Support per-agent model selection in workflow scripts

[Michaelliv]

Problem

The agent() global accepts a model option, and phase metadata accepts a model field — but neither actually switches the model the subagent runs on. Today, buildAgentInstructions in src/workflow.ts just stringifies it into the prompt:

if (options.model) lines.push(`Requested model: ${options.model}`);

So it's a hint to the subagent, not a real model switch. All subagents end up on whatever model WorkflowAgent was constructed with (inherited from the parent session).

Proposal

Thread options.model from AgentRunOptions into the createAgentSession({ model }) call in WorkflowAgent.run (src/agent.ts). Roughly:

const { session } = await createAgentSession({
  cwd: this.cwd,
  agentDir: getAgentDir(),
  sessionManager: SessionManager.inMemory(),
  settingsManager: SettingsManager.inMemory({ compaction: { enabled: false } }),
  customTools,
  ...this.sessionOptions,
  ...(options.model ? { model: options.model } : {}),
});

And drop the "Requested model: ..." line from buildAgentInstructions — it becomes redundant once the real switch works.

Why

Lets a workflow pick a cheap/fast model for fan-out scans and a stronger model for synthesis, e.g.:

const findings = await parallel(files.map(f => () =>
  agent(`Audit ${f}`, { model: 'haiku', label: f })
))

const summary = await agent(
  'Synthesize these findings:\n' + JSON.stringify(findings),
  { model: 'sonnet', label: 'synthesis' },
)

Out of scope

Per-phase defaults and workflow-level model config — separate issue if wanted.