feat: token/cost tracking per agent session (#43)

[itscooleric]

Summary

• New scripts/token-cost.py parses Claude Code conversation JSONL for token usage
• session-logger.sh runs it during cleanup, adds to session_end event:
  • input_tokens, output_tokens, total_tokens
  • estimated_cost_usd (per-model: Opus $15/$75, Sonnet $3/$15, Haiku $0.80/$4)
  • turns count
• Includes cache token tracking (creation + read)
• Script works standalone: python3 token-cost.py conversation.jsonl

Example output

{"input_tokens": 1207, "output_tokens": 145398, "total_tokens": 159673999,
 "estimated_cost_usd": 343.23, "model": "claude-opus-4-6", "turns": 659}

Closes #43

Test plan

• Tested against real conversation logs from clide sessions on clidesdale
• Model detection works for opus, sonnet, haiku variants
• CI checks pass
• End-to-end: run a claude session, verify session_end event has token data

🤖 Generated with Claude Code