devils-advocatelisted
Install: claude install-skill xoonjaeho/devils-advocate-skill
# /devils-advocate — adversarial review by a separate engine
The orchestrator (you) never writes the review. You build the packet, dispatch one engine, check the result, verify its specific claims, and print it as-is.
## 1. Args
Claude Code: `/devils-advocate [target] [--engine E] [--focus TEXT] [--kind K] [--lang L] [--method M]`. Codex: `$devils-advocate ...` or name the skill; same args.
Host = the product running you. Know it from your own system prompt (`CLAUDECODE=1` in the environment confirms Claude Code); never infer it from which tools exist.
- `target` — a git rev (`HEAD`, `abc123`), a range (`main..HEAD`), or a path. Absent → the working tree: `build` runs `git diff HEAD` itself (lockfiles excluded; empty diff → falls back to `HEAD`). Untracked files are not in that diff — say so in the report, or pass their path as the target. A directory target skips `.gitignore`d files and lockfiles; the packet's Target section lists what was skipped. To review a lockfile, pass its path.
- `--engine` — `claude` · `codex`; default = the host's own engine (claude in Claude Code, codex in Codex) — a fresh session of the same engine is independent. Never switch engines silently; if the chosen one is unavailable, say so and stop.
- `--kind` default `script`; `--lang` default `en`; `--focus` default none; `--method` default `STATIC` (`TEST` only when the repo has a runnable test suite).
### Paths and Python
Shell variables do not persist between the host's tool calls. Resolve