file that declares entities — the same top-level rule as the built-in value modules (see ## Setup → Imports); the resolver does not walk submodules. Parent records then ride .toEntity(...) with no per-field .payload.

// OptTextValue.mo — option → sentinel
module { public func _toRow(self : ?Text) : OQL.Value =
  switch self { case null { #text("") }; case (?t) { #text(t) } }; };

// StatusValue.mo — variant → tag text
module { public func _toRow(self : Status) : OQL.Value =
  #text(switch self { case (#draft) "draft"; case (#published) "published" }); };

// DepartmentValue.mo — nested record → child PK (then .edge the field)
module { public func _toRow(self : Department) : OQL.Value = #text(self.name); };

Always return ONE Value variant, even for null (sentinel "" / 0 / false) — a _toRow that sometimes returns #null_ makes the reported schema type flip-flop by row order. Sentinels keep the field queryable (eq value "" matches the nulls). For a one-off field, inline the same conversion in a .payload instead of a module; lift to a module only when 2+ entities need it. A record used both as an entity and as a nested field just ships its <Type>Value.mo — the structural Row derivation and your Value collapse are distinct types and coexist.

Entity patterns beyond one-row-per-record

The same storage can back several entities — pick what the client should see:

reshaped — flatten Map<K1, Map<K2, V>> into rows; have the flattener emit a flat record (not a tuple) so it still auto-derives, then .edge the promoted keys.
enumerated — derive an entity from index keys (Map<Author, …>.keys()) via OQL.Entity.manual; entries with no rows simply don't appear.
synthetic — project a junction from an array field to make a many-to-many queryable from both sides:

OQL.Entity.manual<(Article, Text)>("articleTag", func () = flattenTags(articles), "Pair", "pair")
  .sample(({ id = 0 }, ""))
  .payload("article", func ((a, _)) = a.id) .edge("article", "article")
  .payload("tag",     func ((_, t)) = t)    .edge("tag", "tag")
  .build()

Larger data — `OQL.Table`

For a table expected to grow large (tens of thousands of rows and up — events, transactions, logs, imported datasets), store it in an OQL.Table. It scales far past what heap collections hold; rows are keyed by their append position.

Declaring one is three decisions — columns, indexes, row function:

import OQL    "mo:caffeineai-oql";
import Expose "mo:caffeineai-oql/Expose";
import Entity "mo:caffeineai-oql/Entity";   // the builder chain — always import
import Table  "mo:caffeineai-oql/Table";

actor {
  type Event = { kind : Nat; amount : Nat; note : Text };

  // 1. Columns: (name, type) pairs. Order matters and the set is FIXED for the
  //    table's life. Types: #nat / #int / #float / #bool (64-bit cells) + #text.
  // 2. Indexes: the columns queries will filter or order by —
  //    #hash for equality, #ordered for ranges / orderBy.
  let events = Table.new(
    [("kind", #nat), ("amount", #nat), ("note", #text)],   // columns
    [("kind", #hash), ("amount", #ordered)],               // indexes
  );

  // 3. Row function: your record → one (name, Value) per column, names matching.
  func eventRow(e : Event) : [(Text, OQL.Value)] =
    [("kind", #nat(e.kind)), ("amount", #nat(e.amount)), ("note", #text(e.note))];

  // Writes: the returned position is the row's primary key.
  // A modify is delete + append.
  public func addEvent(e : Event) : async Nat {
    Table.append(events, e, eventRow);
  };

  // REQUIRED: register the table's entity in the Expose mixin — schema() and
  // execute() only see what Expose registers; a Table that is not in this list
  // exists but is invisible to every query. Same list, same auth levels, and
  // FK edges to other entities, as for any entity. `entity` defaults the type
  // name to the entity name and the primary key to "id";
  // `entityWith(events, "event", "Event", "id")` overrides either.
  include Expose({
    entities = [
      Table.entity(events, "event").public_().build(),
      // ... the app's other entities ...
    ];
  });
};

Rules that matter:

The schema is fixed once data is flushed — no adding, dropping, or retyping columns; a different shape means a new table.
A value's kind must equal its column's type — a mismatch traps loudly at append rather than storing corrupt cells; numeric cells are 64-bit.
A table that will be bulk-loaded is declared with NO indexes — index after the load (next section).
A Table is queryable only through its entity in Expose — declaring the table alone stores data but exposes nothing; schema() / execute() see exactly the entities list.

Beyond what the index serves, a Table answers whole-column sum / avg straight from segment stats — flat in table size — and reads only the columns a query touches.

Bulk upload (`ImportData`) — load existing data into a `Table`

When the user's data already exists (a CSV export from a spreadsheet or the system the app replaces), don't trickle it through append — include the ImportData mixin and load it as pre-built segment images: the loader lays rows out off-chain in exactly a flushed segment's byte layout and the canister validates and copies the bytes, so one message costs O(columns) instead of O(rows), and the heap stays flat throughout.

import Expose     "mo:caffeineai-oql/Expose";
import Entity     "mo:caffeineai-oql/Entity";   // the builder chain — always import
import Table      "mo:caffeineai-oql/Table";
import ImportData "mo:caffeineai-oql/ImportData";

actor {
  // Declare the table INDEX-FREE for the load — loadSegment traps on a table
  // that declares an index (a ready index missing the loaded rows would
  // silently under-fetch). The index is built after the load, in the background.
  let events = Table.new([("kind", #nat), ("amount", #nat), ("note", #text)], []);

  include Expose({
    entities = [
      Table.entity(events, "event").public_().build(),
      // ... the app's other entities ...
    ];
  });
  include ImportData([ events.importTarget("event") ]);
};

The mixin adds controller-only endpoints (layout, rows, putSegment, importFlush, buildIndex, indexStatus) — the image's stats are trusted answers, so the load surface belongs to the principal that could install code anyway.

Running the loader. The tool ships with this skill in scripts/ — plain Node (≥ 20) with no dependencies to install; every call goes through icp canister call, so icp-cli must be set up:

node <this skill's directory>/scripts/ingest.mjs \
  --canister <canister-id> --target event --file events.csv \
  -e ic --index kind:hash

--target is the importTarget name declared in the canister; the canister's layout() is the schema authority — the CSV header is matched to the columns by name, so file column order never matters (extra CSV columns are ignored; a declared column missing from the file is an error).
Connection flags pass through to icp-cli, the same way every other canister call works: -e <environment> (-e ic for the deployed app), -n <network>, --identity <name> (default: your current one). The endpoints are controller-only, so the identity must be a controller of the canister. With none of them, your icp defaults apply — a canister NAME then resolves against the project's default environment, so run the tool from the app's project directory.
--index col:kind (repeatable, kind = hash) builds and uploads that column's index off-chain after the rows. For anything it can't build (#ordered, composites), build on-chain instead through the mixin's endpoint: icp canister call <id> buildIndex '("event", vec {"kind"}, variant {hash})' — queries scan (correctly, just slower) until the build completes, then the index serves; indexStatus '("event")' reports progress.
Verify a load the same way clients will read it: icp canister call <id> execute '("{\"start\":\"event\",\"aggregate\":[{\"fn\":\"count\"}]}")' --query must return the CSV's row count.
CSV cells decode by declared type. An unquoted empty field is a null cell; a quoted "" in a #text column is the empty string; a #bytes(w) column's field is base64 and must decode to exactly w bytes.
Re-running the same command resumes: rows the table already holds are skipped, and putSegment's expect-first-row guard turns any double-send into a loud trap instead of duplicated rows.

Operating rules (violations refuse loudly rather than corrupt):

Finish the first load completely — data and indexes — before the app starts writing. If writes land mid-load, a re-run cannot finish the index; rebuild it in-canister with buildIndex (slower, always works).
After that, load and write in turns. Between loads the app writes freely; the next run of the same command picks up the growth as a delta load, nothing to prepare.
Never both at once — the canister refuses, it does not mis-answer.
Queries keep working throughout, including mid-load; a column whose index is still uploading scans until it finishes.