feat: add configurable effort levels for thinking budget

[guidan-nick]

Summary

Add native Kiro effort parameter support to control thinking/reasoning depth. This fixes the issue where OpenCode's thinking budget settings are ignored because Kiro uses its own effort levels instead of Anthropic's budget_tokens/max_thinking_length.

Problem

When using thinking models (e.g., claude-opus-4-8-thinking) through the Kiro auth plugin, the thinking budget is capped at ~20k tokens regardless of OpenCode's max thinking setting. This is because Kiro ignores Anthropic's native thinking parameters and instead requires effort to be passed via additionalModelRequestFields.output_config.effort.

Solution

Inject the additionalModelRequestFields with the resolved effort level into API requests via SDK middleware.

Changes

• schema.ts: Add effort and auto_effort_mapping config options
• effort.ts: New module with budget-to-effort mapping logic
• sdk-client.ts: Inject additionalModelRequestFields via middleware when effort is set
• request.ts: Resolve effort based on model, thinking state, budget, and config
• request-handler.ts: Pass config effort settings through request pipeline
• types.ts: Add Effort type and extend SdkPreparedRequest

Effort Levels

Level	Description	Approximate Tokens
low	Minimal reasoning	-
medium	Balanced (default)	~20k
high	Deeper reasoning	~50k
xhigh	Extended (opus-4.7/4.8 only)	~80k
max	Maximum depth	~128k

Configuration

// kiro.json - explicit effort level
{
  "effort": "max"
}

// kiro.json - auto-map from OpenCode budget (default behavior)
{
  "auto_effort_mapping": true
}

Testing

• Build passes (npm run build)
• Added unit tests for effort resolution logic

References

• Kiro CLI effort implementation: https://github.com/d-kuro/kirocc
• Kiro effort levels from /effort command: low, medium, high, xhigh, max