132 lines
6.4 KiB
Markdown
132 lines
6.4 KiB
Markdown
---
|
||
summary: "Refactor notes for the macOS gateway client: single shared websocket + follow-ups"
|
||
read_when:
|
||
- Investigating duplicate/stale Gateway WS connections
|
||
- Refactoring macOS gateway client architecture
|
||
- Debugging noisy reconnect storms on gateway restart
|
||
---
|
||
# Gateway Refactor Notes (macOS client)
|
||
|
||
Last updated: 2025-12-12
|
||
|
||
This document captures the rationale and direction for the macOS app’s Gateway client refactor: **one shared websocket connection per app process**, plus follow-up improvements to simplify lifetimes and reduce “hidden” reconnection behavior.
|
||
|
||
Related docs:
|
||
- `docs/refactor/new-arch.md` (overall gateway protocol/server plan)
|
||
- `docs/gateway.md` (gateway operations/runbook)
|
||
- `docs/presence.md` (presence semantics and dedupe)
|
||
- `docs/mac/webchat.md` (WebChat surfaces and debugging)
|
||
|
||
---
|
||
|
||
## Background: what was wrong
|
||
|
||
Symptoms:
|
||
- Restarting the gateway produced a *storm* of reconnects/log spam (`gateway/ws in connect`, `hello`, `hello-ok`) and elevated `clients=` counts.
|
||
- Even with “one panel open”, the mac app could hold tens of websocket connections to `ws://127.0.0.1:18789`.
|
||
|
||
Root cause (historical bug):
|
||
- The mac app was repeatedly “reconfiguring” a gateway client on a timer (via health polling), creating a new websocket owner each time.
|
||
- Old websocket owners were not fully torn down and could keep watchdog/tick tasks alive, leading to **connection accumulation** over time.
|
||
|
||
---
|
||
|
||
## Current architecture (as of 2025-12-12)
|
||
|
||
Goal: enforce the invariant **“one gateway websocket per app process (per effective config)”**.
|
||
|
||
Key elements:
|
||
- `GatewayConnection.shared` owns the one websocket and is the *only* supported entry point for app code that needs gateway RPC.
|
||
- Consumers (e.g. Control UI, Agent RPC, SwiftUI WebChat) call `GatewayConnection.shared.request(...)` and do not create their own sockets.
|
||
- If the effective connection config changes (local ↔ remote tunnel port, token change), `GatewayConnection` replaces the underlying connection.
|
||
- Server-push frames are delivered via `GatewayConnection.shared.subscribe(...) -> AsyncStream<GatewayPush>`, which is the in-process event bus (no `NotificationCenter`).
|
||
|
||
Notes:
|
||
- Remote mode requires an SSH control tunnel. `GatewayConnection` **does not** start tunnels; it consumes the already-established forwarded port (owned by `ConnectionModeCoordinator` / `RemoteTunnelManager`).
|
||
|
||
---
|
||
|
||
## Design constraints / principles
|
||
|
||
- **Single ownership:** Exactly one component owns the actual socket and reconnect policy.
|
||
- **Explicit config changes:** Recreate/reconnect only when config changes, not as a side effect of periodic work.
|
||
- **No implicit fan-out sockets:** Adding new UI features must not accidentally add new persistent gateway connections.
|
||
- **Testable seams:** Connection config and websocket session creation should be overridable in tests.
|
||
|
||
---
|
||
|
||
## Follow-up refactors (recommended)
|
||
|
||
### Status (as of 2025-12-12)
|
||
|
||
- ✅ One shared websocket per app process (per config)
|
||
- ✅ Event streaming moved into `GatewayConnection` (`AsyncStream<GatewayPush>`)
|
||
- ✅ `NotificationCenter` removed for in-process gateway events
|
||
- ✅ `GatewayConnection` no longer implicitly starts the remote control tunnel
|
||
- ⏳ Further separation of concerns (polish/cleanup): push parsing helpers + clearer UI adapters
|
||
- ⏳ Optional: a dedicated “resolved endpoint” publisher for remote mode (to make mode transitions observable)
|
||
|
||
### 1) Move event streaming into `GatewayConnection` (done)
|
||
|
||
Implemented:
|
||
- `GatewayChannelActor` no longer posts global notifications; it forwards pushes to `GatewayConnection` via a callback.
|
||
- `GatewayConnection` fans out pushes via `subscribe(...) -> AsyncStream<GatewayPush>` and replays the latest snapshot to new subscribers.
|
||
|
||
### 2) Replace `NotificationCenter` for in-process events (done)
|
||
|
||
Implemented:
|
||
- `ControlChannel`, `InstancesStore`, and SwiftUI WebChat now subscribe to `GatewayConnection` directly.
|
||
- This removed the risk of leaking `NotificationCenter` observer tokens when views/controllers churn.
|
||
|
||
### 3) Separate control-plane vs chat-plane concerns (partially done)
|
||
|
||
As features grow, split responsibilities:
|
||
- **RPC layer**: request/response, retries, timeouts.
|
||
- **Event bus**: typed gateway events with buffering/backpressure.
|
||
- **UI adapters**: user-facing state and error mapping.
|
||
|
||
This reduces the risk that “a UI refresh” causes connection or tunnel side effects.
|
||
|
||
Notes:
|
||
- The RPC layer and event bus are now centralized in `GatewayConnection`.
|
||
- There’s still room to extract small helpers for decoding specific event payloads (agent/chat/presence) so UI code stays thin.
|
||
|
||
### 4) Centralize tunnel lifecycle (remote mode) (done for GatewayConnection)
|
||
|
||
Previously, “first request wins” could implicitly start/ensure a tunnel (via `GatewayConnection`’s default config provider).
|
||
|
||
Now:
|
||
- `GatewayConnection` uses the already-running forwarded port from `RemoteTunnelManager` and will error if remote mode is enabled but no tunnel is active.
|
||
- Remote tunnel lifecycle is owned by mode/application coordinators (e.g. `ConnectionModeCoordinator`), not by incidental RPC calls.
|
||
|
||
Future improvement:
|
||
- A dedicated coordinator that owns remote tunnel lifecycle and publishes a resolved endpoint.
|
||
- `GatewayConnection` consumes that endpoint rather than calling into tunnel code itself.
|
||
|
||
This makes remote mode behavior easier to reason about (and test).
|
||
|
||
---
|
||
|
||
## Testing strategy (what we want to cover)
|
||
|
||
Minimum invariants:
|
||
- Repeated requests under the same config do **not** create additional websocket tasks.
|
||
- Concurrent requests still create **exactly one** websocket and reuse it.
|
||
- Shutdown prevents any reconnect loop after failures.
|
||
- Config changes (token / endpoint) cancel the old socket and reconnect once.
|
||
|
||
Nice-to-have integration coverage:
|
||
- Multiple “consumers” (Control UI + Agent RPC + SwiftUI WebChat) all call through the shared connection and still produce only one websocket.
|
||
|
||
Additional coverage added (macOS):
|
||
- Subscribing after connect replays the latest snapshot.
|
||
- Sequence gaps emit an explicit `GatewayPush.seqGap(...)` before the corresponding event.
|
||
|
||
---
|
||
|
||
## Debug notes (operational)
|
||
|
||
When diagnosing “too many connections”:
|
||
- Prefer counting actual TCP connections on port 18789 and grouping by PID to see which process is holding sockets.
|
||
- Gateway `--verbose` prints *every* connect/hello and event broadcast; use it only when needed and filter output if you’re just sanity-checking.
|