Problem

Ralphex uses Claude Code's default model for all phases (task execution, review, finalize). There is no way to use different models for different phases. This means users pay Opus-level costs for review iterations where Sonnet would suffice, or can't use Opus for tasks without also using it for the 5-7 parallel review agents.

Proposed Solution

Add two new config parameters:

# model for task execution (empty = CLI default)
claude_model = opus

# model for review phases: first review, fix loops, codex eval, finalize
# falls back to claude_model if not set
review_model = sonnet

Plus corresponding CLI flags: --claude-model and --review-model.

Implementation Details

• Add Model field to ClaudeExecutor, injecting --model <value> into the CLI args when set
• Create a separate review executor in runner.New() when review_model differs from claude_model
• Review phases (runClaudeReview, runClaudeReviewLoop, external review eval, finalize) use the review executor
• Task phase and plan mode continue using the primary executor
• Empty values preserve current behavior (no --model flag passed)
• Agent frontmatter model: field continues to work independently (controls subagent model via prompt text)

Precedence

CLI flags > local config (.ralphex/) > global config (~/.config/ralphex/) > embedded defaults > empty (CLI default)

Use Case

Running with claude_model = opus and review_model = sonnet significantly reduces cost on review-heavy runs (where 5+ parallel review agents iterate multiple times), while keeping task quality high.