claude-reviewlisted

# Claude Review Independent code review using Claude Code, pinned to `claude-opus-5` with `--effort max`. The runner is size-aware: it starts with the richest payload that fits, then automatically falls back to diff-only, per-file batching, or segmented single-file review when the scope is too large. It is also untracked-file aware. Brand-new files that have no `git diff` are reviewed from their current file contents, with content segmentation when needed; the runner must not retry an empty diff-only payload for an untracked file. ## Why This Exists Claude reviews from a different model family than Codex. The value comes from anti-anchoring: pass the code and diff to Claude, never prior reviewer notes or planning context. The runner keeps Claude tools enabled so it can inspect files, run read-only diagnostics, run tests, and use web tools when useful, but the prompt instructs Claude to review rather than modify. Do not add `--bare` for normal use. Claude Code bare mode intentionally ignores OAuth/keychain login and requires `ANTHROPIC_API_KEY` or an API-key helper, so it fails after ordinary `claude auth login`. Isolation comes from independent input selection and disabled session persistence, not from disabling tools. ## Scope Rules - If `$ARGUMENTS` contains a file path or directory: review that path. - If `$ARGUMENTS` contains `pr`, `branch`, or `panel`: review current branch changes versus `main`. - If `$ARGUMENTS` is empty: review all uncommitted changes. ## Work