diff --git a/quartz/plans/2026-07-16-relay-limits.md b/quartz/plans/2026-07-16-relay-limits.md
new file mode 100644
index 0000000000..b0426776b2
--- /dev/null
+++ b/quartz/plans/2026-07-16-relay-limits.md
@@ -0,0 +1,113 @@
+# Relay `LIMITS` message — status & roadmap
+
+Support for the relay-to-client `LIMITS` frame (from
+[nostr-protocol/nips#1434](https://github.com/nostr-protocol/nips/pull/1434)),
+where a relay advertises the current connection's rights and limits — on
+connect and again whenever they change (e.g. after a NIP-42 AUTH flips
+`can_write`). Live-verified against `wss://pipe.imwald.eu/` (which sends `AUTH`
+then `LIMITS`).
+
+Note: `LIMITS` is **not** NIP-22 (that's Comment / kind 1111); it's the #1434
+proposal and is label-based on the wire, so the code keys off the `"LIMITS"`
+label rather than a NIP number.
+
+## Done (branch `claude/limits-message-support-g5plrd`, PR #3597)
+
+- **Parse** — `LimitsMessage` model (`nip01Core/relay/commands/toClient/`) with
+  every #1434 field plus the two non-spec extensions this relay sends
+  (`auth_for_read` / `auth_for_write`). Wired into both codecs:
+  Jackson (`LimitsDeserializer` + `MessageSerializer`, jvmAndroid) and
+  kotlinx (`LimitsKSerializer` + `MessageKSerializer`, iOS/native). Unknown
+  fields are ignored, so a future field can't re-trigger the original
+  "Message LIMITS is not supported" crash.
+- **Cache/expose** — `RelayLimitsTracker` (`nip01Core/relay/client/limits/`),
+  a passive `RelayConnectionListener` accessory (modeled on
+  `RelayAuthenticator`) that caches the latest `LimitsMessage` per
+  `NormalizedRelayUrl` and publishes a Compose-stable
+  `StateFlow<PersistentMap<url, LimitsMessage>>`. Connection-scoped (dropped on
+  disconnect). Wired in `AppModules` as `Amethyst.instance.relayLimits`.
+- **Cleanup** — removed the unused experimental prototype
+  (`experimental/limits/Limits.kt` + `LimitProcessor.kt`); renamed the client
+  accessory to `…Tracker` to avoid colliding with the server-side
+  `relay/server/policies/RelayLimits`.
+
+## Audit fixes to fold in (small, do first)
+
+From the 2026-07-16 review of the branch:
+
+1. **kotlinx parse should be as lenient as Jackson.** `MessageKSerializer` is
+   the *iOS* incoming path; today `LimitsKSerializer.deserializeFromElement`
+   throws on a mistyped field (`(tag as JsonArray)`, `.jsonPrimitive.int`) or a
+   payload-less `["LIMITS"]` (`array[1].jsonObject`), where Jackson coerces.
+   Same frame → parsed on Android, dropped on iOS. Use `intOrNull` /
+   `booleanOrNull` / `longOrNull`, guard the array cast, and default a missing
+   payload object to an empty `LimitsMessage`.
+2. **Drop the stale "NIP-22" labels** in `LimitsKSerializer` KDoc and the
+   `// NIP-22 wire format` comments in `MessageKSerializer` / `MessageSerializer`
+   (already removed from `LimitsMessage.kt`).
+3. **Make `LimitsMessage` a `data class`** so `StateFlow.distinctUntilChanged`
+   suppresses no-op emissions when a relay re-advertises identical limits, and
+   tests get value equality.
+4. Fold the explicit-`null`→`false`/`0` coercion nuance into (1) via the
+   `…OrNull` accessors on both sides.
+
+## Roadmap
+
+### 1. Client-side apply/enforcement (highest user-facing value)
+
+#1434 says clients MUST apply limits when sending. Provide **pure, testable
+helpers on `LimitsMessage`** (mirror the server's `LimitsPolicy` logic; the
+deleted `LimitProcessor` in git history is a starting sketch):
+
+- `clamp(filter)` / `clamp(filters)` → cap each filter `limit` to `max_limit`.
+- `rejectionForPublish(event): String?` → `can_write`, `accepted`/`blocked
+  _event_kinds`, `max_content_length`, `max_event_tags`, `min_pow_difficulty`,
+  `created_at_msecs_ago`/`ahead` window, `required_tags`.
+- `canRead()` / `canWrite()` convenience.
+
+Keep the helpers in quartz (pure, no side effects). Wire them into the send
+path at the app layer (subscription/filter-assembly + publish), reading from
+`RelayLimitsTracker`. Decide the UX for a rejected publish (surface vs
+silently drop) with the maintainer — this is the one opinionated call.
+
+### 2. Server-side emit (`geode`) — the symmetric half
+
+The relay side already **enforces** (`LimitsPolicy` + `RelaySession`) and
+**advertises via NIP-11** (`RelayLimits.toNip11Limitation()`), but never sends
+the dynamic `LIMITS` frame. Add:
+
+- `RelayLimits.toLimitsMessage(canRead, canWrite, authForRead, authForWrite)`
+  next to `toNip11Limitation()`, so one source of truth drives NIP-11 *and* the
+  frame.
+- `RelaySession` / `EventSourceServer` sends `LIMITS` on connect and **re-sends
+  after AUTH** when effective rights change.
+- Then geode ↔ our own client becomes interop-testable (relayBench / the amy
+  serve path).
+
+### 3. NIP-11 ↔ LIMITS bridge
+
+The two overlap ~80% (`RelayInformationLimitation` vs `LimitsMessage`). Add
+`LimitsMessage.toNip11Limitation()` / `RelayInformationLimitation
+.toLimitsMessage()`, and let `RelayLimitsTracker` **seed** from the NIP-11 doc
+before the socket connects, then override with the live frame — one "limits of
+relay X" model regardless of source.
+
+### 4. Tooling, tests, docs
+
+- `amy relay info` already prints the NIP-11 doc; add the live `LIMITS` frame
+  (the interop tool the repo already leans on).
+- Catalog `RelayLimitsTracker` in
+  `nip01Core/relay/client/accessories/README.md`.
+- Test the **AUTH → re-LIMITS** transition (`auth_for_write` flipping after a
+  successful AUTH) — the one dynamic behavior not yet exercised live.
+
+### 5. Spec follow-up
+
+`auth_for_read` / `auth_for_write` are not in #1434. Worth a note on the PR;
+we tolerate them as extras either way (they pair with the `AUTH` this relay
+also sends).
+
+## Suggested sequencing
+
+audit fixes → (1) client apply helpers → (2) server emit → (3) NIP-11 bridge →
+(4) tooling.