feat(agent): add chain mode with {previous} placeholder support

[anh-chu]

Adds a chain parameter to the Agent tool that runs multiple agents sequentially, with each step receiving the previous step's output via a {previous} placeholder.

Usage:

Agent({
  chain: [
    { subagent_type: "Explore", prompt: "Find all auth-related files" },
    { subagent_type: "Plan", prompt: "Using: {previous}, design a JWT migration" },
    { subagent_type: "worker", prompt: "Implement: {previous}" },
  ]
})

Changes: 1 file, +228 lines in src/index.ts

• chain TypeBox parameter in Agent tool schema (Type.Optional Array, min 2 items)
• executeChain() function: sequential spawnAndWait, {previous} substitution, per-step error handling, abort support
• Summary output: "Chain completed in Xms (N steps)" with each step's result

When chain is provided, the single-agent path is skipped entirely. Chain steps inherit top-level params (model, thinking, isolation) as defaults.

[anh-chu]

Just to clarify, this means that users can prompt something like:
explore, then plan -> review on plan, then implement
Which would result in a streamlined E2E flow where each agent pipes their output to the next directly, avoiding polluting the main agent's context.