What problem does it solve? When a coding task needs a second opinion or a different model's reasoning, manually switching to Codex CLI and crafting a well-structured prompt is tedious and error-prone. This Skill standardizes how Claude Code hands off diagnosis, fix, and code review work to codex exec with consistent prompt templates and sandbox settings. ## Core Features & Use Cases - Three execution modes: rescue (diagnose and fix with workspace-write sandbox), review (read-only diff review), and adversarial-review (hostile review that tries to prove the change should not ship). - Model and effort selection: Prompts the user once to pick from tiered Model x Effort combinations (e.g., gpt-5.5 + xhigh for hard problems, gpt-5.3-codex + high for standard reviews). - Structured prompt templates: Ships XML prompt block templates (task, grounding rules, output contracts, verification loops) so Codex receives grounded, contract-driven instructions instead of vague requests. - Use Case: You have a large uncommitted diff and want a harsh second opinion before merging. Invoke the skill, choose adversarial-review with the top model tier, and receive a severity-ordered findings report with file:line evidence. ## Quick Start Ask the agent to run codex-exec in review mode on the current working tree changes and return the findings report.
