codex-adversarial-review-lite

Independent adversarial review

Audit first. The reviewer reports evidence; the builder validates each finding. Do not invent criticism or treat optional polish as a defect. When the evidence supports no material issue, say that no change is needed.

Invocation and settings

• audit (default): changes and their verification evidence.
• plan / feasibility: a proposed plan, constraints, and unresolved assumptions.
• selftest: installed local checks; no model call. Add live-probe to verify the user's selected model/effort and synthetic review, with no repository content (account usage applies).
• reviewer:<model>, reviewer-model:<model>, or model:<model>: explicit model. The user chooses model AND effort; there is no silently selected public default. Reuse an explicit choice from this invocation/session or a saved reviewer preference, announce it, and honor overrides. If no choice exists, ask whether the user knows which reviewer model and effort are available to them: I know my model/effort, show available choices, or automatically choose strongest/highest. For a known choice collect both settings. For uncertainty run inspect and show advertised models/efforts; availability still requires verification after choice. Wait for the answer before any generation or project-content dispatch. If the host cannot ask interactively, return selection_needed with the exact invocation options.
• reviewer:highest reasoning:highest: optional automatic choice. Query the newest compatible installed review CLI's catalog, rank advertised capabilities, and verify access with synthetic content before sending the bundle. Explicit CLI overrides win. Never install/update a CLI, guess a dated model ID, or replace an explicit selection.
• reasoning:<level>: user-selected effort, or explicit highest for the highest advertised level, verified by a synthetic request. This can include Codex ultra and Claude max; use capabilities from the actual CLI rather than assuming those levels exist.
• reviewer:cli-default reasoning:cli-default: opt out of automatic selection. Explicit model/effort choices are verified without automatic replacement. Custom Codex providers require reviewer:configured (or a model ID) plus explicit effort or reasoning:cli-default; the public CLI catalog cannot rank a gateway.
• fallback:<model>: one deliberate availability fallback for explicit model and effort settings in the legacy direct runner; the skill uses verified choices and stops on an unavailable explicit selection. Automatic selection has bounded probes.
• test-spec:<path> / test-data:<path>: focused verification expectations/fixtures.
• rubric:<path>: a nonempty domain checklist, included verbatim in the bundle.
• strict: requires a rubric, human diff review for every change, and sign-off after findings before fixes. This is an explicitly selected review policy.