dsh-subagent-model-visibility

In DeepSeek Harness, configuring a model does not guarantee that every subagent request actually reaches that model. Phoenix or other routing layers may failover, retry, or redistribute mid-session, yet the UI still shows the configured value rather than the one that served the request. dsh-subagent-model-visibility fills that visibility gap.

The plugin wraps two hooks—tools/execute and llm/stream—and uses Node AsyncLocalStorage to safely bind each native tool call to its child session even under concurrent execution. It then records the real provider/model pair dispatched to the LLM runtime and renders it as a compact third line inside the existing tool-call row and subagent catalog. The badge updates live when the route changes, and hovering reveals route history plus request counts. Prompts, routing parameters, tool results, and the subagent lifecycle are never modified.

Designed for DSH users running multi-provider or multi-model stacks, or relying on failover and retry logic, who need to confirm which backend actually served a given request without invasive tracing. Privacy-wise, the browser API exposes only tool and child session IDs, transport type, route names, timestamps, request counts, and terminal status; prompts, model outputs, credentials, and headers never leave the host.