Ralph Review Loop

Overview

Run a tight review/fix loop using Codex’s /review command. Each pass collects findings, applies fixes, and re-runs review until the latest pass produces zero actionable bugs or a stop condition is reached.

Workflow

1. Set loop parameters (confirm before starting).

- Ask the user to confirm: scope and whether to include uncommitted changes. - Ask whether to change the model/effort. If yes, instruct them to run /model and select gpt-5.2-codex with reasoning effort xhigh (or high if xhigh isn’t available). - If the user says they already changed the model, do not ask for exact values; proceed using the current session model.

1. Run /review inside the current Codex session (do not spawn a new Codex process).

- Use the in-session /review command, not codex review in a shell. - If an external codex review command is running, stop it and re-run /review in-session. - When /review prompts for scope, select: - Unpushed commits → diff vs upstream (e.g., origin/main). - Base branch → select the branch you want to compare against. - Single commit → pick the commit from the list. - Uncommitted → include staged/unstaged/untracked. - Capture findings (file/line + severity). - If there are no findings, stop and report success.

1. Triage findings.

- Accept: real, actionable bugs/regressions. - Reject: false positives or out-of-scope. - Defer: requires user input; stop and ask. - Keep a short checklist of accepted/rejected items.

1. Fix accepted items.

- Apply minimal, targeted changes. - Update/add tests if needed. - Summarize fixes briefly.

1. Re-run /review only if code changed.

1. Stop conditions.

- Latest pass has zero accepted findings. - User asks to stop or needs input before continuing.

Checkpoints

• Before starting: ask the user to confirm scope and model/effort.
• After the first pass: ask if they want to change scope or model/effort before continuing. If they already changed the model, do not ask for exact values—just proceed.

Output expectations

• For each pass, report: pass number, accepted/rejected counts, and what changed (files/patches).
• On completion, explicitly state why the loop stopped (clean pass or needs input).

Guardrails

• Do not loop forever. Stop when the review is clean or the user asks to stop.
• Do not re-run /review without code changes.
• Do not broaden scope without user approval.