implement-command

Implement a node-redis Command

Overview

A command in node-redis is a single Command object exported from packages/<pkg>/lib/commands/<NAME>.ts. It declares how to serialize arguments onto the wire (parseCommand) and how to map the RESP reply to a JS value (transformReply). It becomes callable on clients only after it is registered in the package's commands/index.ts. Both the raw name and a camelCase alias are exposed (HSET and hSet).

This skill is for @redis/client core commands and for module packages (@redis/json, @redis/search, @redis/bloom, @redis/time-series). It does not cover RESP codec changes or new client transports.

Before writing, read 2-3 existing commands with a similar shape (simple key read, key + options, variadic, RESP2/3-divergent reply) and mirror them. Consult redis.io/commands for argument order and reply type, but treat existing parseCommand/transformReply files as the source of truth for repo conventions.

File layout

Naming: file name = raw wire name with subcommand _ separators (ACL_CAT, CONFIG_GET, CLUSTER_FORGET). A distinct reply variant gets its own file (HRANDFIELD_COUNT_WITHVALUES). Module commands drop the dotted prefix from the file name (ARRAPPEND.ts → wire JSON.ARRAPPEND).

Step 0 — Gather inputs (ask the user first)

New commands are often implemented before they are publicly released, so redis.io may not document them yet and a default redis:latest may not have them. Before writing any code, ask the user for three things:

1. The command spec. Ask for the redis/redis JSON spec file — one per command under src/commands/<name>.json (subcommands use -, e.g. client-info.json, acl-cat.json). If the command is unreleased, ask the user to paste the spec from their branch. As a fallback on a live server: redis-cli --json COMMAND DOCS <name> and COMMAND INFO <name>.
2. A running Redis instance that has the command. This is the single most useful input — ask for it explicitly and up front. Ask for connection details (host/port, TLS, auth, module loaded). Use it to explore real behavior and confirm the implementation matches the spec — do not rely on the spec alone. Probe every argument branch and diff the real reply against your transformReply. A quick redis-cli session or a throwaway probe script (packages/client is already wired for tsx) is enough; never commit the probe.
3. The first server version that ships the command. Ask which Redis (or module) version introduced it — the spec's since field is the answer when present; otherwise ask the user directly. You need this to (a) write @since in the JSDoc (Step 2) and (b) gate the behavior tests with minimumDockerVersion (Step 3) so they don't run — and fail — on older servers in CI.

If the user cannot provide a spec, derive arguments/reply from redis.io/commands but flag that it is unverified. If they cannot provide a live instance, implement from the spec but state that runtime behavior was not confirmed. If the introducing version is unknown, say so and leave @since/minimumDockerVersion out rather than guessing.

Reading the spec JSON → mapping to a Command

The redis/redis spec drives every part of the Command object. Example (getex.json, trimmed):

{
  "GETEX": {
    "since": "6.2.0", "arity": -2,
    "command_flags": ["WRITE", "FAST"],
    "key_specs": [{ "begin_search": { "index": { "pos": 1 } }, "flags": ["RW", "UPDATE"] }],
    "arguments": [
      { "name": "key", "type": "key", "key_spec_index": 0 },
      { "name": "expiration", "type": "oneof", "optional": true, "arguments": [
        { "name": "seconds", "type": "integer", "token": "EX" },
        { "name": "persist", "type": "pure-token", "token": "PERSIST" }
      ]}
    ],
    "reply_schema": { "oneOf": [ { "type": "string" }, { "type": "null" } ] }
  }
}

After implementing, run the command against the live instance and diff the real reply against reply_schema and your transformReply output.

Step 1 — Write <NAME>.ts

Minimal pass-through command (packages/client/lib/commands/GET.ts):

import { CommandParser } from '../client/parser';
import { RedisArgument, BlobStringReply, NullReply, Command } from '../RESP/types';

export default {
  CACHEABLE: true,
  IS_READ_ONLY: true,
  parseCommand(parser: CommandParser, key: RedisArgument) {
    parser.push('GET');
    parser.pushKey(key);
  },
  transformReply: undefined as unknown as () => BlobStringReply | NullReply
} as const satisfies Command;

as const satisfies Command is mandatory — it preserves the literal arg types for the public API while type-checking the shape.

Command flags (all optional)

• IS_READ_ONLY: true — read command; routable to replicas. Set for reads, omit/false for writes.
• CACHEABLE: true — eligible for client-side caching. Only for pure reads with no side effects.
• NOT_KEYED_COMMAND: true — command takes no key (server/connection level, e.g. PING, CONFIG_GET).
• IS_FORWARD_COMMAND — internal; do not set on new commands.

parseCommand — serialize args via CommandParser

First arg is always parser. Push the wire name first, then args in order. Use the parser helpers — do not hand-build arrays:

• push(...args) — raw args (the command token, flags, stringified numbers).
• pushKey(key) — a key. Registers it for cluster slot routing. Use for every key, never push a key.
• pushKeys(keys) / pushKeysLength(keys) — multiple keys; the Length variant prefixes the count.
• pushVariadic(vals) — a RedisVariadicArgument (one value or array) as flat args.
• pushVariadicWithLength(vals) — same, prefixed with the count (e.g. FIELDS <n> ...).
• pushVariadicNumber(vals) — number or array of numbers, stringified.

Numbers are not auto-stringified by push — call .toString(). Optional trailing args go in an options object; export its interface (see SET.ts's SetOptions). Encode keyword flags conditionally:

parseCommand(parser: CommandParser, key: RedisArgument, value: RedisArgument, options?: SetOptions) {
  parser.push('SET');
  parser.pushKey(key);
  parser.push(value);
  if (options?.condition) parser.push(options.condition); // 'NX' | 'XX'
}

transformReply — map RESP reply to JS

• Pass-through (reply already the right shape): transformReply: undefined as unknown as () => <ReplyType>.
• Function: (reply: <RawType>) => <JsType>. Use UnwrapReply<...> to read the raw RESP container.
• RESP-version keyed: { 2: (reply) => ..., 3: (reply) => ... } when RESP2 and RESP3 shapes differ (e.g. flat array vs map/tuple). See HRANDFIELD_COUNT_WITHVALUES.ts.

Unify RESP2 onto the RESP3 shape

When the server returns different shapes per protocol, the library exposes one return type to callers: the RESP3 shape is the source of truth, and the RESP2 reply is transformed to look like it. So the keyed form is almost always:

• 3: — pass-through (undefined as unknown as () => <ReplyType>), because RESP3 already has the target shape (map, tuple, big-number, double, ...).
• 2: — a function that reshapes the flat/legacy RESP2 reply into that same <ReplyType>. Type its input UnwrapReply<Resp2Reply<ReplyType>> so the raw RESP2 container is visible while the output type still matches RESP3.

Canonical example — HELLO.ts turns the RESP2 flat array ([k, v, k, v, ...]) into the RESP3 map, while RESP3 passes through:

transformReply: {
  2: (reply: UnwrapReply<Resp2Reply<HelloReply>>) => ({
    server: reply[1], version: reply[3], proto: reply[5], /* ... */
  }),
  3: undefined as unknown as () => HelloReply
}

Reuse shared transformers where one exists (HGETALL.ts uses transformTuplesReply for 2:, map pass-through for 3:). Only when RESP3 still needs reshaping does 3: get its own function too. Verify the actual per-protocol shapes against the live instance (Step 0) — connect once with RESP: 2 and once with RESP: 3 and diff.

Reply types live in RESP/types: BlobStringReply, SimpleStringReply<'OK'>, NumberReply, DoubleReply, NullReply, BooleanReply, ArrayReply<T>, TuplesReply<[...]>, MapReply, UnwrapReply.

RESP3 is the default. No separate RESP3 test is needed for a new command; the default test setup already exercises RESP3.

Type-mapping precision caveats

• A BLOB_STRING reply cannot be remapped to Number via type mapping; only RESP3 DOUBLE/BIG_NUMBER are precision-risky.
• If a NumberReply can exceed Number.MAX_SAFE_INTEGER (2^53-1), add a @remarks line to the JSDoc (Step 2) telling users to do client.withTypeMapping({ [RESP_TYPES.NUMBER]: String }). See the ARGREP entries in the client index for the exact wording.

Module package commands

Import from the published client subpath, prefix the wire name, and reuse shared transformers (packages/json/lib/commands/ARRAPPEND.ts):

import { CommandParser } from '@redis/client/dist/lib/client/parser';
import { RedisArgument, NumberReply, Command } from '@redis/client/dist/lib/RESP/types';

export default {
  IS_READ_ONLY: false,
  parseCommand(parser, key, path, value) {
    parser.push('JSON.ARRAPPEND');
    parser.pushKey(key);
    parser.push(path, /* transform */ value);
  },
  transformReply: undefined as unknown as () => NumberReply
} as const satisfies Command;

Step 2 — Register in commands/index.ts

import the command, then add it to the default-export map twice: the raw name (shorthand) and a camelCase alias. Every entry MUST have a JSDoc block directly above it — npm run check:command-jsdoc fails on any registry entry without an attached JSDoc comment (no blank-line gap allowed).

import GET from './GET';
// ...
export default {
  /**
   * Returns the value of a key, or null if the key does not exist
   * @param key - Key to read
   * @since 1.0.0
   */
  GET,
  /**
   * Returns the value of a key, or null if the key does not exist
   * @param key - Key to read
   * @since 1.0.0
   */
  get: GET,
} satisfies RedisCommands;

Keep both JSDoc blocks (raw + alias) in sync. Document every parseCommand param after parser with @param. Add @since <version> with the introducing server version from Step 0 (the spec's since); omit it only if that version is unknown. Add @remarks for the precision caveat above when relevant. For module packages the registry files are packages/<pkg>/lib/commands/index.ts (bloom: per-family .../<family>/index.ts).

Step 3 — Write <NAME>.spec.ts (co-located)

Two layers: arg serialization (no server) + behavior (real server, server + cluster topologies). Mirror GET.spec.ts:

import { strict as assert } from 'node:assert';
import testUtils, { GLOBAL } from '../test-utils';
import { parseArgs } from './generic-transformers';
import GET from './GET';

describe('GET', () => {
  it('transformArguments', () => {
    assert.deepEqual(parseArgs(GET, 'key'), ['GET', 'key']);
  });

  testUtils.testAll('get', async client => {
    assert.equal(await client.get('key'), null);
  }, {
    client: { ...GLOBAL.SERVERS.OPEN, minimumDockerVersion: [8, 8] },
    cluster: { ...GLOBAL.CLUSTERS.OPEN, minimumDockerVersion: [8, 8] }
  });
});

• parseArgs(COMMAND, ...args) asserts the exact wire array — cover each option/flag branch and variadic shapes. Arg tests need no server, so never gate them by version.
• testUtils.testAll(name, fn, { client, cluster }) runs the same body against a standalone server and a cluster. Use it so cluster key routing (pushKey) is exercised. Drop cluster only when the command is genuinely cluster-incompatible.
• Gate every behavior test by the introducing version (Step 0). Spread minimumDockerVersion: [major, minor] into both the client and cluster options (as above). CI runs multiple server versions; without the gate the test runs on older servers that lack the command and fails. [8, 8] = "8.8 and newer". Apply the same to testWithClient/testWithCluster by spreading it into their single options object. Omit only if the version is genuinely unknown.
• Pick the right GLOBAL.SERVERS.* / GLOBAL.CLUSTERS.* setup (see test-utils.ts); OPEN is the default.
• Docker is required — test-utils starts real Redis containers.

Step 4 — Regenerate static command metadata

Cluster/sentinel routing (replica-safety, keyedness, CSC eligibility) is resolved from a generated table: packages/client/lib/command-metadata/command-metadata-data.ts. The file is auto-generated — never edit it manually. The table lives in @redis/client, but the COMMAND dump includes module commands (ft, json, bf, ts, ...), so a new command in any package needs a regenerate:

npm run generate:metadata --workspace=packages/client -- redis://localhost:6379

The script rebuilds the entire table from a single live server's COMMAND reply and overwrites the file — entries the server doesn't report are silently dropped. Run it only against a server with all bundled modules loaded and a current core command set (e.g. the CI image redislabs/client-libs-test or a full Redis 8.8+ build); reuse the Step 0 instance only if it meets that bar — a server with just the new command's module would wipe every other module's metadata. The script then applies the curation in packages/client/scripts/command-metadata-overrides.ts: hand-curated excludes (internal, deprecated and cluster-admin commands) plus per-command routing overrides. After regenerating, verify the new command has an entry with the expected flags, and check git diff on command-metadata-data.ts: it must contain only the intended additions/changes. Deletions of other modules' or core entries mean the source server was incomplete — revert and rerun against a full build.

How the table and the command object interact (override-first — see lib/command-metadata/predicates.ts):

• The table sets defaults only, for both replica routing (cluster and sentinel) and client-side caching eligibility. A keyed entry without the write flag is already replica-safe, so a keyed read command with a correct table entry needs no IS_READ_ONLY on the command object — bloom and json omit it everywhere.
• IS_READ_ONLY/CACHEABLE on the command object win over the table. Set them as deliberate corrections, not to restate the table. The main case is keyless reads, which default to master routing: PING/INFO in the client, and module reads whose args are not keys — search and time-series set IS_READ_ONLY: true on exactly their keyless commands (FT.SEARCH takes an index name, TS.MGET/TS.MRANGE take filters, ...).
• Table-shape fixes (wrong, missing or excluded entries, routing policies) belong in command-metadata-overrides.ts; value intent (IS_READ_ONLY, CACHEABLE) belongs in the command definition, never in the overrides file.

Step 5 — Build, verify, lint

npm run build                                   # tsc --build (project references)
npm run check:command-jsdoc                     # registry JSDoc gate
npm run test-single -- packages/<pkg>/lib/commands/<NAME>.spec.ts
npm run lint                                     # changed files

If the build fails on stale dist/ from project references:

find packages -type d -name "dist" -exec rm -rf {} + && npm run build

For module packages, build the client first (or whole repo) — they import from @redis/client/dist.

Completion checklist

• Asked the user for spec, a live instance with the command, and the introducing server version (Step 0); probed real behavior against the live instance.
• <NAME>.ts created with parseCommand + transformReply, as const satisfies Command.
• Flags set correctly (IS_READ_ONLY for reads, CACHEABLE only for side-effect-free reads, NOT_KEYED_COMMAND if no key; omit flags the metadata table already implies — Step 4).
• Every key uses pushKey/pushKeys; numbers stringified; options behind an exported interface.
• RESP2/3 divergence handled via keyed transformReply: RESP3 is the target shape (usually 3: pass-through), RESP2 transformed to match it; both shapes verified against the live instance.
• Registered in commands/index.ts: import + raw entry + camelCase alias, each with JSDoc (@param per arg; @since for the introducing version; @remarks for >2^53 precision).
• Static command metadata regenerated (npm run generate:metadata) against a server with all bundled modules; the new command's entry verified and the diff contains no dropped entries (Step 4).
• <NAME>.spec.ts: parseArgs covers all branches; testUtils.testAll covers server + cluster; behavior tests gated with minimumDockerVersion on both client and cluster.
• npm run build, npm run check:command-jsdoc, the spec, and npm run lint all pass.
• Commit message uses Conventional Commits; no company-internal refs.