Agent Teams: per-teammate LLM route is invisible to the roster and unpinnable from spawn_teammate

[ZackZhaii]

Feature/behavior request. Note: this repository's issue tracker is disabled (HTTP 410), so this is filed as a Discussion.

Summary (EN)

Two related gaps in the Agent Teams stack, both verified against
@deepseek-ai/dsh-desktop@0.2.0-rc.2:

1. spawn_teammate cannot pin a teammate's LLM route. The tool schema takes
   only name/description/prompt/context
   (@deepseek-ai/dsh-experimental-tool-agent-team/lib/index.js:245-266) and the
   Team runtime builds its subagent request as { prompt, parent } only
   (@deepseek-ai/dsh-experimental-agent-team/lib/index.js, spawnAdmitted,
   around lines 544-582). The child route is resolved inside
   startContinuable via resolveChildAgentOptions(parent, request.agentOptions, depth) before any provider hook (@deepseek-ai/dsh-subagent/lib/types/ continuation.js:101-170), so agentRouteDefaults cannot reach this path
   either (it is consumed only by dsh-tool-subagent). Net effect: every
   teammate inherits the Lead's route, and heterogeneous teams are impossible
   without switching the Lead's own model between spawns.

2. The roster shows the wrong model for idle teammates. roster.list()
   computes const model = live?.options.model ?? root.options.model
   (@deepseek-ai/dsh-experimental-agent-team/lib/index.js:448-449), so once a
   teammate has no live activation, list_agents displays the Lead's model —
   even though the teammate's real route is durable in its session's
   subagent/descriptor (agentProvider/agentModel/agentReasoningEffort), and
   cold resume correctly restores it from there
   (dsh-subagent/lib/types/continuation.js:355-374). Verified with real
   sessions: a teammate spawned pinned to grok/grok-4.7@xhigh had both its
   initial and resume requests on that route while list_agents showed the
   Lead's deepseek-flash.

Reproduction

1. In a Team-enabled session (Lead on model A), spawn a teammate whose child
   session runs on model B (e.g. via the local workaround plugin below, or by
   switching the Lead's model around the spawn).
2. Let the teammate finish its turn (status inactive).
3. list_agents shows the teammate on model A. The child session's
   request/header events (both initial and resume) prove model B.

Suggested fixes

• Route visibility: have startContinuable return the resolved child route
  (or let the Team runtime read it back), store it on the member snapshot, and
  make roster.list() fall back live?.options.model ?? member.model ?? root.options.model. Alternatively resolve the model from the member session's
  subagent/descriptor on demand.
• Route pinning: add optional provider / model / reasoning_effort
  parameters to spawn_teammate and forward them as agentOptions into the
  spawn request. The runtime side already honors request.agentOptions; only
  the tool/service layer needs the pass-through. spawn_teammate_modeled in the
  workaround below is a working reference implementation of the semantics
  (including a fork context that must keep using the fork backend).

Workaround in use today (local plugin, not a vendor patch)

dsh-teammate-model-routes (local profile bundle): wraps the raw
subagents.startContinuable (resolved via ctx.reflect._getImpl(...).value,
because ctx.get() returns a traceable Proxy that swallows property writes) and
injects agentOptions only inside an AsyncLocalStorage scope opened by its
spawn_teammate_modeled tool for exactly that caller/label (consume-once); it
also records childId -> pinned route and wraps agentTeams.listMembers to
correct the idle display. If DSH adopts the native fixes above, the plugin can
be retired.

Environment

• @deepseek-ai/dsh-desktop@0.2.0-rc.2 (Electron desktop profile, Windows)
• Evidence artifacts (request headers, descriptors, source excerpts with line
  numbers) available on request.

---

中文摘要

Agent Teams 两个相关缺口（针对 0.2.0-rc.2，均有源码行号与会话级证据）：

1. spawn_teammate 无法按席指定模型：工具层无路由参数，Team runtime 只向
   startContinuable 传 {prompt, parent}，且路由在 provider 钩子之前就已
   按“继承父会话”定死——队友永远继承 Lead 的模型。
2. 名单显示错误：roster.list() 在队友无驻留进程时回退显示 Lead 的模型，
   而真实路由其实持久在子会话 subagent/descriptor 里（冷恢复也据此还原），
   造成 list_agents 与实际请求路由不一致。

建议：(a) startContinuable 回传解析后的路由（或 Team 落入成员快照），
list() 三级回退；(b) spawn_teammate 增加可选 provider/model/ reasoning_effort 并透传 agentOptions（runtime 已支持）。本地已有可退役的
参考实现（dsh-teammate-model-routes）。

[argszero]

Both gaps are confirmed at source in master @ d743267388, and one detail of your workaround has a cleaner form.

Gap 1 — no route pinning. packages/experimental/tool-agent-team/src/index.ts registers spawn_teammate with exactly name / description / prompt / context, and the spawn request carries no route. Matches your reading.

Gap 2 — the roster shows the Lead's model. The source is packages/experimental/agent-team/src/roster.ts:139 (you cited the bundled index.js line):

const model = live?.options.model ?? root.options.model

so an idle member falls back to the root's route exactly as you describe, while the durable subagent/descriptor still holds the truth. Your third fallback (... ?? member.model ?? root...) is the minimal fix.

One correction to the workaround. ctx.reflect._getImpl(...).value is a private path. ctx.get(name) returns a traceable view: property writes on the view do not reach the underlying service instance, but the instance itself is reachable through the global symbol the wrapper carries — ctx.get(name)[Symbol.for('cordis.original')]. That is the same object as _getImpl().value, without touching a private module.

A seam that avoids startContinuable entirely. The route is resolved before the hooks, as you say, but it is consumed by a request-time waterfall — agent/request — which runs after resolveChildAgentOptions and can rewrite the LlmCallConfig for that request. @argszero/cordis-plugin-subagent-model already uses that seam to re-resolve subagent routes per request. Applied to teams, it means: keep a session-id → route map (recorded when the member is created, or read back from the durable subagent/descriptor) and apply it on every request from that session. That removes the AsyncLocalStorage consume-once discipline your wrapper needs — the pin is per session, not per call — and the same map is what the roster correction reads, so both gaps are served by one source of truth.

Caveat worth stating: pinning at request time applies to every request the teammate makes, so adding provider / model / reasoning_effort to spawn_teammate (your option b) is still the better end state. The plugin is the stand-in until the tool layer passes it through.

[ZackZhaii]

Cross-reference: the same gap is tracked in #6985 "Model selection in agent-teams" (2026-09-17), where two community plugins are already listed as workarounds — probably worth consolidating there so maintainers see one thread.

Confirming argszero's readings independently against the packaged runtime (0.2.0-rc.2, unpacked app.asar mirror):

• dsh-experimental-tool-agent-team/lib/index.js:16-19 and :290 — spawn_teammate takes exactly name/description/prompt/context, and the backend is configurable: provider: context === "fork" ? config.forkProvider : config.freshProvider.
• dsh-experimental-agent-team/lib/index.js:447-461 — const model = live?.options.model ?? root.options.model, with status = failed | provisioning | running | inactive.

One safety note for whoever wires the roster fallback: that fallback is not merely cosmetic. For a running member the vendor reports the live model; only the no-live-agent case (status: 'inactive') falls back to the root route. So a client-side patch that rewrites the field unconditionally will also hide genuine drift (provider failover / manual switch) — we hit exactly that in a local workaround and had to scope it to status === 'inactive'. Adding ?? member.model to the chain keeps that distinction intact; a display-layer cache does not.

On ctx.reflect._getImpl(...).value: agreed that it is private. We use it deliberately as a last-resort seam, verify the install against the raw object rather than the traceable view, and fail loudly if the path disappears. A supported way to wrap subagents.startContinuable (or to pass agentOptions from the Team path) would make that unnecessary.