Track token usage and cost per agent

[that-github-user]

Summary

Track per-agent token usage and cost by switching the Claude Code runner from --output-format text to --output-format json.

Available data

claude -p --output-format json returns:

• total_cost_usd — total cost for the agent run
• usage.input_tokens, usage.output_tokens, usage.cache_read_input_tokens
• modelUsage — per-model breakdown

Implementation

1. In src/runners/claude-code.ts, change --output-format text to --output-format json
2. Parse the JSON response to extract result (the agent's text output) and cost/usage fields
3. Add costUsd, inputTokens, outputTokens to the AgentResult type in src/types.ts
4. Store in run JSON so thinktank stats can report aggregate cost data
5. Display per-agent cost in the results table

Why

Users need to understand the cost of ensemble runs. Currently we estimate "$1-5 per run" in docs but have no actual tracking. This would let thinktank stats show real cost data.