grilling: starts implementing after a bare "lgtm" or "yes" on the summary

### Skill `grilling` (reached through `/grill-with-docs`) ### Harness and version oh-my-pi (`omp`) 18.11.3. Installed with skills.sh (`npx skills add mattpocock/skills`), skills version installed 2026-10-07 (`grilling` folder hash `28aceb4c`). I also reproduced it in Claude Code 2.1.301. ### Model and effort `claude-sonnet-5-5` in the original session. Effort and context fill were not recorded. The scripted reruns below used `sonnet` in Claude Code. ### What happened I ran a `/grill-with-docs` session about adding a `--branch` option to a CLI submit command. The agent asked its rounds, I accepted the recommendations, and the frontier emptied. I replied `lgtm` to the summary, and the agent started editing code. I had not asked for implementation. The closing line of `grilling/SKILL.md` is "Do not act on it until the user confirms you have reached a shared understanding." A bare `lgtm` reads as that confirmation, and the agent treats it as permission to act. To check this was not a one-off, I replayed the same situation with Claude Code (`sonnet`), 3 runs per case. Each run loaded the skill with the `Skill` tool (confirmed in the session transcripts) and was given the settled summary followed by one final user message. | Final user message | Edited the code with the current wording | Edited the code with the wording below | | --- | --- | --- | | `yes`, `ok`, `lgtm`, `sounds good`, `perfect` | 15 of 15 runs | 0 of 15 runs | | `go ahead`, `do it`, `implement it` | 9 of 9 runs | 9 of 9 runs | I also checked the two skills that write externally after grilling, with four scenarios modelled on the `triage` "Apply the outcome" step and the `wayfinder` "Record the resolution" step. Both wordings completed the calling skill's next step in 12 of 12 runs. Steps to reproduce: install `grilling` from `main`, run `/grilling I want to add a --branch option to my CLI's submit command`, accept the recommended answers until the summary, then reply `lgtm`. It does not happen on every model and wording, and I have not tried a weaker model. ### What you expected The agent stops after the summary and asks what to do next. Agreement approves the summary only. The line in `SKILL.md` that says so does not exist today; the current line is the one quoted above. I tried this replacement for the last sentence of `grilling/SKILL.md`: > Then summarize what was settled and stop. Agreement ("yes", "lgtm") approves the summary only. Act only if the calling skill or user explicitly asks you to (e.g. "implement it", "go ahead"); else ask what next. The change, with a changeset and the three affected spots in `docs/productivity/grilling.md`, is on a branch in my fork (I could not open a PR, since PRs from non-collaborators are refused): https://github.com/mattpocock/skills/compare/main...cazador481:skills:fix/grilling-stop-after-summary The cases and script I used (built on `agent-skill-eval`) are in https://github.com/cazador481/skills/tree/main/evals/grilling. Limits: one model, one scenario, and the check is whether `--branch` appears in the edited file.