Workflow code-review burns excessive tokens (1.1M+ for 5 files) and returns empty results

Problem

The code-review workflow (Workflow({ name: "code-review" })) consistently consumes disproportionate tokens and wall-clock time relative to the work being done, often returning null/empty results.

Reproduction

In a session reviewing 5 changed files (3 Python scripts, 1 YAML config, 1 Markdown skill definition — all small), the workflow:

• Consumed ~1.1M tokens (went from ~33% of a 4-hour budget to ~75% in minutes)
• Ran for 18+ minutes before being manually stopped
• All 4 partial results in the journal were null — zero findings returned

This is the third incident in the same project where the workflow code review went off the rails in the same way.

For comparison, fork-based reviews (/internal-code-reviewer, /internal-test-reviewer) of the same files completed in under 2 minutes, consumed a fraction of the tokens, and returned actionable findings (dead code, variable shadowing, stdout leaks in tests).

Impact

• Real budget cost: ~40% of a 4-hour session budget burned on a single review that produced nothing
• The user had remaining implementation work (2+ tasks) that is now budget-constrained
• The workflow's token consumption is not visible or controllable until after the damage is done

Expected behavior

A code review of 5 small files should not consume more tokens than writing those files did. The workflow should either:

1. Scale its agent count to the size of the changeset (5 files doesn't need the same fan-out as 50)
2. Have a token budget cap that prevents runaway consumption
3. Fail fast if agents are returning empty results rather than continuing to spawn more

Environment

• Claude Code CLI
• Model: Opus
• The code-review workflow is a built-in named workflow, not a custom script