dsh-auto-reconnect

English | 中文

Auto-reconnect for the DeepSeek Harness (DSH) web client. The official connection controller has backoff reconnect built in, but it sleeps through two real-world scenarios — this plugin wakes it. Client-only, zero build, no official package touched.

Install / Uninstall

dsh plugin --profile web add github:tqcq/dsh-auto-reconnect

Then restart dsh-web — adding or removing plugins always needs a restart; a browser refresh alone is not enough.

Upgrade (re-running add won't move the resolved commit; update re-resolves):

dsh plugin --profile web update dsh-auto-reconnect

Uninstall:

dsh plugin --profile web remove dsh-auto-reconnect

Runtime tweak, no reinstall needed (then reload the page):

localStorage["dsh-auto-reconnect"] = JSON.stringify({ minIntervalMs: 3000, retryDelayMs: 3000, maxAutoRetries: 20, log: true });

Two blind spots in the official controller (source-audited)

Audited against the actually-served @deepseek-ai/dsh-client-connection 0.1.2-rc.1:

1. Terminal disconnected sleeps forever. Once the official 500ms→10s backoff tops out, the controller enters a terminal state and only wakes on the browser online event or a manual tap. Phone sleep/wake keeps navigator.onLine === true the whole time, so online never fires — every wake-up needs a manual tap.
2. Zombie connected. The official client has no visibilitychange / focus listener and no heartbeat. A silently dropped TCP leaves the WebSocket in OPEN while the state still reports connected.

What this plugin does

Listens to visibilitychange(→visible), focus, pointerdown, keydown, pageshow(persisted), and subscribes to the connection state store:

• Event wakes: call connection.reconnect() only when state ≠ connected (resets backoff, retries immediately) — never interrupts a healthy connection.
• bfcache restore (persisted): sockets are dead by definition, reconnect unconditionally.
• Transition into terminal disconnected: delayed auto-retry (default 3s), budgeted to 20 consecutive attempts, reset on connected. Covers "page already visible and focused, no browser event will ever fire again" (2026-09-11 field report: a hard refresh after a host restart landed in terminal state, still needed a manual tap).

Every trigger logs with a [dsh-auto-reconnect] prefix (DevTools console forensics), throttled to 3s by default.

Compatibility

• Field-verified on host @deepseek-ai/dsh 0.1.5-rc.2 (boot activation assertions all green).
• Client contract audited against dsh-client-connection 0.1.2-rc.1 source.
• Shape mismatch on ctx.connection → warn and stay idle, never breaks the page: graceful degradation on API drift.

Limitations

The general form of zombie connected (background tab NAT-dropped, no OS sleep) cannot be detected client-side: the RPC unary channel is separate HTTP, so a successful probe does not prove the WebSocket is alive. The real fix is a host-side heartbeat + dead-connection sweep (prior art: the deprecated enterhalf/dsh-web-network-optimizer) — a separate plugin if ever needed.

Tests

node tests/client_test.js   # 14 groups, vm-sandboxed DOM/connection stubs, no browser

Forking notes

• The in-repo cordis.patch.yml must stay []: a comment-only YAML file parses to null and kills profile boot.
• Client-code-only changes need just a browser hard refresh, no dsh-web restart.

License: MIT (see LICENSE).