dsh-escalation-advisor

In daily development most turns are handled just fine by a lighter model, but the moment you hit a stuck loop, a failing validation, or are about to ship, you genuinely need a stronger model to weigh in. dsh-escalation-advisor solves exactly that tension: keep the bulk of your work on cheaper models and invoke a stronger DSH model as a visible advisor only when a second opinion adds value, without manually switching models mid-flow.

The plugin offers three modes — manual for on-demand consultation, escalate (default) for automatic triggering on deterministic stuck or failure signals, and continuous for shadow review at natural turn boundaries. A separate completion review checks whether the goal is met, validations are done, and no obligations remain before final delivery. Every advisor run is a visible DSH child session under the exact requesting agent, so you can open the agent tree and inspect the full transcript, tool calls, token usage, and final verdict. You can continue the same child session via consultation_id, or pin a named advisor profile to fix the model route and tool policy.

Tool permissions default to off, and a child session can only see tools the requester itself already has access to, preventing any privilege-escalation path through the advisor. Case packets are capped at 48 KB with structural degradation and truncation logging; delivered verdicts fit within 16 KB. These budgets keep per-consultation cost predictable. Main-agent and subagent coverage are configured independently, and the per-session panel lets you override mode, timeout, and trigger switches without touching global defaults.

It is built for engineers who run multi-turn development in DSH and want to keep token costs down without losing the safety net of a stronger model at critical moments. No provider reconfiguration, no changes to the main conversation model — enable and go.