@deepseek-ai/dsh-model-roles

@deepseek-ai/dsh-model-roles

English | 中文

Settings-backed subagent model routing: a model-roles settings namespace of per-model capability roles plus a per-session pinned subagent model, surfaced to the delegating model as advisory guidance. Nothing routes inside the loop, and no provider adapter schema is touched.

Roles and pinning

The model-roles settings namespace maps provider route → model id → { description, subagent }. description is the capability text used for task-based routing; subagent (default true) toggles whether that model may be auto-assigned when the user names none.

The /subagent-model / command pins the session's subagent model; /subagent-model off clears it. The pin is logged (model-roles/subagent-pin, last wins), so resume and fork restore it, and it is rendered into the model-roles:routing system-prompt section. Routing is soft: the section tells the delegating model to pass agentOptions on each subagent tool call, which requires a subagent provider with the agentOptions capability (in-process spawn/fork).

Config

Key	Meaning
toolName	The model-facing subagent tool the guidance names, default subagent.

Model Experience

Prompt section

What the model sees

When a pin is logged, one short instruction naming the pinned provider/model and the agentOptions to pass. Otherwise, when role candidates exist, a list of provider / model: description rows plus the rule to pick by task fit, pass agentOptions, and re-delegate with a stronger candidate on an unsatisfactory result. No pin and no candidates render nothing.

Token effect

One short section only while a pin or at least one auto-assignable role exists; its length scales with the number of described models.

KV Cache effect

Prefix-stable while roles and the pin are unchanged; a role edit or pin flip changes the section text for subsequent requests.

Command

What the model sees

/subagent-model runs as a slash command and never reaches the model; its success text is the state-change feedback shown to the user.

Token effect

None beyond the logged model-roles/subagent-pin event, which is log-only and not in model history.

KV Cache effect

None; the event is log-only.

Known Limitations and Deferred Work

• Routing is advisory — the pin and candidate list are enforced only through the model's cooperation; a misbehaving model can ignore them. Deterministic routing would require intercepting the subagent tool call.
• Routes are not validated against the LLM registry — a pin or role naming an unregistered provider/model fails at delegation time (NO_ADAPTER), not at pin time.
• Reasoning effort per child is out of scope — agentOptions carries provider/model/maxTokens only; escalating thinking intensity means choosing a stronger model.