[Bug] Agent Team roster reads model from a stale creation-time snapshot (offline teammates get the Lead's model)

[DDDFXYqiming]

Summary

The Agent Team roster (list_agents / the header panel) reports a member's model from Agent.options.model. That value is a creation-time snapshot, which produces two wrong rows:

1. The Lead row can be stale. agent.options is filled once at agents.create() from agentDefaultModel.currentSelection() (packages/api/session-controller). A later model switch only appends a model/selection session event; it never rewrites agent.options. So the Lead keeps displaying the model that was current when the agent was created, even though every request uses the new one.
2. An offline teammate is mislabeled with the Lead's model. Teammate rows use live?.options.model ?? root.options.model. When a teammate becomes inactive there is no live agent, so the fallback borrows the Lead's (possibly stale) model and the row changes.

Environment

• @deepseek-ai/dsh 0.1.6-alpha.2 (npm, profile web)
• Bundles: @deepseek-ai/dsh-experimental-agent-team-profile, @deepseek-ai/dsh-experimental-agent-team-web-profile
• The two models differ here because the session's agent was created while the previous default was still current; the model was switched moments later.

Reproduction

1. Create a session, then change the model in the composer right after, so that agent.options.model differs from the current agent-default-model selection.
2. spawn_teammate({ name: "probe-model", description: "...", prompt: "reply ok" }).
3. Read the roster immediately.
4. Wait until the teammate is inactive.
5. Read the roster again.

Observed

when	roster row
right after spawn	probe-model / teammate / running / model=deepseek-flash
after it finished	probe-model / teammate / inactive / model=qwen3.8-flash
Lead, unchanged	lead / lead / running / model=qwen3.8-flash

The teammate really ran on deepseek-flash (startContinuable passes no model, so the child route is resolved fresh), but once it is unloaded the roster reports the Lead's stale qwen3.8-flash. Two teammates from an earlier run show the same flip.

Expected

• The Lead row should reflect the session's current model route, not a creation-time snapshot.
• An offline teammate should keep its own model, or omit the field — not borrow the Lead's.

Code

packages/experimental/agent-team/src/roster.ts (built lib/index.js):

// lead row
...root.options.model === void 0 ? {} : { model: root.options.model },

// teammate rows
const model = live?.options.model ?? root.options.model;

packages/api/session-controller/src/index.ts:

return (await this.ctx.agents.create({
  sessionId,
  agentOptions: this.agentOptions(),   // snapshot, taken once at creation
  ...
})).agent;

agentOptions() {
  const { provider, model } = this.ctx.agentDefaultModel.currentSelection();
  return { provider, model };
}

Suggested fix

Resolve the displayed model from the live route (the session's latest model/selection, or agentDefaultModel.currentSelection()), and omit model for a teammate with no live agent instead of falling back to root.options.model.

[omnisurge]

Independent confirmation from a downstream shell (omnidsh) that composes Agent Teams, with a small addition that may narrow the fix.

It reproduces exactly as described

A 9-member team, all teammates finished. The panel:

lead             idle      · model: cline-pass/deepseek-v4.1-flash   <- correct (= agent-default-model)
step3-roles      inactive  · model: cline-pass/deepseek-v4.1-flash   <- actually ran tuanjie/codely-core
explorer         inactive  · model: cline-pass/deepseek-v4.1-flash   <- actually ran grok2api/grok-4.6
doc-writer       inactive  · model: cline-pass/deepseek-v4.1-flash   <- actually ran grok2api/grok-4.6

The distinction worth pinning down: it is not a refresh lag

The roster does re-read on every call - this.ctx.agents.get(member.id) is not cached. The value is wrong because there is nothing to read:

for (const member of state.members) {
  const live = this.ctx.agents.get(member.id);
  const model = live?.options.model ?? root.options.model;   // <- inactive => live is undefined
  result.push({ ... member ..., model });
}

So an inactive member does not fall back to a stale snapshot of itself - it falls back to the Lead's model, which is what makes the display actively misleading rather than merely out of date.

Observed transition on the same member: while step3-roles was running, the panel showed codely-core (correct). The moment it settled to inactive, the same row switched to the Lead's model. That is the whole bug in one observation.

Independent evidence that the dispatch was correct

We wrap ctx.subagents.startContinuable in a plugin, so we can record what was actually injected at dispatch time. The durable per-child descriptor (independent of the roster) agrees with our record and disagrees with the panel:

{ "type": "subagent/descriptor",
  "provider": "spawn",
  "agentProvider": "tuanjie",
  "agentModel": "codely-core",
  "agentReasoningEffort": "high" }

The panel showed the Lead's model for that same child.

Two smaller observations

1. reasoningEffort is not surfaced at all. TeamMemberView has model? but no effort field, so a per-child effort selection is invisible in the UI even for a live member. If the roster row is being revisited, it may be worth carrying the effort alongside the model.

2. A member's model is knowable without the live Agent. The value exists in the durable log (subagent/descriptor) and in the member record created at spawn time; only the rendering path depends on ctx.agents.get(...). Reading it from the durable member record instead would make the row correct for offline members too - which is the case that actually matters, since most of a team is offline most of the time.

Environment

dsh                    @deepseek-ai/dsh@0.1.6-alpha.2
dsh-experimental-agent-team       0.1.6-alpha.2
dsh-experimental-tool-agent-team  0.1.6-alpha.2
macOS 27.0 (Apple M1 Pro, arm64)
application-owned profile (a desktop shell launches dsh web)

Happy to run any probe you want against this setup.

[DDDFXYqiming]

@omnisurge Thanks for the independent repro and for narrowing it to the rendering path.

I confirmed the same code path on dsh 0.1.6-alpha.2 / dsh-experimental-agent-team 0.1.6-alpha.2: lib/types/roster.js:117 still reads const model = live?.options.model ?? root.options.model;, so an inactive member falls back to the Lead's model exactly as you describe — misleading rather than merely stale.

On the fix: the per-member record in the projection carries provider, context and phase but no model, so reading the spawn-time value from the durable member record means adding the field to the projection; your alternative of reading it from the subagent/descriptor avoids that. Either way it is a data-model change rather than a one-line render fix, which is why I am not patching it locally.

Both smaller observations hold on this version as well: reasoningEffort is not surfaced at all, and a member's model is not recoverable from ctx.agents.get() once the member is inactive.