feat: support per-call model and thinking_level parameters in Agent and AgentSwarm tools

What feature would you like to see?

With the release of Kimi K3 (which introduces variable thinking efforts: low, high, max) alongside models like K2.7 (kimi-for-coding / highspeed), subagent collaboration via Agent and AgentSwarm tools currently forces all child agents to inherit the parent session's default model and default thinking effort level.

In multi-agent architectures:

1. Cost & Token Efficiency: Running every subagent on high thinking or K3 for simple tasks (such as single-file pattern searches, glob lookups, or minor formatting edits) consumes excessive context and reasoning tokens unnecessarily.
2. Execution Speed: Lightweight tasks (e.g. codebase exploration) benefit from fast execution (thinking_level: "low" or model: "kimi-for-coding-highspeed"), while complex architecture refactoring needs thinking_level: "high" or "max".

Proposed Solution

Add explicit optional parameters to Agent and AgentSwarm tool schemas:

• thinking_level: z.enum(['off', 'low', 'medium', 'high', 'xhigh', 'max']).optional() — Overrides parent session thinking effort for the subagent call.
• model: z.string().optional() — Overrides parent session model / model alias for the subagent call (e.g., "k3", "kimi-for-coding-highspeed").

Precedence Resolution Chain

1. Per-call tool argument: args.thinking_level / args.model (highest precedence)
2. Subagent profile default: profile.thinking_level / profile.model (e.g., explore.yaml setting thinking_level: low)
3. Parent session default: parent.config.thinkingEffort / parent.config.modelAlias (fallback)

Additional information

• This enables coordinator agents to dynamically select model tiers and thinking depth based on task scope and complexity.
• We have validated this pattern in custom profile directory workflows (--agents-dir), drastically reducing token consumption and latency in multi-agent swarms.

Start by locating the Agent and AgentSwarm tool schemas, then trace how subagent profiles and parent session defaults are resolved. Verify that per-call model and thinking_level values take precedence over profile and parent defaults, and confirm the schemas support the listed optional values.