executelisted
Install: claude install-skill metallurgical/ogre-runner
# /ogre:execute
Use this skill after the plan is reviewed and approved.
## Inputs
Accept:
- Issue number, e.g. `107`
- Plan name, e.g. `issue-107`
- Plan path, e.g. `.ai/.ogre/plans/issue-107.md`
- `--job <job-id>` instead of any of the above, e.g. `/ogre:execute --job job-6d7715e4-...`
Optional flags:
- `--executor codex|claude` — **every codex spawn runs fully unsandboxed** (`--dangerously-bypass-approvals-and-sandbox`: no filesystem/shell/network confinement, no approval prompts), unconditional, not just `[BROWSER-CHECK]` steps. There is no opt-in flag for this anymore (`--codex-unsandboxed-browser-check`/`codex_unsandboxed_browser_check` are retired) — it's simply how codex always runs in Ogre now, because codex's own sandbox otherwise blocks things Ogre needs outright (real registry/network access, spawning a real browser). Ogre is a dev-only tool as a result. `claude` isolates fine by default (`--permission-mode bypassPermissions`) and needs no such tradeoff. If you see the `WARNING: codex steps run UNSANDBOXED` log line, that is expected on every codex step now, not a bug or a sign something was misconfigured — do not stop the chain over it.
- `--model MODEL`
- `--reasoning LEVEL` — reasoning effort for the executor (`claude -p` gets `--effort LEVEL`, `codex exec` gets `-c model_reasoning_effort=LEVEL`). Omit it to use the CLI's own default; Ogre never forces one.
- `--task <task-id>` — target one specific seeded step out of order
- `--step <n>` — target step N (