adversarial-reviewlisted

# Adversarial Review Spawn `safecodex` (a wrapped claude CLI) as an adversarial reviewer in a herdr split pane, hand it the work plus its stated intent, wait for it to finish, then interrogate its findings before relaying them. The reviewer's job is to find real problems, not validate the work; your job afterward is to verify which of its findings are real. The reviewer runs **interactively** (no `-p`): the TUI is live in the split pane, so the user can watch it work, approve permission prompts, or jump in and steer. Herdr detects claude-family agents natively, so completion is tracked via agent status, not output scraping. `safecodex` may be a shell alias/function, so launch it with `herdr pane run` (interactive shell) — never `herdr agent start -- argv`, which spawns a raw process where aliases don't resolve. ## 1. Identify the target and the intent Identify what to review from context: recent diffs, referenced plans, or the user's message. Default to the working diff (`git diff HEAD`, plus untracked files worth including); if the user points at a plan or file, review that instead. Then determine the **intent** — what the author is trying to achieve. This is critical: the reviewer challenges whether the work achieves the intent well, not whether the intent is correct. State the intent explicitly (one or two sentences, shown to the user) before proceeding. If you cannot infer it, ask. ## 2. Prepare the handoff files ```bash mkdir -p /tmp/adversarial-review ID=$(openss