dsh-session-resilience

The worst thing for a long-running DSH web session is a host restart: unfinished tool calls, context, and multi-turn dialogue vanish. DSH Session Resilience exists to prevent exactly that. It lets the browser reconnect to a genuinely ready replacement instance, only signalling success once a tokenised launch URL has been confirmed. Whether the cause is a network blip, a token limit, or an agent stuck in a loop, recovery happens locally with zero dependency on remote services.

Core capabilities fall into three layers. Safe restart and stop record active root sessions, launch a replacement host, and hold the browser until the launch URL is live; it never reports readiness prematurely. Token-aware reconnect is paired with four recovery policies (Safe, Balanced, Long task, Manual) that pin the handoff window, retry cap, scan range, and backoff cadence; in Manual mode every field is individually editable. The safety rail includes an idempotency guard that reminds the model to verify unknown results or skip already-succeeded calls before continuing, adaptive backoff that stretches intervals and halts at a configured limit, and a loop guard that cancels turns repeating identical tool calls or assistant text. Permanent errors such as authentication or balance failures are skipped and surfaced via browser notifications.

Best suited for users running builds, large data imports, or agent sessions beyond ten minutes in DSH, who want restart risk turned into a manageable recovery window. All markers, helper logs, and control routes stay local to the host; the model sees only two lightweight tool definitions and gains no fixed system-prompt prefix.