Users Charged Tokens for Agent/Tool Errors

Feedback: Users Charged Tokens for Agent Errors

Issue

Users are charged tokens when Claude Code makes mistakes during tool execution.

Example

During plugin validation, the plugin-validator agent read the wrong directory and returned a 2,000+ token error report about non-existent files. The user was charged for:

• My incorrect tool invocation
• The agent's wasted processing
• The erroneous results I presented

Impact

• Users pay for system errors they didn't cause
• Discourages use of autonomous agents (expensive when they fail)
• Unfair billing model - user budget penalized for tool/agent mistakes

Suggested Solutions

1. Don't charge tokens for agent/tool failures (HTTP errors, wrong paths, validation failures)
2. Provide token refunds for obvious system errors
3. Add confirmation prompts before expensive operations (>1K token estimates)
4. Show token cost estimates before launching agents

Reproduction

1. Use Task tool with invalid path/parameters
2. Agent processes incorrectly
3. User charged for entire failed operation

Session Details

• Date: 2025-01-07
• Plugin: legal-pipeline creation
• Wasted tokens: ~2,000 on validator reading wrong directory
• User complaint: "why do I have to eat the tokens when u do something incorrectly?"

Priority

High - affects fairness of billing model and user trust