serverless-ircd 0.4.0 → 0.6.0

package/docs/Release-Process.md

@@ -0,0 +1,443 @@
+# Release Process
+
+End-to-end runbook for cutting, publishing, and rolling back a
+ServerlessIRCd release. This is the canonical process — every release
+follows this document top-to-bottom.
+
+**Distribution model.** All workspace packages are `"private": true`. A
+release is a **signed git tag** plus a **GitHub Release** whose body is
+copied from `CHANGELOG.md`. There is no npm publish step in 0.x. See
+§10 (Future: npm publish) for the criteria that would flip that on.
+
+**Current cadence.** 0.x is "tag when ready" — there is no fixed
+calendar release. Anyone with push rights may propose a release by
+opening a PR that bumps versions and updates `CHANGELOG.md`; the
+maintainer team cuts the tag.
+
+Cross-reference: `README.md` (Roadmap), `progress.md` / `tickets.md`
+(release content), `docs/deployment-cf.md` (prod deploy step §6),
+`.github/workflows/ci.yml`, `.github/workflows/deploy-cf.yml`.
+
+---
+
+## 1. Versioning
+
+ServerlessIRCd follows [Semantic Versioning](https://semver.org/) with
+the 0.x convention:
+
+| Bump        | When                                                            |
+|-------------|-----------------------------------------------------------------|
+| `0.x.0 → 0.(x+1).0` | Any user-visible behavior change, new command/cap, new adapter, schema change to persisted DO state. |
+| `0.x.y → 0.x.(y+1)`* | Internal-only fixes: tests, docs, refactors that don't change runtime behavior, perf work with identical output. (*Reserved for after 1.0; in 0.x every change is allowed to be a minor.) |
+| `1.0.0`     | The Cloudflare **and** AWS adapters are deployed to staging, the parametrized contract suite passes on both, ≥3 reference clients connect cleanly (PLAN §9 exit criteria). |
+
+**The 0.x disclaimer.** Per SemVer 0.x, anything may change at any
+minor bump. Consumers of `@serverless-ircd/*` packages (today: only
+this repo itself) should pin exact versions.
+
+### 1.1 Workspace version lock
+
+All eight `package.json` files (root + 7 workspace packages) carry the
+**same** version string. They are bumped together in the release
+commit. The pre-release checklist in §4 enforces this.
+
+Rationale: even though the packages are private today, keeping a single
+version simplifies future npm publishing, release notes, and
+bug-report triage ("what version are you on?" has one answer).
+
+### 1.2 Persisted-state schema
+
+A change to `PERSISTED_STATE_VERSION` in
+`packages/cf-adapter/src/serialize.ts` (or any equivalent versioned
+blob in a future adapter) is a **minor bump minimum**, and requires a
+migration entry in `CHANGELOG.md` under a **⚠️ Migration required**
+heading. See §9 (Rollback) for why rollback across such a bump is
+not always possible.
+
+---
+
+## 2. Roles & permissions
+
+| Role                        | Who has it today         | Required for                          |
+|-----------------------------|--------------------------|---------------------------------------|
+| Push to `main`              | Maintainers              | Merging the release PR.               |
+| Push tags (`v*`)            | Maintainers              | `git push origin v0.x.0`.             |
+| Create GitHub Release       | Maintainers              | Publishing the release notes.         |
+| `CLOUDFLARE_API_TOKEN` secret | Repo Settings → Secrets | Local prod deploys. CI staging deploys. |
+| `CF_SMOKE_URL` variable     | Repo Settings → Variables | Staging smoke e2e in CI.              |
+
+There is no separate "release engineer" role. The maintainer who cuts
+the tag owns the release end-to-end, including the production deploy
+and the post-release smoke check.
+
+---
+
+## 3. Branch & tag conventions
+
+- **Releases are cut from `main`.** No release branches in 0.x.
+- **Tags** are annotated, signed where possible, formatted `v<semver>`
+  (e.g. `v0.1.0`, `v0.1.1`, `v1.0.0-rc.1`).
+- **Pre-release tags** use SemVer suffixes: `v0.2.0-rc.1`,
+  `v1.0.0-beta.3`. GitHub marks the release as a pre-release
+  automatically based on the suffix.
+- The release commit message is `chore(release): vX.Y.Z` (matches the
+  tag).
+
+---
+
+## 4. Pre-release checklist
+
+Run every box before tagging. Most are verifiable from the repo root.
+
+```bash
+# 0. Sanity: clean tree on main, up to date with origin.
+git status                       # working tree clean
+git checkout main
+git pull --ff-only origin main
+
+# 1. All desired work merged. No pending tickets targeted for this release.
+#    Cross-check progress.md — every ticket listed under this release is ✅.
+
+# 2. CI is green on the tip of main.
+#    Open: https://github.com/<owner>/<repo>/actions
+#    - CI workflow: lint / typecheck / coverage all green.
+#    - deploy-cf workflow: staging deploy + smoke e2e green.
+
+# 3. Local full verify.
+pnpm install --frozen-lockfile
+pnpm lint                        # biome check .
+pnpm typecheck                   # tsc --noEmit across workspace
+pnpm test                        # full suite
+pnpm build                       # every package produces dist/
+
+# 4. Coverage gate (enforced in CI per package).
+#    The per-package coverage gate (TICKET-048) runs in CI via
+#    `pnpm coverage`; Vitest fails the build on any threshold regression.
+#    Thresholds: 100% (lines/functions/branches/statements) on
+#    irc-core / irc-server / in-memory-runtime, and ≥90% on every
+#    other package (cf-adapter / local-cli / tcp-ws-forwarder). See the
+#    "Coverage gate" step in .github/workflows/ci.yml. CI must already
+#    be green (step 2), so locally re-run `pnpm coverage` as a
+#    belt-and-braces check before tagging.
+
+# 5. Staging smoke e2e passes against the deployed tip-of-main.
+node apps/cf-worker/scripts/smoke.mjs \
+  --url wss://serverless-ircd-staging.<subdomain>.workers.dev
+
+# 6. CHANGELOG.md is complete and accurate (§5).
+# 7. All package.json versions are bumped and lock-step (§6).
+# 8. The Release PR is merged (§7).
+```
+
+If any box fails, **stop**. Fix forward on `main` and restart the
+checklist — do not tag over a broken tip.
+
+### 4.1 What is explicitly out of scope for a 0.x release
+
+- AWS adapter (TICKET-039 through TICKET-044).
+- SASL authentication (TICKET-026).
+- Flood control (TICKET-028).
+- Server password / rate limits / cloaking (TICKET-046).
+- Config loader (TICKET-047).
+
+These belong in a later 0.x release. Their absence is **not** a
+blocker and should be called out under "Known limitations" in the
+GitHub Release body.
+
+---
+
+## 5. Writing `CHANGELOG.md`
+
+Format: [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), newest
+entry first. One `## [X.Y.Z] - YYYY-MM-DD` section per release.
+
+Categories (use only what applies):
+
+- **Added** — new features, commands, caps, adapters.
+- **Changed** — behavior changes in existing functionality.
+- **⚠️ Migration required** — anything that needs operator action on
+  upgrade (DO schema bumps, persisted-state version bumps, config
+  shape changes). One bullet per action item.
+- **Deprecated** — functionality scheduled for removal.
+- **Removed** — functionality removed in this release.
+- **Fixed** — bug fixes.
+- **Security** — vulnerability fixes (also request a CVE).
+- **Known limitations** — explicit "not in this release" list so
+  operators aren't surprised.
+
+Source material for the entry:
+
+1. `progress.md` — the ticket table. Every `✅` row since the previous
+   release tag is a candidate bullet.
+2. `tickets.md` — the per-ticket **Description** field, for the prose.
+3. `git log v<previous>..HEAD --oneline` — for fixes that don't have a
+   ticket.
+
+Group by user-facing theme (Registration, Channel ops, IRCv3,
+Cloudflare adapter, …) rather than by ticket number. Operators read
+themes; developers cross-reference tickets.
+
+Rule of thumb: a new operator reading only `CHANGELOG.md` should
+understand what changed and what they need to do.
+
+---
+
+## 6. Bumping versions
+
+There are eight `package.json` files. They must end up at the same
+version:
+
+```
+./package.json                                          (workspace root)
+packages/irc-core/package.json
+packages/irc-server/package.json
+packages/in-memory-runtime/package.json
+packages/cf-adapter/package.json
+apps/local-cli/package.json
+apps/cf-worker/package.json
+tools/tcp-ws-forwarder/package.json
+```
+
+To cut `v0.X.Y`:
+
+```bash
+# from the repo root, on a fresh branch off main:
+git checkout -b release/v0.X.Y
+
+# bump each file. The version field is the only change.
+# (Optional helper: `pnpm -r exec -- node -e '...'` if you script it.)
+
+# verify all eight files agree:
+rg '"version":' package.json packages/*/package.json apps/*/package.json tools/*/package.json
+```
+
+The lockfile (`pnpm-lock.yaml`) is unaffected — none of the workspace
+dependencies are version-pinned to the workspace root.
+
+Do **not** run `pnpm publish` or `npm version` — those are for npm
+distribution, which is not the current model.
+
+---
+
+## 7. The release PR
+
+The release PR contains **exactly two files changed**:
+
+1. `CHANGELOG.md` (new entry under the version being cut).
+2. Every `package.json` listed in §6 (version bump).
+
+Nothing else. If a code change is needed for the release, it lands in
+a separate PR first and the release PR is opened only after that PR
+merges.
+
+**PR title:** `chore(release): v0.X.Y`
+**PR body:**
+
+```
+This cuts v0.X.Y. See CHANGELOG.md for the full diff vs. v0.W.Z.
+
+Pre-release checklist: docs/release.md §4 — all green.
+Out-of-scope tickets: docs/release.md §4.1.
+
+After merge: tag + GitHub Release + manual prod deploy per
+docs/release.md §8.
+```
+
+The PR is squash-merged with the same title as the commit message.
+
+---
+
+## 8. Cutting the release
+
+Run **immediately** after the release PR merges, so the tag points at
+the exact commit the changelog describes.
+
+```bash
+git checkout main
+git pull --ff-only origin main
+
+# Confirm the tip commit is the squashed release PR.
+git log -1 --oneline              # expect: chore(release): v0.X.Y
+
+# Confirm versions are at the new release.
+rg '"version":' package.json packages/*/package.json apps/*/package.json tools/*/package.json
+```
+
+### 8.1 Tag and push
+
+```bash
+git tag -a v0.X.Y -m "v0.X.Y"
+git push origin v0.X.Y
+```
+
+Annotated (`-a`) tags carry tagger + date + message; use them over
+lightweight tags. Sign with `-s` if you have a GPG / SSH signing key
+configured (`git config tag.gpgSign true`).
+
+### 8.2 Create the GitHub Release
+
+1. <https://github.com/<owner>/<repo>/releases/new> → choose the tag
+   just pushed.
+2. **Title:** `v0.X.Y` (matches the tag, nothing fancier).
+3. **Body:** paste the matching section from `CHANGELOG.md`. Strip the
+   `## [X.Y.Z] - YYYY-MM-DD` heading line (it's redundant with the
+   release title).
+4. **Set as latest release** unless this is a pre-release. Pre-release
+   tags (`-rc.N`, `-beta.N`, etc.) get the "Set as a pre-release"
+   checkbox instead.
+5. **Publish release.**
+
+### 8.3 Production deploy (manual)
+
+There is **no CI auto-deploy to production**. The maintainer runs the
+prod deploy by hand, from a clean checkout on the tag:
+
+```bash
+git checkout v0.X.Y
+
+# Sanity-check the tag matches the released commit.
+git log -1 --oneline              # expect: chore(release): v0.X.Y
+
+pnpm install --frozen-lockfile
+pnpm build
+pnpm typecheck
+pnpm test
+
+# Deploy the Cloudflare Worker to production.
+pnpm deploy:cf:prod
+# → wrangler deploy   (default env = production)
+
+# Tail prod logs while the first connections arrive.
+npx wrangler tail
+
+# Smoke-check prod with the same script CI uses on staging.
+# You must point it at the prod wss:// URL:
+node apps/cf-worker/scripts/smoke.mjs \
+  --url wss://<prod-workers-subdomain>.workers.dev
+```
+
+If smoke fails on prod:
+
+1. **Do not roll forward blindly.** Tail logs, identify the regression.
+2. If a rollback is required, follow §9.
+3. If the regression is in the *deployed Worker code* (not the
+   protocol core), a hotfix follows the same process as a normal
+   release (new patch / minor → tag → GitHub Release → prod deploy).
+
+### 8.4 Announce
+
+- Update `progress.md` Status header (if a "current release" marker
+  was added there) — not currently present, optional.
+- Open an issue with the body "v0.X.Y released" if external tracking
+  is desired. (Not currently wired to any external channel.)
+
+---
+
+## 9. Rollback
+
+There are two independent rollback axes: **code** and **data**.
+
+### 9.1 Code rollback
+
+To redeploy a previous release:
+
+```bash
+git checkout v0.W.Z                 # the previous tag
+pnpm install --frozen-lockfile
+pnpm build
+pnpm deploy:cf:prod
+```
+
+This works as long as the previous release's code is compatible with
+the persisted DO state shape (see 9.2).
+
+There is no in-CI "prod rollback" workflow. The maintainer runs the
+commands above from a checkout.
+
+### 9.2 Data rollback (persisted DO state)
+
+**Persisted-state version is one-way.** Once
+`packages/cf-adapter/src/serialize.ts` ships a new
+`PERSISTED_STATE_VERSION`, old code reading new blobs will reject
+them, and new code reading old blobs may or may not migrate them
+depending on whether a forward migration was shipped alongside.
+
+Rules of thumb:
+
+- If the release's `CHANGELOG.md` has **no ⚠️ Migration required**
+  entry, code rollback to the previous tag is safe — DO storage is
+  forward- and backward-compatible.
+- If it **does** have a migration entry, assume code rollback will
+  not work without losing or transforming state. The migration entry
+  in `CHANGELOG.md` must spell out the recovery procedure.
+
+In a catastrophic prod failure where the new version's migration
+already ran, the path forward is:
+
+1. Freeze new connections (set `MOTD_LINES` to a maintenance banner,
+   or block at the edge with a Route rule).
+2. Decide: fix forward on the new code, or wipe staging-state (NOT
+   prod-state) and re-test the migration offline.
+3. **Production DO namespaces must never be deleted** — the data is
+   gone permanently (per `docs/deployment-cf.md` §12.1).
+
+---
+
+## 10. Future: npm publish
+
+When `@serverless-ircd/irc-core` becomes the first package consumers
+outside this repo depend on (likely: the in-memory runtime is pulled
+into a third-party test harness, or someone ships an alternate
+adapter), the following gates apply before flipping `"private": false`:
+
+1. **Public API audit.** Every export from `src/index.ts` is one we're
+   willing to support across a minor. Anything experimental goes under
+   a `/internal` subpath.
+2. **Stable type surface.** No `any`, no `@ts-expect-error` in the
+   public surface, no `exactOptionalPropertyTypes` surprises.
+3. **README per package**, not just the root.
+4. **`CHANGELOG.md` is already disciplined** (Keep a Changelog format,
+   one section per version, dated).
+5. **CI coverage gate ≥90%** (TICKET-048) is wired and green.
+6. **Provenance & provenanced verification** — npm publish from CI
+   with Sigstore signatures, not a local laptop.
+
+At that point this document grows a §11 "Publishing to npm" with
+`pnpm -r publish --filter @serverless-ircd/irc-core --access public`
+and a new `publish.yml` workflow. The git-tag-plus-GitHub-Release flow
+above stays unchanged — npm publish becomes an *additional* step, not
+a replacement.
+
+---
+
+## 11. Quick reference: cutting v0.X.Y
+
+```bash
+# 1. Pre-release checklist (docs/release.md §4).
+git checkout main && git pull --ff-only origin main
+pnpm install --frozen-lockfile
+pnpm lint && pnpm typecheck && pnpm test && pnpm build
+
+# 2. Branch, bump versions + CHANGELOG, open PR (§6, §7).
+git checkout -b release/v0.X.Y
+# ... edit 8 × package.json + CHANGELOG.md ...
+git commit -am "chore(release): v0.X.Y"
+git push -u origin release/v0.X.Y
+# ... open PR, get it merged ...
+
+# 3. Tag and push (§8.1).
+git checkout main && git pull --ff-only origin main
+git tag -a v0.X.Y -m "v0.X.Y"
+git push origin v0.X.Y
+
+# 4. GitHub Release (§8.2): paste CHANGELOG entry, publish.
+
+# 5. Prod deploy (§8.3).
+git checkout v0.X.Y
+pnpm install --frozen-lockfile && pnpm build
+pnpm deploy:cf:prod
+node apps/cf-worker/scripts/smoke.mjs --url wss://<prod>.workers.dev
+```
+
+Total wall-clock for a routine 0.x release: ~30 minutes including the
+prod deploy and smoke check.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "serverless-ircd",
-  "version": "0.4.0",
+  "version": "0.6.0",
   "private": false,
   "description": "Serverless IRC daemon with a platform-agnostic core and Cloudflare Workers + AWS adapters",
   "license": "BSD-3-Clause",
@@ -17,7 +17,7 @@
     "typescript": "^5.9.3",
     "vite": "^7.3.6",
     "vitest": "^4.1.10",
-    "@serverless-ircd/aws-adapter": "0.4.0"
+    "@serverless-ircd/aws-adapter": "0.6.0"
   },
   "scripts": {
     "build": "turbo run build",

package/packages/aws-adapter/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@serverless-ircd/aws-adapter",
-  "version": "0.4.0",
+  "version": "0.6.0",
   "private": true,
   "description": "AWS Lambda + DynamoDB adapter: AwsRuntime implementing IrcRuntime + $connect/$disconnect/$default handlers",
   "license": "BSD-3-Clause",

package/packages/aws-adapter/src/aws-runtime.ts

@@ -51,6 +51,7 @@ import {
   type Nick,
   type RawLine,
   type RosterEntry,
+  type ServerConfig,
   SystemClock,
   emptyModes,
   toSnapshot as toConnSnapshot,
@@ -119,6 +120,14 @@ export interface AwsRuntimeOptions {
    */
   managementApi: ApiGatewayManagementApi | PostToConnection | null;
   clock?: Clock;
+  /**
+   * Config-reload source for `REHASH`. The Lambda handler binds this to a
+   * closure that re-invokes {@link loadServerConfigFromLambdaEnv} against the
+   * invocation env (Secrets Manager / SSM / env vars). When omitted
+   * `reloadConfig` rejects so the actor applies a graceful error-suffixed
+   * `382` and retains the prior config.
+   */
+  readonly configLoader?: () => Promise<ServerConfig>;
 }
 
 /**
@@ -133,6 +142,7 @@ export class AwsRuntime implements IrcRuntime {
   private readonly handlers: AwsRuntimeHandlers;
   private readonly managementApi: ApiGatewayManagementApi | PostToConnection | null;
   private readonly clock: Clock;
+  private readonly configLoader: (() => Promise<ServerConfig>) | undefined;
 
   constructor(opts: AwsRuntimeOptions) {
     this.dynamo = opts.dynamo;
@@ -141,6 +151,7 @@ export class AwsRuntime implements IrcRuntime {
     this.handlers = opts.handlers;
     this.managementApi = opts.managementApi;
     this.clock = opts.clock ?? SystemClock;
+    this.configLoader = opts.configLoader;
   }
 
   // -------------------------------------------------------------------------
@@ -210,6 +221,51 @@ export class AwsRuntime implements IrcRuntime {
     }
   }
 
+  /**
+   * Global cross-connection fan-out for `WALLOPS`. `Scan`s the `Connections`
+   * table filtered to rows whose `userModes.wallops` is `true`, then
+   * `PostToConnection`s each one (skipping `except`). The bound connection's
+   * own delivery goes through the in-process handlers (no APIGW round-trip),
+   * mirroring `send`.
+   *
+   * Cost: a DynamoDB `Scan` reads the ENTIRE table (chargeable per row read,
+   * not per match), so this is expensive at high connection counts. A fleet
+   * of 10k connections costs one full scan per WALLOPS plus one
+   * `PostToConnection` per `+w` recipient. Deployments should bound the
+   * oper's WALLOPS rate; a projected GSI on `userModes.wallops` would lift
+   * the scan to a targeted query (documented follow-up). The scan uses
+   * `FilterExpression` so only `+w` rows survive to the `PostToConnection`
+   * phase — the APIGW call count is bounded by the genuine recipient set.
+   */
+  async broadcastWallops(lines: RawLine[], except?: ConnId): Promise<void> {
+    let startKey: Record<string, NativeAttributeValue> | undefined;
+    // Paginate the scan: DynamoDB caps a single Scan at 1 MB. Loop until
+    // no `LastEvaluatedKey` remains so every +w connection is reached.
+    do {
+      const result = await this.dynamo.send(
+        new ScanCommand({
+          TableName: this.tables.Connections,
+          FilterExpression: 'userModes.#m = :true',
+          ExpressionAttributeNames: { '#m': 'wallops' },
+          ExpressionAttributeValues: { ':true': true },
+          ...(startKey !== undefined ? { ExclusiveStartKey: startKey } : {}),
+        }),
+      );
+      const items = (result.Items ?? []) as unknown as MarshalledConnection[];
+      for (const row of items) {
+        const connId = row.connectionId;
+        if (connId === except) continue;
+        // `send` short-cuts the bound connection through the in-process
+        // handlers and catches GoneException so vanished sockets are
+        // cleaned up rather than aborting the fan-out mid-loop.
+        if (row.userModes.wallops) {
+          await this.send(connId, lines);
+        }
+      }
+      startKey = result.LastEvaluatedKey as Record<string, NativeAttributeValue> | undefined;
+    } while (startKey !== undefined);
+  }
+
   // -------------------------------------------------------------------------
   // Nick registry
   // -------------------------------------------------------------------------
@@ -365,6 +421,19 @@ export class AwsRuntime implements IrcRuntime {
     return out;
   }
 
+  /**
+   * Re-invokes the bound config loader (Secrets Manager / SSM / env) and
+   * returns the fresh {@link ServerConfig}. The Lambda handler binds the
+   * loader at construction; when no loader was supplied this rejects so the
+   * actor's graceful failure path applies.
+   */
+  async reloadConfig(): Promise<ServerConfig> {
+    if (this.configLoader === undefined) {
+      throw new Error('no config loader bound to AwsRuntime');
+    }
+    return this.configLoader();
+  }
+
   // -------------------------------------------------------------------------
   // Public helpers (re-exported surface)
   // -------------------------------------------------------------------------

package/packages/aws-adapter/src/handlers/connect.ts

@@ -14,7 +14,13 @@
 
 import type { DynamoDBDocumentClient } from '@aws-sdk/lib-dynamodb';
 import { PutCommand, ScanCommand } from '@aws-sdk/lib-dynamodb';
-import type { ParsedServerConfig } from '@serverless-ircd/irc-core';
+import {
+  type ParsedServerConfig,
+  type WsSubprotocol,
+  parseSecWsProtocolOffers,
+  selectSubprotocol,
+  wsFrameModeFor,
+} from '@serverless-ircd/irc-core';
 import { type AdmissionOutcome, decideConnectAdmission } from '../admission.js';
 import { marshalConnection } from '../serialize.js';
 import type { TablesConfig } from '../tables.js';
@@ -29,10 +35,24 @@ export interface ConnectParams {
   now?: number;
   /** Server config — consulted for the `maxClients` admission cap. */
   serverConfig: ParsedServerConfig;
+  /**
+   * Raw `Sec-WebSocket-Protocol` header offered by the client on the
+   * `$connect` upgrade, if present. Parsed into an ordered offer list and
+   * the first supported IRCv3 subprotocol is agreed (legacy fallback when
+   * none match or the header is absent).
+   */
+  secWebSocketProtocol?: string | null;
 }
 
-/** Re-exported so the dispatcher imports the outcome type from one place. */
-export type { AdmissionOutcome as ConnectOutcome } from '../admission.js';
+/**
+ * Outcome of a `$connect`. Mirrors {@link AdmissionOutcome} on the rejection
+ * branch and carries the negotiated IRCv3 {@link WsSubprotocol} on success so
+ * the dispatcher can echo it back to the client in the integration response.
+ * `subprotocol === null` signals a legacy (no-subprotocol) connection.
+ */
+export type ConnectOutcome =
+  | { readonly admitted: true; readonly subprotocol: WsSubprotocol | null }
+  | { readonly admitted: false; readonly statusCode: number; readonly reason: string };
 
 /**
  * Enforces the `maxClients` admission cap, then inserts the Connections
@@ -58,20 +78,31 @@ export type { AdmissionOutcome as ConnectOutcome } from '../admission.js';
  * into {@link decideConnectAdmission}'s `perIp` branch. That schema
  * change is tracked as a follow-up; the pure policy already supports it.
  */
-export async function handleConnect(params: ConnectParams): Promise<AdmissionOutcome> {
+export async function handleConnect(params: ConnectParams): Promise<ConnectOutcome> {
   const now = params.now ?? Date.now();
   const total = await countConnections(params.dynamo, params.tables.Connections);
   const outcome = decideConnectAdmission({ total }, { maxClients: params.serverConfig.maxClients });
   if (!outcome.admitted) return outcome;
+  // IRCv3 WebSocket subprotocol negotiation: parse the offered
+  // `Sec-WebSocket-Protocol` list, select the first supported entry, and
+  // persist the resulting frame mode so every subsequent `$default`
+  // invocation recovers it (a Lambda `$connect` has no live socket to
+  // tag, unlike the CF adapter's hibernation tags). `null` means no
+  // supported subprotocol was offered → legacy framing, no mode column.
+  const chosen = selectSubprotocol(parseSecWsProtocolOffers(params.secWebSocketProtocol));
+  const mode = wsFrameModeFor(chosen ?? undefined);
   const state = createInitialConnectionState(params.connectionId, now);
   const row = marshalConnection(state, now);
+  if (mode !== 'legacy') {
+    row.wsMode = mode;
+  }
   await params.dynamo.send(
     new PutCommand({
       TableName: params.tables.Connections,
       Item: row,
     }),
   );
-  return outcome;
+  return { admitted: true, subprotocol: chosen };
 }
 
 /**