Cost and token usage tracking

[adamdaum]

Summary

Each loop iteration makes API calls to Claude (or Codex), but there's no visibility into how many tokens are consumed or what the cost is. A long-running loop could become unexpectedly expensive.

Current Behavior

• ralph.sh invokes claude or codex CLI with no cost instrumentation
• No per-iteration or cumulative cost summary
• No way to set a budget limit or cost alert

Proposal

• After each iteration, capture and log token usage (input/output tokens, cache hits)
• Maintain a running total in a structured file (e.g., usage.json or appended to progress log)
• Print a summary at the end of ralph.sh (total tokens, estimated cost, iterations run)
• Optional: add a --budget flag to ralph.sh that halts the loop if estimated cost exceeds the limit
• Optional: add a --dry-run flag that estimates cost without executing

Example Output

Loop complete: 8 iterations
  Total input tokens:  245,000
  Total output tokens:  89,000
  Estimated cost:       $4.12
  Features completed:   6/10

Considerations

• Token counting depends on the CLI tool exposing usage data — check what claude and codex CLIs output
• Cost estimates need model-specific pricing (Opus vs Sonnet vs Haiku)
• Could integrate with Anthropic's usage API for accurate billing data