thinking-budget

What problem does it solve?

Centralizes and documents per-agent thinking budget configuration so teams can control how much chain-of-thought reasoning each agent is allowed, avoid zero-budget errors for unspecified agents, and allocate token budget according to role importance.

Core Features & Use Cases

• Centralized JSON configuration stored at ~/.claude/config/thinking-budgets.json for system-wide budget defaults.
• Tiered budgets with recommended values for common roles: 0 for mechanical reviewers, 1024 as a safe default, 2048 for light reasoning, 4096 for moderate analysis, and 8192 for deep reasoning roles.
• Per-agent override via the agent frontmatter thinking_budget field to customize budgets for planners, orchestrators, code-writers, debuggers, and reviewers.
• Use case: increase the planner and orchestrator budgets for multi-file planning tasks while keeping haiku reviewers at zero to save tokens.

Quick Start

Update the agent's thinking_budget frontmatter or edit the JSON file at ~/.claude/config/thinking-budgets.json to set the numeric token budget for the target agents.