effort

type Level ¶

type Level string

Level is the provider-neutral reasoning-effort scale shared by every executor backend. It controls how deeply a model thinks and its overall token spend. The scale follows the Anthropic effort levels — the superset across providers — and each backend maps it onto its native control:

• Claude: output_config.effort, an identity mapping (claudeexecutor.WithEffort).
• Gemini: thinkingLevel on Gemini 3.x models, thinkingBudget tiers on earlier models (googleexecutor.WithEffort).
• OpenAI-compatible: reasoning_effort, where XHigh and Max clamp to "high" (openaiexecutor.WithEffort).

The zero value means "not configured": backends keep their model default, and Validate rejects it — express "no effort" by not passing the option.

const (
	// Low suits short, scoped tasks and latency-sensitive work that is not
	// intelligence-sensitive.
	Low Level = "low"
	// Medium trades some depth for cost; a step down from each backend's
	// default.
	Medium Level = "medium"
	// High balances depth and token spend; it is the Claude model default.
	High Level = "high"
	// XHigh is the recommended setting for hard coding/agentic work on
	// Claude. Backends without an equivalent clamp it to their maximum.
	XHigh Level = "xhigh"
	// Max spends whatever it takes; reserve it for correctness-critical work.
	// Backends without an equivalent clamp it to their maximum.
	Max Level = "max"
)

func (l Level) Validate() error

package main

import (
	"fmt"

	"chainguard.dev/driftlessaf/agents/effort"
)

func main() {
	if err := effort.XHigh.Validate(); err != nil {
		fmt.Println("unexpected:", err)
	}

	if err := effort.Level("extreme").Validate(); err != nil {
		fmt.Println(err)
	}
}

Output:
invalid effort level "extreme" (want low|medium|high|xhigh|max)