dsh-subagent-model-visibility

A small DeepSeek Harness plugin that shows the actual provider/model used by a subagent directly inside the existing native subagent tool-call row.

It does not guess from the configured model. The host observes the child's real llm/stream GenerateOptions, so the displayed route reflects the model that actually received the request after DSH routing/retry/failover logic.

Example

Tool call · subagent · Implement ATR position sizing & risk limits

  MODEL  brainz / claude-opus-4.7   · 2 model requests · subagent: spawn

OUT
started subagent 616c26ce-...

If a subagent changes routes during recovery:

MODEL  brainz / qwen3-coder   3 routes · 4 model requests

Hover the model line to see the route history and request counts.

How it works

1. tools/execute is wrapped with Node AsyncLocalStorage to associate one native DSH tool call with the subagent/start event it creates, safely even with concurrent tool calls.
2. subagent/start records the child session id and the DSH subagent transport (spawn, fork, ACP, etc.).
3. llm/stream observes actual child requests by sessionId and records the final provider/model dispatched to the LLM runtime.
4. subagent/end records the terminal state.
5. The Web client uses DSH's stable [data-chat-call-id] row identity and adds a compact model line inside the existing row. It updates live when the route changes.

The plugin does not modify prompts, model routing, tool arguments, tool results, or the subagent lifecycle.

Install

From the folder containing the extracted plugin:

dsh plugin --profile web add .\dsh-subagent-model-visibility
dsh web

Or install the packaged .tgz.

Configuration

- id: subagent-model-visibility
  name: dsh-subagent-model-visibility
  config:
    retentionMs: 86400000
    maxEntries: 1000
    uiRefreshMs: 1000

• retentionMs: how long finished observations remain available to the UI.
• maxEntries: memory bound for observed subagent sessions.
• uiRefreshMs: browser refresh cadence for model/failover changes.

Accuracy semantics

• Actual model means a provider/model pair observed on a real ctx.llm.stream() call whose sessionId is the subagent child.
• A newly started background child may briefly display Resolving actual model… before its first model request begins.
• If Phoenix or another router changes model, the badge changes to the latest route and exposes the route count/history.
• Remote subagent transports that do not issue their model request through the parent DSH ctx.llm cannot expose an authoritative underlying model; those remain at Resolving actual model… rather than inventing a value.
• Historical rows created before this plugin started cannot be reconstructed reliably and are intentionally left untouched.

Privacy

The browser API contains only tool/child ids, transport name, provider/model route names, timestamps, request counts and terminal status. Prompts, model output, credentials, headers, filesystem paths and tool arguments are never included.

License

MIT