verify

# /verify — run the checks, show the evidence Claude stops when work *looks* done. Without a check it can actually run, "looks done" is the only signal available, and the human ends up being the verification loop. This skill is the check. **This skill never modifies source code.** It runs verification commands and reports what happened. It is also the **manual path for the Stop gate**. `dotagents-verify-gate.sh` runs these same checks at the end of a turn on both agents — but only Claude Code's Stop hook can actually refuse to finish. **Cursor's `stop` hook cannot block**; it only auto-submits a follow-up message asking the agent to keep going, and Cursor caps how many times that can happen. So in Cursor the gate is a nudge, and running `/verify` explicitly is how you actually know. Do not assume parity. ## Preconditions | Condition | If unmet | |---|---| | The working directory is inside a git repository with an `origin` remote | Stop and report it. The profile is resolved by remote. | | The dotagents checkout is locatable (`~/.claude/.dotagents-managed.json` → `repo`) | Report it, and run the checks without arming the gate. Say that the gate is not active. | | A profile matches this repository | **Stop.** Report that no profile exists, show the repository's remote, and offer to write one. **Never guess commands** — running an invented `npm test` in an unfamiliar repository is how a verification tool loses trust. | ## Position in the workflow | Upstream | This skill |...