Roles: refresh the live selection when re-pointing a reused Session

Deploying dee76b1 proved a raw durable append insufficient: the 07:54 dispatch recorded the wandering selection, yet the request assembled half a second later still carried the deepseek header, because prompt assembly reads the Session Controller's in-memory selection and the internal Agent controller caches one per Agent. Call its own selectForNextRequest primitive — append plus cache refresh, still without saving the deployment default — and fall back to the public selectModel if the surface moves. Fixture gains the Controller; two cases pin fast-path and fallback.

src/app/role-models.ts

@@ -2,6 +2,7 @@
 import type { Context } from '@deepseek-ai/cordis'
 import type { Agent, ModelSelection } from '@deepseek-ai/dsh-agent'
 import { ReasoningEffortId } from '@deepseek-ai/dsh-llm'
+import type { SessionId } from '@deepseek-ai/dsh-session/types'
 import type { RoleModel } from '../policy.ts'
 
 /** Resolve a configured route through the live provider; catalog membership is not a restriction. */
@@ -27,9 +28,16 @@ export async function alignRoleModel(ctx: Context, agent: Agent, model: RoleMode
   const selected = await resolveRoleModel(ctx, model)
   if (current && current.provider === selected.provider && current.model === selected.model
     && (selected.reasoningEffort === undefined || current.reasoningEffort === selected.reasoningEffort)) return
-  // Appending the durable event is what a Session-local selection is; the model-selection projection feeds it to the next
-  // request, and unlike selectModel it leaves the deployment-wide default alone. A mid-session switch is ordinary DSH usage.
-  agent.session.append('model/selection', selected)
+  // A durable event alone cannot re-point a live Agent: prompt assembly reads the Session Controller's in-memory selection,
+  // which a raw append leaves stale. `selectForNextRequest` is that Controller's own primitive — append plus cache refresh —
+  // and unlike the public selectModel it leaves the deployment-wide default alone. The fallback keeps correctness, at the cost
+  // of saving the default, if the internal surface ever moves; a mid-session switch itself is ordinary DSH usage.
+  const installSelection = (ctx.sessionController as unknown as {
+    agents?: { selectForNextRequest?(agent: Agent, selection: ModelSelection): void }
+  }).agents?.selectForNextRequest?.bind((ctx.sessionController as unknown as { agents: object }).agents)
+  if (installSelection) installSelection(agent, selected)
+  else await ctx.sessionController.selectModel({ sessionId: agent.id as SessionId, provider: selected.provider, model: selected.model,
+    ...(selected.reasoningEffort ? { reasoningEffort: selected.reasoningEffort } : {}) })
   await ctx.sessionPersistence.flush()
   ctx.logger.info('dsh-superagent re-pointed Session %s from %s/%s to %s/%s', agent.id,
     current?.provider ?? 'unset', current?.model ?? 'unset', selected.provider, selected.model)