Design: Gateway stability, hot config apply, and IM decoupling redesign

[hazeone]

Summary

Research-backed redesign proposal for the three coupled stability problems in ClawX today:

1. Config changes and model switches frequently force a full Gateway restart (minutes of downtime; always on Windows).
2. IM channels run inside the Gateway process, so every restart drops all IM sessions — while the Gateway is also the process we restart most.
3. Restarts are slow because unconditional plugin sync, config sanitation, and uv/Python warmup sit on every spawn path, followed by an ~8-minute silent readiness loop.

What's included

• harness/reference/gateway-stability-and-hot-reload-redesign.md — full design reference:
  • Current-state analysis with exact file/function citations (scheduleGatewayRefresh, FORCE_RESTART_CHANNELS, prepareGatewayLaunchContext, waitForGatewayReady, Windows reload degradation, etc.).
  • Comparative research of NousResearch/hermes-agent (backendPool, single shared runtime), netease-youdao/LobsterAI (Cowork product/runtime split, Main-owned IM gateways), OpenAI Codex App Server (per-turn model/policy parameters, long-lived JSON-RPC server), WorkBuddy (sidecar supervisor fleet), and the closest public matches for "codecode" (Claude Code desktop multi-session daemon, Tencent CodeBuddy-style workspaces).
  • Three-phase target architecture: (1) hot config apply via a single Main-side apply planner + reload RPC for Windows parity, (2) fast/invisible restarts via manifest-hashed prelaunch, background uv warmup, staged readiness milestones, and blue/green swap, (3) IM decoupling via a separately supervised Channel Host process (WorkBuddy/LobsterAI-informed, Option B recommended).
  • Migration steps mapped to concrete files, success metrics, and risks/open questions.
• harness/specs/tasks/gateway-hot-config-and-im-decoupling.md — spec-driven anchor for the implementation phases under the gateway-backend-communication scenario (per the repo's spec-driven harness rule).

Verification

• pnpm harness validate --spec harness/specs/tasks/gateway-hot-config-and-im-decoupling.md (with diff, --since origin/main): valid.
• pnpm run harness:ci: all baseline checks and harness unit tests pass (12/12).

Docs-only change; no runtime behavior is modified in this PR.